From 3b6cfe7c927df178ca3c11643c3ec93e143471c9 Mon Sep 17 00:00:00 2001 From: Alex Petenchea Date: Tue, 22 Aug 2023 21:58:16 +0300 Subject: [PATCH 001/147] convert.py : clarifying error message (#2718) --- convert.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert.py b/convert.py index 71978d671..e720889fd 100644 --- a/convert.py +++ b/convert.py @@ -964,7 +964,7 @@ def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, Sentence path = path3 else: raise FileNotFoundError( - f"Could not find tokenizer.model in {path} or its parent; " + f"Could not find {vocab_file} in {path} or its parent; " "if it's in another directory, pass the directory as --vocab-dir") print(f"Loading vocab file '{path}', type '{vocabtype}'") From c63bb1d16a70c03440671b76954bb767513cead8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Tue, 22 Aug 2023 22:47:05 +0200 Subject: [PATCH 002/147] CUDA: use mul_mat_q kernels by default (#2683) --- common/common.cpp | 16 ++++++++-------- common/common.h | 2 +- examples/server/server.cpp | 13 ++++++------- ggml-cuda.cu | 2 +- 4 files changed, 16 insertions(+), 17 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 1623ba21f..2a83b379e 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -387,11 +387,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { #else fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n"); #endif // GGML_USE_CUBLAS - } else if (arg == "--mul-mat-q" || arg == "-mmq") { + } else if (arg == "--no-mul-mat-q" || arg == "-nommq") { #ifdef GGML_USE_CUBLAS - params.mul_mat_q = true; + params.mul_mat_q = false; #else - fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n"); + fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n"); #endif // GGML_USE_CUBLAS } else if (arg == "--low-vram" || arg == "-lv") { #ifdef GGML_USE_CUBLAS @@ -599,11 +599,11 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stdout, " number of layers to store in VRAM\n"); fprintf(stdout, " -ts SPLIT --tensor-split SPLIT\n"); fprintf(stdout, " how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n"); - fprintf(stdout, " -mg i, --main-gpu i the GPU to use for scratch and small tensors\n" ); - fprintf(stdout, " -lv, --low-vram don't allocate VRAM scratch buffer\n" ); - fprintf(stdout, " -mmq, --mul-mat-q use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" ); - fprintf(stdout, " Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" ); - fprintf(stdout, " is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" ); + fprintf(stdout, " -mg i, --main-gpu i the GPU to use for scratch and small tensors\n"); + fprintf(stdout, " -lv, --low-vram don't allocate VRAM scratch buffer\n"); + fprintf(stdout, " -nommq, --no-mul-mat-q\n"); + fprintf(stdout, " use cuBLAS instead of custom mul_mat_q CUDA kernels.\n"); + fprintf(stdout, " Not recommended since this is both slower and uses more VRAM.\n"); #endif fprintf(stdout, " --mtest compute maximum memory usage\n"); fprintf(stdout, " --export export the computation graph to 'llama.ggml'\n"); diff --git a/common/common.h b/common/common.h index c50a6edfc..18fd951ea 100644 --- a/common/common.h +++ b/common/common.h @@ -68,7 +68,7 @@ struct gpt_params { size_t hellaswag_tasks = 400; // number of tasks to use when computing the HellaSwag score bool low_vram = false; // if true, reduce VRAM usage at the cost of performance - bool mul_mat_q = false; // if true, use experimental mul_mat_q kernels + bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS bool memory_f16 = true; // use f16 instead of f32 for memory kv bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 39fdf3307..e5bc52cd0 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -671,12 +671,11 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, fprintf(stdout, " number of layers to store in VRAM\n"); fprintf(stdout, " -ts SPLIT --tensor-split SPLIT\n"); fprintf(stdout, " how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n"); - fprintf(stdout, " how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n"); fprintf(stdout, " -mg i, --main-gpu i the GPU to use for scratch and small tensors\n"); fprintf(stdout, " -lv, --low-vram don't allocate VRAM scratch buffer\n"); - fprintf(stdout, " -mmq, --mul-mat-q use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" ); - fprintf(stdout, " Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" ); - fprintf(stdout, " is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" ); + fprintf(stdout, " -nommq, --no-mul-mat-q\n"); + fprintf(stdout, " use cuBLAS instead of custom mul_mat_q CUDA kernels.\n"); + fprintf(stdout, " Not recommended since this is both slower and uses more VRAM.\n"); #endif fprintf(stdout, " -m FNAME, --model FNAME\n"); fprintf(stdout, " model path (default: %s)\n", params.model.c_str()); @@ -867,12 +866,12 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {}); #endif // GGML_USE_CUBLAS } - else if (arg == "--mul-mat-q" || arg == "-mmq") + else if (arg == "--no-mul-mat-q" || arg == "-nommq") { #ifdef GGML_USE_CUBLAS - params.mul_mat_q = true; + params.mul_mat_q = false; #else - LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n", {}); + LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n", {}); #endif // GGML_USE_CUBLAS } else if (arg == "--main-gpu" || arg == "-mg") diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 4fe378c21..70a950bb5 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -287,7 +287,7 @@ static int g_device_count = -1; static int g_main_device = 0; static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES]; static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0}; -static bool g_mul_mat_q = false; +static bool g_mul_mat_q = true; static void * g_scratch_buffer = nullptr; static size_t g_scratch_size = 1024*1024*1024; // 1 GB by default From 46ef5b5fcf4c366e1fb27726b6394adbbf8fd0ea Mon Sep 17 00:00:00 2001 From: goerch Date: Tue, 22 Aug 2023 23:10:42 +0200 Subject: [PATCH 003/147] llama : fix whitespace escaping in tokenizer (#2724) --- llama.cpp | 13 +++---------- tests/test-tokenizer-0.cpp | 11 ++++++++++- tests/test-tokenizer-1.cpp | 13 +++---------- 3 files changed, 16 insertions(+), 21 deletions(-) diff --git a/llama.cpp b/llama.cpp index 6abdc44f2..6c5da1309 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2253,18 +2253,11 @@ static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) { } static std::string llama_escape_whitespace(const std::string& text) { - std::string result; - bool escaping = false; - result += "\xe2\x96\x81"; + std::string result = "\xe2\x96\x81"; for (size_t offs = 0; offs < text.length(); ++offs) { if (text[offs] == ' ') { - if (!escaping) { - result += "\xe2\x96\x81"; - escaping = true; - } - } - else { - escaping = false; + result += "\xe2\x96\x81"; + } else { result += text[offs]; } } diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp index 81764565b..f3ee851a3 100644 --- a/tests/test-tokenizer-0.cpp +++ b/tests/test-tokenizer-0.cpp @@ -17,6 +17,8 @@ static std::string unescape_whitespace(llama_context* ctx, const std::vector> & k_tests() { static std::map> _k_tests = { { " ", {1, 259, }, }, + { " ", { 1, 1678, }, }, + { " ", { 1, 268, }, }, { "\t", { 1, 29871, 12, }, }, { "\n", { 1, 29871, 13, }, }, { "\t\n", { 1, 29871, 12, 13, }, }, @@ -38,6 +40,12 @@ static const std::map> & k_tests() { 243, 162, 155, 185, 30722, 243, 162, 143, 174, 30598, 313, 20787, 953, 3848, 275, 16125, 630, 29897, 29871, 31681, 313, 6194, 953, 29877, 2397, 393, 756, 967, 1914, 5993, 29897, }, }, + { "Hello", { 1, 15043 }, }, + { " Hello", { 1, 29871, 15043 }, }, + { " Hello", { 1, 259, 15043 }, }, + { " Hello", { 1, 1678, 15043 }, }, + { " Hello", { 1, 268, 15043 }, }, + { " Hello\n Hello", { 1, 268, 15043, 13, 1678, 15043 }, }, }; return _k_tests; @@ -106,7 +114,8 @@ int main(int argc, char **argv) { if (!correct) { fprintf(stderr, "%s : failed test: '%s'\n", __func__, test_kv.first.c_str()); - fprintf(stderr, "%s : detokenized to: '%s'\n", __func__, unescape_whitespace(ctx, test_kv.second).c_str()); + fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__, + unescape_whitespace(ctx, res).c_str(), unescape_whitespace(ctx, test_kv.second).c_str()); fprintf(stderr, "%s : expected tokens: ", __func__); for (const auto & t : test_kv.second) { fprintf(stderr, "%6d, ", t); diff --git a/tests/test-tokenizer-1.cpp b/tests/test-tokenizer-1.cpp index d8db7cd96..993d17f18 100644 --- a/tests/test-tokenizer-1.cpp +++ b/tests/test-tokenizer-1.cpp @@ -11,18 +11,11 @@ #include static std::string escape_whitespace(const std::string& text) { - std::string result; - bool escaping = false; - result += "\xe2\x96\x81"; + std::string result = "\xe2\x96\x81"; for (size_t offs = 0; offs < text.length(); ++offs) { if (text[offs] == ' ') { - if (!escaping) { - result += "\xe2\x96\x81"; - escaping = true; - } - } - else { - escaping = false; + result += "\xe2\x96\x81"; + } else { result += text[offs]; } } From 777f42ba18b29f25c71ff8de3ecf97b8017304c0 Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Tue, 22 Aug 2023 17:39:39 -0600 Subject: [PATCH 004/147] Improve handling of special tokens in GGML to GGUF converter (#2725) * Improve UNK, BOS, EOS token handling when converting without metadata. * Allow importing as a module. * Remove some obsolete code and minor cleanups. * Set default UNK token mapping from -1 to 0 in llama.cpp * Try to handle overflow due to buggy Windows Python with a better error message --- convert-llama-ggmlv3-to-gguf.py | 43 +++++++++++++++++++++++---------- llama.cpp | 2 +- 2 files changed, 31 insertions(+), 14 deletions(-) diff --git a/convert-llama-ggmlv3-to-gguf.py b/convert-llama-ggmlv3-to-gguf.py index fa4a044ca..5b038fc0a 100644 --- a/convert-llama-ggmlv3-to-gguf.py +++ b/convert-llama-ggmlv3-to-gguf.py @@ -1,10 +1,12 @@ -import sys, struct, math, argparse +import sys, struct, math, argparse, warnings from pathlib import Path import numpy as np import gguf +warnings.filterwarnings('error') + # Note: Does not support GGML_QKK_64 QK_K = 256 # Items here are (block size, type size) @@ -215,15 +217,10 @@ class GGMLToGGUF: if self.vocab_override is not None: vo = self.vocab_override print('* Adding vocab item(s)') - for (idx, vitem) in enumerate(vo.all_tokens()): - if len(vitem) == 3: - tokens.append(vitem[0]) - scores.append(vitem[1]) - toktypes.append(vitem[2]) - else: - # Maybe try to guess the token type here? - tokens.append(vitem[0]) - scores.append(vitem[1]) + for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()): + tokens.append(vbytes) + scores.append(score) + toktypes.append(ttype) assert len(tokens) == hp.n_vocab, f'Override vocab has a different number of items than hyperparameters - override = {len(tokens)} but n_vocab={hp.n_vocab}' gguf_writer.add_token_list(tokens) gguf_writer.add_token_scores(scores) @@ -231,9 +228,21 @@ class GGMLToGGUF: gguf_writer.add_token_types(toktypes) return print(f'* Adding {hp.n_vocab} vocab item(s)') + assert len(self.model.vocab.items) >= 3, 'Cannot handle unexpectedly short model vocab' for (tokid, (vbytes, vscore)) in enumerate(self.model.vocab.items): tt = 1 # Normal - if len(vbytes) == 0: + # Special handling for UNK, BOS, EOS tokens. + if tokid <= 2: + if tokid == 0: + vbytes = b'' + tt = 2 + elif tokid == 1: + vbytes = b'' + tt = 3 + else: + vbytes = b'' + tt = 3 + elif len(vbytes) == 0: tt = 3 # Control elif tokid >= 3 and tokid <= 258 and len(vbytes) == 1: vbytes = bytes(f'<0x{vbytes[0]:02X}>', encoding = 'UTF-8') @@ -246,6 +255,9 @@ class GGMLToGGUF: gguf_writer.add_token_list(tokens) gguf_writer.add_token_scores(scores) gguf_writer.add_token_types(toktypes) + gguf_writer.add_unk_token_id(0) + gguf_writer.add_bos_token_id(1) + gguf_writer.add_eos_token_id(2) def add_tensors(self, gguf_writer): nm = self.name_map @@ -315,7 +327,11 @@ def main(): data = np.memmap(cfg.input, mode = 'r') model = GGMLV3Model() print('* Scanning GGML input file') - offset = model.load(data, 0) + try: + offset = model.load(data, 0) + except OverflowError: + print(f'!!! Caught overflow loading tensors. The most likely issue is running on Windows but not in WSL. Try running in WSL if possible.', file = sys.stderr) + raise print(f'* GGML model hyperparameters: {model.hyperparameters}') vocab_override = None params_override = None @@ -330,4 +346,5 @@ def main(): converter.save() print(f'* Successful completion. Output saved to: {cfg.output}') -main() +if __name__ == '__main__': + main() diff --git a/llama.cpp b/llama.cpp index 6c5da1309..fd8eaa180 100644 --- a/llama.cpp +++ b/llama.cpp @@ -703,7 +703,7 @@ struct llama_vocab { // default LLaMA special tokens id special_bos_id = 1; id special_eos_id = 2; - id special_unk_id = -1; + id special_unk_id = 0; id special_sep_id = -1; id special_pad_id = -1; From f5fe98d11bdf9e7797bcfb05c0c3601ffc4b9d26 Mon Sep 17 00:00:00 2001 From: Evan Jones Date: Tue, 22 Aug 2023 21:01:57 -0400 Subject: [PATCH 005/147] docs : add grammar docs (#2701) * docs : add grammar docs * tweaks to grammar guide * rework GBNF example to be a commented grammar --- README.md | 12 ++++++ examples/main/README.md | 4 ++ grammars/README.md | 91 +++++++++++++++++++++++++++++++++++++++++ 3 files changed, 107 insertions(+) create mode 100644 grammars/README.md diff --git a/README.md b/README.md index 82e070ac3..f746c49eb 100644 --- a/README.md +++ b/README.md @@ -39,6 +39,7 @@ Last revision compatible with the old format: [dadbed9](https://github.com/ggerg
  • Memory/Disk Requirements
  • Quantization
  • Interactive mode
    • +
  • Constrained output with grammars
  • Instruction mode with Alpaca
  • Using OpenLLaMA
  • Using GPT4All
    • @@ -604,6 +605,16 @@ PROMPT_TEMPLATE=./prompts/chat-with-bob.txt PROMPT_CACHE_FILE=bob.prompt.bin \ CHAT_SAVE_DIR=./chat/bob ./examples/chat-persistent.sh ``` +### Constrained output with grammars + +`llama.cpp` supports grammars to constrain model output. For example, you can force the model to output JSON only: + +```bash +./main -m ./models/13B/ggml-model-q4_0.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:' +``` + +The `grammars/` folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](./grammars/README.md). + ### Instruction mode with Alpaca 1. First, download the `ggml` Alpaca model into the `./models` folder @@ -885,3 +896,4 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m / - [BLIS](./docs/BLIS.md) - [Performance troubleshooting](./docs/token_generation_performance_tips.md) - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks) +- [GBNF grammars](./grammars/README.md) diff --git a/examples/main/README.md b/examples/main/README.md index 60e3907d5..d555afdcc 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -288,6 +288,10 @@ These options help improve the performance and memory usage of the LLaMA models. - `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation. +### Grammars + +- `--grammar GRAMMAR`, `--grammar-file FILE`: Specify a grammar (defined inline or in a file) to constrain model output to a specific format. For example, you could force the model to output JSON or to speak only in emojis. See the [GBNF guide](../../grammars/README.md) for details on the syntax. + ### Quantization For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-data--run). diff --git a/grammars/README.md b/grammars/README.md new file mode 100644 index 000000000..7f3b11ca5 --- /dev/null +++ b/grammars/README.md @@ -0,0 +1,91 @@ +# GBNF Guide + +GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `examples/main` and `examples/server`. + +## Background + +[Bakus-Naur Form (BNF)](https://en.wikipedia.org/wiki/Backus%E2%80%93Naur_form) is a notation for describing the syntax of formal languages like programming languages, file formats, and protocols. GBNF is an extension of BNF that primarily adds a few modern regex-like features. + +## Basics + +In GBNF, we define *production rules* that specify how a *non-terminal* (rule name) can be replaced with sequences of *terminals* (characters, specifically Unicode [code points](https://en.wikipedia.org/wiki/Code_point)) and other non-terminals. The basic format of a production rule is `nonterminal ::= sequence...`. + +## Example + +Before going deeper, let's look at some of the features demonstrated in `grammars/chess.gbnf`, a small chess notation grammar: +``` +# `root` specifies the pattern for the overall output +root ::= ( + # it must start with the characters "1. " followed by a sequence + # of characters that match the `move` rule, followed by a space, followed + # by another move, and then a newline + "1. " move " " move "\n" + + # it's followed by one or more subsequent moves, numbered with one or two digits + ([1-9] [0-9]? ". " move " " move "\n")+ +) + +# `move` is an abstract representation, which can be a pawn, nonpawn, or castle. +# The `[+#]?` denotes the possibility of checking or mate signs after moves +move ::= (pawn | nonpawn | castle) [+#]? + +pawn ::= ... +nonpawn ::= ... +castle ::= ... +``` + +## Non-Terminals and Terminals + +Non-terminal symbols (rule names) stand for a pattern of terminals and other non-terminals. They are required to be a dashed lowercase word, like `move`, `castle`, or `check-mate`. + +Terminals are actual characters ([code points](https://en.wikipedia.org/wiki/Code_point)). They can be specified as a sequence like `"1"` or `"O-O"` or as ranges like `[1-9]` or `[NBKQR]`. + +## Characters and character ranges + +Terminals support the full range of Unicode. Unicode characters can be specified directly in the grammar, for example `hiragana ::= [ぁ-ゟ]`, or with escapes: 8-bit (`\xXX`), 16-bit (`\uXXXX`) or 32-bit (`\UXXXXXXXX`). + +Character ranges can be negated with `^`: +``` +single-line ::= [^\n]+ "\n"` +``` + +## Sequences and Alternatives + +The order of symbols in a sequence matter. For example, in `"1. " move " " move "\n"`, the `"1. "` must come before the first `move`, etc. + +Alternatives, denoted by `|`, give different sequences that are acceptable. For example, in `move ::= pawn | nonpawn | castle`, `move` can be a `pawn` move, a `nonpawn` move, or a `castle`. + +Parentheses `()` can be used to group sequences, which allows for embedding alternatives in a larger rule or applying repetition and optptional symbols (below) to a sequence. + +## Repetition and Optional Symbols + +- `*` after a symbol or sequence means that it can be repeated zero or more times. +- `+` denotes that the symbol or sequence should appear one or more times. +- `?` makes the preceding symbol or sequence optional. + +## Comments and newlines + +Comments can be specified with `#`: +``` +# defines optional whitspace +ws ::= [ \t\n]+ +``` + +Newlines are allowed between rules and between symbols or sequences nested inside parentheses. Additionally, a newline after an alternate marker `|` will continue the current rule, even outside of parentheses. + +## The root rule + +In a full grammar, the `root` rule always defines the starting point of the grammar. In other words, it specifies what the entire output must match. + +``` +# a grammar for lists +root ::= ("- " item)+ +item ::= [^\n]+ "\n" +``` + +## Next steps + +This guide provides a brief overview. Check out the GBNF files in this directory (`grammars/`) for examples of full grammars. You can try them out with: +``` +./main -m --grammar-file grammars/some-grammar.gbnf -p 'Some prompt' +``` From b8ad1b66b23f9b2e6e4531e9a62753323036a556 Mon Sep 17 00:00:00 2001 From: Xiao-Yong Jin Date: Wed, 23 Aug 2023 02:12:12 -0500 Subject: [PATCH 006/147] server : allow json array in prompt or content for direct token input (#2306) * server: allow json array in prompt or content We accept an array of strings and numbers representing tokens, in addition to the current string valued prompt or content. This allows direct token input, so that any special tokens can be processed and used at the frontend during the construction of the json data, before sending to the server. And the server does not need to know or parse special tokens from textual input. With this, we can use EOS and BOS used in llama-2-chat models. * server: use tokenizePrompt(json) and default "" if empty prompt * server: fix prompt check * server: tokenize endpoint no longer adds BOS --- examples/server/README.md | 2 +- examples/server/server.cpp | 80 ++++++++++++++++++++++++++++++++++---- 2 files changed, 74 insertions(+), 8 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index 4d97db2e4..77997f98d 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -126,7 +126,7 @@ node . `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`. - `prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. A space is inserted in the front like main.cpp does. + `prompt`: Provide a prompt as a string, or as an array of strings and numbers representing tokens. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. If the prompt is a string, or an array with the first element given as a string, a space is inserted in the front like main.cpp does. `stop`: Specify a JSON array of stopping strings. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []). diff --git a/examples/server/server.cpp b/examples/server/server.cpp index e5bc52cd0..1e6d10c1d 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -190,6 +190,7 @@ struct llama_server_context size_t n_past = 0; size_t n_remain = 0; + json prompt; std::vector embd; std::vector last_n_tokens; @@ -267,6 +268,53 @@ struct llama_server_context return true; } + std::vector tokenize(json json_prompt, bool add_bos) + { + // If `add_bos` is true, we only add BOS, when json_prompt is a string, + // or the first element of the json_prompt array is a string. + std::vector prompt_tokens; + + if (json_prompt.is_array()) + { + bool first = true; + for (const auto& p : json_prompt) + { + if (p.is_string()) + { + auto s = p.template get(); + std::vector p; + if (first) + { + s.insert(0, 1, ' '); // add a space if it's the first + p = ::llama_tokenize(ctx, s, add_bos); + first = false; + } + else + { + p = ::llama_tokenize(ctx, s, false); + } + prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end()); + } + else + { + if (first) + { + first = false; + } + prompt_tokens.push_back(p.template get()); + } + } + } + else + { + auto s = json_prompt.template get(); + s.insert(0, 1, ' '); // always add a first space + prompt_tokens = ::llama_tokenize(ctx, s, add_bos); + } + + return prompt_tokens; + } + bool loadGrammar() { if (!params.grammar.empty()) { @@ -294,8 +342,8 @@ struct llama_server_context void loadPrompt() { - params.prompt.insert(0, 1, ' '); // always add a first space - std::vector prompt_tokens = ::llama_tokenize(ctx, params.prompt, true); + auto prompt_tokens = tokenize(prompt, true); // always add BOS + num_prompt_tokens = prompt_tokens.size(); if (params.n_keep < 0) @@ -1016,7 +1064,7 @@ static json format_final_response(llama_server_context &llama, const std::string {"tokens_predicted", llama.num_tokens_predicted}, {"tokens_evaluated", llama.num_prompt_tokens}, {"generation_settings", format_generation_settings(llama)}, - {"prompt", llama.params.prompt}, + {"prompt", llama.prompt}, {"truncated", llama.truncated}, {"stopped_eos", llama.stopped_eos}, {"stopped_word", llama.stopped_word}, @@ -1085,10 +1133,18 @@ static void parse_options_completion(const json &body, llama_server_context &lla llama.params.penalize_nl = json_value(body, "penalize_nl", default_params.penalize_nl); llama.params.n_keep = json_value(body, "n_keep", default_params.n_keep); llama.params.seed = json_value(body, "seed", default_params.seed); - llama.params.prompt = json_value(body, "prompt", default_params.prompt); llama.params.grammar = json_value(body, "grammar", default_params.grammar); llama.params.n_probs = json_value(body, "n_probs", default_params.n_probs); + if (body.count("prompt") != 0) + { + llama.prompt = body["prompt"]; + } + else + { + llama.prompt = ""; + } + llama.params.logit_bias.clear(); if (json_value(body, "ignore_eos", false)) { @@ -1345,8 +1401,11 @@ int main(int argc, char **argv) auto lock = llama.lock(); const json body = json::parse(req.body); - const std::string content = json_value(body, "content", ""); - const std::vector tokens = llama_tokenize(llama.ctx, content, false); + std::vector tokens; + if (body.count("content") != 0) + { + tokens = llama.tokenize(body["content"], false); + } const json data = format_tokenizer_response(tokens); return res.set_content(data.dump(), "application/json"); }); @@ -1358,7 +1417,14 @@ int main(int argc, char **argv) llama.rewind(); llama_reset_timings(llama.ctx); - llama.params.prompt = json_value(body, "content", ""); + if (body.count("content") != 0) + { + llama.prompt = body["content"]; + } + else + { + llama.prompt = ""; + } llama.params.n_predict = 0; llama.loadPrompt(); llama.beginCompletion(); From 7f7ddd5002040804e33fcdbde44aa22f8635f57d Mon Sep 17 00:00:00 2001 From: IgnacioFDM Date: Wed, 23 Aug 2023 06:31:09 -0300 Subject: [PATCH 007/147] Fix ggml to gguf conversion on Windows (#2733) This fixes `RuntimeWarning: overflow encountered in long_scalars` Credit: anon (not mine) --- convert-llama-ggmlv3-to-gguf.py | 12 +++--------- 1 file changed, 3 insertions(+), 9 deletions(-) diff --git a/convert-llama-ggmlv3-to-gguf.py b/convert-llama-ggmlv3-to-gguf.py index 5b038fc0a..86d459680 100644 --- a/convert-llama-ggmlv3-to-gguf.py +++ b/convert-llama-ggmlv3-to-gguf.py @@ -1,12 +1,10 @@ -import sys, struct, math, argparse, warnings +import sys, struct, math, argparse from pathlib import Path import numpy as np import gguf -warnings.filterwarnings('error') - # Note: Does not support GGML_QKK_64 QK_K = 256 # Items here are (block size, type size) @@ -95,7 +93,7 @@ class Tensor: pad = ((offset + 31) & ~31) - offset offset += pad n_elems = np.prod(self.dims) - n_bytes = (n_elems * tysize) // blksize + n_bytes = np.int64(np.int64(n_elems) * np.int64(tysize)) // np.int64(blksize) self.start_offset = offset self.len_bytes = n_bytes offset += n_bytes @@ -327,11 +325,7 @@ def main(): data = np.memmap(cfg.input, mode = 'r') model = GGMLV3Model() print('* Scanning GGML input file') - try: - offset = model.load(data, 0) - except OverflowError: - print(f'!!! Caught overflow loading tensors. The most likely issue is running on Windows but not in WSL. Try running in WSL if possible.', file = sys.stderr) - raise + offset = model.load(data, 0) print(f'* GGML model hyperparameters: {model.hyperparameters}') vocab_override = None params_override = None From 62959e740e8759d246ac8d09036950efde09981c Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Wed, 23 Aug 2023 12:56:42 +0300 Subject: [PATCH 008/147] Strided perplexity (#2714) * Implementing strided computation of perplexity * Alternative way to output PPL results --------- Co-authored-by: Iwan Kawrakow --- common/common.cpp | 12 +++ common/common.h | 4 + examples/perplexity/perplexity.cpp | 126 ++++++++++++++++++++++++++++- 3 files changed, 141 insertions(+), 1 deletion(-) diff --git a/common/common.cpp b/common/common.cpp index 2a83b379e..88a962ae3 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -417,6 +417,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { params.antiprompt.push_back(argv[i]); } else if (arg == "--perplexity") { params.perplexity = true; + } else if (arg == "--ppl-stride") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.ppl_stride = std::stoi(argv[i]); + } else if (arg == "--ppl-output-type") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.ppl_output_type = std::stoi(argv[i]); } else if (arg == "--hellaswag") { params.hellaswag = true; } else if (arg == "--hellaswag-tasks") { diff --git a/common/common.h b/common/common.h index 18fd951ea..d68a8ef88 100644 --- a/common/common.h +++ b/common/common.h @@ -64,6 +64,10 @@ struct gpt_params { std::string lora_adapter = ""; // lora adapter path std::string lora_base = ""; // base model path for the lora adapter + int ppl_stride = 0; // stride for perplexity calculations. If left at 0, the pre-existing approach will be used. + int ppl_output_type = 0; // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line + // (which is more convenient to use for plotting) + // bool hellaswag = false; // compute HellaSwag score over random tasks from datafile supplied in prompt size_t hellaswag_tasks = 400; // number of tasks to use when computing the HellaSwag score diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index f3c045aec..e89725efc 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -27,7 +27,121 @@ std::vector softmax(const std::vector& logits) { return probs; } +void perplexity_v2(llama_context * ctx, const gpt_params & params) { + + // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research + // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` + // Output: `perplexity: 13.5106 [114/114]` + // BOS tokens will be added for each chunk before eval + + if (params.ppl_stride <= 0) { + fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride); + return; + } + auto tokens = ::llama_tokenize(ctx, params.prompt, true); + + const int calc_chunk = params.n_ctx; + + fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk); + + if (int(tokens.size()) <= calc_chunk) { + fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__, + tokens.size(), params.n_ctx, params.ppl_stride); + return; + } + + const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1) / params.ppl_stride; + + const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max); + const int n_vocab = llama_n_vocab(ctx); + const int n_batch = params.n_batch; + + int count = 0; + double nll = 0.0; + + fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch); + + for (int i = 0; i < n_chunk; ++i) { + const int start = i * params.ppl_stride; + const int end = start + calc_chunk; + + const int num_batches = (calc_chunk + n_batch - 1) / n_batch; + //fprintf(stderr, "%s: evaluating %d...%d using %d batches\n", __func__, start, end, num_batches); + + std::vector logits; + + const auto t_start = std::chrono::high_resolution_clock::now(); + + for (int j = 0; j < num_batches; ++j) { + const int batch_start = start + j * n_batch; + const int batch_size = std::min(end - batch_start, n_batch); + + //fprintf(stderr, " Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch); + if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) { + //fprintf(stderr, "%s : failed to eval\n", __func__); + return; + } + + // save original token and restore it after eval + const auto token_org = tokens[batch_start]; + + // add BOS token for the first batch of each chunk + if (j == 0) { + tokens[batch_start] = llama_token_bos(ctx); + } + + const auto batch_logits = llama_get_logits(ctx); + logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + + if (j == 0) { + tokens[batch_start] = token_org; + } + } + + const auto t_end = std::chrono::high_resolution_clock::now(); + + if (i == 0) { + const float t_total = std::chrono::duration(t_end - t_start).count(); + fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total); + int total_seconds = (int)(t_total * n_chunk); + if (total_seconds >= 60*60) { + fprintf(stderr, "%d hours ", total_seconds / (60*60)); + total_seconds = total_seconds % (60*60); + } + fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); + } + + //fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start); + for (int j = params.n_ctx - params.ppl_stride - 1; j < params.n_ctx - 1; ++j) { + + // Calculate probability of next token, given the previous ones. + const std::vector tok_logits( + logits.begin() + (j + 0) * n_vocab, + logits.begin() + (j + 1) * n_vocab); + + const float prob = softmax(tok_logits)[tokens[start + j + 1]]; + + nll += -std::log(prob); + ++count; + } + // perplexity is e^(average negative log-likelihood) + if (params.ppl_output_type == 0) { + printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); + } else { + printf("%8d %.4lf\n", i*params.ppl_stride, std::exp(nll / count)); + } + fflush(stdout); + } + printf("\n"); +} + void perplexity(llama_context * ctx, const gpt_params & params) { + + if (params.ppl_stride > 0) { + perplexity_v2(ctx, params); + return; + } + // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` // Output: `perplexity: 13.5106 [114/114]` @@ -116,7 +230,11 @@ void perplexity(llama_context * ctx, const gpt_params & params) { ++count; } // perplexity is e^(average negative log-likelihood) - printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); + if (params.ppl_output_type == 0) { + printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); + } else { + printf("%8d %.4lf\n", i*params.n_ctx, std::exp(nll / count)); + } fflush(stdout); } printf("\n"); @@ -369,6 +487,12 @@ int main(int argc, char ** argv) { params.perplexity = true; params.n_batch = std::min(params.n_batch, params.n_ctx); + if (params.ppl_stride > 0) { + fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n", + params.n_ctx, params.n_ctx + params.ppl_stride/2); + params.n_ctx += params.ppl_stride/2; + } + if (params.n_ctx > 2048) { fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);" "expect poor results\n", __func__, params.n_ctx); From 8207214b6a37a46526cee9e72d4c9092b9d1872f Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Wed, 23 Aug 2023 12:57:12 +0300 Subject: [PATCH 009/147] Fix values shown in the quantize tool help (#2735) Co-authored-by: Iwan Kawrakow --- examples/quantize/quantize.cpp | 28 ++++++++++++++-------------- 1 file changed, 14 insertions(+), 14 deletions(-) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index f628d0642..d172f645a 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -14,25 +14,25 @@ struct quant_option { }; static const std::vector QUANT_OPTIONS = { - { "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 3.50G, +0.2499 ppl @ 7B", }, - { "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 3.90G, +0.1846 ppl @ 7B", }, - { "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 4.30G, +0.0796 ppl @ 7B", }, - { "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0415 ppl @ 7B", }, + { "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 3.56G, +0.2166 ppl @ LLaMA-v1-7B", }, + { "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", }, + { "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", }, + { "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", }, #ifdef GGML_USE_K_QUANTS - { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.67G, +0.8698 ppl @ 7B", }, + { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", }, { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, - { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5505 ppl @ 7B", }, - { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.06G, +0.2437 ppl @ 7B", }, - { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1803 ppl @ 7B", }, + { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", }, + { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", }, + { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", }, { "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", }, - { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.56G, +0.1149 ppl @ 7B", }, - { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0535 ppl @ 7B", }, + { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", }, + { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", }, { "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", }, - { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0353 ppl @ 7B", }, - { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0142 ppl @ 7B", }, - { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 5.15G, +0.0044 ppl @ 7B", }, + { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", }, + { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", }, + { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", }, #endif - { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 6.70G, +0.0004 ppl @ 7B", }, + { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", }, { "F16", LLAMA_FTYPE_MOSTLY_F16, "13.00G @ 7B", }, { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, }; From f19dca04ea5fbf9a0b2753091d93464585d5c73b Mon Sep 17 00:00:00 2001 From: JohnnyB Date: Wed, 23 Aug 2023 15:28:22 +0100 Subject: [PATCH 010/147] devops : RPM Specs (#2723) * Create llama-cpp.srpm * Rename llama-cpp.srpm to llama-cpp.srpm.spec Correcting extension. * Tested spec success. * Update llama-cpp.srpm.spec * Create lamma-cpp-cublas.srpm.spec * Create lamma-cpp-clblast.srpm.spec * Update lamma-cpp-cublas.srpm.spec Added BuildRequires * Moved to devops dir --- .devops/lamma-cpp-clblast.srpm.spec | 58 ++++++++++++++++++++++++++++ .devops/lamma-cpp-cublas.srpm.spec | 59 +++++++++++++++++++++++++++++ .devops/llama-cpp.srpm.spec | 58 ++++++++++++++++++++++++++++ 3 files changed, 175 insertions(+) create mode 100644 .devops/lamma-cpp-clblast.srpm.spec create mode 100644 .devops/lamma-cpp-cublas.srpm.spec create mode 100644 .devops/llama-cpp.srpm.spec diff --git a/.devops/lamma-cpp-clblast.srpm.spec b/.devops/lamma-cpp-clblast.srpm.spec new file mode 100644 index 000000000..739c68281 --- /dev/null +++ b/.devops/lamma-cpp-clblast.srpm.spec @@ -0,0 +1,58 @@ +# SRPM for building from source and packaging an RPM for RPM-based distros. +# https://fedoraproject.org/wiki/How_to_create_an_RPM_package +# Built and maintained by John Boero - boeroboy@gmail.com +# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal + +# Notes for llama.cpp: +# 1. Tags are currently based on hash - which will not sort asciibetically. +# We need to declare standard versioning if people want to sort latest releases. +# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies. +# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed. +# Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo +# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries. +# It is up to the user to install the correct vendor-specific support. + +Name: llama.cpp-clblast +Version: master +Release: 1%{?dist} +Summary: OpenCL Inference of LLaMA model in pure C/C++ +License: MIT +Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz +BuildRequires: coreutils make gcc-c++ git mesa-libOpenCL-devel +URL: https://github.com/ggerganov/llama.cpp + +%define debug_package %{nil} +%define source_date_epoch_from_changelog 0 + +%description +CPU inference for Meta's Lllama2 models using default options. + +%prep +%setup -n llama.cpp-master + +%build +make -j LLAMA_CLBLAST=1 + +%install +mkdir -p %{buildroot}%{_bindir}/ +cp -p main %{buildroot}%{_bindir}/llamacppclblast +cp -p server %{buildroot}%{_bindir}/llamacppclblastserver +cp -p simple %{buildroot}%{_bindir}/llamacppclblastsimple + +%clean +rm -rf %{buildroot} +rm -rf %{_builddir}/* + +%files +%{_bindir}/llamacppclblast +%{_bindir}/llamacppclblastserver +%{_bindir}/llamacppclblastsimple + +%pre + +%post + +%preun +%postun + +%changelog diff --git a/.devops/lamma-cpp-cublas.srpm.spec b/.devops/lamma-cpp-cublas.srpm.spec new file mode 100644 index 000000000..75d32fbe7 --- /dev/null +++ b/.devops/lamma-cpp-cublas.srpm.spec @@ -0,0 +1,59 @@ +# SRPM for building from source and packaging an RPM for RPM-based distros. +# https://fedoraproject.org/wiki/How_to_create_an_RPM_package +# Built and maintained by John Boero - boeroboy@gmail.com +# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal + +# Notes for llama.cpp: +# 1. Tags are currently based on hash - which will not sort asciibetically. +# We need to declare standard versioning if people want to sort latest releases. +# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies. +# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed. +# Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo +# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries. +# It is up to the user to install the correct vendor-specific support. + +Name: llama.cpp-cublas +Version: master +Release: 1%{?dist} +Summary: CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL) +License: MIT +Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz +BuildRequires: coreutils make gcc-c++ git cuda-toolkit +Requires: cuda-toolkit +URL: https://github.com/ggerganov/llama.cpp + +%define debug_package %{nil} +%define source_date_epoch_from_changelog 0 + +%description +CPU inference for Meta's Lllama2 models using default options. + +%prep +%setup -n llama.cpp-master + +%build +make -j LLAMA_CUBLAS=1 + +%install +mkdir -p %{buildroot}%{_bindir}/ +cp -p main %{buildroot}%{_bindir}/llamacppcublas +cp -p server %{buildroot}%{_bindir}/llamacppcublasserver +cp -p simple %{buildroot}%{_bindir}/llamacppcublassimple + +%clean +rm -rf %{buildroot} +rm -rf %{_builddir}/* + +%files +%{_bindir}/llamacppcublas +%{_bindir}/llamacppcublasserver +%{_bindir}/llamacppcublassimple + +%pre + +%post + +%preun +%postun + +%changelog diff --git a/.devops/llama-cpp.srpm.spec b/.devops/llama-cpp.srpm.spec new file mode 100644 index 000000000..c65251a5a --- /dev/null +++ b/.devops/llama-cpp.srpm.spec @@ -0,0 +1,58 @@ +# SRPM for building from source and packaging an RPM for RPM-based distros. +# https://fedoraproject.org/wiki/How_to_create_an_RPM_package +# Built and maintained by John Boero - boeroboy@gmail.com +# In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal + +# Notes for llama.cpp: +# 1. Tags are currently based on hash - which will not sort asciibetically. +# We need to declare standard versioning if people want to sort latest releases. +# 2. Builds for CUDA/OpenCL support are separate, with different depenedencies. +# 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed. +# Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo +# 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries. +# It is up to the user to install the correct vendor-specific support. + +Name: llama.cpp +Version: master +Release: 1%{?dist} +Summary: CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL) +License: MIT +Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz +BuildRequires: coreutils make gcc-c++ git +URL: https://github.com/ggerganov/llama.cpp + +%define debug_package %{nil} +%define source_date_epoch_from_changelog 0 + +%description +CPU inference for Meta's Lllama2 models using default options. + +%prep +%autosetup + +%build +make -j + +%install +mkdir -p %{buildroot}%{_bindir}/ +cp -p main %{buildroot}%{_bindir}/llamacpp +cp -p server %{buildroot}%{_bindir}/llamacppserver +cp -p simple %{buildroot}%{_bindir}/llamacppsimple + +%clean +rm -rf %{buildroot} +rm -rf %{_builddir}/* + +%files +%{_bindir}/llamacpp +%{_bindir}/llamacppserver +%{_bindir}/llamacppsimple + +%pre + +%post + +%preun +%postun + +%changelog From 7c2227a1972a4add4b5c118e4914c086513d0382 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Wed, 23 Aug 2023 10:29:09 -0400 Subject: [PATCH 011/147] chmod : make scripts executable (#2675) --- ci/run.sh | 0 convert-falcon-hf-to-gguf.py | 1 + convert-gptneox-hf-to-gguf.py | 1 + convert-llama-7b-pth-to-gguf.py | 1 + convert-llama-ggmlv3-to-gguf.py | 1 + convert-llama-hf-to-gguf.py | 1 + convert-lora-to-ggml.py | 2 +- convert.py | 2 +- examples/embd-input/embd_input.py | 1 + examples/embd-input/llava.py | 1 + examples/embd-input/minigpt4.py | 1 + examples/embd-input/panda_gpt.py | 1 + examples/jeopardy/graph.py | 1 + examples/jeopardy/jeopardy.sh | 0 examples/json-schema-to-grammar.py | 1 + examples/make-ggml.py | 1 + examples/reason-act.sh | 1 - examples/server-llama2-13B.sh | 0 examples/server/api_like_OAI.py | 1 + examples/server/chat-llama2.sh | 0 examples/server/chat.sh | 0 gguf.py | 1 + scripts/get-wikitext-2.sh | 0 23 files changed, 16 insertions(+), 3 deletions(-) mode change 100644 => 100755 ci/run.sh mode change 100644 => 100755 convert-falcon-hf-to-gguf.py mode change 100644 => 100755 convert-gptneox-hf-to-gguf.py mode change 100644 => 100755 convert-llama-7b-pth-to-gguf.py mode change 100644 => 100755 convert-llama-ggmlv3-to-gguf.py mode change 100644 => 100755 convert-llama-hf-to-gguf.py mode change 100644 => 100755 convert.py mode change 100644 => 100755 examples/embd-input/embd_input.py mode change 100644 => 100755 examples/embd-input/llava.py mode change 100644 => 100755 examples/embd-input/minigpt4.py mode change 100644 => 100755 examples/embd-input/panda_gpt.py mode change 100644 => 100755 examples/jeopardy/graph.py mode change 100644 => 100755 examples/jeopardy/jeopardy.sh mode change 100644 => 100755 examples/json-schema-to-grammar.py mode change 100644 => 100755 examples/make-ggml.py mode change 100644 => 100755 examples/server-llama2-13B.sh mode change 100644 => 100755 examples/server/chat-llama2.sh mode change 100644 => 100755 examples/server/chat.sh mode change 100644 => 100755 gguf.py mode change 100644 => 100755 scripts/get-wikitext-2.sh diff --git a/ci/run.sh b/ci/run.sh old mode 100644 new mode 100755 diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py old mode 100644 new mode 100755 index b3e190a0f..50069db56 --- a/convert-falcon-hf-to-gguf.py +++ b/convert-falcon-hf-to-gguf.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 # HF falcon--> gguf conversion import gguf diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py old mode 100644 new mode 100755 index a7cefc6f3..6eeff5bb1 --- a/convert-gptneox-hf-to-gguf.py +++ b/convert-gptneox-hf-to-gguf.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 # HF gptneox--> gguf conversion import gguf diff --git a/convert-llama-7b-pth-to-gguf.py b/convert-llama-7b-pth-to-gguf.py old mode 100644 new mode 100755 index ab5c80b69..f103f5f61 --- a/convert-llama-7b-pth-to-gguf.py +++ b/convert-llama-7b-pth-to-gguf.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 # 7b pth llama --> gguf conversion # Only models with a single datafile are supported, like 7B # HF files required in the model dir: config.json tokenizer_config.json tokenizer.json tokenizer.model diff --git a/convert-llama-ggmlv3-to-gguf.py b/convert-llama-ggmlv3-to-gguf.py old mode 100644 new mode 100755 index 86d459680..3bf93627d --- a/convert-llama-ggmlv3-to-gguf.py +++ b/convert-llama-ggmlv3-to-gguf.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import sys, struct, math, argparse from pathlib import Path diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py old mode 100644 new mode 100755 index f8cfdaa80..08fde238b --- a/convert-llama-hf-to-gguf.py +++ b/convert-llama-hf-to-gguf.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 # HF llama --> gguf conversion import gguf diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index b4999ff5a..04a7b8bbf 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -1,4 +1,4 @@ -#!/usr/bin/env python +#!/usr/bin/env python3 import json import os import re diff --git a/convert.py b/convert.py old mode 100644 new mode 100755 index e720889fd..a701ab41b --- a/convert.py +++ b/convert.py @@ -1,4 +1,4 @@ -#!/usr/bin/env python +#!/usr/bin/env python3 import gguf import argparse diff --git a/examples/embd-input/embd_input.py b/examples/embd-input/embd_input.py old mode 100644 new mode 100755 index be2896614..f146acdc1 --- a/examples/embd-input/embd_input.py +++ b/examples/embd-input/embd_input.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import ctypes from ctypes import cdll, c_char_p, c_void_p, POINTER, c_float, c_int import numpy as np diff --git a/examples/embd-input/llava.py b/examples/embd-input/llava.py old mode 100644 new mode 100755 index bcbdd2bed..06fad55f4 --- a/examples/embd-input/llava.py +++ b/examples/embd-input/llava.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import sys import os sys.path.insert(0, os.path.dirname(__file__)) diff --git a/examples/embd-input/minigpt4.py b/examples/embd-input/minigpt4.py old mode 100644 new mode 100755 index 15c9b77c0..7b13e4a5c --- a/examples/embd-input/minigpt4.py +++ b/examples/embd-input/minigpt4.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import sys import os sys.path.insert(0, os.path.dirname(__file__)) diff --git a/examples/embd-input/panda_gpt.py b/examples/embd-input/panda_gpt.py old mode 100644 new mode 100755 index 0cfac5f32..891ad7cc9 --- a/examples/embd-input/panda_gpt.py +++ b/examples/embd-input/panda_gpt.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import sys import os sys.path.insert(0, os.path.dirname(__file__)) diff --git a/examples/jeopardy/graph.py b/examples/jeopardy/graph.py old mode 100644 new mode 100755 index 1b6c54bff..8bc0706b8 --- a/examples/jeopardy/graph.py +++ b/examples/jeopardy/graph.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import matplotlib.pyplot as plt import os import csv diff --git a/examples/jeopardy/jeopardy.sh b/examples/jeopardy/jeopardy.sh old mode 100644 new mode 100755 diff --git a/examples/json-schema-to-grammar.py b/examples/json-schema-to-grammar.py old mode 100644 new mode 100755 index 2dccc118a..2a4cb65bc --- a/examples/json-schema-to-grammar.py +++ b/examples/json-schema-to-grammar.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import argparse import json import re diff --git a/examples/make-ggml.py b/examples/make-ggml.py old mode 100644 new mode 100755 index f63d9fc22..6a34eeac5 --- a/examples/make-ggml.py +++ b/examples/make-ggml.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 """ This script converts Hugging Face llama models to GGML and quantizes them. diff --git a/examples/reason-act.sh b/examples/reason-act.sh index e7fe655db..046c48db5 100755 --- a/examples/reason-act.sh +++ b/examples/reason-act.sh @@ -1,4 +1,3 @@ - #!/bin/bash cd `dirname $0` diff --git a/examples/server-llama2-13B.sh b/examples/server-llama2-13B.sh old mode 100644 new mode 100755 diff --git a/examples/server/api_like_OAI.py b/examples/server/api_like_OAI.py index aa325a03e..ed19237b0 100755 --- a/examples/server/api_like_OAI.py +++ b/examples/server/api_like_OAI.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import argparse from flask import Flask, jsonify, request, Response import urllib.parse diff --git a/examples/server/chat-llama2.sh b/examples/server/chat-llama2.sh old mode 100644 new mode 100755 diff --git a/examples/server/chat.sh b/examples/server/chat.sh old mode 100644 new mode 100755 diff --git a/gguf.py b/gguf.py old mode 100644 new mode 100755 index 465746718..9421080b8 --- a/gguf.py +++ b/gguf.py @@ -1,3 +1,4 @@ +#!/usr/bin/env python3 import shutil import sys import struct diff --git a/scripts/get-wikitext-2.sh b/scripts/get-wikitext-2.sh old mode 100644 new mode 100755 From cc34dbda9681418a2b18382446b90cdcec398d82 Mon Sep 17 00:00:00 2001 From: akawrykow <142945436+akawrykow@users.noreply.github.com> Date: Wed, 23 Aug 2023 07:31:34 -0700 Subject: [PATCH 012/147] gitignore : fix for windows (#2729) --- .gitignore | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/.gitignore b/.gitignore index a4df837a4..f3121794a 100644 --- a/.gitignore +++ b/.gitignore @@ -3,6 +3,8 @@ *.so *.gguf *.bin +*.exe +*.dll .DS_Store .build/ .cache/ @@ -81,4 +83,3 @@ tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0 - From 5290c38e6e9b66ee2b543e560e301c1a1a90929c Mon Sep 17 00:00:00 2001 From: klosax <131523366+klosax@users.noreply.github.com> Date: Wed, 23 Aug 2023 16:46:03 +0200 Subject: [PATCH 013/147] main : insert bos if no tokens (#2727) * main.cpp : insert bos if no tokens * Update examples/main/main.cpp * Update examples/main/main.cpp --------- Co-authored-by: Georgi Gerganov --- examples/main/main.cpp | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 388e1f7d7..0a22f3c25 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -197,6 +197,11 @@ int main(int argc, char ** argv) { embd_inp = session_tokens; } + // Should not run without any tokens + if (embd_inp.empty()) { + embd_inp.push_back(llama_token_bos(ctx)); + } + // Tokenize negative prompt std::vector guidance_inp; int guidance_offset = 0; From 335acd2ffd7b04501c6d8773ab9fcee6e7bf8639 Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 23 Aug 2023 16:46:54 +0200 Subject: [PATCH 014/147] fix convert-lora-to-ggml.py (#2738) --- convert-lora-to-ggml.py | 33 +++++++++++++++++---------------- 1 file changed, 17 insertions(+), 16 deletions(-) diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index 04a7b8bbf..a94a7d0af 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -6,23 +6,22 @@ import struct import sys from typing import Any, Dict, Sequence, TextIO +import numpy as np import torch -from convert import DATA_TYPE_TO_FTYPE, NUMPY_TYPE_TO_DATA_TYPE, DataType +NUMPY_TYPE_TO_FTYPE: Dict[str, int] = {"float32": 0, "float16": 1} + HF_SUBLAYER_TO_GGML = { - "self_attn.q_proj": "attention.wq", - "self_attn.k_proj": "attention.wk", - "self_attn.v_proj": "attention.wv", - "self_attn.o_proj": "attention.wo", - "mlp.gate_proj": "feed_forward.w1", - "mlp.down_proj": "feed_forward.w2", - "mlp.up_proj": "feed_forward.w3", - "input_layernorm": "attention_norm", + "self_attn.q_proj": "attn_q", + "self_attn.k_proj": "attn_k", + "self_attn.v_proj": "attn_v", + "self_attn.o_proj": "attn_output", + "mlp.gate_proj": "ffn_gate", + "mlp.down_proj": "ffn_down", + "mlp.up_proj": "ffn_up", + "input_layernorm": "attn_norm", "post_attention_layernorm": "ffn_norm", - # "norm": "norm", - # "embed_tokens": "tok_embeddings", - # "lm_head": "output", } @@ -39,7 +38,7 @@ def translate_tensor_name(t: str) -> str: sys.exit(1) output_string = ( - f"layers.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}" + f"blk.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}" ) return output_string else: @@ -54,12 +53,14 @@ def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None: # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int # but some models ship a float value instead # let's convert to int, but fail if lossless conversion is not possible - assert int(params["lora_alpha"]) == params["lora_alpha"], "cannot convert float to int losslessly" + assert ( + int(params["lora_alpha"]) == params["lora_alpha"] + ), "cannot convert float to int losslessly" fout.write(struct.pack("i", int(params["lora_alpha"]))) def write_tensor_header( - self, name: str, shape: Sequence[int], data_type: DataType + self, name: str, shape: Sequence[int], data_type: np.dtype ) -> None: sname = name.encode("utf-8") fout.write( @@ -67,7 +68,7 @@ def write_tensor_header( "iii", len(shape), len(sname), - DATA_TYPE_TO_FTYPE[NUMPY_TYPE_TO_DATA_TYPE[data_type]], + NUMPY_TYPE_TO_FTYPE[data_type.name], ) ) fout.write(struct.pack("i" * len(shape), *shape[::-1])) From 95385241a91a616788a3bb76d12c9b7b2379ca2d Mon Sep 17 00:00:00 2001 From: Olivier Chafik Date: Wed, 23 Aug 2023 20:33:05 +0100 Subject: [PATCH 015/147] examples : restore the functionality to import llama2.c models (#2685) * Fix import of llama2.c models that don't share weights between embedding layers * llama2c: reinstate ggmlv3 conversion output + update readme w/ gguf conv * llama2.c: comment out legacy "load from ggml model" logic * llama2.c: convert special-cased "<0xXX>" single byte tokens from tokenizer.bin --- examples/convert-llama2c-to-ggml/README.md | 14 +- .../convert-llama2c-to-ggml.cpp | 234 ++++++++++-------- 2 files changed, 144 insertions(+), 104 deletions(-) diff --git a/examples/convert-llama2c-to-ggml/README.md b/examples/convert-llama2c-to-ggml/README.md index 868f57d6d..fd561fcbc 100644 --- a/examples/convert-llama2c-to-ggml/README.md +++ b/examples/convert-llama2c-to-ggml/README.md @@ -12,15 +12,19 @@ usage: ./convert-llama2c-to-ggml [options] options: -h, --help show this help message and exit - --copy-vocab-from-model FNAME model path from which to copy vocab (default 'models/ggml-vocab.bin') + --copy-vocab-from-model FNAME model path from which to copy vocab (default 'tokenizer.bin') --llama2c-model FNAME [REQUIRED] model path from which to load Karpathy's llama2.c model --llama2c-output-model FNAME model path to save the converted llama2.c model (default ak_llama_model.bin') ``` -An example command is as follows: +An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows: -`$ ./convert-llama2c-to-ggml --copy-vocab-from-model --llama2c-model --llama2c-output-model ` +`$ ./convert-llama2c-to-ggml --copy-vocab-from-model ../llama2.c/tokenizer.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.ggmlv3.bin` -Now you can use the model with command like: +For now the generated model is in the legacy GGJTv3 format, so you need to convert it to gguf manually: -`$ ./main -m -p "One day, Lily met a Shoggoth" -n 500 -c 256 -eps 1e-5` +`$ python ./convert-llama-ggmlv3-to-gguf.py --eps 1e-5 --input stories42M.ggmlv3.bin --output stories42M.gguf.bin` + +Now you can use the model with a command like: + +`$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256` diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index 469d6e3de..1551a85cd 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -17,6 +17,9 @@ #pragma warning(disable: 4244 4267) // possible loss of data #endif +#define LLAMA_FILE_MAGIC_GGJT 0x67676a74u // 'ggjt' +#define LLAMA_FILE_VERSION_GGJT_V3 3 + //////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc. typedef struct { int dim; // transformer dimension @@ -49,10 +52,10 @@ typedef struct { // float* freq_cis_real; // (seq_len, dim/2) // float* freq_cis_imag; // (seq_len, dim/2) // (optional) classifier weights for the logits, on the last layer - //float* wcls; + float* wcls; } TransformerWeights; -void malloc_weights(TransformerWeights* w, Config* p) { +void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) { // we calloc instead of malloc to keep valgrind happy w->token_embedding_table = new float[p->vocab_size * p->dim](); printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim); @@ -86,9 +89,16 @@ void malloc_weights(TransformerWeights* w, Config* p) { w->rms_final_weight = new float[p->dim](); printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim); + + if (shared_weights) { + w->wcls = NULL; + } else { + w->wcls = new float[p->vocab_size * p->dim](); + printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim); + } } -int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) { +int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) { if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast(p->vocab_size * p->dim)) return 1; if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast(p->n_layers * p->dim)) return 1; if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast(p->n_layers * p->dim * p->dim)) return 1; @@ -100,6 +110,22 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f) { if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast(p->n_layers * p->hidden_dim * p->dim)) return 1; if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast(p->n_layers * p->dim * p->hidden_dim)) return 1; if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast(p->dim)) return 1; + + // Skip freq_cis_real & freq_cis_imag + int head_size = p->dim / p->n_heads; + fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR); + + if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast(p->vocab_size * p->dim)) return 1; + + // Check we didn't forget to read anything + auto curr = ftell(f); + fseek(f, 0, SEEK_END); + auto end = ftell(f); + if (curr != end) { + printf("Error: failed to read the checkpoint file to the end (curr = %ld, end = %ld)\n", curr, end); + return 1; + } + return 0; } @@ -115,6 +141,7 @@ void free_weights(TransformerWeights* w) { delete w->w2; delete w->w3; delete w->rms_final_weight; + if (w->wcls) delete w->wcls; } void print_sample_weights(TransformerWeights *w){ @@ -131,6 +158,7 @@ void print_sample_weights(TransformerWeights *w){ printf("%f\n", w->w2[0]); printf("%f\n", w->w3[0]); printf("%f\n", w->rms_att_weight[0]); + if (w->wcls) printf("%f\n", w->wcls[0]); } //////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -509,26 +537,28 @@ bool is_ggml_file(const char *filename) { } void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) { - // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary - if (is_ggml_file(filename)) { - - struct llama_context_params llama_params = llama_context_default_params(); - llama_params.vocab_only = true; - - struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params); - struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params); - - const int n_vocab = llama_n_vocab(lctx); - vocab->id_to_token.resize(n_vocab); - for (int i=0; iid_to_token[i].text = llama_token_get_text(lctx, i); - vocab->id_to_token[i].score = llama_token_get_score(lctx, i); - vocab->id_to_token[i].type = llama_token_get_type(lctx, i); - vocab->token_to_id.emplace(vocab->id_to_token[i].text, i); - } - llama_free(lctx); - llama_free_model(lmodel); - } else { // assume llama2.c vocabulary +#pragma message("TODO: implement reading vocabulary using gguf") +// // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary +// if (is_ggml_file(filename)) { +// +// struct llama_context_params llama_params = llama_context_default_params(); +// llama_params.vocab_only = true; +// +// struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params); +// struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params); +// +// const int n_vocab = llama_n_vocab(lctx); +// vocab->id_to_token.resize(n_vocab); +// for (int i=0; iid_to_token[i].text = llama_token_get_text(lctx, i); +// vocab->id_to_token[i].score = llama_token_get_score(lctx, i); +// vocab->id_to_token[i].type = llama_token_get_type(lctx, i); +// vocab->token_to_id.emplace(vocab->id_to_token[i].text, i); +// } +// llama_free(lctx); +// llama_free_model(lmodel); +// } else + { // assume llama2.c vocabulary printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename); llama_file file(filename, "rb"); const int n_vocab = config->vocab_size; @@ -538,6 +568,12 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) float_t score = file.read_f32(); uint32_t len = file.read_u32(); std::string text = file.read_string(len); + // Special-case handling of <0xXX> single byte tokens. + char byte_val; + if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) { + char cstr[2] = { byte_val, 0 }; + text = cstr; + } vocab->id_to_token[i].text = text; vocab->id_to_token[i].score = score; vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED; @@ -589,83 +625,80 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod } #pragma message("TODO: implement file saving using gguf") - (void) vocab; - (void) model; - (void) w; -// // write_magic -// file.write_u32(LLAMA_FILE_MAGIC); // magic -// file.write_u32(LLAMA_FILE_VERSION); // version -// // write_hparams -// file.write_u32(model->hparams.n_vocab); -// file.write_u32(model->hparams.n_embd); -// file.write_u32(model->hparams.n_mult); -// file.write_u32(model->hparams.n_head); -// file.write_u32(model->hparams.n_layer); -// file.write_u32(model->hparams.n_rot); -// file.write_u32(LLAMA_FTYPE_ALL_F32); -// -// // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk. -// uint32_t n_vocab = model->hparams.n_vocab; -// for (uint32_t i = 0; i < n_vocab; i++) { -// const auto & token_data = vocab->id_to_token.at(i); -// file.write_u32((uint32_t) token_data.tok.size()); -// file.write_raw(token_data.tok.data(), token_data.tok.size()); -// file.write_raw(&token_data.score, sizeof(token_data.score)); -// } -// -// // stuff AK weights into GG weights one by one. -// // w->token_embedding_table -> model->tok_embeddings -// // float* -> struct ggml_tensor -// stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table); -// stuff_karpathy_weights_into_gg(model->output, w->token_embedding_table); -// -// stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight); -// //print_row(model->norm, 0); -// -// // for rms-att-weight -// int row_length = model->hparams.n_embd; -// const auto & hparams = model->hparams; -// //int n_ff = model->hparams.n_embd; -// int n_ff = get_n_ff(&hparams); -// -// for (uint32_t i = 0; i < model->hparams.n_layer; ++i){ -// auto & layer = model->layers[i]; -// // 1d -// stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]); -// stuff_karpathy_weights_into_gg(layer.ffn_norm , &w->rms_ffn_weight[i*row_length]); -// -// // from 3d matrix layer x dim x dim to 2d matrix dim x dim -// stuff_karpathy_weights_into_gg(layer.wq , &w->wq[i*row_length*row_length]); -// stuff_karpathy_weights_into_gg(layer.wk , &w->wk[i*row_length*row_length]); -// stuff_karpathy_weights_into_gg(layer.wv , &w->wv[i*row_length*row_length]); -// stuff_karpathy_weights_into_gg(layer.wo , &w->wo[i*row_length*row_length]); -// -// stuff_karpathy_weights_into_gg(layer.w1 , &w->w1[i*row_length*n_ff]); -// stuff_karpathy_weights_into_gg(layer.w2 , &w->w2[i*n_ff*row_length]); -// stuff_karpathy_weights_into_gg(layer.w3 , &w->w3[i*row_length*n_ff]); -// } -// // write tensors -// write_tensor(&file, model->tok_embeddings); -// write_tensor(&file, model->norm); -// write_tensor(&file, model->output); // ? -// for (uint32_t i = 0; i < model->hparams.n_layer; ++i) { -// auto & layer = model->layers[i]; -// -// write_tensor(&file, layer.attention_norm); -// write_tensor(&file, layer.wq); -// write_tensor(&file, layer.wk); -// write_tensor(&file, layer.wv); -// write_tensor(&file, layer.wo); -// write_tensor(&file, layer.ffn_norm); -// write_tensor(&file, layer.w1); -// write_tensor(&file, layer.w2); -// write_tensor(&file, layer.w3); -// } + // write_magic + file.write_u32(LLAMA_FILE_MAGIC_GGJT); // magic + file.write_u32(LLAMA_FILE_VERSION_GGJT_V3); // version + // write_hparams + file.write_u32(model->hparams.n_vocab); + file.write_u32(model->hparams.n_embd); + file.write_u32(model->hparams.n_mult); + file.write_u32(model->hparams.n_head); + file.write_u32(model->hparams.n_layer); + file.write_u32(model->hparams.n_rot); + file.write_u32(LLAMA_FTYPE_ALL_F32); + + // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk. + uint32_t n_vocab = model->hparams.n_vocab; + for (uint32_t i = 0; i < n_vocab; i++) { + const auto & token_data = vocab->id_to_token.at(i); + file.write_u32((uint32_t) token_data.text.size()); + file.write_raw(token_data.text.data(), token_data.text.size()); + file.write_raw(&token_data.score, sizeof(token_data.score)); + } + + // stuff AK weights into GG weights one by one. + // w->token_embedding_table -> model->tok_embeddings + // float* -> struct ggml_tensor + stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table); + stuff_karpathy_weights_into_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table); + + stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight); + //print_row(model->norm, 0); + + // for rms-att-weight + int row_length = model->hparams.n_embd; + const auto & hparams = model->hparams; + //int n_ff = model->hparams.n_embd; + int n_ff = get_n_ff(&hparams); + + for (uint32_t i = 0; i < model->hparams.n_layer; ++i){ + auto & layer = model->layers[i]; + // 1d + stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]); + stuff_karpathy_weights_into_gg(layer.ffn_norm , &w->rms_ffn_weight[i*row_length]); + + // from 3d matrix layer x dim x dim to 2d matrix dim x dim + stuff_karpathy_weights_into_gg(layer.wq , &w->wq[i*row_length*row_length]); + stuff_karpathy_weights_into_gg(layer.wk , &w->wk[i*row_length*row_length]); + stuff_karpathy_weights_into_gg(layer.wv , &w->wv[i*row_length*row_length]); + stuff_karpathy_weights_into_gg(layer.wo , &w->wo[i*row_length*row_length]); + + stuff_karpathy_weights_into_gg(layer.w1 , &w->w1[i*row_length*n_ff]); + stuff_karpathy_weights_into_gg(layer.w2 , &w->w2[i*n_ff*row_length]); + stuff_karpathy_weights_into_gg(layer.w3 , &w->w3[i*row_length*n_ff]); + } + // write tensors + write_tensor(&file, model->tok_embeddings); + write_tensor(&file, model->norm); + write_tensor(&file, model->output); // ? + for (uint32_t i = 0; i < model->hparams.n_layer; ++i) { + auto & layer = model->layers[i]; + + write_tensor(&file, layer.attention_norm); + write_tensor(&file, layer.wq); + write_tensor(&file, layer.wk); + write_tensor(&file, layer.wv); + write_tensor(&file, layer.wo); + write_tensor(&file, layer.ffn_norm); + write_tensor(&file, layer.w1); + write_tensor(&file, layer.w2); + write_tensor(&file, layer.w3); + } } struct train_params get_default_train_params() { struct train_params params; - params.fn_vocab_model = "models/ggml-vocab.bin"; + params.fn_vocab_model = "tokenizer.bin"; params.fn_llama2c_output_model = "ak_llama_model.bin"; params.fn_train_data = "shakespeare.txt"; params.fn_checkpoint_in = "checkpoint.bin"; @@ -718,7 +751,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params) fprintf(stderr, "\n"); fprintf(stderr, "options:\n"); fprintf(stderr, " -h, --help show this help message and exit\n"); - fprintf(stderr, " --copy-vocab-from-model FNAME llama2.c vocabulary or ggml model path from which to copy vocab (default '%s')\n", params->fn_vocab_model); + fprintf(stderr, " --copy-vocab-from-model FNAME llama2.c vocabulary or ggmlv3 model path from which to copy vocab (default '%s')\n", params->fn_vocab_model); fprintf(stderr, " --llama2c-model FNAME [REQUIRED] model path from which to load Karpathy's llama2.c model\n"); fprintf(stderr, " --llama2c-output-model FNAME model path to save the converted llama2.c model (default %s')\n", params->fn_llama2c_output_model); fprintf(stderr, "\n"); @@ -791,9 +824,12 @@ int main(int argc, char ** argv) { if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; } // read in the config header if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; } + auto shared_weights = config.vocab_size > 0; + config.vocab_size = abs(config.vocab_size); + // read in the Transformer weights - malloc_weights(&weights, &config); - if(checkpoint_init_weights(&weights, &config, file)) { return 1; } + malloc_weights(&weights, &config, shared_weights); + if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; } fclose(file); } From a192860cfec89a38d59a943623bf595b1fe4495b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 23 Aug 2023 22:37:39 +0300 Subject: [PATCH 016/147] minor : fix trailing whitespace --- examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index 1551a85cd..f8a58dc7a 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -557,7 +557,7 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) // } // llama_free(lctx); // llama_free_model(lmodel); -// } else +// } else { // assume llama2.c vocabulary printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename); llama_file file(filename, "rb"); From cf658adc832badaaa2ca119fe86070e5a830f8f6 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 23 Aug 2023 23:08:04 +0300 Subject: [PATCH 017/147] llm : add Falcon support (#2717) * llama : refactor GGUF constants into static maps * llama : check if model architecture is known * llama : refactor llama_model_load_internal() * gguf : add KV constant maps * llm : read arch-specific KVs * convert : add dummy scores + types * falcon : load tensor data (CPU only) * llama : fix loading progress bar * llama : add arch member to llama_model * falcon : CPU inference working * falcon : support non-40B models * falcon : minor * llama : minor updates ggml-ci * convert-falcon-hf-to-gguf.py : fix special token mapping * llama.cpp : llama default UNK token = id 0 * llama.cpp : fix bpe tokenizer * llama.cpp : fix the fix of bpe tokenizer * ggml : pass eps to ggml_norm * metal : implement RoPE (mode = 2) + avoid ggml_repeat * ggml : ggml_repeat always creates new tensor * falcon : copy-paste self-attention from LLaMA * metal : print extra compute pipeline info * falcon : minor changes (still chasing the Metal problem) * llama.cpp : fix linefeed token * metal : fix GELU kernel numerical stability by using precise::tanh * metal : temporary workaround for the concurrency optimization bug * falcon : add CUDA offloading (#2739) * llama : better model naming and size reporting * llama : prep new tokenizer support * llama : advanced BPE tokenizer based on ggllm.cpp imlpementation * llama : remove oboslete comment ggml-ci * common : remove obsolete BPE API + disable test-tokenizer-1 * llama : revert BPE special-case in llama_byte_to_token() * cuda : add TODOs for RoPE NeoX implementation * llama : default special tokens based on vocab type * perplexity : add log for start of tokenization --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> Co-authored-by: slaren --- common/common.cpp | 32 - common/common.h | 9 - convert-falcon-hf-to-gguf.py | 55 +- convert.py | 6 +- examples/main/main.cpp | 14 +- examples/perplexity/perplexity.cpp | 31 +- ggml-alloc.c | 4 +- ggml-alloc.h | 2 +- ggml-cuda.cu | 29 +- ggml-metal.m | 132 +- ggml-metal.metal | 27 +- ggml.c | 30 +- ggml.h | 7 +- gguf.py | 26 +- llama.cpp | 1826 +++++++++++++++++++++------- llama.h | 15 +- tests/CMakeLists.txt | 3 +- tests/test-tokenizer-1.cpp | 16 +- 18 files changed, 1596 insertions(+), 668 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 88a962ae3..53002ba30 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -744,35 +744,3 @@ std::string llama_token_to_str(const struct llama_context * ctx, llama_token tok return std::string(result.data(), result.size()); } - -std::vector llama_tokenize_bpe( - struct llama_context * ctx, - const std::string & text, - bool add_bos) { - int n_tokens = text.length() + add_bos; - std::vector result(n_tokens); - n_tokens = llama_tokenize_bpe(ctx, text.c_str(), result.data(), result.size(), add_bos); - if (n_tokens < 0) { - result.resize(-n_tokens); - int check = llama_tokenize_bpe(ctx, text.c_str(), result.data(), result.size(), add_bos); - GGML_ASSERT(check == -n_tokens); - } else { - result.resize(n_tokens); - } - return result; -} - -std::string llama_token_to_str_bpe(const struct llama_context * ctx, llama_token token) { - std::vector result(8, 0); - const int n_tokens = llama_token_to_str_bpe(ctx, token, result.data(), result.size()); - if (n_tokens < 0) { - result.resize(-n_tokens); - const int check = llama_token_to_str_bpe(ctx, token, result.data(), result.size()); - GGML_ASSERT(check == -n_tokens); - } else { - result.resize(n_tokens); - } - - return std::string(result.data(), result.size()); -} - diff --git a/common/common.h b/common/common.h index d68a8ef88..17d271e67 100644 --- a/common/common.h +++ b/common/common.h @@ -120,15 +120,6 @@ std::vector llama_tokenize( const std::string & text, bool add_bos); -std::vector llama_tokenize_bpe( - struct llama_context * ctx, - const std::string & text, - bool add_bos); - std::string llama_token_to_str( const struct llama_context * ctx, llama_token token); - -std::string llama_token_to_str_bpe( - const struct llama_context * ctx, - llama_token token); diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py index 50069db56..43e208497 100755 --- a/convert-falcon-hf-to-gguf.py +++ b/convert-falcon-hf-to-gguf.py @@ -95,14 +95,17 @@ print("gguf: get model metadata") block_count = hparams["n_layer"] -gguf_writer.add_name(last_dir) +gguf_writer.add_name("Falcon") gguf_writer.add_context_length(2048) # not in config.json gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform gguf_writer.add_embedding_length(hparams["hidden_size"]) gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"]) gguf_writer.add_block_count(block_count) gguf_writer.add_head_count(hparams["n_head"]) -if "n_head_kv" in hparams: gguf_writer.add_head_count_kv(hparams["n_head_kv"]) +if "n_head_kv" in hparams: + gguf_writer.add_head_count_kv(hparams["n_head_kv"]) +else: + gguf_writer.add_head_count_kv(1) gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) # TOKENIZATION @@ -110,6 +113,8 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) print("gguf: get tokenizer metadata") tokens: List[str] = [] +scores: List[float] = [] +toktypes: List[int] = [] merges: List[str] = [] @@ -153,41 +158,30 @@ if Path(dir_model + "/tokenizer.json").is_file(): text = bytearray(pad_token) tokens.append(text) + scores.append(0.0) # dymmy + toktypes.append(gguf.TokenType.NORMAL) # dummy gguf_writer.add_token_list(tokens) + gguf_writer.add_token_scores(scores) + gguf_writer.add_token_types(toktypes) - if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file(): - print("gguf: get special token ids") +print("gguf: get special token ids") +# Look for special tokens in config.json - with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f: - tokenizer_config = json.load(f) +if "bos_token_id" in hparams and hparams["bos_token_id"] != None: + gguf_writer.add_bos_token_id(hparams["bos_token_id"]) - # find special token ids +if "eos_token_id" in hparams and hparams["eos_token_id"] != None: + gguf_writer.add_eos_token_id(hparams["eos_token_id"]) - if "bos_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["bos_token"]: - gguf_writer.add_bos_token_id(key["id"]) +if "unk_token_id" in hparams and hparams["unk_token_id"] != None: + gguf_writer.add_unk_token_id(hparams["unk_token_id"]) - if "eos_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["eos_token"]: - gguf_writer.add_eos_token_id(key["id"]) +if "sep_token_id" in hparams and hparams["sep_token_id"] != None: + gguf_writer.add_sep_token_id(hparams["sep_token_id"]) - if "unk_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["unk_token"]: - gguf_writer.add_unk_token_id(key["id"]) - - if "sep_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["sep_token"]: - gguf_writer.add_sep_token_id(key["id"]) - - if "pad_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["pad_token"]: - gguf_writer.add_pad_token_id(key["id"]) +if "pad_token_id" in hparams and hparams["pad_token_id"] != None: + gguf_writer.add_pad_token_id(hparams["pad_token_id"]) # TENSORS @@ -195,8 +189,9 @@ if Path(dir_model + "/tokenizer.json").is_file(): tensor_map = gguf.get_tensor_name_map(ARCH,block_count) # params for qkv transform -n_head = hparams["n_head"] +n_head = hparams["n_head"] n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1 + head_dim = hparams["hidden_size"] // n_head # tensor info diff --git a/convert.py b/convert.py index a701ab41b..8d34d5f29 100755 --- a/convert.py +++ b/convert.py @@ -733,7 +733,11 @@ class OutputFile: self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) def add_meta_arch(self, params: Params) -> None: - self.gguf.add_name ("LLaMA") + ver = None + if (params.n_ctx == 4096): + ver = "v2" + + self.gguf.add_name ("LLaMA" if ver == None else "LLaMA " + ver) self.gguf.add_context_length (params.n_ctx) self.gguf.add_embedding_length (params.n_embd) self.gguf.add_block_count (params.n_layer) diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 0a22f3c25..1393f0b08 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -43,7 +43,7 @@ static bool is_interacting = false; void sigint_handler(int signo) { if (signo == SIGINT) { if (!is_interacting) { - is_interacting=true; + is_interacting = true; } else { console::cleanup(); printf("\n"); @@ -189,10 +189,12 @@ int main(int argc, char ** argv) { } } + const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; + // tokenize the prompt std::vector embd_inp; if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) { - embd_inp = ::llama_tokenize(ctx, params.prompt, true); + embd_inp = ::llama_tokenize(ctx, params.prompt, is_spm); } else { embd_inp = session_tokens; } @@ -208,9 +210,9 @@ int main(int argc, char ** argv) { int original_prompt_len = 0; if (ctx_guidance) { params.cfg_negative_prompt.insert(0, 1, ' '); - guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, true); + guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, is_spm); - std::vector original_inp = ::llama_tokenize(ctx, params.prompt, true); + std::vector original_inp = ::llama_tokenize(ctx, params.prompt, is_spm); original_prompt_len = original_inp.size(); guidance_offset = (int)guidance_inp.size() - original_prompt_len; } @@ -257,8 +259,8 @@ int main(int argc, char ** argv) { } // prefix & suffix for instruct mode - const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true); - const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false); + const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", is_spm); + const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false); // in instruct mode, we inject a prefix and a suffix to each input by the user if (params.instruct) { diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index e89725efc..a7bd9db2a 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -28,7 +28,6 @@ std::vector softmax(const std::vector& logits) { } void perplexity_v2(llama_context * ctx, const gpt_params & params) { - // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` // Output: `perplexity: 13.5106 [114/114]` @@ -38,7 +37,13 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) { fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride); return; } - auto tokens = ::llama_tokenize(ctx, params.prompt, true); + + const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; + const bool add_bos = is_spm; + + fprintf(stderr, "%s: tokenizing the input ..\n", __func__); + + auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos); const int calc_chunk = params.n_ctx; @@ -86,7 +91,7 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) { const auto token_org = tokens[batch_start]; // add BOS token for the first batch of each chunk - if (j == 0) { + if (add_bos && j == 0) { tokens[batch_start] = llama_token_bos(ctx); } @@ -136,7 +141,6 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) { } void perplexity(llama_context * ctx, const gpt_params & params) { - if (params.ppl_stride > 0) { perplexity_v2(ctx, params); return; @@ -146,7 +150,13 @@ void perplexity(llama_context * ctx, const gpt_params & params) { // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` // Output: `perplexity: 13.5106 [114/114]` // BOS tokens will be added for each chunk before eval - auto tokens = ::llama_tokenize(ctx, params.prompt, true); + + const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; + const bool add_bos = is_spm; + + fprintf(stderr, "%s: tokenizing the input ..\n", __func__); + + auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos); const int n_chunk_max = tokens.size() / params.n_ctx; @@ -177,7 +187,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) { const auto token_org = tokens[batch_start]; // add BOS token for the first batch of each chunk - if (j == 0) { + if (add_bos && j == 0) { tokens[batch_start] = llama_token_bos(ctx); } @@ -295,8 +305,10 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) { size_t hs_task_count = prompt_lines.size()/6; fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count); + const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; + // This is needed as usual for LLaMA models - bool prepend_bos = true; + const bool add_bos = is_spm; // Number of tasks to use when computing the score if ( params.hellaswag_tasks < hs_task_count ) { @@ -352,14 +364,13 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) { std::vector tok_logits(n_vocab); for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) { - // Tokenize the context to count tokens - std::vector context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, prepend_bos); + std::vector context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos); size_t context_size = context_embd.size(); // Do the 1st ending // In this case we include the context when evaluating - auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], prepend_bos); + auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos); auto query_size = query_embd.size(); //printf("First query: %d\n",(int)query_size); diff --git a/ggml-alloc.c b/ggml-alloc.c index f06f9a3c1..547ec0399 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -238,7 +238,7 @@ static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_t alloc->n_free_blocks++; } -void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n) { +void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) { int pos = 0; for (int i = 0; i < n; i++) { if (list[i] != -1) { @@ -547,7 +547,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n( struct ggml_tensor * view_src = get_view_source(parent); struct hash_node * view_src_hn = hash_get(ht, view_src); view_src_hn->n_views -= 1; - AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src->n_children, view_src->n_views); + AT_PRINTF("view_src %s\n", view_src->name); if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) { ggml_allocator_free_tensor(alloc, view_src); } diff --git a/ggml-alloc.h b/ggml-alloc.h index 14a4350ac..9559da758 100644 --- a/ggml-alloc.h +++ b/ggml-alloc.h @@ -12,7 +12,7 @@ GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment); // tell the allocator to parse nodes following the order described in the list // you should call this if your graph are optimized to execute out-of-order -GGML_API void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n); +GGML_API void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n); GGML_API void ggml_allocr_free(struct ggml_allocr * alloc); GGML_API bool ggml_allocr_is_measure(struct ggml_allocr * alloc); diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 70a950bb5..868b7a7b9 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -3907,6 +3907,29 @@ static __global__ void rope_f32(const float * x, float * dst, const int ncols, c dst[i + 1] = x0*sin_theta + x1*cos_theta; } +// TODO: this implementation is wrong! +//static __global__ void rope_neox_f32(const float * x, float * dst, const int ncols, const float p0, +// const float p_delta, const int p_delta_rows, const float theta_scale) { +// const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); +// +// if (col >= ncols) { +// return; +// } +// +// const int row = blockDim.x*blockIdx.x + threadIdx.x; +// const int i = row*ncols + col/2; +// +// const float theta = (p0 + p_delta * (row/p_delta_rows))*powf(theta_scale, col/2); +// const float sin_theta = sinf(theta); +// const float cos_theta = cosf(theta); +// +// const float x0 = x[i + 0]; +// const float x1 = x[i + ncols/2]; +// +// dst[i + 0] = x0*cos_theta - x1*sin_theta; +// dst[i + ncols/2] = x0*sin_theta + x1*cos_theta; +//} + static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const float p, const float block_p, const float theta_scale) { const int col = blockDim.x*blockIdx.x + threadIdx.x; const int half_n_dims = ncols/4; @@ -5515,7 +5538,8 @@ inline void ggml_cuda_op_rope( const float theta_scale = powf(freq_base, -2.0f/n_dims); - const bool is_glm = mode & 4; + const bool is_neox = mode & 2; + const bool is_glm = mode & 4; // compute if (is_glm) { @@ -5523,6 +5547,9 @@ inline void ggml_cuda_op_rope( const float id_p = min(p, n_ctx - 2.f); const float block_p = max(p - (n_ctx - 2.f), 0.f); rope_glm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, id_p, block_p, theta_scale, cudaStream_main); + } else if (is_neox) { + GGML_ASSERT(false && "RoPE NeoX not implemented yet"); +#pragma message("TODO: implement RoPE NeoX for CUDA") } else { const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale; rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main); diff --git a/ggml-metal.m b/ggml-metal.m index 835c5f297..969cf7daa 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -167,7 +167,9 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { #define GGML_METAL_ADD_KERNEL(name) \ ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \ ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \ - fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name); \ + fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \ + (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \ + (int) ctx->pipeline_##name.threadExecutionWidth); \ if (error) { \ fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ return NULL; \ @@ -218,12 +220,12 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { #undef GGML_METAL_ADD_KERNEL } - fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); - fprintf(stderr, "%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false"); + fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); + fprintf(stderr, "%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false"); if (ctx->device.maxTransferRate != 0) { - fprintf(stderr, "%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0); + fprintf(stderr, "%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0); } else { - fprintf(stderr, "%s: maxTransferRate = built-in GPU\n", __func__); + fprintf(stderr, "%s: maxTransferRate = built-in GPU\n", __func__); } return ctx; @@ -537,8 +539,8 @@ void ggml_metal_graph_compute( id encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc]; - const int node_start = (cb_idx + 0) * n_nodes_per_cb; - const int node_end = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb; + const int node_start = (cb_idx + 0) * n_nodes_per_cb; + const int node_end = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes); for (int ind = node_start; ind < node_end; ++ind) { const int i = has_concur ? ctx->concur_list[ind] : ind; @@ -744,32 +746,31 @@ void ggml_metal_graph_compute( [ctx->device supportsFamily:MTLGPUFamilyApple7] && ne00%32 == 0 && ne11 > 1) { - switch (src0->type) { - case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break; - case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break; - case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break; - case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break; - case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break; - case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break; - case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break; - case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break; - default: GGML_ASSERT(false && "MUL MAT-MAT not implemented"); - } - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9]; - [encoder setBytes:&gqa length:sizeof(gqa) atIndex:10]; - [encoder setThreadgroupMemoryLength:8192 atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; + switch (src0->type) { + case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break; + case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break; + case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break; + case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break; + case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break; + case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break; + case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break; + case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break; + default: GGML_ASSERT(false && "MUL MAT-MAT not implemented"); } - else { + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6]; + [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9]; + [encoder setBytes:&gqa length:sizeof(gqa) atIndex:10]; + [encoder setThreadgroupMemoryLength:8192 atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; + } else { int nth0 = 32; int nth1 = 1; @@ -868,24 +869,24 @@ void ggml_metal_graph_compute( [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14]; [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15]; [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16]; - [encoder setBytes:&gqa length:sizeof(gqa) atIndex:17]; + [encoder setBytes:&gqa length:sizeof(gqa) atIndex:17]; if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else if (src0t == GGML_TYPE_Q3_K) { #ifdef GGML_QKK_64 - [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; #else - [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; #endif } else if (src0t == GGML_TYPE_Q5_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else if (src0t == GGML_TYPE_Q6_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else { [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; @@ -938,16 +939,17 @@ void ggml_metal_graph_compute( } break; case GGML_OP_NORM: { - const float eps = 1e-5f; + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); const int nth = 256; [encoder setComputePipelineState:ctx->pipeline_norm]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; - [encoder setBytes:&eps length:sizeof( float) atIndex:4]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; + [encoder setBytes:&eps length:sizeof( float) atIndex:4]; [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0]; const int64_t nrows = ggml_nrows(src0); @@ -990,7 +992,9 @@ void ggml_metal_graph_compute( [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16]; [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17]; [encoder setBytes:&m0 length:sizeof( float) atIndex:18]; + const int nth = 32; + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; case GGML_OP_ROPE: @@ -1005,8 +1009,8 @@ void ggml_metal_graph_compute( memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); [encoder setComputePipelineState:ctx->pipeline_rope]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; @@ -1057,24 +1061,24 @@ void ggml_metal_graph_compute( default: GGML_ASSERT(false && "not implemented"); } - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12]; + [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14]; + [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15]; + [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16]; + [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; diff --git a/ggml-metal.metal b/ggml-metal.metal index ce3541f4b..7bc3fdf37 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -87,7 +87,12 @@ kernel void kernel_gelu( device float * dst, uint tpig[[thread_position_in_grid]]) { float x = src0[tpig]; - dst[tpig] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); + + // BEWARE !!! + // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs! + // This was observed with Falcon 7B and 40B models + // + dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); } kernel void kernel_soft_max( @@ -571,7 +576,25 @@ kernel void kernel_rope( dst_data[1] = x0*sin_theta + x1*cos_theta; } } else { - // TODO: implement + for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { + for (int64_t ic = 0; ic < n_dims; ic += 2) { + const float cos_theta = cos(theta); + const float sin_theta = sin(theta); + + theta *= theta_scale; + + const int64_t i0 = ib*n_dims + ic/2; + + device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + device float * dst_data = (device float *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta; + } + } } } diff --git a/ggml.c b/ggml.c index dffb97731..8cb5c404f 100644 --- a/ggml.c +++ b/ggml.c @@ -3554,9 +3554,9 @@ inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; } inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } -static const float GELU_COEF_A = 0.044715f; -static const float GELU_QUICK_COEF = -1.702f; -static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; +static const float GELU_COEF_A = 0.044715f; +static const float GELU_QUICK_COEF = -1.702f; +static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; inline static float ggml_gelu_f32(float x) { return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); @@ -5555,10 +5555,6 @@ struct ggml_tensor * ggml_repeat( is_node = true; } - if (ggml_are_same_shape(a, b) && !is_node) { - return a; - } - struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne); result->op = GGML_OP_REPEAT; @@ -5789,6 +5785,7 @@ struct ggml_tensor * ggml_silu_back( static struct ggml_tensor * ggml_norm_impl( struct ggml_context * ctx, struct ggml_tensor * a, + float eps, bool inplace) { bool is_node = false; @@ -5799,7 +5796,7 @@ static struct ggml_tensor * ggml_norm_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - // TODO: maybe store epsilon here? + ggml_set_op_params(result, &eps, sizeof(eps)); result->op = GGML_OP_NORM; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -5810,14 +5807,16 @@ static struct ggml_tensor * ggml_norm_impl( struct ggml_tensor * ggml_norm( struct ggml_context * ctx, - struct ggml_tensor * a) { - return ggml_norm_impl(ctx, a, false); + struct ggml_tensor * a, + float eps) { + return ggml_norm_impl(ctx, a, eps, false); } struct ggml_tensor * ggml_norm_inplace( struct ggml_context * ctx, - struct ggml_tensor * a) { - return ggml_norm_impl(ctx, a, true); + struct ggml_tensor * a, + float eps) { + return ggml_norm_impl(ctx, a, eps, true); } // ggml_rms_norm @@ -10619,7 +10618,8 @@ static void ggml_compute_forward_norm_f32( GGML_TENSOR_UNARY_OP_LOCALS; - const float eps = 1e-5f; // TODO: make this a parameter + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); // TODO: optimize for (int64_t i03 = 0; i03 < ne03; i03++) { @@ -12537,7 +12537,7 @@ static void ggml_compute_forward_rope_f32( dst_data[1] = x0*sin_theta*zeta + x1*cos_theta*zeta; } } else { - // TODO: this is probably wrong, but I can't figure it out .. + // TODO: this might be wrong for ne0 != n_dims - need double check // ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28 for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { for (int64_t ic = 0; ic < n_dims; ic += 2) { @@ -12666,7 +12666,7 @@ static void ggml_compute_forward_rope_f16( dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); } } else { - // TODO: this is probably wrong, but I can't figure it out .. + // TODO: this might be wrong for ne0 != n_dims - need double check // ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28 for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { for (int64_t ic = 0; ic < n_dims; ic += 2) { diff --git a/ggml.h b/ggml.h index 3c48fd27f..421c0df60 100644 --- a/ggml.h +++ b/ggml.h @@ -909,14 +909,15 @@ extern "C" { struct ggml_tensor * b); // normalize along rows - // TODO: eps is hardcoded to 1e-5 for now GGML_API struct ggml_tensor * ggml_norm( struct ggml_context * ctx, - struct ggml_tensor * a); + struct ggml_tensor * a, + float eps); GGML_API struct ggml_tensor * ggml_norm_inplace( struct ggml_context * ctx, - struct ggml_tensor * a); + struct ggml_tensor * a, + float eps); GGML_API struct ggml_tensor * ggml_rms_norm( struct ggml_context * ctx, diff --git a/gguf.py b/gguf.py index 9421080b8..5c37f0f0b 100755 --- a/gguf.py +++ b/gguf.py @@ -30,12 +30,12 @@ KEY_GENERAL_SOURCE_HF_REPO = "general.source.hugginface.repository" KEY_GENERAL_FILE_TYPE = "general.file_type" # LLM -KEY_LLM_CONTEXT_LENGTH = "{arch}.context_length" -KEY_LLM_EMBEDDING_LENGTH = "{arch}.embedding_length" -KEY_LLM_BLOCK_COUNT = "{arch}.block_count" -KEY_LLM_FEED_FORWARD_LENGTH = "{arch}.feed_forward_length" -KEY_LLM_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual" -KEY_LLM_TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout" +KEY_CONTEXT_LENGTH = "{arch}.context_length" +KEY_EMBEDDING_LENGTH = "{arch}.embedding_length" +KEY_BLOCK_COUNT = "{arch}.block_count" +KEY_FEED_FORWARD_LENGTH = "{arch}.feed_forward_length" +KEY_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual" +KEY_TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout" # attention KEY_ATTENTION_HEAD_COUNT = "{arch}.attention.head_count" @@ -583,7 +583,7 @@ class GGUFWriter: self.add_string(KEY_GENERAL_AUTHOR, author) def add_tensor_data_layout(self, layout: str): - self.add_string(KEY_LLM_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout) + self.add_string(KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout) def add_url(self, url: str): self.add_string(KEY_GENERAL_URL, url) @@ -613,27 +613,27 @@ class GGUFWriter: def add_context_length(self, length: int): self.add_uint32( - KEY_LLM_CONTEXT_LENGTH.format(arch=self.arch), length) + KEY_CONTEXT_LENGTH.format(arch=self.arch), length) def add_embedding_length(self, length: int): self.add_uint32( - KEY_LLM_EMBEDDING_LENGTH.format(arch=self.arch), length) + KEY_EMBEDDING_LENGTH.format(arch=self.arch), length) def add_block_count(self, length: int): self.add_uint32( - KEY_LLM_BLOCK_COUNT.format(arch=self.arch), length) + KEY_BLOCK_COUNT.format(arch=self.arch), length) def add_feed_forward_length(self, length: int): self.add_uint32( - KEY_LLM_FEED_FORWARD_LENGTH.format(arch=self.arch), length) + KEY_FEED_FORWARD_LENGTH.format(arch=self.arch), length) def add_parallel_residual(self, use: bool): self.add_bool( - KEY_LLM_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use) + KEY_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use) def add_tensor_data_layout(self, layout: str): self.add_string( - KEY_LLM_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout) + KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout) def add_head_count(self, count: int): self.add_uint32( diff --git a/llama.cpp b/llama.cpp index fd8eaa180..f2dc4da1d 100644 --- a/llama.cpp +++ b/llama.cpp @@ -72,6 +72,7 @@ #include #include #include +#include #include #include #include @@ -80,20 +81,6 @@ #pragma warning(disable: 4244 4267) // possible loss of data #endif -// tensor names -#define TN_TOKEN_EMBD "token_embd.weight" -#define TN_OUTPUT_NORM "output_norm.weight" -#define TN_OUTPUT "output.weight" -#define TN_ATTN_NORM "blk.%d.attn_norm.weight" -#define TN_ATTN_Q "blk.%d.attn_q.weight" -#define TN_ATTN_K "blk.%d.attn_k.weight" -#define TN_ATTN_V "blk.%d.attn_v.weight" -#define TN_ATTN_OUTPUT "blk.%d.attn_output.weight" -#define TN_FFN_NORM "blk.%d.ffn_norm.weight" -#define TN_FFN_GATE "blk.%d.ffn_gate.weight" -#define TN_FFN_DOWN "blk.%d.ffn_down.weight" -#define TN_FFN_UP "blk.%d.ffn_up.weight" - #ifdef __GNUC__ #ifdef __MINGW32__ #define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) @@ -107,6 +94,7 @@ // // logging // + LLAMA_ATTRIBUTE_FORMAT(2, 3) static void llama_log_internal (llama_log_level level, const char* format, ...); static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data); @@ -119,6 +107,21 @@ static void llama_log_callback_default(llama_log_level level, const char * text, // helpers // +static size_t utf8_len(char src) { + const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 }; + uint8_t highbits = static_cast(src) >> 4; + return lookup[highbits]; +} + +void replace_all(std::string & s, const std::string & search, const std::string & replace) { + for (size_t pos = 0; ; pos += replace.length()) { + pos = s.find(search, pos); + if (pos == std::string::npos) break; + s.erase(pos, search.length()); + s.insert(pos, replace); + } +} + static void zeros(std::ofstream & file, size_t n) { char zero = 0; for (size_t i = 0; i < n; ++i) { @@ -142,6 +145,241 @@ static std::string format(const char * fmt, ...) { return std::string(buf.data(), size); } +// +// gguf constants (sync with gguf.py) +// + +enum llm_arch { + LLM_ARCH_LLAMA, + LLM_ARCH_FALCON, + LLM_ARCH_GPT2, + LLM_ARCH_GPTJ, + LLM_ARCH_GPTNEOX, + LLM_ARCH_MPT, + LLM_ARCH_UNKNOWN, +}; + +static std::map LLM_ARCH_NAMES = { + { LLM_ARCH_LLAMA, "llama" }, + { LLM_ARCH_FALCON, "falcon" }, + { LLM_ARCH_GPT2, "gpt2" }, + { LLM_ARCH_GPTJ, "gptj" }, + { LLM_ARCH_GPTNEOX, "gptneox" }, + { LLM_ARCH_MPT, "mpt" }, +}; + +enum llm_kv { + LLM_KV_GENERAL_ARCHITECTURE, + LLM_KV_GENERAL_QUANTIZATION_VERSION, + LLM_KV_GENERAL_ALIGNMENT, + LLM_KV_GENERAL_NAME, + LLM_KV_GENERAL_AUTHOR, + LLM_KV_GENERAL_URL, + LLM_KV_GENERAL_DESCRIPTION, + LLM_KV_GENERAL_LICENSE, + LLM_KV_GENERAL_SOURCE_URL, + LLM_KV_GENERAL_SOURCE_HF_REPO, + + LLM_KV_CONTEXT_LENGTH, + LLM_KV_EMBEDDING_LENGTH, + LLM_KV_BLOCK_COUNT, + LLM_KV_FEED_FORWARD_LENGTH, + LLM_KV_USE_PARALLEL_RESIDUAL, + LLM_KV_TENSOR_DATA_LAYOUT, + + LLM_KV_ATTENTION_HEAD_COUNT, + LLM_KV_ATTENTION_HEAD_COUNT_KV, + LLM_KV_ATTENTION_MAX_ALIBI_BIAS, + LLM_KV_ATTENTION_CLAMP_KQV, + LLM_KV_ATTENTION_LAYERNORM_EPS, + LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, + + LLM_KV_ROPE_DIMENSION_COUNT, + LLM_KV_ROPE_SCALE_LINEAR, + + LLM_KV_TOKENIZER_MODEL, + LLM_KV_TOKENIZER_LIST, + LLM_KV_TOKENIZER_TOKEN_TYPE, + LLM_KV_TOKENIZER_SCORES, + LLM_KV_TOKENIZER_MERGES, + LLM_KV_TOKENIZER_BOS_ID, + LLM_KV_TOKENIZER_EOS_ID, + LLM_KV_TOKENIZER_UNK_ID, + LLM_KV_TOKENIZER_SEP_ID, + LLM_KV_TOKENIZER_PAD_ID, + LLM_KV_TOKENIZER_HF_JSON, + LLM_KV_TOKENIZER_RWKV, +}; + +static std::map LLM_KV_NAMES = { + { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, + { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, + { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, + { LLM_KV_GENERAL_NAME, "general.name" }, + { LLM_KV_GENERAL_AUTHOR, "general.author" }, + { LLM_KV_GENERAL_URL, "general.url" }, + { LLM_KV_GENERAL_DESCRIPTION, "general.description" }, + { LLM_KV_GENERAL_LICENSE, "general.license" }, + { LLM_KV_GENERAL_SOURCE_URL, "general.source_url" }, + { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source_hf_repo" }, + + { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, + { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" }, + { LLM_KV_BLOCK_COUNT, "%s.block_count" }, + { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" }, + { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" }, + { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" }, + + { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, + { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, + { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, + { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, + { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, + { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, + + { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, + { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, + + { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, + { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, + { LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" }, + { LLM_KV_TOKENIZER_SCORES, "tokenizer.ggml.scores" }, + { LLM_KV_TOKENIZER_MERGES, "tokenizer.ggml.merges" }, + { LLM_KV_TOKENIZER_BOS_ID, "tokenizer.ggml.bos_token_id" }, + { LLM_KV_TOKENIZER_EOS_ID, "tokenizer.ggml.eos_token_id" }, + { LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" }, + { LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" }, + { LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" }, + { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" }, + { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" }, +}; + +struct LLM_KV { + LLM_KV(llm_arch arch) : arch(arch) {} + + llm_arch arch; + + std::string operator()(llm_kv kv) const { + return ::format(LLM_KV_NAMES[kv].c_str(), LLM_ARCH_NAMES[arch].c_str()); + } +}; + +enum llm_tensor { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_POS_EMBD, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_ROPE_FREQS, + LLM_TENSOR_ATTN_Q, + LLM_TENSOR_ATTN_K, + LLM_TENSOR_ATTN_V, + LLM_TENSOR_ATTN_QKV, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_NORM_2, + LLM_TENSOR_ATTN_ROT_EMBD, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_UP, + LLM_TENSOR_FFN_NORM, +}; + +static std::map> LLM_TENSOR_NAMES = { + { + LLM_ARCH_LLAMA, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_FALCON, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, +}; + +static llm_arch llm_arch_from_string(const std::string & name) { + for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT + if (kv.second == name) { + return kv.first; + } + } + + return LLM_ARCH_UNKNOWN; +} + +// helper to handle gguf constants +// usage: +// +// const auto tn = LLM_TN(LLM_ARCH_LLAMA); +// +// std::string name = tn(LLM_TENSOR_OUTPUT); -> "output" +// std::string name = tn(LLM_TENSOR_TOKEN_EMBD, "bias"); -> "token_embd.bias" +// std::string name = tn(LLM_TENSOR_ATTN_NORM, "weight", 3); -> "blk.3.attn_norm.weight" +// +struct LLM_TN { + LLM_TN(llm_arch arch) : arch(arch) {} + + llm_arch arch; + + std::string operator()(llm_tensor tensor) const { + return LLM_TENSOR_NAMES[arch].at(tensor); + } + + std::string operator()(llm_tensor tensor, const std::string & suffix) const { + return LLM_TENSOR_NAMES[arch].at(tensor) + "." + suffix; + } + + std::string operator()(llm_tensor tensor, int bid) const { + return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid); + } + + std::string operator()(llm_tensor tensor, const std::string & suffix, int bid) const { + return ::format(LLM_TENSOR_NAMES[arch].at(tensor).c_str(), bid) + "." + suffix; + } +}; + +// +// gguf helpers +// + +#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \ +{ \ + const std::string skey(key); \ + const int kid = gguf_find_key(ctx, skey.c_str()); \ + if (kid >= 0) { \ + enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \ + if (ktype != (type)) { \ + throw std::runtime_error(format("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype))); \ + } \ + (dst) = func(ctx, kid); \ + } else if (req) { \ + throw std::runtime_error(format("key not found in model: %s", skey.c_str())); \ + } \ +} + // // ggml helpers // @@ -589,12 +827,13 @@ enum e_model { MODEL_7B, MODEL_13B, MODEL_30B, + MODEL_40B, MODEL_65B, MODEL_70B, }; static const size_t kB = 1024; -static const size_t MB = 1024*1024; +static const size_t MB = kB*kB; // default hparams (LLaMA 7B) struct llama_hparams { @@ -608,6 +847,7 @@ struct llama_hparams { uint32_t n_rot = 64; uint32_t n_ff = 11008; + float f_norm_eps = 1e-5; float f_norm_rms_eps = 1e-5; float rope_freq_base = 10000.0f; @@ -641,21 +881,25 @@ struct llama_hparams { struct llama_layer { // normalization - struct ggml_tensor * attention_norm; + struct ggml_tensor * attn_norm; + struct ggml_tensor * attn_norm_b; + struct ggml_tensor * attn_norm_2; + struct ggml_tensor * attn_norm_2_b; // attention struct ggml_tensor * wq; struct ggml_tensor * wk; struct ggml_tensor * wv; struct ggml_tensor * wo; + struct ggml_tensor * wqkv; // normalization struct ggml_tensor * ffn_norm; // ff - struct ggml_tensor * w1; - struct ggml_tensor * w2; - struct ggml_tensor * w3; + struct ggml_tensor * w1; // ffn_gate + struct ggml_tensor * w2; // ffn_down + struct ggml_tensor * w3; // ffn_up }; struct llama_kv_cache { @@ -681,10 +925,6 @@ struct llama_kv_cache { }; struct llama_vocab { - // TODO: - // - add a vector of merges - // so that we can pass it to different types of tokenizers with a common interface - using id = int32_t; using token = std::string; using ttype = llama_token_type; @@ -695,11 +935,13 @@ struct llama_vocab { ttype type; }; - llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM; + enum llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM; std::unordered_map token_to_id; std::vector id_to_token; + std::map, int> bpe_ranks; + // default LLaMA special tokens id special_bos_id = 1; id special_eos_id = 2; @@ -708,21 +950,40 @@ struct llama_vocab { id special_pad_id = -1; id linefeed_id = 13; + + int find_bpe_rank(std::string token_left, std::string token_right) const { + replace_all(token_left, " ", "Ġ"); + replace_all(token_left, "\n", "Ċ"); + replace_all(token_right, " ", "Ġ"); + replace_all(token_right, "\n", "Ċ"); + + auto it = bpe_ranks.find(std::make_pair(token_left, token_right)); + if (it == bpe_ranks.end()) { + return -1; + } + + return it->second; + } }; struct llama_model { e_model type = MODEL_UNKNOWN; + llm_arch arch = LLM_ARCH_UNKNOWN; llama_ftype ftype = LLAMA_FTYPE_ALL_F32; + std::string name = "n/a"; + llama_hparams hparams; llama_vocab vocab; struct ggml_tensor * tok_embeddings; - struct ggml_tensor * norm; + struct ggml_tensor * output_norm; + struct ggml_tensor * output_norm_b; struct ggml_tensor * output; std::vector layers; + int n_gpu_layers; // context @@ -800,8 +1061,6 @@ struct llama_context { // key + value cache for the self attention struct llama_kv_cache kv_self; - size_t mem_per_token = 0; - // decode output (2-dimensional array: [n_tokens][n_vocab]) std::vector logits; bool logits_all = false; @@ -880,11 +1139,11 @@ static bool llama_kv_cache_init( // model loading and saving // -enum llama_file_version { +enum llama_fver { GGUF_FILE_VERSION_V1 = 1, }; -static const char * llama_file_version_name(llama_file_version version) { +static const char * llama_file_version_name(llama_fver version) { switch (version) { case GGUF_FILE_VERSION_V1: return "GGUF V1 (latest)"; } @@ -892,11 +1151,11 @@ static const char * llama_file_version_name(llama_file_version version) { return "unknown"; } -static std::string llama_format_tensor_shape(const std::vector & ne) { +static std::string llama_format_tensor_shape(const std::vector & ne) { char buf[256]; - snprintf(buf, sizeof(buf), "%5u", ne.at(0)); + snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0)); for (size_t i = 1; i < ne.size(); i++) { - snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5u", ne.at(i)); + snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i)); } return buf; } @@ -919,9 +1178,9 @@ struct llama_model_loader { bool use_mmap = false; - llama_file file; + llama_file file; llama_ftype ftype; - llama_file_version fver; + llama_fver fver; std::unique_ptr mapping; @@ -942,7 +1201,7 @@ struct llama_model_loader { n_kv = gguf_get_n_kv(ctx_gguf); n_tensors = gguf_get_n_tensors(ctx_gguf); - fver = (enum llama_file_version) gguf_get_version(ctx_gguf); + fver = (enum llama_fver ) gguf_get_version(ctx_gguf); for (int i = 0; i < n_tensors; i++) { const char * name = gguf_get_tensor_name(ctx_gguf, i); @@ -1039,6 +1298,21 @@ struct llama_model_loader { } } + std::string get_arch_name() const { + const auto kv = LLM_KV(LLM_ARCH_UNKNOWN); + + std::string arch_name; + GGUF_GET_KEY(ctx_gguf, arch_name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_ARCHITECTURE)); + + return arch_name; + } + + enum llm_arch get_arch() const { + const std::string arch_name = get_arch_name(); + + return llm_arch_from_string(arch_name); + } + const char * get_tensor_name(int i) const { return gguf_get_tensor_name(ctx_gguf, i); } @@ -1076,7 +1350,7 @@ struct llama_model_loader { return tensor; } - struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector & ne, ggml_backend backend) { + struct ggml_tensor * create_tensor(struct ggml_context * ctx, const std::string & name, const std::vector & ne, ggml_backend backend) { struct ggml_tensor * cur = ggml_get_tensor(ctx_meta, name.c_str()); if (cur == NULL) { @@ -1244,228 +1518,279 @@ static const char * llama_model_type_name(e_model type) { case MODEL_7B: return "7B"; case MODEL_13B: return "13B"; case MODEL_30B: return "30B"; + case MODEL_40B: return "40B"; case MODEL_65B: return "65B"; case MODEL_70B: return "70B"; - default: GGML_ASSERT(false); + default: return "?B"; } } -static void llama_model_load_internal( - const std::string & fname, +static void llm_load_arch(llama_model_loader & ml, llama_model & model) { + model.arch = ml.get_arch(); + if (model.arch == LLM_ARCH_UNKNOWN) { + throw std::runtime_error("unknown model architecture: '" + ml.get_arch_name() + "'"); + } +} + +static void llm_load_hparams( + llama_model_loader & ml, llama_model & model, - llama_vocab & vocab, int n_ctx, + float rope_freq_base, + float rope_freq_scale) { + struct gguf_context * ctx = ml.ctx_gguf; + + const auto kv = LLM_KV(model.arch); + + auto & hparams = model.hparams; + + // get general kv + GGUF_GET_KEY(ctx, model.name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_NAME)); + + // get hparams kv + GGUF_GET_KEY(ctx, hparams.n_vocab, gguf_get_arr_n, GGUF_TYPE_ARRAY, true, kv(LLM_KV_TOKENIZER_LIST)); + GGUF_GET_KEY(ctx, hparams.n_ctx_train, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_CONTEXT_LENGTH)); + GGUF_GET_KEY(ctx, hparams.n_embd, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH)); + GGUF_GET_KEY(ctx, hparams.n_ff, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH)); + GGUF_GET_KEY(ctx, hparams.n_head, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT)); + GGUF_GET_KEY(ctx, hparams.n_layer, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT)); + + // n_head_kv is optional, default to n_head + hparams.n_head_kv = hparams.n_head; + GGUF_GET_KEY(ctx, hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV)); + + // TODO: manually setting rope scale should override this + // rope_freq_scale (inverse of the kv) is optional + { + float ropescale = 1.0f; + GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); + if (ropescale != 1.0f) { + rope_freq_scale = 1.0f/ropescale; + } + } + + // sanity check for n_rot (optional) + { + hparams.n_rot = hparams.n_embd / hparams.n_head; + + GGUF_GET_KEY(ctx, hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT)); + + if (hparams.n_rot != hparams.n_embd / hparams.n_head) { + throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head)); + } + } + + // arch-specific KVs + switch (model.arch) { + case LLM_ARCH_LLAMA: + { + GGUF_GET_KEY(ctx, hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + + switch (hparams.n_layer) { + case 26: model.type = e_model::MODEL_3B; break; + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + case 60: model.type = e_model::MODEL_30B; break; + case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_FALCON: + { + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 60: model.type = e_model::MODEL_40B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + default: (void)0; + }; + + model.ftype = ml.ftype; + + hparams.n_ctx = n_ctx; + hparams.rope_freq_base = rope_freq_base; + hparams.rope_freq_scale = rope_freq_scale; +} + +// TODO: This should probably be in llama.h +static std::vector llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos, bool escape); + +static void llm_load_vocab( + llama_model_loader & ml, + llama_model & model) { + auto & vocab = model.vocab; + + struct gguf_context * ctx = ml.ctx_gguf; + + const auto kv = LLM_KV(model.arch); + + const int token_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_LIST).c_str()); + if (token_idx == -1) { + throw std::runtime_error("cannot find tokenizer vocab in model file\n"); + } + + const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str()); + if (score_idx == -1) { + throw std::runtime_error("cannot find tokenizer scores in model file\n"); + } + + const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx); + + const int toktype_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE).c_str()); + if (toktype_idx == -1) { + throw std::runtime_error("cannot find token type list in GGUF file\n"); + } + + const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx); + + // determine vocab type + { + std::string tokenizer_name; + + GGUF_GET_KEY(ctx, tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, kv(LLM_KV_TOKENIZER_MODEL)); + + if (tokenizer_name == "llama") { + vocab.type = LLAMA_VOCAB_TYPE_SPM; + + // default special tokens + vocab.special_bos_id = 1; + vocab.special_eos_id = 2; + vocab.special_unk_id = 0; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + } else if (tokenizer_name == "gpt2") { + vocab.type = LLAMA_VOCAB_TYPE_BPE; + + // read bpe merges and populate bpe ranks + const int merges_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_MERGES).c_str()); + if (merges_keyidx == -1) { + throw std::runtime_error("cannot find tokenizer merges in model file\n"); + } + + const int n_merges = gguf_get_arr_n(ctx, merges_keyidx); + + for (int i = 0; i < n_merges; i++) { + const std::string word = gguf_get_arr_str(ctx, merges_keyidx, i); + + std::string first; + std::string second; + + const size_t pos = word.find(' ', 1); + + if (pos != std::string::npos) { + first = word.substr(0, pos); + second = word.substr(pos + 1); + } + + vocab.bpe_ranks.emplace(std::make_pair(first, second), i); + } + + // default special tokens + vocab.special_bos_id = 11; + vocab.special_eos_id = 11; + vocab.special_unk_id = -1; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; + } else { + LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str()); + LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__); + + vocab.type = LLAMA_VOCAB_TYPE_SPM; + } + } + + const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx); + + vocab.id_to_token.resize(n_vocab); + + for (uint32_t i = 0; i < n_vocab; i++) { + std::string word = gguf_get_arr_str(ctx, token_idx, i); + + vocab.token_to_id[word] = i; + + auto & token_data = vocab.id_to_token[i]; + token_data.text = std::move(word); + token_data.score = scores[i]; + token_data.type = (llama_token_type) toktypes[i]; + } + + // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n' + vocab.linefeed_id = llama_tokenize_internal(vocab, "\n", false, false)[0]; + + // special tokens + GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID)); + GGUF_GET_KEY(ctx, vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID)); + GGUF_GET_KEY(ctx, vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID)); + GGUF_GET_KEY(ctx, vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID)); + GGUF_GET_KEY(ctx, vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID)); +} + +static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { + const auto & hparams = model.hparams; + const auto & vocab = model.vocab; + + // hparams + LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml.fver)); + LLAMA_LOG_INFO("%s: arch = %s\n", __func__, LLM_ARCH_NAMES.at(model.arch).c_str()); + LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, vocab.type == LLAMA_VOCAB_TYPE_SPM ? "SPM" : "BPE"); // TODO: fix + LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab); + LLAMA_LOG_INFO("%s: n_merges = %u\n", __func__, (int) vocab.bpe_ranks.size()); + LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train); + LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, hparams.n_ctx); + LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd); + LLAMA_LOG_INFO("%s: n_head = %u\n", __func__, hparams.n_head); + LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv); + LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); + LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim + LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa()); + LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_eps); + LLAMA_LOG_INFO("%s: f_norm_rms_eps = %.1e\n", __func__, hparams.f_norm_rms_eps); + LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff); + LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base); + LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale); + LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); + LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); + LLAMA_LOG_INFO("%s: model size = %.2f B\n", __func__, ml.n_elements*1e-9); + + // general kv + LLAMA_LOG_INFO("%s: general.name = %s\n", __func__, model.name.c_str()); + + // special tokens + if (vocab.special_bos_id != -1) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].text.c_str() ); } + if (vocab.special_eos_id != -1) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].text.c_str() ); } + if (vocab.special_unk_id != -1) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].text.c_str() ); } + if (vocab.special_sep_id != -1) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].text.c_str() ); } + if (vocab.special_pad_id != -1) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].text.c_str() ); } + if (vocab.linefeed_id != -1) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, vocab.linefeed_id, vocab.id_to_token[vocab.linefeed_id].text.c_str() ); } +} + +static void llm_load_tensors( + llama_model_loader & ml, + llama_model & model, int n_batch, int n_gpu_layers, int main_gpu, const float * tensor_split, const bool mul_mat_q, - float rope_freq_base, - float rope_freq_scale, bool low_vram, ggml_type memory_type, - bool use_mmap, bool use_mlock, - bool vocab_only, llama_progress_callback progress_callback, void * progress_callback_user_data) { model.t_start_us = ggml_time_us(); - std::unique_ptr ml(new llama_model_loader(fname, use_mmap)); - - model.n_gpu_layers = n_gpu_layers; - + auto & ctx = model.ctx; auto & hparams = model.hparams; - std::string general_name = "n/a"; - std::string general_arch = "n/a"; - - // read hparams - { - struct gguf_context * ctx = ml->ctx_gguf; - -#define GGUF_GET(dst, func, type, req, key) \ - { \ - const int kid = gguf_find_key(ctx, key); \ - if (kid >= 0) { \ - enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \ - if (ktype != (type)) { \ - throw std::runtime_error(format("key %s has wrong type: %s", key, gguf_type_name(ktype))); \ - } \ - (dst) = func(ctx, kid); \ - } else if (req) { \ - throw std::runtime_error(format("key not found in model: %s", key)); \ - } \ - } - - std::string tokenizer_name; - GGUF_GET(tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, "tokenizer.ggml.model"); - - if (tokenizer_name == "llama") { - vocab.type = LLAMA_VOCAB_TYPE_SPM; - } else if (tokenizer_name == "gpt2") { - vocab.type = LLAMA_VOCAB_TYPE_BPE; - } else { - LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str()); - LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__); - vocab.type = LLAMA_VOCAB_TYPE_SPM; - } - - // get hparams kv - GGUF_GET(hparams.n_vocab, gguf_get_arr_n, GGUF_TYPE_ARRAY, true, "tokenizer.ggml.tokens"); - GGUF_GET(hparams.n_ctx_train, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.context_length"); - GGUF_GET(hparams.n_embd, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.embedding_length"); - GGUF_GET(hparams.n_ff, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.feed_forward_length"); - GGUF_GET(hparams.n_head, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.attention.head_count"); - GGUF_GET(hparams.n_layer, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.block_count"); - GGUF_GET(hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.rope.dimension_count"); - GGUF_GET(hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, "llama.attention.layer_norm_rms_epsilon"); - - // n_head_kv is optional, default to n_head - hparams.n_head_kv = hparams.n_head; - GGUF_GET(hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "llama.attention.head_count_kv"); - - // TODO: manually setting rope scale should override this - // rope_freq_scale (inverse of the kv) is optional - float ropescale = 1.0f; - GGUF_GET(ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, "llama.rope.scale_linear"); - if (ropescale != 1.0f) { - rope_freq_scale = 1.0f/ropescale; - } - - // get general kv - GGUF_GET(general_name, gguf_get_val_str, GGUF_TYPE_STRING, false, "general.name"); - GGUF_GET(general_arch, gguf_get_val_str, GGUF_TYPE_STRING, false, "general.architecture"); - - // special tokens - GGUF_GET(vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.bos_token_id"); - GGUF_GET(vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.eos_token_id"); - GGUF_GET(vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.unknown_token_id"); - GGUF_GET(vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.separator_token_id"); - GGUF_GET(vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "tokenizer.ggml.padding_token_id"); - -#undef GGUF_GET - - switch (hparams.n_layer) { - case 26: model.type = e_model::MODEL_3B; break; - case 32: model.type = e_model::MODEL_7B; break; - case 40: model.type = e_model::MODEL_13B; break; - case 60: model.type = e_model::MODEL_30B; break; - case 80: model.type = e_model::MODEL_65B; break; - default: - { - if (hparams.n_layer < 32) { - model.type = e_model::MODEL_7B; - } - } break; - } - - model.ftype = ml->ftype; - - hparams.n_ctx = n_ctx; - - // LLaMAv2 - // TODO: probably not needed - { - const auto n_gqa = hparams.n_gqa(); - - if (model.type == e_model::MODEL_65B && n_gqa == 8) { - LLAMA_LOG_WARN("%s: assuming 70B model based on GQA == %d\n", __func__, n_gqa); - model.type = e_model::MODEL_70B; - } - } - - hparams.rope_freq_base = rope_freq_base; - hparams.rope_freq_scale = rope_freq_scale; - } - - // read vocab - { - struct gguf_context * ctx = ml->ctx_gguf; - - vocab.id_to_token.resize(hparams.n_vocab); - - const int token_idx = gguf_find_key(ctx, "tokenizer.ggml.tokens"); - if (token_idx == -1) { - throw std::runtime_error("cannot find tokenizer vocab in model file\n"); - } - - const int score_idx = gguf_find_key(ctx, "tokenizer.ggml.scores"); - if (score_idx == -1) { - throw std::runtime_error("cannot find tokenizer scores in model file\n"); - } - - const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx); - - const int toktype_idx = gguf_find_key(ctx, "tokenizer.ggml.token_type"); - if (toktype_idx == -1) { - throw std::runtime_error("cannot find token type list in GGUF file\n"); - } - - const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx); - - for (uint32_t i = 0; i < hparams.n_vocab; i++) { - std::string word = gguf_get_arr_str(ctx, token_idx, i); - - vocab.token_to_id[word] = i; - - auto & token_data = vocab.id_to_token[i]; - token_data.text = std::move(word); - token_data.score = scores[i]; - token_data.type = (llama_token_type) toktypes[i]; - - // determine the newline token: 0x0A == 10 == '\n' - if (token_data.text == "<0x0A>") { - vocab.linefeed_id = i; - } - } - } - - { - // hparams - LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml->fver)); - LLAMA_LOG_INFO("%s: arch = %s\n", __func__, general_arch.c_str()); - LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, vocab.type == LLAMA_VOCAB_TYPE_SPM ? "SPM" : "BPE"); // TODO: fix - LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab); - LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train); - LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, hparams.n_ctx); - LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd); - LLAMA_LOG_INFO("%s: n_head = %u\n", __func__, hparams.n_head); - LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv); - LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); - LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim - LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa()); - LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_rms_eps); - LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff); - LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base); - LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale); - LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); - LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); - LLAMA_LOG_INFO("%s: model size = %.2f B\n", __func__, ml->n_elements*1e-9); - - // general kv - LLAMA_LOG_INFO("%s: general.name = %s\n", __func__, general_name.c_str()); - - // special tokens - if (vocab.special_bos_id != -1) { LLAMA_LOG_INFO( "%s: BOS token = %d '%s'\n", __func__, vocab.special_bos_id, vocab.id_to_token[vocab.special_bos_id].text.c_str() ); } - if (vocab.special_eos_id != -1) { LLAMA_LOG_INFO( "%s: EOS token = %d '%s'\n", __func__, vocab.special_eos_id, vocab.id_to_token[vocab.special_eos_id].text.c_str() ); } - if (vocab.special_unk_id != -1) { LLAMA_LOG_INFO( "%s: UNK token = %d '%s'\n", __func__, vocab.special_unk_id, vocab.id_to_token[vocab.special_unk_id].text.c_str() ); } - if (vocab.special_sep_id != -1) { LLAMA_LOG_INFO( "%s: SEP token = %d '%s'\n", __func__, vocab.special_sep_id, vocab.id_to_token[vocab.special_sep_id].text.c_str() ); } - if (vocab.special_pad_id != -1) { LLAMA_LOG_INFO( "%s: PAD token = %d '%s'\n", __func__, vocab.special_pad_id, vocab.id_to_token[vocab.special_pad_id].text.c_str() ); } - if (vocab.linefeed_id != -1) { LLAMA_LOG_INFO( "%s: LF token = %d '%s'\n", __func__, vocab.linefeed_id, vocab.id_to_token[vocab.linefeed_id].text.c_str() ); } - } - - if (vocab_only) { - LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__); - return; - } - - auto & ctx = model.ctx; + model.n_gpu_layers = n_gpu_layers; size_t ctx_size; size_t mmapped_size; - ml->calc_sizes(ctx_size, mmapped_size); + ml.calc_sizes(ctx_size, mmapped_size); LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0); @@ -1480,7 +1805,7 @@ static void llama_model_load_internal( struct ggml_init_params params = { /*.mem_size =*/ model.buf.size, /*.mem_buffer =*/ model.buf.data, - /*.no_alloc =*/ ml->use_mmap, + /*.no_alloc =*/ ml.use_mmap, }; model.ctx = ggml_init(params); @@ -1509,75 +1834,146 @@ static void llama_model_load_internal( // prepare memory for the weights size_t vram_weights = 0; { - const uint32_t n_embd = hparams.n_embd; - const uint32_t n_embd_gqa = hparams.n_embd_gqa(); - const uint32_t n_layer = hparams.n_layer; - const uint32_t n_vocab = hparams.n_vocab; + const int64_t n_embd = hparams.n_embd; + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + const int64_t n_layer = hparams.n_layer; + const int64_t n_vocab = hparams.n_vocab; - model.tok_embeddings = ml->create_tensor(ctx, TN_TOKEN_EMBD, {n_embd, n_vocab}, GGML_BACKEND_CPU); + const auto tn = LLM_TN(model.arch); - // "output" tensor - { - ggml_backend backend_norm; - ggml_backend backend_output; - if (n_gpu_layers > int(n_layer)) { // NOLINT - // norm is not performance relevant on its own but keeping it in VRAM reduces data copying - // on Windows however this is detrimental unless everything is on the GPU + switch (model.arch) { + case LLM_ARCH_LLAMA: + { + model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output + { + ggml_backend backend_norm; + ggml_backend backend_output; + + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; #else - backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; - } else { - backend_norm = GGML_BACKEND_CPU; - backend_output = GGML_BACKEND_CPU; - } + backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } - model.norm = ml->create_tensor(ctx, TN_OUTPUT_NORM, {n_embd}, backend_norm); - model.output = ml->create_tensor(ctx, TN_OUTPUT, {n_embd, n_vocab}, backend_output); - if (backend_norm == GGML_BACKEND_GPU) { - vram_weights += ggml_nbytes(model.norm); - } - if (backend_output == GGML_BACKEND_GPU_SPLIT) { - vram_weights += ggml_nbytes(model.output); - } - } + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); - const uint32_t n_ff = hparams.n_ff; + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } - const int i_gpu_start = n_layer - n_gpu_layers; + const uint32_t n_ff = hparams.n_ff; - model.layers.resize(n_layer); - for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const int i_gpu_start = n_layer - n_gpu_layers; - auto & layer = model.layers[i]; - layer.attention_norm = ml->create_tensor(ctx, format(TN_ATTN_NORM, i), {n_embd}, backend); + model.layers.resize(n_layer); - layer.wq = ml->create_tensor(ctx, format(TN_ATTN_Q, i), {n_embd, n_embd}, backend_split); - layer.wk = ml->create_tensor(ctx, format(TN_ATTN_K, i), {n_embd, n_embd_gqa}, backend_split); - layer.wv = ml->create_tensor(ctx, format(TN_ATTN_V, i), {n_embd, n_embd_gqa}, backend_split); - layer.wo = ml->create_tensor(ctx, format(TN_ATTN_OUTPUT, i), {n_embd, n_embd}, backend_split); + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT + const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT - layer.ffn_norm = ml->create_tensor(ctx, format(TN_FFN_NORM, i), {n_embd}, backend); + auto & layer = model.layers[i]; - layer.w1 = ml->create_tensor(ctx, format(TN_FFN_GATE, i), {n_embd, n_ff}, backend_split); - layer.w2 = ml->create_tensor(ctx, format(TN_FFN_DOWN, i), { n_ff, n_embd}, backend_split); - layer.w3 = ml->create_tensor(ctx, format(TN_FFN_UP, i), {n_embd, n_ff}, backend_split); + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - if (backend == GGML_BACKEND_GPU) { - vram_weights += - ggml_nbytes(layer.attention_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + - ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.w1) + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); - } - } + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + + layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.w1) + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + } + } + } break; + case LLM_ARCH_FALCON: + { + // TODO: CPU-only for now + + model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output + { + ggml_backend backend_norm; + ggml_backend backend_output; + + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU +#ifndef _WIN32 + backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; +#else + backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; +#endif // _WIN32 + + backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } + + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT + const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + + if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) { + layer.attn_norm_2 = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, backend); + layer.attn_norm_2_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, backend); + } + + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + + layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + } + } break; + default: + throw std::runtime_error("unknown architecture"); + }; } - ml->done_getting_tensors(); + ml.done_getting_tensors(); // print memory requirements { @@ -1589,8 +1985,7 @@ static void llama_model_load_internal( mmapped_size - vram_weights; // weights in VRAM not in memory // this is the memory required by one llama_state - const size_t mem_required_state = - scale*hparams.kv_size(); + const size_t mem_required_state = scale*hparams.kv_size(); LLAMA_LOG_INFO("%s: mem required = %7.2f MB (+ %7.2f MB per state)\n", __func__, mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0); @@ -1640,8 +2035,8 @@ static void llama_model_load_internal( } // populate `tensors_by_name` - for (int i = 0; i < ml->n_tensors; ++i) { - struct ggml_tensor * cur = ggml_get_tensor(ctx, ml->get_tensor_name(i)); + for (int i = 0; i < ml.n_tensors; ++i) { + struct ggml_tensor * cur = ggml_get_tensor(ctx, ml.get_tensor_name(i)); model.tensors_by_name.emplace_back(ggml_get_name(cur), cur); } @@ -1652,13 +2047,13 @@ static void llama_model_load_internal( } #endif - ml->load_all_data(ctx, progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL); + ml.load_all_data(ctx, progress_callback, progress_callback_user_data, use_mlock ? &model.mlock_mmap : NULL); if (progress_callback) { progress_callback(1.0f, progress_callback_user_data); } - model.mapping = std::move(ml->mapping); + model.mapping = std::move(ml.mapping); // loading time will be recalculate after the first eval, so // we take page faults deferred by mmap() into consideration @@ -1668,7 +2063,6 @@ static void llama_model_load_internal( static bool llama_model_load( const std::string & fname, llama_model & model, - llama_vocab & vocab, int n_ctx, int n_batch, int n_gpu_layers, @@ -1685,17 +2079,36 @@ static bool llama_model_load( llama_progress_callback progress_callback, void *progress_callback_user_data) { try { - llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gpu_layers, - main_gpu, tensor_split, mul_mat_q, rope_freq_base, rope_freq_scale, low_vram, memory_type, - use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data); - return true; + std::unique_ptr ml(new llama_model_loader(fname, use_mmap)); + + llm_load_arch (*ml, model); + llm_load_hparams(*ml, model, n_ctx, rope_freq_base, rope_freq_scale); + llm_load_vocab (*ml, model); + + llm_load_print_meta(*ml, model); + + if (model.hparams.n_vocab != model.vocab.id_to_token.size()) { + throw std::runtime_error("vocab size mismatch"); + } + + if (vocab_only) { + LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__); + return true; + } + + llm_load_tensors( + *ml, model, n_batch, n_gpu_layers, + main_gpu, tensor_split, mul_mat_q, low_vram, memory_type, + use_mlock, progress_callback, progress_callback_user_data); } catch (const std::exception & err) { LLAMA_LOG_ERROR("error loading model: %s\n", err.what()); return false; } + + return true; } -static struct ggml_cgraph * llama_build_graph( +static struct ggml_cgraph * llm_build_llama( llama_context & lctx, const llama_token * tokens, const float * embd, @@ -1729,8 +2142,7 @@ static struct ggml_cgraph * llama_build_graph( const int n_gpu_layers = model.n_gpu_layers; - auto & mem_per_token = lctx.mem_per_token; - auto & buf_compute = lctx.buf_compute; + auto & buf_compute = lctx.buf_compute; struct ggml_init_params params = { /*.mem_size =*/ buf_compute.size, @@ -1820,8 +2232,8 @@ static struct ggml_cgraph * llama_build_graph( offload_func(cur); ggml_set_name(cur, "rms_norm_0"); - // cur = cur*attention_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].attention_norm); + // cur = cur*attn_norm(broadcasted) + cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); offload_func(cur); ggml_set_name(cur, "attention_norm_0"); } @@ -1872,10 +2284,7 @@ static struct ggml_cgraph * llama_build_graph( ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); } - struct ggml_tensor * Q = - ggml_permute(ctx0, - Qcur, - 0, 2, 1, 3); + struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); offload_func_kq(Q); ggml_set_name(Q, "Q"); @@ -2005,14 +2414,16 @@ static struct ggml_cgraph * llama_build_graph( inpL = cur; } + cur = inpL; + // norm { - cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps); + cur = ggml_rms_norm(ctx0, cur, norm_rms_eps); offload_func_nr(cur); ggml_set_name(cur, "rms_norm_2"); // cur = cur*norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.norm); + cur = ggml_mul(ctx0, cur, model.output_norm); // offload_func_nr(cur); // TODO CPU + GPU mirrored backend ggml_set_name(cur, "result_norm"); } @@ -2021,20 +2432,349 @@ static struct ggml_cgraph * llama_build_graph( cur = ggml_mul_mat(ctx0, model.output, cur); ggml_set_name(cur, "result_output"); - // logits -> probs - //cur = ggml_soft_max_inplace(ctx0, cur); - ggml_build_forward_expand(gf, cur); - if (mem_per_token == 0) { - mem_per_token = ggml_used_mem(ctx0)/N; - } - ggml_free(ctx0); return gf; } +static struct ggml_cgraph * llm_build_falcon( + llama_context & lctx, + const llama_token * tokens, + const float * embd, + int n_tokens, + int n_past) { + + GGML_ASSERT((!tokens && embd) || (tokens && !embd)); // NOLINT + + const int N = n_tokens; + + const auto & model = lctx.model; + const auto & hparams = model.hparams; + + const auto & kv_self = lctx.kv_self; + + GGML_ASSERT(!!kv_self.ctx); + + const int64_t n_embd = hparams.n_embd; + const int64_t n_layer = hparams.n_layer; + const int64_t n_ctx = hparams.n_ctx; + const int64_t n_head = hparams.n_head; + const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_embd_head = hparams.n_embd_head(); + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_rot); + + const float freq_base = hparams.rope_freq_base; + const float freq_scale = hparams.rope_freq_scale; + const float norm_eps = hparams.f_norm_eps; + + const int n_gpu_layers = model.n_gpu_layers; + + auto & buf_compute = lctx.buf_compute; + + struct ggml_init_params params = { + /*.mem_size =*/ buf_compute.size, + /*.mem_buffer =*/ buf_compute.data, + /*.no_alloc =*/ false, + }; + + params.no_alloc = true; + + struct ggml_context * ctx0 = ggml_init(params); + + ggml_cgraph * gf = ggml_new_graph(ctx0); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + if (tokens) { + struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + + ggml_allocr_alloc(lctx.alloc, inp_tokens); + if (!ggml_allocr_is_measure(lctx.alloc)) { + memcpy(inp_tokens->data, tokens, N*ggml_element_size(inp_tokens)); + } + ggml_set_name(inp_tokens, "inp_tokens"); + + inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); + } else { +#ifdef GGML_USE_MPI + GGML_ASSERT(false && "not implemented"); +#endif + + inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N); + + ggml_allocr_alloc(lctx.alloc, inpL); + if (!ggml_allocr_is_measure(lctx.alloc)) { + memcpy(inpL->data, embd, N * n_embd * ggml_element_size(inpL)); + } + } + + const int i_gpu_start = n_layer - n_gpu_layers; + (void) i_gpu_start; + + // offload functions set the tensor output backend to GPU + // tensors are GPU-accelerated if any input or the output has been offloaded + // + // with the low VRAM option VRAM scratch is disabled in llama_load_model_internal + // in that case ggml_cuda_assign_buffers has no effect + offload_func_t offload_func_nr = llama_nop; // nr = non-repeating + offload_func_t offload_func_kq = llama_nop; + offload_func_t offload_func_v = llama_nop; + +#ifdef GGML_USE_CUBLAS + if (n_gpu_layers > n_layer) { + offload_func_nr = ggml_cuda_assign_buffers_no_alloc; + } + if (n_gpu_layers > n_layer + 1) { + offload_func_v = ggml_cuda_assign_buffers_no_alloc; + } + if (n_gpu_layers > n_layer + 2) { + offload_func_kq = ggml_cuda_assign_buffers_no_alloc; + } +#endif // GGML_USE_CUBLAS + + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + ggml_allocr_alloc(lctx.alloc, KQ_scale); + if (!ggml_allocr_is_measure(lctx.alloc)) { + ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); + } + ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * attn_norm; + + offload_func_t offload_func = llama_nop; + +#ifdef GGML_USE_CUBLAS + if (il >= i_gpu_start) { + offload_func = ggml_cuda_assign_buffers_no_alloc; + } +#endif // GGML_USE_CUBLAS + + // self-attention + // TODO: refactor into common function (shared with LLaMA) + { + attn_norm = ggml_norm(ctx0, inpL, norm_eps); + offload_func(attn_norm); + + attn_norm = ggml_add(ctx0, + ggml_mul(ctx0, attn_norm, model.layers[il].attn_norm), + model.layers[il].attn_norm_b); + offload_func(attn_norm->src[0]); + offload_func(attn_norm); + + if (model.layers[il].attn_norm_2) { // Falcon-40B + cur = ggml_norm(ctx0, inpL, norm_eps); + offload_func(cur); + + cur = ggml_add(ctx0, + ggml_mul(ctx0, cur, model.layers[il].attn_norm_2), + model.layers[il].attn_norm_2_b); + offload_func(cur->src[0]); + offload_func(cur); + } else { // Falcon 7B + cur = attn_norm; + } + + // compute QKV + + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + offload_func_kq(cur); + + // Note that the strides for Kcur, Vcur are set up so that the + // resulting views are misaligned with the tensor's storage + // (by applying the K/V offset we shift the tensor's original + // view to stick out behind the viewed QKV tensor's allocated + // memory, so to say). This is ok because no actual accesses + // happen to that out-of-range memory, but it can require some + // trickery when trying to accurately dump these views for + // debugging. + + const size_t wsize = ggml_type_size(cur->type); + + struct ggml_tensor * tmpq = ggml_view_3d( + ctx0, cur, n_embd_head, n_head, N, + wsize * n_embd_head, + wsize * n_embd_head * (n_head + 2 * n_head_kv), + 0); + offload_func_kq(tmpq); + + struct ggml_tensor * tmpk = ggml_view_3d( + ctx0, cur, n_embd_head, n_head_kv, N, + wsize * n_embd_head, + wsize * n_embd_head * (n_head + 2 * n_head_kv), + wsize * n_embd_head * n_head); + offload_func_kq(tmpk); + + struct ggml_tensor * tmpv = ggml_view_3d( + ctx0, cur, n_embd_head, n_head_kv, N, + wsize * n_embd_head, + wsize * n_embd_head * (n_head + 2 * n_head_kv), + wsize * n_embd_head * (n_head + n_head_kv)); + offload_func_v(tmpv); + + // using mode = 2 for neox mode + struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, tmpq, n_past, n_embd_head, 2, 0, freq_base, freq_scale); + offload_func_kq(Qcur); + struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, tmpk, n_past, n_embd_head, 2, 0, freq_base, freq_scale); + offload_func_kq(Kcur); + + { + struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, N)); + offload_func_v(Vcur); + offload_func_v(Vcur->src[0]->src[0]); + ggml_set_name(Vcur, "Vcur"); + + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + n_past)); + offload_func_kq(k); + ggml_set_name(k, "k"); + + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd_gqa, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + n_past*ggml_element_size(kv_self.v)); + offload_func_v(v); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); + ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); + } + + struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + offload_func_kq(Q); + ggml_set_name(Q, "Q"); + + struct ggml_tensor * K = + ggml_view_3d(ctx0, kv_self.k, + n_embd_head, n_past + N, n_head_kv, + ggml_element_size(kv_self.k)*n_embd_gqa, + ggml_element_size(kv_self.k)*n_embd_head, + ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); + offload_func_kq(K); + ggml_set_name(K, "K"); + + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + offload_func_kq(KQ); + ggml_set_name(KQ, "KQ"); + + struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); + offload_func_kq(KQ_scaled); + ggml_set_name(KQ_scaled, "KQ_scaled"); + + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); + offload_func_kq(KQ_masked); + ggml_set_name(KQ_masked, "KQ_masked"); + + struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); + offload_func_v(KQ_soft_max); + ggml_set_name(KQ_soft_max, "KQ_soft_max"); + + struct ggml_tensor * V = + ggml_view_3d(ctx0, kv_self.v, + n_past + N, n_embd_head, n_head_kv, + ggml_element_size(kv_self.v)*n_ctx, + ggml_element_size(kv_self.v)*n_ctx*n_embd_head, + ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); + offload_func_v(V); + ggml_set_name(V, "V"); + + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); + offload_func_v(KQV); + ggml_set_name(KQV, "KQV"); + + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + offload_func_v(KQV_merged); + ggml_set_name(KQV_merged, "KQV_merged"); + + cur = ggml_cpy(ctx0, KQV_merged, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + offload_func_v(cur); + ggml_set_name(cur, "KQV_merged_contiguous"); + + cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); + offload_func(cur); + ggml_set_name(cur, "result_wo"); + } + + struct ggml_tensor * attn_out = cur; + + // feed forward + { + struct ggml_tensor * inpFF = attn_norm; + + cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF); + + // TODO: this is temporary needed to introduce artificial dependency between FF and ATTN + // adding this, because there seems to be a bug in the Metal concurrency optimization + // without this line, the results are non-deterministic and wrong + cur->src[2] = attn_out; + offload_func(cur); + + cur = ggml_gelu(ctx0, cur); + offload_func(cur); + cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); + offload_func(cur); + } + + cur = ggml_add(ctx0, cur, attn_out); + offload_func(cur); + cur = ggml_add(ctx0, cur, inpL); + offload_func(cur); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + // norm + { + cur = ggml_norm(ctx0, cur, norm_eps); + offload_func_nr(cur); + + cur = ggml_add(ctx0, + ggml_mul(ctx0, cur, model.output_norm), + model.output_norm_b); + ggml_set_name(cur, "result_norm"); + } + + cur = ggml_mul_mat(ctx0, model.output, cur); + ggml_set_name(cur, "result_output"); + + ggml_build_forward_expand(gf, cur); + + ggml_free(ctx0); + + return gf; +} + +static struct ggml_cgraph * llama_build_graph( + llama_context & lctx, + const llama_token * tokens, + const float * embd, + int n_tokens, + int n_past) { + const auto & model = lctx.model; + + struct ggml_cgraph * result = NULL; + + switch (model.arch) { + case LLM_ARCH_LLAMA: + { + result = llm_build_llama(lctx, tokens, embd, n_tokens, n_past); + } break; + case LLM_ARCH_FALCON: + { + result = llm_build_falcon(lctx, tokens, embd, n_tokens, n_past); + } break; + default: + GGML_ASSERT(false); + }; + + return result; +} + // evaluate the transformer // // - lctx: llama context @@ -2077,8 +2817,8 @@ static bool llama_eval_internal( GGML_ASSERT(!!kv_self.ctx); - const int64_t n_embd = hparams.n_embd; - const int64_t n_vocab = hparams.n_vocab; + const int64_t n_embd = hparams.n_embd; + const int64_t n_vocab = hparams.n_vocab; ggml_allocr_reset(lctx.alloc); @@ -2108,11 +2848,11 @@ static bool llama_eval_internal( // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads; - struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1]; + struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1]; struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2]; - GGML_ASSERT(strcmp(res->name, "result_output") == 0); - GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0); + GGML_ASSERT(strcmp(res->name, "result_output") == 0); + GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0); #if GGML_USE_MPI const int64_t n_layer = hparams.n_layer; @@ -2271,13 +3011,7 @@ static std::string llama_unescape_whitespace(const std::string& word) { return word; } -static size_t utf8_len(char src) { - const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 }; - uint8_t highbits = static_cast(src) >> 4; - return lookup[highbits]; -} - -struct llama_sp_symbol { +struct llm_symbol { using index = int; index prev; index next; @@ -2285,33 +3019,35 @@ struct llama_sp_symbol { size_t n; }; -static_assert(std::is_trivially_copyable::value, "llama_sp_symbol is not trivially copyable"); +static_assert(std::is_trivially_copyable::value, "llm_symbol is not trivially copyable"); -struct llama_sp_bigram { +// SPM tokenizer +// original implementation: +// https://github.com/ggerganov/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4 + +struct llm_bigram_spm { struct comparator { - bool operator()(llama_sp_bigram & l, llama_sp_bigram & r) { + bool operator()(llm_bigram_spm & l, llm_bigram_spm & r) { return (l.score < r.score) || (l.score == r.score && l.left > r.left); } }; - using queue_storage = std::vector; - using queue = std::priority_queue; - llama_sp_symbol::index left; - llama_sp_symbol::index right; + using queue_storage = std::vector; + using queue = std::priority_queue; + llm_symbol::index left; + llm_symbol::index right; float score; size_t size; }; -// original implementation: -// https://github.com/ggerganov/llama.cpp/commit/074bea2eb1f1349a0118239c4152914aecaa1be4 -struct llama_tokenizer { - llama_tokenizer(const llama_vocab & vocab): vocab_(vocab) {} +struct llm_tokenizer_spm { + llm_tokenizer_spm(const llama_vocab & vocab): vocab(vocab) {} void tokenize(const std::string & text, std::vector & output) { // split string into utf8 chars int index = 0; size_t offs = 0; while (offs < text.size()) { - llama_sp_symbol sym; + llm_symbol sym; size_t len = utf8_len(text[offs]); GGML_ASSERT(offs + len <= text.size()); sym.text = text.c_str() + offs; @@ -2320,21 +3056,21 @@ struct llama_tokenizer { sym.prev = index - 1; sym.next = offs == text.size() ? -1 : index + 1; index++; - symbols_.emplace_back(sym); + symbols.emplace_back(sym); } // seed the work queue with all possible 2-character tokens. - for (size_t i = 1; i < symbols_.size(); ++i) { + for (size_t i = 1; i < symbols.size(); ++i) { try_add_bigram(i - 1, i); } // keep substituting the highest frequency pairs for as long as we can. - while (!work_queue_.empty()) { - auto bigram = work_queue_.top(); - work_queue_.pop(); + while (!work_queue.empty()) { + auto bigram = work_queue.top(); + work_queue.pop(); - auto & left_sym = symbols_[bigram.left]; - auto & right_sym = symbols_[bigram.right]; + auto & left_sym = symbols[bigram.left]; + auto & right_sym = symbols[bigram.right]; // if one of the symbols already got merged, skip it. if (left_sym.n == 0 || right_sym.n == 0 || @@ -2351,7 +3087,7 @@ struct llama_tokenizer { // remove the right sym from the chain left_sym.next = right_sym.next; if (right_sym.next >= 0) { - symbols_[right_sym.next].prev = bigram.left; + symbols[right_sym.next].prev = bigram.left; } // find more substitutions @@ -2359,19 +3095,19 @@ struct llama_tokenizer { try_add_bigram(bigram.left, left_sym.next); } - for (int i = 0; i != -1; i = symbols_[i].next) { - auto & symbol = symbols_[i]; + for (int i = 0; i != -1; i = symbols[i].next) { + auto & symbol = symbols[i]; resegment(symbol, output); } } private: - void resegment(llama_sp_symbol &symbol, std::vector &output) { + void resegment(llm_symbol & symbol, std::vector & output) { auto text = std::string(symbol.text, symbol.n); - auto token = vocab_.token_to_id.find(text); + auto token = vocab.token_to_id.find(text); // Do we need to support is_unused? - if (token != vocab_.token_to_id.end()) { + if (token != vocab.token_to_id.end()) { output.push_back((*token).second); return; } @@ -2381,14 +3117,14 @@ private: if (p == rev_merge.end()) { // output any symbols that did not form tokens as bytes. for (int j = 0; j < (int)symbol.n; ++j) { - llama_vocab::id token_id = llama_byte_to_token(vocab_, symbol.text[j]); + llama_vocab::id token_id = llama_byte_to_token(vocab, symbol.text[j]); output.push_back(token_id); } return; } - resegment(symbols_[p->second.first], output); - resegment(symbols_[p->second.second], output); + resegment(symbols[p->second.first], output); + resegment(symbols[p->second.second], output); } void try_add_bigram(int left, int right) { @@ -2396,56 +3132,261 @@ private: return; } - const std::string text = std::string(symbols_[left].text, symbols_[left].n + symbols_[right].n); - auto token = vocab_.token_to_id.find(text); + const std::string text = std::string(symbols[left].text, symbols[left].n + symbols[right].n); + auto token = vocab.token_to_id.find(text); - if (token == vocab_.token_to_id.end()) { + if (token == vocab.token_to_id.end()) { return; } - if (static_cast((*token).second) >= vocab_.id_to_token.size()) { + if (static_cast((*token).second) >= vocab.id_to_token.size()) { return; } - const auto &tok_data = vocab_.id_to_token[(*token).second]; + const auto & tok_data = vocab.id_to_token[(*token).second]; - llama_sp_bigram bigram; - bigram.left = left; + llm_bigram_spm bigram; + bigram.left = left; bigram.right = right; bigram.score = tok_data.score; - bigram.size = text.size(); - work_queue_.push(bigram); + bigram.size = text.size(); + + work_queue.push(bigram); // Do we need to support is_unused? rev_merge[text] = std::make_pair(left, right); } - const llama_vocab & vocab_; - std::vector symbols_; - llama_sp_bigram::queue work_queue_; - std::map > rev_merge; + const llama_vocab & vocab; + + std::vector symbols; + llm_bigram_spm::queue work_queue; + + std::map> rev_merge; +}; + +// BPE tokenizer +// adapted from https://github.com/cmp-nct/ggllm.cpp [MIT License] +// tried to simplify unicode stuff, so most likely does not work 100% correctly! + +// TODO: there are a lot of common parts between spm and bpe tokenizers, should be refactored and reused + +struct llm_bigram_bpe { + struct comparator { + bool operator()(llm_bigram_bpe & l, llm_bigram_bpe & r) { + return l.rank > r.rank || (l.rank == r.rank && l.left > r.left); + } + }; + + using queue_storage = std::vector; + using queue = std::priority_queue; + llm_symbol::index left; + llm_symbol::index right; + std::string text; + int rank; + size_t size; +}; + +struct llm_tokenizer_bpe { + llm_tokenizer_bpe(const llama_vocab & vocab, bool g2ws): vocab(vocab) { flag_g2ws = g2ws; } + + void tokenize(const std::string & text, std::vector & output) { + int final_prev_index = -1; + auto word_collection = bpe_gpt2_preprocess(text); + + symbols_final.clear(); + + for (auto & word : word_collection) { + work_queue = llm_bigram_bpe::queue(); + symbols.clear(); + + int index = 0; + size_t offset = 0; + + while (offset < word.size()) { + llm_symbol sym; + size_t char_len = std::min(word.size() - offset, (size_t) ::utf8_len(word[offset])); + sym.text = word.c_str() + offset; + sym.n = 1; + sym.n = char_len; + offset += sym.n; + sym.prev = index - 1; + sym.next = offset == word.size() ? -1 : index + 1; + index++; + symbols.emplace_back(sym); + } + for (size_t i = 1; i < symbols.size(); ++i) { + add_new_bigram(i - 1, i); + } + + // build token(s) + while (!work_queue.empty()) { + auto bigram = work_queue.top(); + work_queue.pop(); + + auto & left_symbol = symbols[bigram.left]; + auto & right_symbol = symbols[bigram.right]; + + if (left_symbol.n == 0 || right_symbol.n == 0) { + continue; + } + std::string left_token = std::string(left_symbol.text, left_symbol.n); + std::string right_token = std::string(right_symbol.text, right_symbol.n); + if (left_token + right_token != bigram.text) { + continue; // Skip this bigram if it's outdated + } + + // merge the right sym into the left one + left_symbol.n += right_symbol.n; + right_symbol.n = 0; + + // remove the right sym from the chain + left_symbol.next = right_symbol.next; + if (right_symbol.next >= 0) { + symbols[right_symbol.next].prev = bigram.left; + } + + add_new_bigram(left_symbol.prev, bigram.left); // left side of current symbol + add_new_bigram(bigram.left, left_symbol.next); // right side of current symbol + } + + // add the fnished tokens to the final list keeping correct order for next and prev + for (auto & sym : symbols) { + if (sym.n > 0) { + sym.prev = final_prev_index; + sym.next = -1; + if (final_prev_index != -1) { + symbols_final[final_prev_index].next = symbols_final.size(); + } + symbols_final.emplace_back(sym); + final_prev_index = symbols_final.size() - 1; + } + } + } + + symbols = symbols_final; + + if (!symbols.empty()) { + for (int i = 0; i != -1; i = symbols[i].next) { + auto & symbol = symbols[i]; + if (symbol.n == 0) { + continue; + } + + const std::string str = std::string(symbol.text, symbol.n); + const auto token = vocab.token_to_id.find(str); + + if (token == vocab.token_to_id.end()) { + for (auto j = str.begin(); j != str.end(); ++j) { + std::string byte_str(1, *j); + auto token_multibyte = vocab.token_to_id.find(byte_str); + if (token_multibyte == vocab.token_to_id.end()) { + fprintf(stderr,"ERROR: byte not found in vocab: '%s'\n", byte_str.c_str()); + } + output.push_back((*token_multibyte).second); + } + } else { + output.push_back((*token).second); + } + } + } + } + +private: + void add_new_bigram(int left, int right) { + if (left == -1 || right == -1) { + return; + } + + std::string left_token = std::string(symbols[left].text, symbols[left].n); + std::string right_token = std::string(symbols[right].text, symbols[right].n); + + int rank_found = -1; + + rank_found = vocab.find_bpe_rank(left_token, right_token); + + if (rank_found < 0) { + return; + } + + llm_bigram_bpe bigram; + + bigram.left = left; + bigram.right = right; + bigram.text = left_token + right_token; + bigram.size = left_token.size() + right_token.size(); + bigram.rank = rank_found; + + work_queue.push(bigram); + } + + // probably not 100% correct + // TODO: this is quite slow - how to make it more efficient? + static std::vector bpe_gpt2_preprocess(std::string text) { + std::vector words; + + // ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53 + const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"; + const std::regex re(pattern); + std::smatch m; + + while (std::regex_search(text, m, re)) { + for (auto x : m) { + words.push_back(x); + } + text = m.suffix(); + } + + return words; + } + + bool flag_g2ws = false; + + const llama_vocab & vocab; + + std::vector symbols; + std::vector symbols_final; + + llm_bigram_bpe::queue work_queue; }; static std::vector llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos, bool escape) { - llama_tokenizer tokenizer(vocab); std::vector output; if (raw_text.empty()) { return output; } - if (bos) { - output.push_back(vocab.special_bos_id); - } + switch (vocab.type) { + case LLAMA_VOCAB_TYPE_SPM: + { + llm_tokenizer_spm tokenizer(vocab); - std::string text; - if (escape) { - text = llama_escape_whitespace(raw_text); - } else { - text = raw_text; - } + if (bos) { + output.push_back(vocab.special_bos_id); + } + + std::string text; + if (escape) { + text = llama_escape_whitespace(raw_text); + } else { + text = raw_text; + } + + tokenizer.tokenize(text, output); + } break; + case LLAMA_VOCAB_TYPE_BPE: + { + llm_tokenizer_bpe tokenizer(vocab, escape); + + if (bos && vocab.special_bos_id != -1) { + output.push_back(vocab.special_bos_id); + } + + tokenizer.tokenize(raw_text, output); + } break; + }; - tokenizer.tokenize(text, output); return output; } @@ -3449,13 +4390,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s nthread = std::thread::hardware_concurrency(); } - std::unique_ptr model_loader(new llama_model_loader(fname_inp, /*use_mmap*/ false)); + std::unique_ptr ml(new llama_model_loader(fname_inp, /*use_mmap*/ false)); const size_t align = GGUF_DEFAULT_ALIGNMENT; struct gguf_context * ctx_out = gguf_init_empty(); // copy the KV pairs from the input file - gguf_set_kv (ctx_out, model_loader->ctx_gguf); + gguf_set_kv (ctx_out, ml->ctx_gguf); gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION); gguf_set_val_u32(ctx_out, "general.file_type", ftype); @@ -3463,8 +4404,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s int n_attention_wv = 0; int n_feed_forward_w2 = 0; - for (int i = 0; i < model_loader->n_tensors; ++i) { - struct ggml_tensor * meta = model_loader->get_tensor_meta(i); + for (int i = 0; i < ml->n_tensors; ++i) { + struct ggml_tensor * meta = ml->get_tensor_meta(i); const std::string name = ggml_get_name(meta); @@ -3498,8 +4439,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s std::vector work; // populate the original tensors so we get an initial meta data - for (int i = 0; i < model_loader->n_tensors; ++i) { - struct ggml_tensor * meta = model_loader->get_tensor_meta(i); + for (int i = 0; i < ml->n_tensors; ++i) { + struct ggml_tensor * meta = ml->get_tensor_meta(i); gguf_add_tensor(ctx_out, meta); } @@ -3512,17 +4453,17 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // placeholder for the meta data ::zeros(fout, meta_size); - for (int i = 0; i < model_loader->n_tensors; ++i) { - struct ggml_tensor * tensor = model_loader->get_tensor_meta(i); + for (int i = 0; i < ml->n_tensors; ++i) { + struct ggml_tensor * tensor = ml->get_tensor_meta(i); const std::string name = ggml_get_name(tensor); read_data.resize(ggml_nbytes(tensor)); tensor->data = read_data.data(); - model_loader->load_data_for(tensor); + ml->load_data_for(tensor); LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ", - ++idx, model_loader->n_tensors, + ++idx, ml->n_tensors, ggml_get_name(tensor), llama_format_tensor_shape(tensor).c_str(), ggml_type_name(tensor->type)); @@ -3548,7 +4489,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s new_type = quantized_type; #ifdef GGML_USE_K_QUANTS // TODO: avoid hardcoded tensor names - use the TN_* constants - if (name == TN_OUTPUT) { + const auto tn = LLM_TN(ml->get_arch()); + + if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K == 0 && ny % QK_K == 0) { @@ -3600,10 +4543,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } } if (convert_incompatible_tensor) { - if (name == TN_OUTPUT) { + if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing. LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n"); - } else if (name == TN_TOKEN_EMBD) { + } else if (name == tn(LLM_TENSOR_TOKEN_EMBD, "weight")) { new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing. LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n"); } else { @@ -3785,28 +4728,28 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const } // load base model - std::unique_ptr model_loader; + std::unique_ptr ml; ggml_context * base_ctx = NULL; std::vector base_buf; if (path_base_model) { LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model); - model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true)); + ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true)); size_t ctx_size; size_t mmapped_size; - model_loader->calc_sizes(ctx_size, mmapped_size); + ml->calc_sizes(ctx_size, mmapped_size); base_buf.resize(ctx_size); ggml_init_params base_params; base_params.mem_size = base_buf.size(); base_params.mem_buffer = base_buf.data(); - base_params.no_alloc = model_loader->use_mmap; + base_params.no_alloc = ml->use_mmap; base_ctx = ggml_init(base_params); // maybe this should in llama_model_loader - if (model_loader->use_mmap) { - model_loader->mapping.reset(new llama_mmap(&model_loader->file, /* prefetch */ 0, ggml_is_numa())); + if (ml->use_mmap) { + ml->mapping.reset(new llama_mmap(&ml->file, /* prefetch */ 0, ggml_is_numa())); } } @@ -3910,18 +4853,19 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const #endif // GGML_USE_CUBLAS ggml_tensor * base_t; - if (model_loader) { - struct gguf_context * ctx_gguf = model_loader->ctx_gguf; + if (ml) { + struct gguf_context * ctx_gguf = ml->ctx_gguf; // load from base model if (gguf_find_tensor(ctx_gguf, base_name.c_str()) < 0) { + // TODO: throw LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str()); return 1; } // TODO: not tested!! maybe not working! - base_t = model_loader->create_tensor(base_ctx, base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] }, GGML_BACKEND_CPU); - model_loader->load_data_for(base_t); + base_t = ml->create_tensor(base_ctx, base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] }, GGML_BACKEND_CPU); + ml->load_data_for(base_t); } else { base_t = dest_t; } @@ -4096,7 +5040,23 @@ struct llama_model * llama_load_model_from_file( ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32; - if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gpu_layers, + unsigned cur_percentage = 0; + if (params.progress_callback == NULL) { + params.progress_callback_user_data = &cur_percentage; + params.progress_callback = [](float progress, void * ctx) { + unsigned * cur_percentage_p = (unsigned *) ctx; + unsigned percentage = (unsigned) (100 * progress); + while (percentage > *cur_percentage_p) { + *cur_percentage_p = percentage; + LLAMA_LOG_INFO("."); + if (percentage >= 100) { + LLAMA_LOG_INFO("\n"); + } + } + }; + } + + if (!llama_model_load(path_model, *model, params.n_ctx, params.n_batch, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base, params.rope_freq_scale, params.low_vram, memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback, params.progress_callback_user_data)) { @@ -4126,22 +5086,6 @@ struct llama_context * llama_new_context_with_model( params.seed = time(NULL); } - unsigned cur_percentage = 0; - if (params.progress_callback == NULL) { - params.progress_callback_user_data = &cur_percentage; - params.progress_callback = [](float progress, void * ctx) { - unsigned * cur_percentage_p = (unsigned *) ctx; - unsigned percentage = (unsigned) (100 * progress); - while (percentage > *cur_percentage_p) { - *cur_percentage_p = percentage; - LLAMA_LOG_INFO("."); - if (percentage >= 100) { - LLAMA_LOG_INFO("\n"); - } - } - }; - } - ctx->rng = std::mt19937(params.seed); ctx->logits_all = params.logits_all; @@ -4279,13 +5223,14 @@ struct llama_context * llama_new_context_with_model( struct llama_context * llama_init_from_file( const char * path_model, struct llama_context_params params) { - struct llama_model * model = llama_load_model_from_file(path_model, params); if (!model) { return nullptr; } + struct llama_context * ctx = llama_new_context_with_model(model, params); ctx->model_owner = true; + return ctx; } @@ -4305,6 +5250,10 @@ int llama_n_embd(const struct llama_context * ctx) { return ctx->model.hparams.n_embd; } +enum llama_vocab_type llama_vocab_type(const struct llama_context * ctx) { + return ctx->model.vocab.type; +} + int llama_model_n_vocab(const struct llama_model * model) { return model->vocab.id_to_token.size(); } @@ -4318,7 +5267,10 @@ int llama_model_n_embd(const struct llama_model * model) { } int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size) { - return snprintf(buf, buf_size, "LLaMA %s %s", llama_model_type_name(model->type), llama_model_ftype_name(model->ftype).c_str()); + return snprintf(buf, buf_size, "%s %s %s", + model->name.c_str(), + llama_model_type_name(model->type), + llama_model_ftype_name(model->ftype).c_str()); } int llama_model_quantize( @@ -4839,26 +5791,6 @@ int llama_tokenize( return llama_tokenize_with_model(&ctx->model, text, tokens, n_max_tokens, add_bos); } -int llama_tokenize_bpe( - struct llama_context * ctx, - const char * text, - llama_token * tokens, - int n_max_tokens, - bool add_bos) { - auto res = llama_tokenize_internal(ctx->model.vocab, text, add_bos, false); - - if (n_max_tokens < (int) res.size()) { - LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); - return -((int) res.size()); - } - - for (size_t i = 0; i < res.size(); i++) { - tokens[i] = res[i]; - } - - return res.size(); -} - int llama_tokenize_with_model( const struct llama_model * model, const char * text, @@ -4884,18 +5816,6 @@ int llama_token_to_str(const struct llama_context * ctx, llama_token token, char return llama_token_to_str_with_model(&ctx->model, token, buf, length); } -int llama_token_to_str_bpe(const struct llama_context * ctx, llama_token token, char * buf, int length) { - if (0 <= token && token < llama_model_n_vocab(&ctx->model)) { - std::string result = ctx->model.vocab.id_to_token[token].text; - if (length < (int) result.length()) { - return -result.length(); - } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } - return 0; -} - // does not write null-terminator to str int llama_token_to_str_with_model(const struct llama_model * model, llama_token token, char * buf, int length) { if (0 <= token && token < llama_model_n_vocab(model)) { diff --git a/llama.h b/llama.h index 7ce478d54..4e7638c04 100644 --- a/llama.h +++ b/llama.h @@ -247,6 +247,8 @@ extern "C" { LLAMA_API int llama_n_ctx (const struct llama_context * ctx); LLAMA_API int llama_n_embd (const struct llama_context * ctx); + LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_context * ctx); + LLAMA_API int llama_model_n_vocab(const struct llama_model * model); LLAMA_API int llama_model_n_ctx (const struct llama_model * model); LLAMA_API int llama_model_n_embd (const struct llama_model * model); @@ -368,13 +370,6 @@ extern "C" { int n_max_tokens, bool add_bos); - LLAMA_API int llama_tokenize_bpe( - struct llama_context * ctx, - const char * text, - llama_token * tokens, - int n_max_tokens, - bool add_bos); - LLAMA_API int llama_tokenize_with_model( const struct llama_model * model, const char * text, @@ -390,12 +385,6 @@ extern "C" { char * buf, int length); - LLAMA_API int llama_token_to_str_bpe( - const struct llama_context * ctx, - llama_token token, - char * buf, - int length); - LLAMA_API int llama_token_to_str_with_model( const struct llama_model * model, llama_token token, diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 4ccefe932..2afaf86b1 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -28,7 +28,8 @@ llama_build_and_test_executable(test-sampling.cpp) llama_build_executable(test-tokenizer-0.cpp) llama_test_executable (test-tokenizer-0.llama test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) llama_build_executable(test-tokenizer-1.cpp) -llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) +# test-tokenizer-1 requires a BPE vocab. re-enable when we have one. +#llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) #llama_test_executable(test-tokenizer-1.aquila test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf) llama_build_and_test_executable(test-grammar-parser.cpp) llama_build_and_test_executable(test-llama-grammar.cpp) diff --git a/tests/test-tokenizer-1.cpp b/tests/test-tokenizer-1.cpp index 993d17f18..bd607d12b 100644 --- a/tests/test-tokenizer-1.cpp +++ b/tests/test-tokenizer-1.cpp @@ -67,11 +67,13 @@ int main(int argc, char **argv) { } } + GGML_ASSERT(llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_BPE); + const int n_vocab = llama_n_vocab(ctx); for (int i = 0; i < n_vocab; ++i) { - std::string forward = llama_token_to_str_bpe(ctx, i); - std::vector tokens = llama_tokenize_bpe(ctx, forward, false); + std::string forward = llama_token_to_str(ctx, i); + std::vector tokens = llama_tokenize(ctx, forward, false); if (tokens.size() == 1) { if (i != tokens[0]) { std::string backward = llama_token_to_str(ctx, tokens[0]); @@ -79,16 +81,6 @@ int main(int argc, char **argv) { __func__, i, llama_token_to_str(ctx, i).c_str(), tokens[0], backward.c_str()); return 2; } - } else { - llama_token_type type = llama_token_get_type(ctx, i); - if (type == LLAMA_TOKEN_TYPE_UNKNOWN || type == LLAMA_TOKEN_TYPE_CONTROL || type == LLAMA_TOKEN_TYPE_BYTE) { - fprintf(stderr, "%s : info: token %d is string %s and bpe returns tokens %s\n", - __func__, i, llama_token_to_str(ctx, i).c_str(), unescape_whitespace(ctx, tokens).c_str()); - } else { - fprintf(stderr, "%s : error: token %d is string %s but bpe returns tokens %s\n", - __func__, i, llama_token_to_str(ctx, i).c_str(), unescape_whitespace(ctx, tokens).c_str()); - return 2; - } } } From 79da24b58c1ea72340e64f799a4717d372207676 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 23 Aug 2023 23:41:16 +0300 Subject: [PATCH 018/147] readme : update hot topics --- README.md | 162 +++++++++++++++++++++++++++--------------------------- 1 file changed, 80 insertions(+), 82 deletions(-) diff --git a/README.md b/README.md index f746c49eb..669aa7c02 100644 --- a/README.md +++ b/README.md @@ -11,15 +11,17 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ ### Hot topics -A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398) +- Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717#issuecomment-1690568032 -Last revision compatible with the old format: [dadbed9](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa) +- A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398) -### Current `master` should be considered in Beta - expect some issues for a few days! + Last revision compatible with the old format: [dadbed9](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa) -### Be prepared to re-convert and / or re-quantize your GGUF models while this notice is up! + ### Current `master` should be considered in Beta - expect some issues for a few days! -### Issues with non-GGUF models will be considered with low priority! + ### Be prepared to re-convert and / or re-quantize your GGUF models while this notice is up! + + ### Issues with non-GGUF models will be considered with low priority! ---- @@ -66,12 +68,11 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant - Apple silicon first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks - AVX, AVX2 and AVX512 support for x86 architectures - Mixed F16 / F32 precision -- 4-bit, 5-bit and 8-bit integer quantization support -- Supports OpenBLAS/Apple BLAS/ARM Performance Lib/ATLAS/BLIS/Intel MKL/NVHPC/ACML/SCSL/SGIMATH and [more](https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors) in BLAS -- cuBLAS and CLBlast support +- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support +- CUDA, Metal and OpenCL GPU backend support The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022). -Since then, the project has improved significantly thanks to many contributions. This project is for educational purposes and serves +Since then, the project has improved significantly thanks to many contributions. This project is mainly for educational purposes and serves as the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library. **Supported platforms:** @@ -85,6 +86,7 @@ as the main playground for developing new features for the [ggml](https://github - [X] LLaMA 🦙 - [x] LLaMA 2 🦙🦙 +- [X] Falcon - [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca) - [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all) - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2) @@ -115,90 +117,84 @@ as the main playground for developing new features for the [ggml](https://github --- -Here is a typical run using LLaMA-7B: +Here is a typical run using LLaMA v2 13B on M2 Ultra: ```java -make -j && ./main -m ./models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 -I llama.cpp build info: +$ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e +I llama.cpp build info: I UNAME_S: Darwin I UNAME_P: arm I UNAME_M: arm64 -I CFLAGS: -I. -O3 -DNDEBUG -std=c11 -fPIC -pthread -DGGML_USE_ACCELERATE -I CXXFLAGS: -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -pthread +I CFLAGS: -I. -O3 -std=c11 -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes -pthread -DGGML_USE_K_QUANTS -DGGML_USE_ACCELERATE +I CXXFLAGS: -I. -I./common -O3 -std=c++11 -fPIC -DNDEBUG -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar -pthread -DGGML_USE_K_QUANTS I LDFLAGS: -framework Accelerate -I CC: Apple clang version 14.0.0 (clang-1400.0.29.202) -I CXX: Apple clang version 14.0.0 (clang-1400.0.29.202) +I CC: Apple clang version 14.0.3 (clang-1403.0.22.14.1) +I CXX: Apple clang version 14.0.3 (clang-1403.0.22.14.1) make: Nothing to be done for `default'. -main: seed = 1678486056 -llama_model_load: loading model from './models/7B/ggml-model-q4_0.bin' - please wait ... -llama_model_load: n_vocab = 32000 -llama_model_load: n_ctx = 512 -llama_model_load: n_embd = 4096 -llama_model_load: n_mult = 256 -llama_model_load: n_head = 32 -llama_model_load: n_layer = 32 -llama_model_load: n_rot = 128 -llama_model_load: f16 = 2 -llama_model_load: n_ff = 11008 -llama_model_load: ggml ctx size = 4529.34 MB -llama_model_load: memory_size = 512.00 MB, n_mem = 16384 -llama_model_load: .................................... done -llama_model_load: model size = 4017.27 MB / num tensors = 291 +main: build = 1041 (cf658ad) +main: seed = 1692823051 +llama_model_loader: loaded meta data with 16 key-value pairs and 363 tensors from models/llama-13b-v2/ggml-model-q4_0.gguf (version GGUF V1 (latest)) +llama_model_loader: - type f32: 81 tensors +llama_model_loader: - type q4_0: 281 tensors +llama_model_loader: - type q6_K: 1 tensors +llm_load_print_meta: format = GGUF V1 (latest) +llm_load_print_meta: arch = llama +llm_load_print_meta: vocab type = SPM +llm_load_print_meta: n_vocab = 32000 +llm_load_print_meta: n_merges = 0 +llm_load_print_meta: n_ctx_train = 4096 +llm_load_print_meta: n_ctx = 512 +llm_load_print_meta: n_embd = 5120 +llm_load_print_meta: n_head = 40 +llm_load_print_meta: n_head_kv = 40 +llm_load_print_meta: n_layer = 40 +llm_load_print_meta: n_rot = 128 +llm_load_print_meta: n_gqa = 1 +llm_load_print_meta: f_norm_eps = 1.0e-05 +llm_load_print_meta: f_norm_rms_eps = 1.0e-05 +llm_load_print_meta: n_ff = 13824 +llm_load_print_meta: freq_base = 10000.0 +llm_load_print_meta: freq_scale = 1 +llm_load_print_meta: model type = 13B +llm_load_print_meta: model ftype = mostly Q4_0 +llm_load_print_meta: model size = 13.02 B +llm_load_print_meta: general.name = LLaMA v2 +llm_load_print_meta: BOS token = 1 '' +llm_load_print_meta: EOS token = 2 '' +llm_load_print_meta: UNK token = 0 '' +llm_load_print_meta: LF token = 13 '<0x0A>' +llm_load_tensors: ggml ctx size = 0.11 MB +llm_load_tensors: mem required = 7024.01 MB (+ 400.00 MB per state) +................................................................................................... +llama_new_context_with_model: kv self size = 400.00 MB +llama_new_context_with_model: compute buffer total size = 75.41 MB -main: prompt: 'Building a website can be done in 10 simple steps:' -main: number of tokens in prompt = 15 - 1 -> '' - 8893 -> 'Build' - 292 -> 'ing' - 263 -> ' a' - 4700 -> ' website' - 508 -> ' can' - 367 -> ' be' - 2309 -> ' done' - 297 -> ' in' - 29871 -> ' ' - 29896 -> '1' - 29900 -> '0' - 2560 -> ' simple' - 6576 -> ' steps' - 29901 -> ':' - -sampling parameters: temp = 0.800000, top_k = 40, top_p = 0.950000 +system_info: n_threads = 16 / 24 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | +sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000 +generate: n_ctx = 512, n_batch = 512, n_predict = 400, n_keep = 0 -Building a website can be done in 10 simple steps: -1) Select a domain name and web hosting plan -2) Complete a sitemap -3) List your products -4) Write product descriptions -5) Create a user account -6) Build the template -7) Start building the website -8) Advertise the website -9) Provide email support -10) Submit the website to search engines -A website is a collection of web pages that are formatted with HTML. HTML is the code that defines what the website looks like and how it behaves. -The HTML code is formatted into a template or a format. Once this is done, it is displayed on the user's browser. -The web pages are stored in a web server. The web server is also called a host. When the website is accessed, it is retrieved from the server and displayed on the user's computer. -A website is known as a website when it is hosted. This means that it is displayed on a host. The host is usually a web server. -A website can be displayed on different browsers. The browsers are basically the software that renders the website on the user's screen. -A website can also be viewed on different devices such as desktops, tablets and smartphones. -Hence, to have a website displayed on a browser, the website must be hosted. -A domain name is an address of a website. It is the name of the website. -The website is known as a website when it is hosted. This means that it is displayed on a host. The host is usually a web server. -A website can be displayed on different browsers. The browsers are basically the software that renders the website on the user’s screen. -A website can also be viewed on different devices such as desktops, tablets and smartphones. Hence, to have a website displayed on a browser, the website must be hosted. -A domain name is an address of a website. It is the name of the website. -A website is an address of a website. It is a collection of web pages that are formatted with HTML. HTML is the code that defines what the website looks like and how it behaves. -The HTML code is formatted into a template or a format. Once this is done, it is displayed on the user’s browser. -A website is known as a website when it is hosted - -main: mem per token = 14434244 bytes -main: load time = 1332.48 ms -main: sample time = 1081.40 ms -main: predict time = 31378.77 ms / 61.41 ms per token -main: total time = 34036.74 ms + Building a website can be done in 10 simple steps: +Step 1: Find the right website platform. +Step 2: Choose your domain name and hosting plan. +Step 3: Design your website layout. +Step 4: Write your website content and add images. +Step 5: Install security features to protect your site from hackers or spammers +Step 6: Test your website on multiple browsers, mobile devices, operating systems etc… +Step 7: Test it again with people who are not related to you personally – friends or family members will work just fine! +Step 8: Start marketing and promoting the website via social media channels or paid ads +Step 9: Analyze how many visitors have come to your site so far, what type of people visit more often than others (e.g., men vs women) etc… +Step 10: Continue to improve upon all aspects mentioned above by following trends in web design and staying up-to-date on new technologies that can enhance user experience even further! +How does a Website Work? +A website works by having pages, which are made of HTML code. This code tells your computer how to display the content on each page you visit – whether it’s an image or text file (like PDFs). In order for someone else’s browser not only be able but also want those same results when accessing any given URL; some additional steps need taken by way of programming scripts that will add functionality such as making links clickable! +The most common type is called static HTML pages because they remain unchanged over time unless modified manually (either through editing files directly or using an interface such as WordPress). They are usually served up via HTTP protocols – this means anyone can access them without having any special privileges like being part of a group who is allowed into restricted areas online; however, there may still exist some limitations depending upon where one lives geographically speaking. +How to +llama_print_timings: load time = 576.45 ms +llama_print_timings: sample time = 283.10 ms / 400 runs ( 0.71 ms per token, 1412.91 tokens per second) +llama_print_timings: prompt eval time = 599.83 ms / 19 tokens ( 31.57 ms per token, 31.68 tokens per second) +llama_print_timings: eval time = 24513.59 ms / 399 runs ( 61.44 ms per token, 16.28 tokens per second) +llama_print_timings: total time = 25431.49 ms ``` And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook: @@ -543,6 +539,8 @@ As the models are currently fully loaded into memory, you will need adequate dis Several quantization methods are supported. They differ in the resulting model disk size and inference speed. +*(outdated)* + | Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 | |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:| | 7B | perplexity | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 | From c7868b075377c8c3fa916ea7c1aca600f44bed55 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 23 Aug 2023 23:43:00 +0300 Subject: [PATCH 019/147] minor : fix trailing whitespace --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 669aa7c02..0328d60cf 100644 --- a/README.md +++ b/README.md @@ -121,7 +121,7 @@ Here is a typical run using LLaMA v2 13B on M2 Ultra: ```java $ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -I llama.cpp build info: +I llama.cpp build info: I UNAME_S: Darwin I UNAME_P: arm I UNAME_M: arm64 @@ -170,7 +170,7 @@ llm_load_tensors: mem required = 7024.01 MB (+ 400.00 MB per state) llama_new_context_with_model: kv self size = 400.00 MB llama_new_context_with_model: compute buffer total size = 75.41 MB -system_info: n_threads = 16 / 24 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | +system_info: n_threads = 16 / 24 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | sampling: repeat_last_n = 64, repeat_penalty = 1.100000, presence_penalty = 0.000000, frequency_penalty = 0.000000, top_k = 40, tfs_z = 1.000000, top_p = 0.950000, typical_p = 1.000000, temp = 0.800000, mirostat = 0, mirostat_lr = 0.100000, mirostat_ent = 5.000000 generate: n_ctx = 512, n_batch = 512, n_predict = 400, n_keep = 0 From 44d5462b5cddc1c5cbcd7647646f7b55b175b01f Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 23 Aug 2023 23:44:19 +0300 Subject: [PATCH 020/147] readme : fix link --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 0328d60cf..eebb11392 100644 --- a/README.md +++ b/README.md @@ -11,7 +11,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ ### Hot topics -- Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717#issuecomment-1690568032 +- Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717 - A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398) From 6e91a1b0706c2e0e52b9d9be7ee82d3c1e7a33c1 Mon Sep 17 00:00:00 2001 From: Evan Jones Date: Thu, 24 Aug 2023 00:07:13 -0400 Subject: [PATCH 021/147] llama : fix grammar sometimes generating null char (#2756) --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index f2dc4da1d..7cac8a1ce 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4074,7 +4074,7 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c if (!allow_eos) { candidates->data[i].logit = -INFINITY; } - } else if (text.empty()) { + } else if (text.empty() || text[0] == 0) { candidates->data[i].logit = -INFINITY; } else { candidates_decoded.push_back(decode_utf8(text.c_str(), grammar->partial_utf8)); From c3e53b421a9910548be0345f85712c535f467a98 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 24 Aug 2023 12:26:01 +0300 Subject: [PATCH 022/147] llama : escape all U+2581 in a string (#2750) --- llama.cpp | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) diff --git a/llama.cpp b/llama.cpp index 7cac8a1ce..f5526e302 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3004,11 +3004,8 @@ static std::string llama_escape_whitespace(const std::string& text) { return result; } -static std::string llama_unescape_whitespace(const std::string& word) { - if (word.length() >= 3 && word.substr(0, 3) == "\xe2\x96\x81") { - return std::string(" ") + word.substr(3); - } - return word; +static void llama_unescape_whitespace(std::string & word) { + replace_all(word, "\xe2\x96\x81", " "); } struct llm_symbol { @@ -5822,7 +5819,7 @@ int llama_token_to_str_with_model(const struct llama_model * model, llama_token if (llama_is_normal_token(model->vocab, token)) { std::string result = model->vocab.id_to_token[token].text; if (llama_vocab_get_type(model->vocab) == LLAMA_VOCAB_TYPE_SPM) { - result = llama_unescape_whitespace(result); + llama_unescape_whitespace(result); } if (length < (int) result.length()) { return -result.length(); From d67777c202c03bcb74372690599ef3c03affb3ba Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 24 Aug 2023 16:19:57 +0300 Subject: [PATCH 023/147] metal : add Q8_0 support (#2763) * metal : add dequantize_q8_0 kernel * metal : add mul_mat_q8_0_f32 kernel * metal : add Q8_0 mul_mm kernel --- ggml-metal.m | 23 ++++++++++-- ggml-metal.metal | 93 ++++++++++++++++++++++++++++++++++++++++++++---- 2 files changed, 106 insertions(+), 10 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 969cf7daa..06eb3872e 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -63,6 +63,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(get_rows_f16); GGML_METAL_DECL_KERNEL(get_rows_q4_0); GGML_METAL_DECL_KERNEL(get_rows_q4_1); + GGML_METAL_DECL_KERNEL(get_rows_q8_0); GGML_METAL_DECL_KERNEL(get_rows_q2_K); GGML_METAL_DECL_KERNEL(get_rows_q3_K); GGML_METAL_DECL_KERNEL(get_rows_q4_K); @@ -73,6 +74,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(mul_mat_f16_f32); GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32); GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32); + GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32); GGML_METAL_DECL_KERNEL(mul_mat_q2_K_f32); GGML_METAL_DECL_KERNEL(mul_mat_q3_K_f32); GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32); @@ -81,6 +83,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(mul_mm_f16_f32); GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32); GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32); + GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32); GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32); GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32); GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32); @@ -188,6 +191,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(get_rows_f16); GGML_METAL_ADD_KERNEL(get_rows_q4_0); GGML_METAL_ADD_KERNEL(get_rows_q4_1); + GGML_METAL_ADD_KERNEL(get_rows_q8_0); GGML_METAL_ADD_KERNEL(get_rows_q2_K); GGML_METAL_ADD_KERNEL(get_rows_q3_K); GGML_METAL_ADD_KERNEL(get_rows_q4_K); @@ -198,6 +202,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(mul_mat_f16_f32); GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32); GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32); + GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32); GGML_METAL_ADD_KERNEL(mul_mat_q2_K_f32); GGML_METAL_ADD_KERNEL(mul_mat_q3_K_f32); GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32); @@ -205,6 +210,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32); GGML_METAL_ADD_KERNEL(mul_mm_f16_f32); GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32); + GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32); GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32); GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32); GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32); @@ -747,9 +753,10 @@ void ggml_metal_graph_compute( ne00%32 == 0 && ne11 > 1) { switch (src0->type) { - case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break; + case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break; case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break; case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break; + case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break; case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break; case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break; case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break; @@ -800,6 +807,15 @@ void ggml_metal_graph_compute( nth1 = 8; [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32]; } break; + case GGML_TYPE_Q8_0: + { + GGML_ASSERT(ne02 == 1); + GGML_ASSERT(ne12 == 1); + + nth0 = 8; + nth1 = 8; + [encoder setComputePipelineState:ctx->pipeline_mul_mat_q8_0_f32]; + } break; case GGML_TYPE_Q2_K: { GGML_ASSERT(ne02 == 1); @@ -871,7 +887,7 @@ void ggml_metal_graph_compute( [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16]; [encoder setBytes:&gqa length:sizeof(gqa) atIndex:17]; - if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || + if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } @@ -896,9 +912,10 @@ void ggml_metal_graph_compute( case GGML_OP_GET_ROWS: { switch (src0->type) { - case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break; + case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break; case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break; case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break; + case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break; case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break; case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break; case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break; diff --git a/ggml-metal.metal b/ggml-metal.metal index 7bc3fdf37..82e1a0c7a 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -18,6 +18,12 @@ typedef struct { uint8_t qs[QK4_1 / 2]; // nibbles / quants } block_q4_1; +#define QK8_0 32 +typedef struct { + half d; // delta + int8_t qs[QK8_0]; // quants +} block_q8_0; + kernel void kernel_add( device const float * src0, device const float * src1, @@ -357,7 +363,7 @@ void mul_vec_q_n_f32(device const void * src0, device const float * src1, device const int first_row = (r0 * nsg + sgitg) * nr; const uint offset0 = first_row * nb + im/gqa*(nb*ne0); device const block_q_type * x = (device const block_q_type *) src0 + offset0; - device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1; + device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1; float yl[16]; // src1 vector cache float sumf[nr]={0.f}; @@ -429,6 +435,68 @@ kernel void kernel_mul_mat_q4_1_f32( mul_vec_q_n_f32(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg); } +kernel void kernel_mul_mat_q8_0_f32( + device const void * src0, + device const float * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01[[buffer(4)]], + constant int64_t & ne02[[buffer(5)]], + constant int64_t & ne10[[buffer(9)]], + constant int64_t & ne12[[buffer(11)]], + constant int64_t & ne0[[buffer(15)]], + constant int64_t & ne1[[buffer(16)]], + constant uint & gqa[[buffer(17)]], + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiisg[[thread_index_in_simdgroup]], + uint sgitg[[simdgroup_index_in_threadgroup]]) { + const int nr = N_DST; + const int nsg = N_SIMDGROUP; + const int nw = N_SIMDWIDTH; + + const int nb = ne00/QK8_0; + const int r0 = tgpig.x; + const int r1 = tgpig.y; + const int im = tgpig.z; + const int first_row = (r0 * nsg + sgitg) * nr; + const uint offset0 = first_row * nb + im/gqa*(nb*ne0); + device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0; + device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1; + + float yl[16]; + float sumf[nr]={0.f}; + + const int ix = tiisg/2; + const int il = tiisg%2; + + device const float * yb = y + ix * QK8_0 + 16*il; + + // each thread in a SIMD group deals with half a block. + for (int ib = ix; ib < nb; ib += nw/2) { + for (int i = 0; i < 16; ++i) { + yl[i] = yb[i]; + } + + for (int row = 0; row < nr; row++) { + device const int8_t * qs = x[ib+row*nb].qs + 16*il; + float sumq = 0.f; + for (int iq = 0; iq < 16; ++iq) { + sumq += qs[iq] * yl[iq]; + } + sumf[row] += sumq*x[ib+row*nb].d; + } + + yb += QK8_0 * 16; + } + + for (int row = 0; row < nr; ++row) { + const float tot = simd_sum(sumf[row]); + if (tiisg == 0 && first_row + row < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot; + } + } +} + kernel void kernel_mul_mat_f16_f32( device const char * src0, device const char * src1, @@ -480,7 +548,6 @@ kernel void kernel_mul_mat_f16_f32( } } - kernel void kernel_alibi_f32( device const float * src0, device float * dst, @@ -1621,12 +1688,12 @@ template void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) { device const uint16_t * qs = ((device const uint16_t *)xb + 1); const half d = il ? (xb->d / 16.h) : xb->d; - const half m = il ? (-8.h * 16.h) : -8.h; + const half m = il ? ( -8.h * 16.h) : -8.h; const ushort mask0 = il ? 0x00F0 : 0x000F; const ushort mask1 = il ? 0xF000 : 0x0F00; for (int i=0;i<8;i++) { - reg[i/2][2*(i%2)] = (((qs[i] & mask0)) + m) * d; + reg[i/2][2*(i%2)] = (((qs[i] & mask0) ) + m) * d; reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) + m) * d; } } @@ -1640,11 +1707,21 @@ void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg const ushort mask1 = il ? 0xF000 : 0x0F00; for (int i=0;i<8;i++) { - reg[i/2][2*(i%2)] = (((qs[i] & mask0)) * d) + m; + reg[i/2][2*(i%2)] = (((qs[i] & mask0) ) * d) + m; reg[i/2][2*(i%2)+1] = (((qs[i] & mask1) >> 8) * d) + m; } } +template +void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) { + device const int8_t * qs = ((device const int8_t *)xb->qs); + const half d = xb->d; + + for (int i=0;i<16;i++) { + reg[i/4][i%4] = (qs[i + 16*il] * d); + } +} + template void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) { const half d = xb->d; @@ -1947,9 +2024,10 @@ kernel void kernel_mul_mm(device const uchar * src0, typedef void (get_rows_t)(device const void *, device const int *, device float *, constant int64_t &, \ constant uint64_t &, constant uint64_t &, uint, uint, uint); -template [[host_name("kernel_get_rows_f16")]] kernel get_rows_t kernel_get_rows; +template [[host_name("kernel_get_rows_f16")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows; +template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows; template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows; @@ -1960,9 +2038,10 @@ typedef void (mat_mm_t)(device const uchar *, device const float *, device float constant int64_t &, constant int64_t &, constant int64_t &, constant int64_t &, \ constant int64_t &, constant int64_t &, constant uint &, threadgroup uchar *, uint3, uint, uint); -template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm; template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm; From ef955fbd230c571cc1cda0d19baaeec347523175 Mon Sep 17 00:00:00 2001 From: DannyDaemonic Date: Thu, 24 Aug 2023 06:58:02 -0700 Subject: [PATCH 024/147] Tag release with build number (#2732) * Modified build.yml to use build number for release * Add the short hash back into the tag * Prefix the build number with b --- .github/workflows/build.yml | 63 +++++++++++++++++++++++++++---------- 1 file changed, 46 insertions(+), 17 deletions(-) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 84faad37a..10320ad1f 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -291,24 +291,32 @@ jobs: cd build ctest -C Release --verbose --timeout 900 - - name: Get commit hash - id: commit - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - uses: pr-mpt/actions-commit-hash@v2 + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi - name: Pack artifacts id: pack_artifacts if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} run: | Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt - 7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\* + 7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\* - name: Upload artifacts if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} uses: actions/upload-artifact@v3 with: path: | - llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip + llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip windows-latest-cmake-cublas: runs-on: windows-latest @@ -338,23 +346,31 @@ jobs: cmake .. -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON cmake --build . --config Release - - name: Get commit hash - id: commit - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - uses: pr-mpt/actions-commit-hash@v2 + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi - name: Pack artifacts id: pack_artifacts if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} run: | - 7z a llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\* + 7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\* - name: Upload artifacts if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} uses: actions/upload-artifact@v3 with: path: | - llama-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip + llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip - name: Copy and pack Cuda runtime if: ${{ matrix.cuda == '12.1.0' }} @@ -400,21 +416,34 @@ jobs: - windows-latest-cmake-cublas steps: + - name: Clone + id: checkout + uses: actions/checkout@v1 + + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi + - name: Download artifacts id: download-artifact uses: actions/download-artifact@v3 - - name: Get commit hash - id: commit - uses: pr-mpt/actions-commit-hash@v2 - - name: Create release id: create_release uses: anzz1/action-create-release@v1 env: GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} with: - tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }} + tag_name: ${{ steps.tag.outputs.name }} - name: Upload release id: upload_release From fea95c682d0028fdd25853bea58035794a0c964d Mon Sep 17 00:00:00 2001 From: slaren Date: Thu, 24 Aug 2023 17:44:11 +0200 Subject: [PATCH 025/147] fix convert.py for codellama, add llama 34B to the list of recognized models (#2768) --- convert.py | 2 +- llama.cpp | 3 +++ 2 files changed, 4 insertions(+), 1 deletion(-) diff --git a/convert.py b/convert.py index 8d34d5f29..d3776a4da 100755 --- a/convert.py +++ b/convert.py @@ -191,7 +191,7 @@ class Params: def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params': config = json.load(open(config_path)) - n_vocab = config["vocab_size"] + n_vocab = config["vocab_size"] if "vocab_size" in config else -1 n_embd = config["dim"] n_layer = config["n_layers"] n_mult = config["multiple_of"] diff --git a/llama.cpp b/llama.cpp index f5526e302..7ee6bcdae 100644 --- a/llama.cpp +++ b/llama.cpp @@ -827,6 +827,7 @@ enum e_model { MODEL_7B, MODEL_13B, MODEL_30B, + MODEL_34B, MODEL_40B, MODEL_65B, MODEL_70B, @@ -1518,6 +1519,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_7B: return "7B"; case MODEL_13B: return "13B"; case MODEL_30B: return "30B"; + case MODEL_34B: return "34B"; case MODEL_40B: return "40B"; case MODEL_65B: return "65B"; case MODEL_70B: return "70B"; @@ -1590,6 +1592,7 @@ static void llm_load_hparams( case 26: model.type = e_model::MODEL_3B; break; case 32: model.type = e_model::MODEL_7B; break; case 40: model.type = e_model::MODEL_13B; break; + case 48: model.type = e_model::MODEL_34B; break; case 60: model.type = e_model::MODEL_30B; break; case 80: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_65B : e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; From 7694adda8d1111b3cf758ad6c91d754a0a4cacff Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Thu, 24 Aug 2023 10:11:13 -0600 Subject: [PATCH 026/147] Fix for main example getting stuck when -n -2 and --interactive (#2767) * Fix for main example getting stuck when -n -2 and --interactive * Add a comment so future generations may suffer less. --- examples/main/main.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 1393f0b08..cb8747c2b 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -798,7 +798,8 @@ int main(int argc, char ** argv) { } // In interactive mode, respect the maximum number of tokens and drop back to user input when reached. - if (params.interactive && n_remain <= 0 && params.n_predict != -1) { + // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size). + if (params.interactive && n_remain <= 0 && params.n_predict >= 0) { n_remain = params.n_predict; is_interacting = true; } From 8f8c28e89cb9531211783da697d6e7c445e2af1d Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 24 Aug 2023 19:26:19 +0300 Subject: [PATCH 027/147] convert : auto-determine model name based on dir + scripts update --- convert.py | 13 ++++-- scripts/perf-run-all.sh | 93 ----------------------------------------- scripts/ppl-run-all.sh | 39 ----------------- scripts/qnt-all.sh | 27 ++++++++++++ scripts/run-all-perf.sh | 31 ++++++++++++++ scripts/run-all-ppl.sh | 27 ++++++++++++ 6 files changed, 95 insertions(+), 135 deletions(-) delete mode 100755 scripts/perf-run-all.sh delete mode 100755 scripts/ppl-run-all.sh create mode 100755 scripts/qnt-all.sh create mode 100755 scripts/run-all-perf.sh create mode 100755 scripts/run-all-ppl.sh diff --git a/convert.py b/convert.py index d3776a4da..b7c626d84 100755 --- a/convert.py +++ b/convert.py @@ -106,6 +106,9 @@ class Params: ftype: Optional[GGMLFileType] = None + # path to the directory containing the model files + path_model: Optional['Path'] = None + @staticmethod def find_n_mult(n_ff: int, n_embd: int) -> int: # hardcoded magic range @@ -231,6 +234,8 @@ class Params: else: params = Params.guessed(model_plus.model) + params.path_model = model_plus.paths[0].parent + return params @@ -733,11 +738,13 @@ class OutputFile: self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) def add_meta_arch(self, params: Params) -> None: - ver = None + name = "LLaMA" if (params.n_ctx == 4096): - ver = "v2" + name = "LLaMA v2" + if params.path_model: + name = str(params.path_model.parent).split('/')[-1] - self.gguf.add_name ("LLaMA" if ver == None else "LLaMA " + ver) + self.gguf.add_name (name) self.gguf.add_context_length (params.n_ctx) self.gguf.add_embedding_length (params.n_embd) self.gguf.add_block_count (params.n_layer) diff --git a/scripts/perf-run-all.sh b/scripts/perf-run-all.sh deleted file mode 100755 index 7dbfc7c20..000000000 --- a/scripts/perf-run-all.sh +++ /dev/null @@ -1,93 +0,0 @@ -#!/bin/bash -# -# Measure the performance (time per token) of the various quantization techniques -# - -QUANTIZE=0 -if [ "$1" != "" ]; then - echo "Quantizing" - QUANTIZE=1 -fi - -if [ "$QUANTIZE" != "0" ]; then - # - # quantize - # - - # 7B - time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-7b-q4_0.txt - time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-7b-q4_1.txt - time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-7b-q5_0.txt - time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-7b-q5_1.txt - time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-7b-q8_0.txt - - # 13B - time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-13b-q4_0.txt - time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-13b-q4_1.txt - time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-13b-q5_0.txt - time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-13b-q5_1.txt - time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-13b-q8_0.txt -fi - -# -# perf -# run each command twice -# - -set -x - -# 7B - 4 threads - ./bin/main -m ../models/7B/ggml-model-f16.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-f16.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-f16.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q4_0.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q4_1.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q5_0.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q5_1.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-7b-q8_0.txt | grep llama_print_timings - -# 7B - 8 threads - ./bin/main -m ../models/7B/ggml-model-f16.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-f16.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-f16.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q4_0.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q4_1.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q5_0.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q5_1.txt | grep llama_print_timings - ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/7B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-7b-q8_0.txt | grep llama_print_timings - -# 13B - 4 threads - ./bin/main -m ../models/13B/ggml-model-f16.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-f16.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-f16.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q4_0.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q4_1.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q5_0.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q5_1.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 4 2>&1 | tee ../perf-13b-q8_0.txt | grep llama_print_timings - -# 13B - 8 threads - ./bin/main -m ../models/13B/ggml-model-f16.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-f16.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-f16.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q4_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q4_0.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q4_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q4_1.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q5_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q5_0.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q5_1.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q5_1.txt | grep llama_print_timings - ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | grep "I believe" -time ./bin/main -m ../models/13B/ggml-model-q8_0.bin -p "I believe the meaning of life is" --no-mmap -c 2048 --ignore-eos -s 1 -n 64 -t 8 2>&1 | tee ../perf-13b-q8_0.txt | grep llama_print_timings diff --git a/scripts/ppl-run-all.sh b/scripts/ppl-run-all.sh deleted file mode 100755 index c59e3075d..000000000 --- a/scripts/ppl-run-all.sh +++ /dev/null @@ -1,39 +0,0 @@ -#!/bin/bash - -# -# quantize -# - -# 7B -time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-7b-q4_0.txt -time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-7b-q4_1.txt -time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-7b-q5_0.txt -time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-7b-q5_1.txt -time ./bin/quantize ../models/7B/ggml-model-f16.bin ../models/7B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-7b-q8_0.txt - -# 13B -time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_0.bin q4_0 2>&1 | tee ../qnt-13b-q4_0.txt -time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q4_1.bin q4_1 2>&1 | tee ../qnt-13b-q4_1.txt -time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_0.bin q5_0 2>&1 | tee ../qnt-13b-q5_0.txt -time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q5_1.bin q5_1 2>&1 | tee ../qnt-13b-q5_1.txt -time ./bin/quantize ../models/13B/ggml-model-f16.bin ../models/13B/ggml-model-q8_0.bin q8_0 2>&1 | tee ../qnt-13b-q8_0.txt - -# -# perplexity -# - -# 7B -time ./bin/perplexity -m ../models/7B/ggml-model-f16.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-f16.txt -time ./bin/perplexity -m ../models/7B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_0.txt -time ./bin/perplexity -m ../models/7B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q4_1.txt -time ./bin/perplexity -m ../models/7B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_0.txt -time ./bin/perplexity -m ../models/7B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q5_1.txt -time ./bin/perplexity -m ../models/7B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-7b-q8_0.txt - -# 13B -time ./bin/perplexity -m ../models/13B/ggml-model-f16.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-f16.txt -time ./bin/perplexity -m ../models/13B/ggml-model-q4_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_0.txt -time ./bin/perplexity -m ../models/13B/ggml-model-q4_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q4_1.txt -time ./bin/perplexity -m ../models/13B/ggml-model-q5_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_0.txt -time ./bin/perplexity -m ../models/13B/ggml-model-q5_1.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q5_1.txt -time ./bin/perplexity -m ../models/13B/ggml-model-q8_0.bin -f ./wiki.test.raw --no-mmap -t 12 2>&1 | tee ../ppl-13b-q8_0.txt diff --git a/scripts/qnt-all.sh b/scripts/qnt-all.sh new file mode 100755 index 000000000..1b3d07da5 --- /dev/null +++ b/scripts/qnt-all.sh @@ -0,0 +1,27 @@ +#!/bin/bash + +qnt=(q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k) +args="" + +if [ -z "$1" ]; then + echo "usage: $0 [qnt] [args]" + echo "default: $0 \"${qnt[@]}\" \"${args}\"" + exit 1 +fi + +if [ ! -z "$2" ]; then + qnt=($2) +fi + +if [ ! -z "$3" ]; then + args="$3" +fi + +model="$1" +out="../tmp/results-${model}" + +mkdir -p ${out} + +for q in ${qnt[@]}; do + time ./bin/quantize ../models/${model}/ggml-model-f16.gguf ../models/${model}/ggml-model-${q}.gguf ${q} 2>&1 ${args} | tee ${out}/qnt-${q}.txt +done diff --git a/scripts/run-all-perf.sh b/scripts/run-all-perf.sh new file mode 100755 index 000000000..91a6d853f --- /dev/null +++ b/scripts/run-all-perf.sh @@ -0,0 +1,31 @@ +#!/bin/bash + +qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k) +args="-ngl 999 -n 64 -p 512" + +if [ -z "$1" ]; then + echo "usage: $0 [qnt] [args]" + echo "default: $0 \"${qnt[@]}\" \"${args}\"" + exit 1 +fi + +if [ ! -z "$2" ]; then + qnt=($2) +fi + +if [ ! -z "$3" ]; then + args="$3" +fi + +model="$1" +out="../tmp/results-${model}" + +mkdir -p ${out} + +mstr="" + +for q in ${qnt[@]}; do + mstr="${mstr} -m ../models/${model}/ggml-model-${q}.gguf" +done + +./bin/llama-bench ${mstr} ${args} 2> /dev/null diff --git a/scripts/run-all-ppl.sh b/scripts/run-all-ppl.sh new file mode 100755 index 000000000..bdbbb19c9 --- /dev/null +++ b/scripts/run-all-ppl.sh @@ -0,0 +1,27 @@ +#!/bin/bash + +qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k) +args="--no-mmap -ngl 999 -t 8" + +if [ -z "$1" ]; then + echo "usage: $0 [qnt] [args]" + echo "default: $0 \"${qnt[@]}\" \"${args}\"" + exit 1 +fi + +if [ ! -z "$2" ]; then + qnt=($2) +fi + +if [ ! -z "$3" ]; then + args="$3" +fi + +model="$1" +out="../tmp/results-${model}" + +mkdir -p ${out} + +for q in ${qnt[@]}; do + time ./bin/perplexity -m ../models/${model}/ggml-model-f16.gguf -f ./wiki.test.raw ${args} 2>&1 | tee ${out}/ppl-${q}.txt +done From 38b16dfca6e5032e6cfb90c1653bf1ba4cf647b4 Mon Sep 17 00:00:00 2001 From: Shouzheng Liu Date: Thu, 24 Aug 2023 12:27:25 -0400 Subject: [PATCH 028/147] metal : bug-fix when enable ggml-alloc (#2757) * metal: better memory alloc w/ concurrency dispatch The ggml-alloc should only free tensors at memory barriers. * ggml-alloc: avoid return silently In certain cases, the allocate_node() function may silently return without performing any memory allocation. --- ggml-alloc.c | 141 ++++++++++++++++++++++++++++----------------------- llama.cpp | 5 -- 2 files changed, 77 insertions(+), 69 deletions(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index 547ec0399..af4affa4e 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -68,7 +68,7 @@ struct ggml_allocr { size_t max_size; bool measure; int parse_seq[GGML_MAX_NODES]; - bool has_parse_seq; + int parse_seq_len; #ifdef GGML_ALLOCATOR_DEBUG struct ggml_tensor * allocated_tensors[1024]; @@ -239,14 +239,10 @@ static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_t } void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) { - int pos = 0; for (int i = 0; i < n; i++) { - if (list[i] != -1) { - alloc->parse_seq[pos] = list[i]; - pos++; - } + alloc->parse_seq[i] = list[i]; } - alloc->has_parse_seq = true; + alloc->parse_seq_len = n; } void ggml_allocr_reset(struct ggml_allocr * alloc) { @@ -269,7 +265,7 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) /*.max_size = */ 0, /*.measure = */ false, /*.parse_seq = */ {0}, - /*.has_parse_seq = */ false, + /*.parse_seq_len = */ 0, #ifdef GGML_ALLOCATOR_DEBUG /*.allocated_tensors = */ = {0}, #endif @@ -298,7 +294,7 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) { /*.max_size = */ 0, /*.measure = */ true, /*.parse_seq = */ {0}, - /*.has_parse_seq = */ false, + /*.parse_seq_len = */ 0, #ifdef GGML_ALLOCATOR_DEBUG /*.allocated_tensors = */ = {0}, #endif @@ -445,8 +441,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) else { AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name); node->data = parent->data; + return; } - return; } } } @@ -497,69 +493,86 @@ static size_t ggml_allocator_alloc_graph_tensors_n( allocate_node(alloc, input); } } - for (int ind = 0; ind < gf->n_nodes; ind++) { - int i; - if (alloc->has_parse_seq) { - i = alloc->parse_seq[ind]; - } else { - i = ind; - } - struct ggml_tensor * node = gf->nodes[i]; + // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers + int last_barrier_pos = 0; + int n_nodes = alloc->parse_seq_len ? alloc->parse_seq_len : gf->n_nodes; - // allocate parents (leafs) - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * parent = node->src[j]; - if (parent == NULL) { - break; + for (int ind = 0; ind < n_nodes; ind++) { + // allocate a node if there is no parse_seq or this is not a barrier + if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] != -1) { + int i = alloc->parse_seq_len ? alloc->parse_seq[ind] : ind; + struct ggml_tensor * node = gf->nodes[i]; + + // allocate parents (leafs) + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * parent = node->src[j]; + if (parent == NULL) { + break; + } + allocate_node(alloc, parent); } - allocate_node(alloc, parent); + + // allocate node + allocate_node(alloc, node); + + AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name); + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * parent = node->src[j]; + if (parent == NULL) { + break; + } + AT_PRINTF("%s", parent->name); + if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) { + AT_PRINTF(", "); + } + } + AT_PRINTF("\n"); } - // allocate node - allocate_node(alloc, node); - - AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name); - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * parent = node->src[j]; - if (parent == NULL) { - break; - } - AT_PRINTF("%s", parent->name); - if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) { - AT_PRINTF(", "); - } - } - AT_PRINTF("\n"); // update parents - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * parent = node->src[j]; - if (parent == NULL) { - break; + // update immediately if there is no parse_seq + // update only at barriers if there is parse_seq + if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] == -1) { + int update_start = alloc->parse_seq_len ? last_barrier_pos : ind; + int update_end = alloc->parse_seq_len ? ind : ind + 1; + for (int i = update_start; i < update_end; i++) { + int node_i = alloc->parse_seq_len ? alloc->parse_seq[i] : i; + struct ggml_tensor * node = gf->nodes[node_i]; + + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * parent = node->src[j]; + if (parent == NULL) { + break; + } + struct hash_node * p_hn = hash_get(ht, parent); + p_hn->n_children -= 1; + + //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views); + + if (p_hn->n_children == 0 && p_hn->n_views == 0) { + if (ggml_is_view(parent)) { + struct ggml_tensor * view_src = get_view_source(parent); + struct hash_node * view_src_hn = hash_get(ht, view_src); + view_src_hn->n_views -= 1; + AT_PRINTF("view_src %s\n", view_src->name); + if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) { + ggml_allocator_free_tensor(alloc, view_src); + } + } + else { + if (parent->data != node->data) { + ggml_allocator_free_tensor(alloc, parent); + } + } + } + } } - struct hash_node * p_hn = hash_get(ht, parent); - p_hn->n_children -= 1; - - //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views); - - if (p_hn->n_children == 0 && p_hn->n_views == 0) { - if (ggml_is_view(parent)) { - struct ggml_tensor * view_src = get_view_source(parent); - struct hash_node * view_src_hn = hash_get(ht, view_src); - view_src_hn->n_views -= 1; - AT_PRINTF("view_src %s\n", view_src->name); - if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) { - ggml_allocator_free_tensor(alloc, view_src); - } - } - else { - if (parent->data != node->data) { - ggml_allocator_free_tensor(alloc, parent); - } - } + AT_PRINTF("\n"); + if (alloc->parse_seq_len) { + last_barrier_pos = ind + 1; } } - AT_PRINTF("\n"); } // free graph outputs here that wouldn't be freed otherwise because they have no children if (outputs != NULL && outputs[g] != NULL) { diff --git a/llama.cpp b/llama.cpp index 7ee6bcdae..b5266c1e1 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2707,11 +2707,6 @@ static struct ggml_cgraph * llm_build_falcon( struct ggml_tensor * inpFF = attn_norm; cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF); - - // TODO: this is temporary needed to introduce artificial dependency between FF and ATTN - // adding this, because there seems to be a bug in the Metal concurrency optimization - // without this line, the results are non-deterministic and wrong - cur->src[2] = attn_out; offload_func(cur); cur = ggml_gelu(ctx0, cur); From 01f2224682b08185af609b28b1268b95c8b4cfa2 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 24 Aug 2023 19:58:30 +0300 Subject: [PATCH 029/147] falcon : write file type --- convert-falcon-hf-to-gguf.py | 1 + scripts/run-all-ppl.sh | 2 +- 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py index 43e208497..411cbf682 100755 --- a/convert-falcon-hf-to-gguf.py +++ b/convert-falcon-hf-to-gguf.py @@ -107,6 +107,7 @@ if "n_head_kv" in hparams: else: gguf_writer.add_head_count_kv(1) gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) +gguf_writer.add_file_type(ftype) # TOKENIZATION diff --git a/scripts/run-all-ppl.sh b/scripts/run-all-ppl.sh index bdbbb19c9..366d0866c 100755 --- a/scripts/run-all-ppl.sh +++ b/scripts/run-all-ppl.sh @@ -1,7 +1,7 @@ #!/bin/bash qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k) -args="--no-mmap -ngl 999 -t 8" +args="-ngl 999 -t 8" if [ -z "$1" ]; then echo "usage: $0 [qnt] [args]" From 0d3094f0c742ce61f84feb6e4f0b59beee6194d7 Mon Sep 17 00:00:00 2001 From: slaren Date: Thu, 24 Aug 2023 20:04:05 +0200 Subject: [PATCH 030/147] gguf : add rope_freq_base parameter for CodeLlama (#2769) --- convert.py | 43 +++++++++++++++++++++++++------------------ gguf.py | 6 +++++- llama.cpp | 20 ++++++++++++++++++-- 3 files changed, 48 insertions(+), 21 deletions(-) diff --git a/convert.py b/convert.py index b7c626d84..f335d0085 100755 --- a/convert.py +++ b/convert.py @@ -104,6 +104,8 @@ class Params: n_head_kv: int f_norm_eps: float + f_rope_freq_base: Optional[float] = None + ftype: Optional[GGMLFileType] = None # path to the directory containing the model files @@ -194,15 +196,16 @@ class Params: def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params': config = json.load(open(config_path)) - n_vocab = config["vocab_size"] if "vocab_size" in config else -1 - n_embd = config["dim"] - n_layer = config["n_layers"] - n_mult = config["multiple_of"] - n_ctx = 2048 if config["norm_eps"] == 1e-06 else 4096 # hack to determine LLaMA v1 vs v2 - n_ff = -1 - n_head = config["n_heads"] - n_head_kv = config["n_kv_heads"] if "n_kv_heads" in config else n_head - f_norm_eps = config["norm_eps"] + n_vocab = config["vocab_size"] if "vocab_size" in config else -1 + n_embd = config["dim"] + n_layer = config["n_layers"] + n_mult = config["multiple_of"] + n_ctx = 2048 if config["norm_eps"] == 1e-06 else 4096 # hack to determine LLaMA v1 vs v2 + n_ff = -1 + n_head = config["n_heads"] + n_head_kv = config["n_kv_heads"] if "n_kv_heads" in config else n_head + f_norm_eps = config["norm_eps"] + f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None if n_vocab == -1: n_vocab = model["tok_embeddings.weight"].shape[0] @@ -211,15 +214,16 @@ class Params: n_ff = model["layers.0.feed_forward.w1.weight"].shape[0] return Params( - n_vocab = n_vocab, - n_embd = n_embd, - n_mult = n_mult, - n_layer = n_layer, - n_ctx = n_ctx, - n_ff = n_ff, - n_head = n_head, - n_head_kv = n_head_kv, - f_norm_eps = f_norm_eps, + n_vocab = n_vocab, + n_embd = n_embd, + n_mult = n_mult, + n_layer = n_layer, + n_ctx = n_ctx, + n_ff = n_ff, + n_head = n_head, + n_head_kv = n_head_kv, + f_norm_eps = f_norm_eps, + f_rope_freq_base = f_rope_freq_base, ) @staticmethod @@ -754,6 +758,9 @@ class OutputFile: self.gguf.add_head_count_kv (params.n_head_kv) self.gguf.add_layer_norm_rms_eps (params.f_norm_eps) + if params.f_rope_freq_base: + self.gguf.add_rope_freq_base(params.f_rope_freq_base) + if params.ftype: self.gguf.add_file_type(params.ftype) diff --git a/gguf.py b/gguf.py index 5c37f0f0b..f4db7001b 100755 --- a/gguf.py +++ b/gguf.py @@ -47,6 +47,7 @@ KEY_ATTENTION_LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" # RoPE KEY_ROPE_DIMENSION_COUNT = "{arch}.rope.dimension_count" +KEY_ROPE_FREQ_BASE = "{arch}.rope.freq_base" KEY_ROPE_SCALE_LINEAR = "{arch}.rope.scale_linear" # tokenization @@ -663,7 +664,10 @@ class GGUFWriter: self.add_uint32( KEY_ROPE_DIMENSION_COUNT.format(arch=self.arch), count) - def add_rope_scale_linear(self, value: float): + def add_rope_freq_base(self, value: float): + self.add_float32(KEY_ROPE_FREQ_BASE.format(arch=self.arch), value) + + def add_rope_scale_linear(self, value: float): self.add_float32(KEY_ROPE_SCALE_LINEAR.format(arch=self.arch), value) def add_tokenizer_model(self, model: str): diff --git a/llama.cpp b/llama.cpp index b5266c1e1..67319396e 100644 --- a/llama.cpp +++ b/llama.cpp @@ -195,6 +195,7 @@ enum llm_kv { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, LLM_KV_ROPE_DIMENSION_COUNT, + LLM_KV_ROPE_FREQ_BASE, LLM_KV_ROPE_SCALE_LINEAR, LLM_KV_TOKENIZER_MODEL, @@ -238,6 +239,7 @@ static std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, + { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, @@ -1561,12 +1563,26 @@ static void llm_load_hparams( hparams.n_head_kv = hparams.n_head; GGUF_GET_KEY(ctx, hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV)); - // TODO: manually setting rope scale should override this + // TODO: manually setting rope freq base and scale should override this + // FIXME: partial fix when the param specified is not the default value, but + // will not work for overriding the model value to the params default + + llama_context_params defaults = llama_context_default_params(); + + // rope_freq_base + { + float ropebase = 10000.0f; + GGUF_GET_KEY(ctx, ropebase, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE)); + if (ropebase != 10000.0f && rope_freq_base == defaults.rope_freq_base) { + rope_freq_base = ropebase; + } + } + // rope_freq_scale (inverse of the kv) is optional { float ropescale = 1.0f; GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); - if (ropescale != 1.0f) { + if (ropescale != 1.0f && rope_freq_scale == defaults.rope_freq_scale) { rope_freq_scale = 1.0f/ropescale; } } From d0f77b1353fc820d1ff1e6b87bc6bedde315938d Mon Sep 17 00:00:00 2001 From: slaren Date: Thu, 24 Aug 2023 21:10:39 +0200 Subject: [PATCH 031/147] convert.py : try to determine n_ctx automatically for CodeLlama (#2770) --- convert.py | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-) diff --git a/convert.py b/convert.py index f335d0085..10276bf63 100755 --- a/convert.py +++ b/convert.py @@ -200,13 +200,23 @@ class Params: n_embd = config["dim"] n_layer = config["n_layers"] n_mult = config["multiple_of"] - n_ctx = 2048 if config["norm_eps"] == 1e-06 else 4096 # hack to determine LLaMA v1 vs v2 n_ff = -1 n_head = config["n_heads"] n_head_kv = config["n_kv_heads"] if "n_kv_heads" in config else n_head f_norm_eps = config["norm_eps"] f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None + # hack to determine LLaMA v1 vs v2 vs CodeLlama + if f_rope_freq_base and f_rope_freq_base == 1000000: + # CodeLlama + n_ctx = 16384 + elif config["norm_eps"] == 1e-05: + # LLaMA v2 + n_ctx = 4096 + else: + # LLaMA v1 + n_ctx = 2048 + if n_vocab == -1: n_vocab = model["tok_embeddings.weight"].shape[0] From 2e5f70a25fc4576e9ed78603fe493eb7702c37a3 Mon Sep 17 00:00:00 2001 From: Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com> Date: Thu, 24 Aug 2023 14:49:30 -0700 Subject: [PATCH 032/147] Added `enum` to `llama_token_get_type` return type (#2774) --- llama.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.h b/llama.h index 4e7638c04..2bcf94e0f 100644 --- a/llama.h +++ b/llama.h @@ -348,7 +348,7 @@ extern "C" { LLAMA_API float llama_token_get_score(const struct llama_context * ctx, llama_token token); - LLAMA_API llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token); + LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_context * ctx, llama_token token); // Special tokens LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx); // beginning-of-sentence From b91ad7f46134d0d051dc516eb59a76f402de55c2 Mon Sep 17 00:00:00 2001 From: Shouzheng Liu Date: Fri, 25 Aug 2023 01:58:00 -0400 Subject: [PATCH 033/147] ggml-alloc : enlarge size of parse_seq (#2776) Since we also store barriers in this array, we need to double its size. --- ggml-alloc.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index af4affa4e..1ef011654 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -8,6 +8,7 @@ #define UNUSED(x) (void)(x) #define MAX(a, b) ((a) > (b) ? (a) : (b)) +#define GGML_MAX_CONCUR (2*GGML_MAX_NODES) //#define GGML_ALLOCATOR_DEBUG @@ -67,7 +68,7 @@ struct ggml_allocr { struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE]; size_t max_size; bool measure; - int parse_seq[GGML_MAX_NODES]; + int parse_seq[GGML_MAX_CONCUR]; int parse_seq_len; #ifdef GGML_ALLOCATOR_DEBUG From 87e3733f24a85d894cc16e1cbdfa1ea1e81a76f3 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=2E=20Yusuf=20Sar=C4=B1g=C3=B6z?= Date: Fri, 25 Aug 2023 09:26:05 +0300 Subject: [PATCH 034/147] gguf : make gguf pip-installable * gitignore : add dist and rm pyproject.toml * gguf: prepare as Pip package * gguf: prepare as Pip package * gguf : fix line endings * requirements : add gguf * gguf : update readme with build notes * gguf : update readme with build notes * gguf : add notes for tests --- .gitignore | 2 +- gguf-py/LICENSE | 21 +++++++++++++ gguf-py/README.md | 55 +++++++++++++++++++++++++++++++++ gguf-py/gguf/__init__.py | 1 + gguf.py => gguf-py/gguf/gguf.py | 0 gguf-py/pyproject.toml | 28 +++++++++++++++++ gguf-py/tests/test_gguf.py | 7 +++++ requirements.txt | 1 + 8 files changed, 114 insertions(+), 1 deletion(-) create mode 100644 gguf-py/LICENSE create mode 100644 gguf-py/README.md create mode 100644 gguf-py/gguf/__init__.py rename gguf.py => gguf-py/gguf/gguf.py (100%) mode change 100755 => 100644 create mode 100644 gguf-py/pyproject.toml create mode 100644 gguf-py/tests/test_gguf.py diff --git a/.gitignore b/.gitignore index f3121794a..6cb7d9bc6 100644 --- a/.gitignore +++ b/.gitignore @@ -60,6 +60,7 @@ compile_commands.json CMakeSettings.json __pycache__ +dist zig-out/ zig-cache/ @@ -70,7 +71,6 @@ perf-*.txt examples/jeopardy/results.txt -pyproject.toml poetry.lock poetry.toml diff --git a/gguf-py/LICENSE b/gguf-py/LICENSE new file mode 100644 index 000000000..76f67efdc --- /dev/null +++ b/gguf-py/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Georgi Gerganov + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/gguf-py/README.md b/gguf-py/README.md new file mode 100644 index 000000000..03ad306ec --- /dev/null +++ b/gguf-py/README.md @@ -0,0 +1,55 @@ +## gguf + +This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302) +(GGML Universal File) format. + +See [convert-llama-hf-to-gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert-llama-hf-to-gguf.py) +as an example for its usage. + +## Installation +```sh +pip install gguf +``` + +## Development +Maintainers who participate in development of this package are advised to install it in editable mode: + +```sh +cd /path/to/llama.cpp/gguf-py + +pip install --editable . +``` + +**Note**: This may require to upgrade your Pip installation, with a message saying that editable installation currently requires `setup.py`. +In this case, upgrade Pip to the latest: + +```sh +pip install --upgrade pip +``` + +## Publishing +To publish the package, you need to have `twine` and `build` installed: + +```sh +pip install build twine +``` + +Then, folow these steps to release a new version: + +1. Update the version in `pyproject.toml`. +2. Build the package: + +```sh +python -m build +``` + +3. Upload the generated distribution archives: + +```sh +python -m twine upload dist/* +``` + +## TODO +- [ ] Add tests +- [ ] Include conversion scripts as command line entry points in this package. +- Add CI workflow for releasing the package. diff --git a/gguf-py/gguf/__init__.py b/gguf-py/gguf/__init__.py new file mode 100644 index 000000000..718ea71eb --- /dev/null +++ b/gguf-py/gguf/__init__.py @@ -0,0 +1 @@ +from .gguf import GGUFWriter diff --git a/gguf.py b/gguf-py/gguf/gguf.py old mode 100755 new mode 100644 similarity index 100% rename from gguf.py rename to gguf-py/gguf/gguf.py diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml new file mode 100644 index 000000000..a6bce9460 --- /dev/null +++ b/gguf-py/pyproject.toml @@ -0,0 +1,28 @@ +[tool.poetry] +name = "gguf" +version = "0.2.0" +description = "Write ML models in GGUF for GGML" +authors = ["GGML "] +packages = [ + {include = "gguf"}, +] +readme = "README.md" +homepage = "https://ggml.ai" +repository = "https://github.com/ggerganov/llama.cpp" +keywords = ["ggml", "gguf", "llama.cpp"] +classifiers = [ + "Programming Language :: Python :: 3", + "License :: OSI Approved :: MIT License", + "Operating System :: OS Independent", +] + +[tool.poetry.dependencies] +python = ">=3.8" +numpy = ">=1.17" + +[tool.poetry.dev-dependencies] +pytest = "^5.2" + +[build-system] +requires = ["poetry-core>=1.0.0"] +build-backend = "poetry.core.masonry.api" diff --git a/gguf-py/tests/test_gguf.py b/gguf-py/tests/test_gguf.py new file mode 100644 index 000000000..512531dd2 --- /dev/null +++ b/gguf-py/tests/test_gguf.py @@ -0,0 +1,7 @@ +import gguf + +# TODO: add tests + + +def test_write_gguf(): + pass diff --git a/requirements.txt b/requirements.txt index 6c32cbd04..7dc51edb1 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,2 +1,3 @@ numpy==1.24 sentencepiece==0.1.98 +gguf>=0.1.0 From 3f460a2b723c8b936ac29ecfd02f244b3adeba55 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 25 Aug 2023 11:55:59 +0300 Subject: [PATCH 035/147] cuda : add RoPE kernel for mode == 2 (NeoX) (#2760) * cuda : add RoPE kernel for mode == 2 (NeoX) * falcon : do not offload the embeddings layer --- ggml-cuda.cu | 58 ++++++++++++++++++++++++++++++---------------------- llama.cpp | 22 +++++++++++++++++++- 2 files changed, 54 insertions(+), 26 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 868b7a7b9..3bd1caf23 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -3907,28 +3907,27 @@ static __global__ void rope_f32(const float * x, float * dst, const int ncols, c dst[i + 1] = x0*sin_theta + x1*cos_theta; } -// TODO: this implementation is wrong! -//static __global__ void rope_neox_f32(const float * x, float * dst, const int ncols, const float p0, -// const float p_delta, const int p_delta_rows, const float theta_scale) { -// const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); -// -// if (col >= ncols) { -// return; -// } -// -// const int row = blockDim.x*blockIdx.x + threadIdx.x; -// const int i = row*ncols + col/2; -// -// const float theta = (p0 + p_delta * (row/p_delta_rows))*powf(theta_scale, col/2); -// const float sin_theta = sinf(theta); -// const float cos_theta = cosf(theta); -// -// const float x0 = x[i + 0]; -// const float x1 = x[i + ncols/2]; -// -// dst[i + 0] = x0*cos_theta - x1*sin_theta; -// dst[i + ncols/2] = x0*sin_theta + x1*cos_theta; -//} +static __global__ void rope_neox_f32(const float * x, float * dst, const int ncols, const float p0, + const float p_delta, const int p_delta_rows, const float theta_scale) { + const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); + + if (col >= ncols) { + return; + } + + const int row = blockDim.x*blockIdx.x + threadIdx.x; + const int i = row*ncols + col/2; + + const float theta = (p0 + p_delta * (row/p_delta_rows))*powf(theta_scale, col/2); + const float sin_theta = sinf(theta); + const float cos_theta = cosf(theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + ncols/2]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + ncols/2] = x0*sin_theta + x1*cos_theta; +} static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const float p, const float block_p, const float theta_scale) { const int col = blockDim.x*blockIdx.x + threadIdx.x; @@ -4799,13 +4798,21 @@ static void scale_f32_cuda(const float * x, float * dst, const float scale, cons static void rope_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p0, const float p_delta, const int p_delta_rows, const float theta_scale, cudaStream_t stream) { - GGML_ASSERT(nrows % 2 == 0); + GGML_ASSERT(nrows % 2 == 0); // GG: is this assert really needed? I don't see why const dim3 block_dims(1, 2*CUDA_ROPE_BLOCK_SIZE, 1); const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); const dim3 block_nums(nrows, num_blocks_x, 1); rope_f32<<>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale); } +static void rope_neox_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p0, + const float p_delta, const int p_delta_rows, const float theta_scale, cudaStream_t stream) { + const dim3 block_dims(1, 2*CUDA_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); + const dim3 block_nums(nrows, num_blocks_x, 1); + rope_neox_f32<<>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale); +} + static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p, const float block_p, const float theta_scale, cudaStream_t stream) { GGML_ASSERT(nrows % 4 == 0); const dim3 block_dims(4*CUDA_ROPE_BLOCK_SIZE, 1, 1); @@ -5548,8 +5555,9 @@ inline void ggml_cuda_op_rope( const float block_p = max(p - (n_ctx - 2.f), 0.f); rope_glm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, id_p, block_p, theta_scale, cudaStream_main); } else if (is_neox) { - GGML_ASSERT(false && "RoPE NeoX not implemented yet"); -#pragma message("TODO: implement RoPE NeoX for CUDA") + GGML_ASSERT(ne00 == n_dims && "ne00 != n_dims is not implemented for CUDA yet"); + const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale; + rope_neox_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main); } else { const float p0 = (((mode & 1) == 0 ? n_past : 0)) * freq_scale; rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p0, freq_scale, ne01, theta_scale, cudaStream_main); diff --git a/llama.cpp b/llama.cpp index 67319396e..52ba31d79 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1958,6 +1958,14 @@ static void llm_load_tensors( model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + vram_weights += ggml_nbytes(model.output_norm_b); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } } const uint32_t n_ff = hparams.n_ff; @@ -1967,7 +1975,7 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT + const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT auto & layer = model.layers[i]; @@ -1978,6 +1986,11 @@ static void llm_load_tensors( if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i).c_str()) >= 0) { layer.attn_norm_2 = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, backend); layer.attn_norm_2_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, backend); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(layer.attn_norm_2); + vram_weights += ggml_nbytes(layer.attn_norm_2_b); + } } layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); @@ -1985,6 +1998,13 @@ static void llm_load_tensors( layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + } } } break; default: From 6bbc598a632560cb45dd2c51ad403bda8723b629 Mon Sep 17 00:00:00 2001 From: Henri Vasserman Date: Fri, 25 Aug 2023 12:09:42 +0300 Subject: [PATCH 036/147] ROCm Port (#1087) * use hipblas based on cublas * Update Makefile for the Cuda kernels * Expand arch list and make it overrideable * Fix multi GPU on multiple amd architectures with rocblas_initialize() (#5) * add hipBLAS to README * new build arg LLAMA_CUDA_MMQ_Y * fix half2 decomposition * Add intrinsics polyfills for AMD * AMD assembly optimized __dp4a * Allow overriding CC_TURING * use "ROCm" instead of "CUDA" * ignore all build dirs * Add Dockerfiles * fix llama-bench * fix -nommq help for non CUDA/HIP --------- Co-authored-by: YellowRoseCx <80486540+YellowRoseCx@users.noreply.github.com> Co-authored-by: ardfork <134447697+ardfork@users.noreply.github.com> Co-authored-by: funnbot <22226942+funnbot@users.noreply.github.com> Co-authored-by: Engininja2 <139037756+Engininja2@users.noreply.github.com> Co-authored-by: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Co-authored-by: jammm <2500920+jammm@users.noreply.github.com> Co-authored-by: jdecourval <7315817+jdecourval@users.noreply.github.com> --- .devops/full-rocm.Dockerfile | 44 +++++++ .devops/main-rocm.Dockerfile | 44 +++++++ .dockerignore | 9 +- .gitignore | 15 +-- CMakeLists.txt | 38 ++++++ Makefile | 24 ++++ README.md | 29 +++++ common/common.cpp | 4 +- examples/llama-bench/llama-bench.cpp | 4 +- ggml-cuda.cu | 173 ++++++++++++++++++++++----- ggml-cuda.h | 8 ++ llama.cpp | 2 +- 12 files changed, 335 insertions(+), 59 deletions(-) create mode 100644 .devops/full-rocm.Dockerfile create mode 100644 .devops/main-rocm.Dockerfile diff --git a/.devops/full-rocm.Dockerfile b/.devops/full-rocm.Dockerfile new file mode 100644 index 000000000..6c521e9b4 --- /dev/null +++ b/.devops/full-rocm.Dockerfile @@ -0,0 +1,44 @@ +ARG UBUNTU_VERSION=22.04 + +# This needs to generally match the container host's environment. +ARG ROCM_VERSION=5.6 + +# Target the CUDA build image +ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete + +FROM ${BASE_ROCM_DEV_CONTAINER} as build + +# Unless otherwise specified, we make a fat build. +# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 +# This is mostly tied to rocBLAS supported archs. +ARG ROCM_DOCKER_ARCH=\ + gfx803 \ + gfx900 \ + gfx906 \ + gfx908 \ + gfx90a \ + gfx1010 \ + gfx1030 \ + gfx1100 \ + gfx1101 \ + gfx1102 + +COPY requirements.txt requirements.txt + +RUN pip install --upgrade pip setuptools wheel \ + && pip install -r requirements.txt + +WORKDIR /app + +COPY . . + +# Set nvcc architecture +ENV GPU_TARGETS=${ROCM_DOCKER_ARCH} +# Enable ROCm +ENV LLAMA_HIPBLAS=1 +ENV CC=/opt/rocm/llvm/bin/clang +ENV CXX=/opt/rocm/llvm/bin/clang++ + +RUN make + +ENTRYPOINT ["/app/.devops/tools.sh"] diff --git a/.devops/main-rocm.Dockerfile b/.devops/main-rocm.Dockerfile new file mode 100644 index 000000000..789deff6d --- /dev/null +++ b/.devops/main-rocm.Dockerfile @@ -0,0 +1,44 @@ +ARG UBUNTU_VERSION=22.04 + +# This needs to generally match the container host's environment. +ARG ROCM_VERSION=5.6 + +# Target the CUDA build image +ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete + +FROM ${BASE_ROCM_DEV_CONTAINER} as build + +# Unless otherwise specified, we make a fat build. +# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 +# This is mostly tied to rocBLAS supported archs. +ARG ROCM_DOCKER_ARCH=\ + gfx803 \ + gfx900 \ + gfx906 \ + gfx908 \ + gfx90a \ + gfx1010 \ + gfx1030 \ + gfx1100 \ + gfx1101 \ + gfx1102 + +COPY requirements.txt requirements.txt + +RUN pip install --upgrade pip setuptools wheel \ + && pip install -r requirements.txt + +WORKDIR /app + +COPY . . + +# Set nvcc architecture +ENV GPU_TARGETS=${ROCM_DOCKER_ARCH} +# Enable ROCm +ENV LLAMA_HIPBLAS=1 +ENV CC=/opt/rocm/llvm/bin/clang +ENV CXX=/opt/rocm/llvm/bin/clang++ + +RUN make + +ENTRYPOINT [ "/app/main" ] diff --git a/.dockerignore b/.dockerignore index 462fac23a..c6ef6c86c 100644 --- a/.dockerignore +++ b/.dockerignore @@ -5,14 +5,7 @@ .vscode/ .DS_Store -build/ -build-em/ -build-debug/ -build-release/ -build-static/ -build-no-accel/ -build-sanitize-addr/ -build-sanitize-thread/ +build*/ models/* diff --git a/.gitignore b/.gitignore index 6cb7d9bc6..e5faab774 100644 --- a/.gitignore +++ b/.gitignore @@ -16,20 +16,7 @@ .vs/ .vscode/ -build/ -build-em/ -build-debug/ -build-release/ -build-ci-debug/ -build-ci-release/ -build-static/ -build-cublas/ -build-opencl/ -build-metal/ -build-mpi/ -build-no-accel/ -build-sanitize-addr/ -build-sanitize-thread/ +build*/ out/ tmp/ diff --git a/CMakeLists.txt b/CMakeLists.txt index bb63ef98e..ba008bcc6 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -74,6 +74,7 @@ set(LLAMA_CUDA_DMMV_X "32" CACHE STRING "llama: x stride for dmmv CUDA kern set(LLAMA_CUDA_MMV_Y "1" CACHE STRING "llama: y block size for mmv CUDA kernels") option(LLAMA_CUDA_F16 "llama: use 16 bit floats for some calculations" OFF) set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K") +option(LLAMA_HIPBLAS "llama: use hipBLAS" OFF) option(LLAMA_CLBLAST "llama: use CLBlast" OFF) option(LLAMA_METAL "llama: use Metal" OFF) option(LLAMA_MPI "llama: use MPI" OFF) @@ -352,6 +353,43 @@ if (LLAMA_CLBLAST) endif() endif() +if (LLAMA_HIPBLAS) + list(APPEND CMAKE_PREFIX_PATH /opt/rocm) + + if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang") + message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang") + endif() + if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang") + message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++") + endif() + + find_package(hip) + find_package(hipblas) + find_package(rocblas) + + if (${hipblas_FOUND} AND ${hip_FOUND}) + message(STATUS "HIP and hipBLAS found") + add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS) + add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h) + if (LLAMA_CUDA_FORCE_DMMV) + target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV) + endif() + target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X}) + target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y}) + target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER}) + target_compile_definitions(ggml-rocm PRIVATE CC_TURING=1000000000) + set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX) + target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas) + + if (LLAMA_STATIC) + message(FATAL_ERROR "Static linking not supported for HIP/ROCm") + endif() + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm) + else() + message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm") + endif() +endif() + if (LLAMA_ALL_WARNINGS) if (NOT MSVC) set(c_flags diff --git a/Makefile b/Makefile index d31acc450..a3400e491 100644 --- a/Makefile +++ b/Makefile @@ -280,6 +280,30 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h $(CXX) $(CXXFLAGS) -c $< -o $@ endif # LLAMA_CLBLAST +ifdef LLAMA_HIPBLAS + ROCM_PATH ?= /opt/rocm + HIPCC ?= $(ROCM_PATH)/bin/hipcc + GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch) + LLAMA_CUDA_DMMV_X ?= 32 + LLAMA_CUDA_MMV_Y ?= 1 + LLAMA_CUDA_KQUANTS_ITER ?= 2 + CFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS + CXXFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS + LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib + LDFLAGS += -lhipblas -lamdhip64 -lrocblas + HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) + HIPFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) + HIPFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) + HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER) + HIPFLAGS += -DCC_TURING=1000000000 +ifdef LLAMA_CUDA_FORCE_DMMV + HIPFLAGS += -DGGML_CUDA_FORCE_DMMV +endif # LLAMA_CUDA_FORCE_DMMV + OBJS += ggml-cuda.o +ggml-cuda.o: ggml-cuda.cu ggml-cuda.h + $(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $< +endif # LLAMA_HIPBLAS + ifdef LLAMA_METAL CFLAGS += -DGGML_USE_METAL -DGGML_METAL_NDEBUG CXXFLAGS += -DGGML_USE_METAL diff --git a/README.md b/README.md index eebb11392..95471fdbb 100644 --- a/README.md +++ b/README.md @@ -422,6 +422,35 @@ Building the program with BLAS support may lead to some performance improvements | LLAMA_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. | | LLAMA_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. | +- #### hipBLAS + + This provide BLAS acceleation on HIP supported GPU like AMD GPU. + Make sure to have ROCm installed. + You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html). + Windows support is coming soon... + + - Using `make`: + ```bash + make LLAMA_HIPBLAS=1 + ``` + - Using `CMake`: + ```bash + mkdir build + cd build + CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON + cmake --build . + ``` + + The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used. + If your GPU is not officialy supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3. + The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above): + + | Option | Legal values | Default | Description | + |-------------------------|------------------------|---------|-------------| + | LLAMA_CUDA_DMMV_X | Positive integer >= 32 | 32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. | + | LLAMA_CUDA_MMV_Y | Positive integer | 1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. | + | LLAMA_CUDA_KQUANTS_ITER | 1 or 2 | 2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. | + - #### CLBlast OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU. diff --git a/common/common.cpp b/common/common.cpp index 53002ba30..ff19ec4e5 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -613,9 +613,11 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stdout, " how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n"); fprintf(stdout, " -mg i, --main-gpu i the GPU to use for scratch and small tensors\n"); fprintf(stdout, " -lv, --low-vram don't allocate VRAM scratch buffer\n"); +#ifdef GGML_USE_CUBLAS fprintf(stdout, " -nommq, --no-mul-mat-q\n"); - fprintf(stdout, " use cuBLAS instead of custom mul_mat_q CUDA kernels.\n"); + fprintf(stdout, " use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n"); fprintf(stdout, " Not recommended since this is both slower and uses more VRAM.\n"); +#endif // GGML_USE_CUBLAS #endif fprintf(stdout, " --mtest compute maximum memory usage\n"); fprintf(stdout, " --export export the computation graph to 'llama.ggml'\n"); diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 36057bfca..7a2811584 100755 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -18,9 +18,7 @@ #include "llama.h" #include "common.h" #include "build-info.h" -#ifdef GGML_USE_CUBLAS #include "ggml-cuda.h" -#endif // utils static uint64_t get_time_ns() { @@ -504,7 +502,7 @@ struct test { static std::string get_backend() { if (cuda) { - return "CUDA"; + return GGML_CUDA_NAME; } if (opencl) { return "OpenCL"; diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 3bd1caf23..83d53c13c 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -6,15 +6,116 @@ #include #include +#if defined(GGML_USE_HIPBLAS) +#include +#include +#include +#ifdef __HIP_PLATFORM_AMD__ +// for rocblas_initialize() +#include "rocblas/rocblas.h" +#endif +#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F +#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F +#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT +#define CUBLAS_OP_N HIPBLAS_OP_N +#define CUBLAS_OP_T HIPBLAS_OP_T +#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS +#define CUBLAS_TF32_TENSOR_OP_MATH 0 +#define CUDA_R_16F HIPBLAS_R_16F +#define CUDA_R_32F HIPBLAS_R_32F +#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width) +#define cublasCreate hipblasCreate +#define cublasGemmEx hipblasGemmEx +#define cublasHandle_t hipblasHandle_t +#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS +#define cublasSetStream hipblasSetStream +#define cublasSgemm hipblasSgemm +#define cublasStatus_t hipblasStatus_t +#define cudaDeviceProp hipDeviceProp_t +#define cudaDeviceSynchronize hipDeviceSynchronize +#define cudaError_t hipError_t +#define cudaEventCreateWithFlags hipEventCreateWithFlags +#define cudaEventDisableTiming hipEventDisableTiming +#define cudaEventRecord hipEventRecord +#define cudaEvent_t hipEvent_t +#define cudaEventDestroy hipEventDestroy +#define cudaFree hipFree +#define cudaFreeHost hipHostFree +#define cudaGetDevice hipGetDevice +#define cudaGetDeviceCount hipGetDeviceCount +#define cudaGetDeviceProperties hipGetDeviceProperties +#define cudaGetErrorString hipGetErrorString +#define cudaGetLastError hipGetLastError +#define cudaMalloc hipMalloc +#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault) +#define cudaMemcpy hipMemcpy +#define cudaMemcpy2DAsync hipMemcpy2DAsync +#define cudaMemcpyAsync hipMemcpyAsync +#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice +#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost +#define cudaMemcpyHostToDevice hipMemcpyHostToDevice +#define cudaMemcpyKind hipMemcpyKind +#define cudaMemset hipMemset +#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize +#define cudaSetDevice hipSetDevice +#define cudaStreamCreateWithFlags hipStreamCreateWithFlags +#define cudaStreamNonBlocking hipStreamNonBlocking +#define cudaStreamSynchronize hipStreamSynchronize +#define cudaStreamWaitEvent(stream, event) hipStreamWaitEvent(stream, event, 0) +#define cudaStream_t hipStream_t +#define cudaSuccess hipSuccess +#else #include #include #include +#endif #include "ggml-cuda.h" #include "ggml.h" #define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products +#ifndef CC_TURING #define CC_TURING 700 +#endif + +#if defined(GGML_USE_HIPBLAS) +#define __CUDA_ARCH__ 1300 + +typedef int8_t int8x4_t __attribute__((ext_vector_type(4))); +static __device__ __forceinline__ int __vsubss4(const int a, const int b) { + const int8x4_t va = reinterpret_cast(a); + const int8x4_t vb = reinterpret_cast(b); + const int8x4_t c = __builtin_elementwise_sub_sat(va, vb); + return reinterpret_cast(c); +} + +static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { +#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) + c = __builtin_amdgcn_sdot4(a, b, c, false); +#elif defined(__gfx1100__) + c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); +#elif defined(__gfx1010__) || defined(__gfx900__) + int tmp1; + int tmp2; + asm("\n \ + v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \ + v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \ + v_add3_u32 %0, %1, %2, %0 \n \ + v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \ + v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \ + v_add3_u32 %0, %1, %2, %0 \n \ + " + : "+v"(c), "=&v"(tmp1), "=&v"(tmp2) + : "v"(a), "v"(b) + ); +#else + const int8x4_t va = reinterpret_cast(a); + const int8x4_t vb = reinterpret_cast(b); + c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3]; +#endif + return c; +} +#endif #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data @@ -424,8 +525,8 @@ static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const in static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){ const block_q4_1 * x = (const block_q4_1 *) vx; - const dfloat d = x[ib].dm.x; - const dfloat m = x[ib].dm.y; + const dfloat d = __low2half(x[ib].dm); + const dfloat m = __high2half(x[ib].dm); const int vui = x[ib].qs[iqs]; @@ -467,8 +568,8 @@ static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const in static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){ const block_q5_1 * x = (const block_q5_1 *) vx; - const dfloat d = x[ib].dm.x; - const dfloat m = x[ib].dm.y; + const dfloat d = __low2half(x[ib].dm); + const dfloat m = __high2half(x[ib].dm); uint32_t qh; memcpy(&qh, x[ib].qh, sizeof(qh)); @@ -520,8 +621,8 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float const uint8_t q = x[i].qs[32*n + l]; float * y = yy + i*QK_K + 128*n; - float dall = x[i].dm.x; - float dmin = x[i].dm.y; + float dall = __low2half(x[i].dm); + float dmin = __high2half(x[i].dm); y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4); y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); @@ -531,8 +632,8 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, float const int il = tid%16; // 0...15 const uint8_t q = x[i].qs[il] >> (2*is); float * y = yy + i*QK_K + 16*is + il; - float dall = x[i].dm.x; - float dmin = x[i].dm.y; + float dall = __low2half(x[i].dm); + float dmin = __high2half(x[i].dm); y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4); #endif @@ -618,8 +719,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, float float * y = yy + i*QK_K + 64*il + n*ir; - const float dall = x[i].dm.x; - const float dmin = x[i].dm.y; + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); const uint8_t * q = x[i].qs + 32*il + n*ir; @@ -657,8 +758,8 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, float float * y = yy + i*QK_K + 64*il + 2*ir; - const float dall = x[i].dm.x; - const float dmin = x[i].dm.y; + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); const uint8_t * ql = x[i].qs + 32*il + 2*ir; const uint8_t * qh = x[i].qh + 2*ir; @@ -770,8 +871,8 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * y = yy + i * QK_K + y_offset; const uint8_t * q = x[i].qs + q_offset; - const float dall = x[i].dm.x; - const float dmin = x[i].dm.y; + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset); aux[0] = a[0] & 0x0f0f0f0f; @@ -991,8 +1092,8 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * y1 = yy + i*QK_K + y_offset; const float * y2 = y1 + 128; - const float dall = x[i].dm.x; - const float dmin = x[i].dm.y; + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); const uint16_t * a = (const uint16_t *)x[i].scales; aux[0] = a[im+0] & kmask1; @@ -1124,8 +1225,8 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * y1 = yy + i*QK_K + y_offset; const float * y2 = y1 + 128; - const float dall = x[i].dm.x; - const float dmin = x[i].dm.y; + const float dall = __low2half(x[i].dm); + const float dmin = __high2half(x[i].dm); const uint16_t * a = (const uint16_t *)x[i].scales; aux[0] = a[im+0] & kmask1; @@ -1348,8 +1449,8 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest return; } - y[ib].ds.x = d; - y[ib].ds.y = sum; + reinterpret_cast(y[ib].ds.x) = d; + reinterpret_cast(y[ib].ds.y) = sum; } template @@ -2346,7 +2447,7 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1( u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); } - return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, bq8_1->ds.x); + return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, __low2half(bq8_1->ds)); } template static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { @@ -2432,7 +2533,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1( #pragma unroll for (int i = 0; i < QR2_K; ++ i) { u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); - d8[i] = bq8_1[bq8_offset + i].ds.x; + d8[i] = __low2half(bq8_1[bq8_offset + i].ds); } return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8); @@ -2551,7 +2652,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1( #pragma unroll for (int i = 0; i < QR3_K; ++i) { u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); - d8[i] = bq8_1[bq8_offset + i].ds.x; + d8[i] = __low2half(bq8_1[bq8_offset + i].ds); } return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8); @@ -2720,7 +2821,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1( for (int i = 0; i < QR4_K; ++i) { const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; - d8[i] = bq8i->ds.x; + d8[i] = __low2half(bq8i->ds); const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); u[2*i+0] = q8[0]; @@ -2747,8 +2848,8 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1( const float dall = bq4_K->d[0]; const float dmin = bq4_K->d[1]; - const float d8_1 = bq8_1[0].ds.x; - const float d8_2 = bq8_1[1].ds.x; + const float d8_1 = __low2float(bq8_1[0].ds); + const float d8_2 = __low2float(bq8_1[1].ds); const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); @@ -2901,7 +3002,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1( #pragma unroll for (int i = 0; i < QR5_K; ++i) { const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; - d8[i] = bq8i->ds.x; + d8[i] = __low2float(bq8i->ds); const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); u[2*i+0] = q8[0]; @@ -2919,8 +3020,8 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1( const float d = bq5_K->d; - const float d8_1 = bq8_1[0].ds.x; - const float d8_2 = bq8_1[1].ds.x; + const float d8_1 = __low2half(bq8_1[0].ds); + const float d8_2 = __low2half(bq8_1[1].ds); const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); @@ -3075,7 +3176,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1( #pragma unroll for (int i = 0; i < QR6_K; ++i) { u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); - d8[i] = bq8_1[bq8_offset + 2*i].ds.x; + d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds); } return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8); @@ -3243,7 +3344,7 @@ static __device__ __forceinline__ void mul_mat_q( *dsi_dst = *dsi_src; } else { float * dfi_dst = (float *) dsi_dst; - *dfi_dst = (*dsi_src).x; + *dfi_dst = __low2half(*dsi_src); } } @@ -4944,10 +5045,18 @@ void ggml_init_cublas() { static bool initialized = false; if (!initialized) { + +#ifdef __HIP_PLATFORM_AMD__ + // Workaround for a rocBLAS bug when using multiple graphics cards: + // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346 + rocblas_initialize(); + CUDA_CHECK(cudaDeviceSynchronize()); +#endif + CUDA_CHECK(cudaGetDeviceCount(&g_device_count)); GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES); int64_t total_vram = 0; - fprintf(stderr, "%s: found %d CUDA devices:\n", __func__, g_device_count); + fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count); for (int id = 0; id < g_device_count; ++id) { cudaDeviceProp prop; CUDA_CHECK(cudaGetDeviceProperties(&prop, id)); diff --git a/ggml-cuda.h b/ggml-cuda.h index f66bb1678..a72e82069 100644 --- a/ggml-cuda.h +++ b/ggml-cuda.h @@ -2,6 +2,14 @@ #include "ggml.h" +#ifdef GGML_USE_HIPBLAS +#define GGML_CUDA_NAME "ROCm" +#define GGML_CUBLAS_NAME "hipBLAS" +#else +#define GGML_CUDA_NAME "CUDA" +#define GGML_CUBLAS_NAME "cuBLAS" +#endif + #ifdef __cplusplus extern "C" { #endif diff --git a/llama.cpp b/llama.cpp index 52ba31d79..d12b6d1cb 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1836,7 +1836,7 @@ static void llm_load_tensors( (void) main_gpu; (void) mul_mat_q; #if defined(GGML_USE_CUBLAS) - LLAMA_LOG_INFO("%s: using CUDA for GPU acceleration\n", __func__); + LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__); ggml_cuda_set_main_device(main_gpu); ggml_cuda_set_mul_mat_q(mul_mat_q); #define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU From 8194cd8772c58b8a43aa07a2fc468f5366d7e320 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=2E=20Yusuf=20Sar=C4=B1g=C3=B6z?= Date: Fri, 25 Aug 2023 12:43:41 +0300 Subject: [PATCH 037/147] gguf : export objects to user code (#2780) * gguf export more objects to user code * gguf export all objects to user code for now * gguf : bump version --- gguf-py/gguf/__init__.py | 2 +- gguf-py/pyproject.toml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/gguf-py/gguf/__init__.py b/gguf-py/gguf/__init__.py index 718ea71eb..f9b70a85b 100644 --- a/gguf-py/gguf/__init__.py +++ b/gguf-py/gguf/__init__.py @@ -1 +1 @@ -from .gguf import GGUFWriter +from .gguf import * diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index a6bce9460..cc70e28b7 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.2.0" +version = "0.2.1" description = "Write ML models in GGUF for GGML" authors = ["GGML "] packages = [ From 5439a0ab57c16b556ffa91a0953df5e46b1e7fb4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 25 Aug 2023 13:03:25 +0300 Subject: [PATCH 038/147] ci : pip install gguf in editable mode (#2782) ggml-ci --- ci/run.sh | 1 + 1 file changed, 1 insertion(+) diff --git a/ci/run.sh b/ci/run.sh index 54ba6d710..e1486e7c1 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -391,6 +391,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then ln -sfn ${mnt_models} ${SRC}/models-mnt python3 -m pip install -r ${SRC}/requirements.txt + python3 -m pip install --editable gguf-py fi ret=0 From 29674ab4e847fcaba60cc6558f0d46d5f74ae279 Mon Sep 17 00:00:00 2001 From: Jhen-Jie Hong Date: Fri, 25 Aug 2023 18:32:45 +0800 Subject: [PATCH 039/147] server : display token probabilities in the UI (#2489) * server : add n_probs param in chat UI * server : keep message data array & show in probabilites component * server : add simple popover component * server : fix completion_probabilities undefined if not set n_probs * server : implement Probabilites * server : handle bytes * server : make n_probs max to 10 for easy scroll * server : adjust for dark/light mode * server : Fix regenerated prompt * server : update index.html.hpp * server : convert prob to percentage + show original value as div title * server : fix Probabilites not used if included empty str * server : skip byte pair in display probabilites * server : remove array check of completion_probabilities in messages * skip empty array or byte pair (> 1) in Probabilites * generate index.html.hpp * fix incorrect prob convert if the str is already a known token * use final response to show probabilities on stop * revert unnecessary change * correct probabilites usage * remove unused function * always send partial response for get correct probs of last to_send * fix typo * fix content of format_final_response * refactor probs render & make pColor transparent if not found * send empty string when got stop_pos in partial * avoid unnecessary empty data event & send rest of partial tokens on stop * use
    for new line * skip -1 tok in loop to avoid send '' on end * trim last new lines on stop * revert unnecessary change --- examples/server/index.html.hpp | 2808 +++++++++++++++++------------ examples/server/public/index.html | 245 ++- examples/server/server.cpp | 86 +- 3 files changed, 1979 insertions(+), 1160 deletions(-) diff --git a/examples/server/index.html.hpp b/examples/server/index.html.hpp index 76851d9c6..84e6f97ba 100644 --- a/examples/server/index.html.hpp +++ b/examples/server/index.html.hpp @@ -152,149 +152,191 @@ unsigned char index_html[] = { 0x6f, 0x70, 0x65, 0x6e, 0x5d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x30, 0x2e, 0x35, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, - 0x65, 0x61, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, - 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x35, 0x70, 0x78, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x65, 0x78, 0x2d, - 0x67, 0x72, 0x6f, 0x77, 0x3a, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x31, 0x30, - 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x70, 0x72, 0x65, 0x20, 0x63, 0x6f, 0x64, 0x65, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, - 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x62, 0x6c, 0x6f, 0x63, 0x6b, 0x3b, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x2d, 0x73, + 0x65, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, + 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x30, 0x2e, 0x33, 0x65, + 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x72, + 0x64, 0x65, 0x72, 0x2d, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x3a, 0x20, + 0x31, 0x70, 0x78, 0x20, 0x73, 0x6f, 0x6c, 0x69, 0x64, 0x20, 0x23, 0x63, + 0x63, 0x63, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, + 0x3a, 0x20, 0x61, 0x62, 0x73, 0x6f, 0x6c, 0x75, 0x74, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, - 0x23, 0x32, 0x32, 0x32, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x64, 0x64, 0x64, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x64, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x66, 0x6f, 0x6e, 0x74, 0x2d, 0x66, 0x61, 0x6d, 0x69, 0x6c, 0x79, 0x3a, - 0x20, 0x6d, 0x6f, 0x6e, 0x6f, 0x73, 0x70, 0x61, 0x63, 0x65, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, - 0x67, 0x3a, 0x20, 0x30, 0x2e, 0x31, 0x65, 0x6d, 0x20, 0x30, 0x2e, 0x33, - 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, - 0x72, 0x64, 0x65, 0x72, 0x2d, 0x72, 0x61, 0x64, 0x69, 0x75, 0x73, 0x3a, - 0x20, 0x33, 0x70, 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, - 0x74, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20, - 0x30, 0x2e, 0x35, 0x65, 0x6d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, - 0x62, 0x6c, 0x6f, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, - 0x2c, 0x20, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2d, 0x61, 0x6c, - 0x69, 0x67, 0x6e, 0x3a, 0x20, 0x63, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x3b, + 0x77, 0x68, 0x69, 0x74, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x30, 0x2e, + 0x32, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, + 0x6f, 0x78, 0x2d, 0x73, 0x68, 0x61, 0x64, 0x6f, 0x77, 0x3a, 0x20, 0x30, + 0x20, 0x30, 0x20, 0x31, 0x30, 0x70, 0x78, 0x20, 0x72, 0x67, 0x62, 0x61, + 0x28, 0x30, 0x2c, 0x20, 0x30, 0x2c, 0x20, 0x30, 0x2c, 0x20, 0x30, 0x2e, + 0x31, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, + 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x35, 0x70, 0x78, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x66, 0x6c, 0x65, 0x78, 0x2d, 0x67, 0x72, 0x6f, + 0x77, 0x3a, 0x20, 0x31, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6e, 0x74, 0x2d, 0x73, 0x69, 0x7a, 0x65, - 0x3a, 0x20, 0x38, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x38, 0x38, 0x38, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x3c, 0x2f, - 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x3c, 0x73, - 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x6d, 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x20, 0x68, 0x2c, - 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x65, 0x66, 0x66, - 0x65, 0x63, 0x74, 0x2c, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, - 0x64, 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x2c, 0x20, 0x75, - 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x75, 0x73, - 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x75, 0x73, 0x65, - 0x52, 0x65, 0x66, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x66, 0x72, - 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x2e, 0x6a, - 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, - 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x20, - 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, - 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x3b, + 0x70, 0x72, 0x65, 0x20, 0x63, 0x6f, 0x64, 0x65, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, + 0x3a, 0x20, 0x62, 0x6c, 0x6f, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, + 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x32, 0x32, + 0x32, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6c, + 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x64, 0x64, 0x64, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x64, 0x65, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6e, + 0x74, 0x2d, 0x66, 0x61, 0x6d, 0x69, 0x6c, 0x79, 0x3a, 0x20, 0x6d, 0x6f, + 0x6e, 0x6f, 0x73, 0x70, 0x61, 0x63, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, + 0x30, 0x2e, 0x31, 0x65, 0x6d, 0x20, 0x30, 0x2e, 0x33, 0x65, 0x6d, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x6f, 0x72, 0x64, 0x65, + 0x72, 0x2d, 0x72, 0x61, 0x64, 0x69, 0x75, 0x73, 0x3a, 0x20, 0x33, 0x70, + 0x78, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x6d, 0x61, 0x72, 0x67, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x35, + 0x65, 0x6d, 0x20, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x62, 0x6c, 0x6f, + 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x2c, 0x20, 0x66, + 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x2d, 0x61, 0x6c, 0x69, 0x67, 0x6e, + 0x3a, 0x20, 0x63, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x6f, + 0x74, 0x65, 0x72, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x66, 0x6f, 0x6e, 0x74, 0x2d, 0x73, 0x69, 0x7a, 0x65, 0x3a, 0x20, 0x38, + 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x23, 0x38, 0x38, 0x38, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x40, 0x6d, + 0x65, 0x64, 0x69, 0x61, 0x20, 0x28, 0x70, 0x72, 0x65, 0x66, 0x65, 0x72, + 0x73, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x73, 0x63, 0x68, 0x65, + 0x6d, 0x65, 0x3a, 0x20, 0x64, 0x61, 0x72, 0x6b, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, + 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, + 0x72, 0x3a, 0x20, 0x62, 0x6c, 0x61, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, + 0x20, 0x20, 0x3c, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, 0x79, + 0x70, 0x65, 0x3d, 0x22, 0x6d, 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, - 0x7b, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76, - 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, - 0x20, 0x27, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, 0x73, 0x63, 0x68, 0x65, - 0x6d, 0x61, 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, - 0x72, 0x2e, 0x6d, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x6d, - 0x70, 0x74, 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73, - 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x73, 0x61, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65, 0x6e, 0x20, - 0x75, 0x73, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6c, 0x6c, 0x61, - 0x6d, 0x61, 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, 0x69, 0x65, 0x6e, 0x64, - 0x6c, 0x79, 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, 0x6f, 0x74, 0x2e, 0x20, - 0x72, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x64, 0x20, 0x69, 0x6e, 0x20, 0x73, - 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, - 0x77, 0x6e, 0x2e, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, - 0x7b, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c, - 0x6e, 0x7b, 0x7b, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d, - 0x5c, 0x6e, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, - 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, - 0x20, 0x22, 0x7b, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20, - 0x7b, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, - 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20, - 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x68, 0x61, 0x72, 0x3a, 0x20, 0x22, 0x6c, 0x6c, 0x61, - 0x6d, 0x61, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, - 0x73, 0x65, 0x72, 0x3a, 0x20, 0x22, 0x55, 0x73, 0x65, 0x72, 0x22, 0x2c, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x65, - 0x64, 0x69, 0x63, 0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, - 0x74, 0x75, 0x72, 0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, - 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, 0x35, 0x36, 0x2c, - 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, - 0x62, 0x6c, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x2c, - 0x20, 0x2d, 0x31, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, - 0x74, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, - 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x31, 0x2e, 0x31, 0x38, 0x2c, 0x20, 0x2f, - 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, - 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, - 0x6f, 0x70, 0x5f, 0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, 0x20, 0x2f, 0x2f, - 0x20, 0x3c, 0x3d, 0x20, 0x30, 0x20, 0x74, 0x6f, 0x20, 0x75, 0x73, 0x65, - 0x20, 0x76, 0x6f, 0x63, 0x61, 0x62, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x3a, - 0x20, 0x30, 0x2e, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, - 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x3a, - 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, - 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, - 0x6c, 0x5f, 0x70, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, + 0x2c, 0x20, 0x68, 0x2c, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, + 0x20, 0x65, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x63, 0x6f, 0x6d, + 0x70, 0x75, 0x74, 0x65, 0x64, 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, + 0x72, 0x2c, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, + 0x2c, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, + 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x2c, 0x20, 0x43, 0x6f, 0x6d, + 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x69, 0x6e, 0x64, 0x65, + 0x78, 0x2e, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, + 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, + 0x73, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, + 0x72, 0x74, 0x20, 0x7b, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, + 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x7d, 0x20, 0x66, + 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, 0x73, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, 0x61, + 0x6d, 0x6d, 0x61, 0x72, 0x2e, 0x6d, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, + 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, + 0x61, 0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, + 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, 0x73, + 0x20, 0x69, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, + 0x73, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, + 0x65, 0x6e, 0x20, 0x55, 0x73, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, 0x20, + 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, 0x69, + 0x65, 0x6e, 0x64, 0x6c, 0x79, 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, 0x6f, + 0x74, 0x2e, 0x20, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x69, 0x73, 0x20, + 0x68, 0x65, 0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, 0x6b, 0x69, 0x6e, + 0x64, 0x2c, 0x20, 0x68, 0x6f, 0x6e, 0x65, 0x73, 0x74, 0x2c, 0x20, 0x67, + 0x6f, 0x6f, 0x64, 0x20, 0x61, 0x74, 0x20, 0x77, 0x72, 0x69, 0x74, 0x69, + 0x6e, 0x67, 0x2c, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6e, 0x65, 0x76, 0x65, + 0x72, 0x20, 0x66, 0x61, 0x69, 0x6c, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x61, + 0x6e, 0x73, 0x77, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x72, 0x65, + 0x71, 0x75, 0x65, 0x73, 0x74, 0x73, 0x20, 0x69, 0x6d, 0x6d, 0x65, 0x64, + 0x69, 0x61, 0x74, 0x65, 0x6c, 0x79, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x77, + 0x69, 0x74, 0x68, 0x20, 0x70, 0x72, 0x65, 0x63, 0x69, 0x73, 0x69, 0x6f, + 0x6e, 0x2e, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, 0x7b, + 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c, 0x6e, + 0x7b, 0x7b, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d, 0x5c, + 0x6e, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x2c, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, + 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, + 0x22, 0x7b, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20, 0x7b, + 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22, 0x2c, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, + 0x63, 0x72, 0x69, 0x70, 0x74, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20, 0x22, + 0x63, 0x68, 0x61, 0x74, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x68, 0x61, 0x72, 0x3a, 0x20, 0x22, 0x4c, 0x6c, 0x61, 0x6d, + 0x61, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x3a, 0x20, 0x22, 0x55, 0x73, 0x65, 0x72, 0x22, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, + 0x69, 0x63, 0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, + 0x75, 0x72, 0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, + 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, 0x35, 0x36, 0x2c, 0x20, + 0x2f, 0x2f, 0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, + 0x6c, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x2c, 0x20, + 0x2d, 0x31, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, + 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, + 0x74, 0x79, 0x3a, 0x20, 0x31, 0x2e, 0x31, 0x38, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, - 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, - 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, - 0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, - 0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, - 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x72, 0x65, - 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, - 0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, - 0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, - 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, - 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, - 0x20, 0x30, 0x2f, 0x31, 0x2f, 0x32, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, - 0x75, 0x3a, 0x20, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x74, 0x61, 0x72, - 0x67, 0x65, 0x74, 0x20, 0x65, 0x6e, 0x74, 0x72, 0x6f, 0x70, 0x79, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, - 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x3a, 0x20, 0x30, 0x2e, 0x31, 0x2c, - 0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x61, 0x72, 0x6e, 0x69, 0x6e, 0x67, - 0x20, 0x72, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x27, 0x27, 0x2c, + 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, + 0x70, 0x5f, 0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, + 0x3c, 0x3d, 0x20, 0x30, 0x20, 0x74, 0x6f, 0x20, 0x75, 0x73, 0x65, 0x20, + 0x76, 0x6f, 0x63, 0x61, 0x62, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x3a, 0x20, + 0x30, 0x2e, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, + 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x3a, 0x20, + 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, + 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, + 0x5f, 0x70, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, + 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, + 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x65, + 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, + 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, + 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, + 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x72, 0x65, 0x71, + 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, + 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, + 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, + 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, + 0x73, 0x74, 0x61, 0x74, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, + 0x30, 0x2f, 0x31, 0x2f, 0x32, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, + 0x3a, 0x20, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x74, 0x61, 0x72, 0x67, + 0x65, 0x74, 0x20, 0x65, 0x6e, 0x74, 0x72, 0x6f, 0x70, 0x79, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, + 0x74, 0x5f, 0x65, 0x74, 0x61, 0x3a, 0x20, 0x30, 0x2e, 0x31, 0x2c, 0x20, + 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x61, 0x72, 0x6e, 0x69, 0x6e, 0x67, 0x20, + 0x72, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, + 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x27, 0x27, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, + 0x73, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, + 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x53, 0x54, 0x41, 0x52, 0x54, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, - 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65, + 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x6f, 0x72, 0x77, 0x73, 0x65, 0x72, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, @@ -668,1031 +710,1563 @@ unsigned char index_html[] = { 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65, - 0x72, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, - 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, - 0x61, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, - 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, - 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, - 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, - 0x74, 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, - 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x63, 0x6f, + 0x72, 0x73, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, + 0x74, 0x61, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, + 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, + 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, + 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, + 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, + 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x63, + 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x20, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, + 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, + 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, + 0x20, 0x3e, 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, + 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, + 0x28, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, + 0x70, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, + 0x74, 0x72, 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, + 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x65, + 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x78, + 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, + 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, + 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x2c, + 0x20, 0x2e, 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, + 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, + 0x28, 0x73, 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x41, 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, 0x5c, 0x7b, 0x28, 0x2e, + 0x2a, 0x3f, 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, 0x67, 0x2c, 0x20, 0x28, + 0x5f, 0x2c, 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x74, 0x74, + 0x69, 0x6e, 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x29, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, 0x20, 0x6d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, + 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x63, 0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, + 0x20, 0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x20, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, - 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x3d, - 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, - 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, - 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, - 0x3e, 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, - 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, - 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, 0x74, - 0x72, 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, - 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x65, 0x74, - 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x78, 0x74, - 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, - 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, - 0x2e, 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x2c, 0x20, - 0x2e, 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, - 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28, - 0x73, 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, - 0x41, 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, 0x5c, 0x7b, 0x28, 0x2e, 0x2a, - 0x3f, 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, 0x67, 0x2c, 0x20, 0x28, 0x5f, - 0x2c, 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x74, 0x74, 0x69, - 0x6e, 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x29, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, - 0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, - 0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, - 0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, - 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, - 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, - 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, - 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, - 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, 0x6f, - 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, - 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, - 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, - 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x75, 0x73, - 0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x5d, - 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2c, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, - 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, - 0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x28, - 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, - 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, - 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x7d, 0x29, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x22, - 0x5c, 0x6e, 0x22, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, - 0x65, 0x74, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, - 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x27, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73, 0x65, + 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, + 0x67, 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, + 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, + 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, + 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, + 0x63, 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x75, + 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, + 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, + 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, - 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73, - 0x3e, 0x22, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x28, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, - 0x29, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, - 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, - 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, - 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, - 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, - 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, - 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, - 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3a, - 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x63, 0x68, 0x75, - 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, - 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x2b, 0x3d, 0x20, 0x64, - 0x61, 0x74, 0x61, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3b, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x20, 0x72, 0x65, 0x6d, 0x6f, 0x76, 0x65, 0x20, 0x6c, 0x65, 0x61, 0x64, - 0x69, 0x6e, 0x67, 0x20, 0x77, 0x68, 0x69, 0x74, 0x65, 0x73, 0x70, 0x61, - 0x63, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x20, 0x3d, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, - 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2b, 0x2f, 0x2c, 0x20, 0x22, - 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, - 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x63, 0x68, - 0x61, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, - 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x5d, 0x5d, 0x29, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, - 0x28, 0x22, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x66, 0x69, 0x6e, 0x69, 0x73, 0x68, 0x65, 0x64, 0x3a, 0x20, 0x27, - 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, - 0x73, 0x73, 0x61, 0x67, 0x65, 0x2c, 0x20, 0x22, 0x27, 0x2c, 0x20, 0x73, - 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x64, - 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, - 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, - 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, - 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, - 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, - 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e, 0x70, - 0x75, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, - 0x61, 0x6c, 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74, 0x6f, 0x70, - 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, - 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, - 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, - 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x65, 0x74, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, + 0x79, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, + 0x69, 0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x28, + 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, + 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, + 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x3a, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, + 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x3f, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, + 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, + 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, + 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2e, 0x72, 0x65, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2f, 0x2c, 0x20, + 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x64, 0x61, 0x74, 0x61, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, + 0x69, 0x6e, 0x28, 0x22, 0x5c, 0x6e, 0x22, 0x29, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x75, 0x72, + 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, + 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, + 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, + 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, + 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, + 0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x2c, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x63, + 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x2c, 0x20, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x75, 0x73, + 0x65, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x5d, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, + 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, + 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, + 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x74, + 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x20, 0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x61, 0x74, + 0x61, 0x20, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, + 0x74, 0x61, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, + 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, + 0x20, 0x30, 0x20, 0x26, 0x26, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, + 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5b, 0x63, 0x75, + 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, + 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, + 0x74, 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x24, 0x2f, 0x29, 0x20, 0x21, + 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, + 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, + 0x2e, 0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, + 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, + 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, + 0x5b, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x22, 0x2c, + 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, + 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x22, 0x43, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x66, 0x69, 0x6e, 0x69, 0x73, 0x68, + 0x65, 0x64, 0x3a, 0x20, 0x27, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, + 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, + 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, + 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, + 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2c, 0x20, 0x22, 0x27, + 0x2c, 0x20, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3a, 0x20, 0x22, + 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x73, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x64, 0x61, 0x74, + 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, + 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, + 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, + 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, + 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, + 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, + 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, + 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, + 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, + 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, + 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, + 0x28, 0x22, 0x22, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x20, 0x3d, + 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, + 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, + 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x61, 0x62, 0x6f, 0x72, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, + 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, + 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, + 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x28, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69, - 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, - 0x70, 0x28, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x68, 0x61, 0x74, 0x28, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x22, - 0x22, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73, - 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x66, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x68, - 0x69, 0x63, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x33, 0x20, 0x26, - 0x26, 0x20, 0x21, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x68, 0x69, - 0x66, 0x74, 0x4b, 0x65, 0x79, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69, - 0x74, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, - 0x6d, 0x20, 0x6f, 0x6e, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x3d, 0x24, - 0x7b, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, - 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, - 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x32, 0x20, 0x6f, 0x6e, 0x6b, 0x65, 0x79, - 0x70, 0x72, 0x65, 0x73, 0x73, 0x3d, 0x24, 0x7b, 0x65, 0x6e, 0x74, 0x65, - 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73, 0x7d, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, - 0x3d, 0x24, 0x7b, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x65, - 0x73, 0x73, 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x65, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, - 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, 0x61, 0x79, 0x20, 0x73, - 0x6f, 0x6d, 0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x22, - 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, - 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, 0x67, 0x68, 0x74, 0x22, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, - 0x70, 0x65, 0x3d, 0x22, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x22, 0x20, - 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x21, - 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x20, 0x3e, 0x53, 0x65, 0x6e, 0x64, 0x3c, - 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, - 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, - 0x3d, 0x24, 0x7b, 0x73, 0x74, 0x6f, 0x70, 0x7d, 0x20, 0x64, 0x69, 0x73, - 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, - 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x7d, 0x3e, 0x53, 0x74, 0x6f, 0x70, - 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, - 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, - 0x6b, 0x3d, 0x24, 0x7b, 0x72, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, - 0x65, 0x73, 0x65, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3d, 0x20, 0x28, 0x70, - 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20, - 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, - 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x20, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x62, - 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x28, 0x69, 0x66, 0x20, 0x6e, 0x65, - 0x65, 0x64, 0x65, 0x64, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x65, - 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, - 0x65, 0x72, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x70, - 0x61, 0x72, 0x65, 0x6e, 0x74, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, - 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, - 0x6c, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, 0x3c, 0x3d, 0x20, 0x70, - 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, - 0x54, 0x6f, 0x70, 0x20, 0x2b, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, - 0x2e, 0x6f, 0x66, 0x66, 0x73, 0x65, 0x74, 0x48, 0x65, 0x69, 0x67, 0x68, - 0x74, 0x20, 0x2b, 0x20, 0x33, 0x30, 0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, - 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f, - 0x28, 0x30, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, - 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x6d, 0x65, 0x73, - 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, - 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x5b, 0x75, 0x73, - 0x65, 0x72, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, - 0x70, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x7d, - 0x3e, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72, - 0x29, 0x7d, 0x3a, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, - 0x20, 0x3c, 0x24, 0x7b, 0x4d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, - 0x69, 0x73, 0x68, 0x7d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x6d, 0x73, 0x67, - 0x29, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, - 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, - 0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, - 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x66, 0x6c, - 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, - 0x6e, 0x65, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x60, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6e, 0x66, - 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, - 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, - 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, - 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, - 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, - 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, - 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, - 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, - 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, - 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, - 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, - 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, - 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, - 0x5d, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, - 0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, - 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, - 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, - 0x5d, 0x3a, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, - 0x72, 0x28, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, - 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x29, 0x29, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x72, - 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, - 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, - 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x27, 0x27, - 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, - 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, - 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x67, 0x72, 0x61, - 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, - 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x6c, 0x2e, 0x74, - 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, - 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x20, - 0x3d, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, - 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, - 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x28, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, - 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, - 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x2c, 0x27, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2e, 0x72, 0x65, 0x64, 0x75, 0x63, 0x65, 0x28, 0x28, 0x61, 0x63, - 0x63, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x28, 0x7b, 0x2e, 0x2e, 0x2e, 0x61, 0x63, 0x63, 0x2c, 0x20, - 0x5b, 0x63, 0x75, 0x72, 0x2e, 0x74, 0x72, 0x69, 0x6d, 0x28, 0x29, 0x5d, - 0x3a, 0x20, 0x69, 0x7d, 0x29, 0x2c, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, - 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x27, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, - 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, - 0x74, 0x65, 0x72, 0x2e, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x47, 0x72, - 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, - 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, 0x28, 0x60, - 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x66, 0x61, 0x69, 0x6c, - 0x65, 0x64, 0x3a, 0x20, 0x24, 0x7b, 0x65, 0x2e, 0x6d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, - 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x68, 0x61, 0x74, 0x28, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x22, 0x22, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x65, 0x6e, + 0x74, 0x65, 0x72, 0x53, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73, 0x20, 0x3d, + 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x77, 0x68, 0x69, 0x63, + 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x31, 0x33, 0x20, 0x26, 0x26, 0x20, + 0x21, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x68, 0x69, 0x66, 0x74, + 0x4b, 0x65, 0x79, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x28, + 0x65, 0x76, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, 0x20, + 0x6f, 0x6e, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x3d, 0x24, 0x7b, 0x73, + 0x75, 0x62, 0x6d, 0x69, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, - 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x74, 0x79, + 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x72, 0x6f, + 0x77, 0x73, 0x3d, 0x32, 0x20, 0x6f, 0x6e, 0x6b, 0x65, 0x79, 0x70, 0x72, + 0x65, 0x73, 0x73, 0x3d, 0x24, 0x7b, 0x65, 0x6e, 0x74, 0x65, 0x72, 0x53, + 0x75, 0x62, 0x6d, 0x69, 0x74, 0x73, 0x7d, 0x20, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, + 0x7b, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x65, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x7d, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, + 0x64, 0x65, 0x72, 0x3d, 0x22, 0x53, 0x61, 0x79, 0x20, 0x73, 0x6f, 0x6d, + 0x65, 0x74, 0x68, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x22, 0x2f, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, + 0x73, 0x73, 0x3d, 0x22, 0x72, 0x69, 0x67, 0x68, 0x74, 0x22, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, + 0x3d, 0x22, 0x73, 0x75, 0x62, 0x6d, 0x69, 0x74, 0x22, 0x20, 0x64, 0x69, + 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x21, 0x67, 0x65, + 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x7d, 0x20, 0x3e, 0x53, 0x65, 0x6e, 0x64, 0x3c, 0x2f, 0x62, + 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, + 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, + 0x7b, 0x73, 0x74, 0x6f, 0x70, 0x7d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, + 0x6c, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, + 0x74, 0x69, 0x6e, 0x67, 0x7d, 0x3e, 0x53, 0x74, 0x6f, 0x70, 0x3c, 0x2f, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, + 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, + 0x24, 0x7b, 0x72, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, + 0x65, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, + 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x68, + 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, + 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6d, 0x65, 0x73, + 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, + 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, + 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x75, 0x73, + 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, + 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, + 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x62, 0x6f, 0x74, + 0x74, 0x6f, 0x6d, 0x20, 0x28, 0x69, 0x66, 0x20, 0x6e, 0x65, 0x65, 0x64, + 0x65, 0x64, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, + 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, + 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x45, 0x6c, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x70, 0x61, + 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x48, + 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, 0x3c, 0x3d, 0x20, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f, + 0x70, 0x20, 0x2b, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x6f, + 0x66, 0x66, 0x73, 0x65, 0x74, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x20, + 0x2b, 0x20, 0x33, 0x30, 0x30, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, + 0x74, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x54, 0x6f, 0x28, 0x30, + 0x2c, 0x20, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x63, 0x72, + 0x6f, 0x6c, 0x6c, 0x48, 0x65, 0x69, 0x67, 0x68, 0x74, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x6d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x73, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x4c, + 0x69, 0x6e, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x5b, 0x75, 0x73, 0x65, 0x72, + 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, 0x2c, 0x20, 0x69, 0x6e, 0x64, + 0x65, 0x78, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6d, 0x65, 0x73, + 0x73, 0x61, 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, + 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, + 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, + 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, + 0x72, 0x6f, 0x62, 0x73, 0x20, 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, 0x20, + 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, + 0x3d, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, + 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, + 0x20, 0x64, 0x61, 0x74, 0x61, 0x3d, 0x24, 0x7b, 0x64, 0x61, 0x74, 0x61, + 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x69, 0x73, + 0x41, 0x72, 0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, + 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, + 0x28, 0x27, 0x27, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2b, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, + 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, + 0x7b, 0x4d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, + 0x7d, 0x20, 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x74, 0x65, 0x78, 0x74, 0x29, 0x7d, + 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, + 0x70, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, + 0x78, 0x7d, 0x3e, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, + 0x7b, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, + 0x65, 0x72, 0x29, 0x7d, 0x3a, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, + 0x67, 0x3e, 0x20, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, + 0x7d, 0x3c, 0x2f, 0x70, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, + 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x68, 0x61, + 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, + 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x65, 0x73, + 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, + 0x70, 0x28, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x29, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, + 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, + 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x65, + 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, + 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, + 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, + 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, + 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, + 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, + 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x28, + 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, + 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, + 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x70, + 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, 0x6c, + 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x29, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x28, + 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, + 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, + 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x4d, + 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x70, 0x61, + 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, 0x6c, 0x2e, + 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x29, 0x29, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, + 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, + 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, + 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, + 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, + 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, + 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, + 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, + 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, + 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x3d, 0x20, 0x28, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53, + 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, + 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, + 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, + 0x65, 0x77, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, + 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, + 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x73, 0x70, 0x6c, 0x69, + 0x74, 0x28, 0x27, 0x2c, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, + 0x64, 0x75, 0x63, 0x65, 0x28, 0x28, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x63, + 0x75, 0x72, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, + 0x2e, 0x2e, 0x2e, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x5b, 0x63, 0x75, 0x72, + 0x2e, 0x74, 0x72, 0x69, 0x6d, 0x28, 0x29, 0x5d, 0x3a, 0x20, 0x69, 0x7d, + 0x29, 0x2c, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, + 0x65, 0x72, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, + 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, + 0x72, 0x28, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, 0x28, 0x60, 0x43, 0x6f, 0x6e, 0x76, + 0x65, 0x72, 0x74, 0x20, 0x66, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x3a, 0x20, + 0x24, 0x7b, 0x65, 0x2e, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, + 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x46, 0x6c, + 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, + 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, + 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, + 0x73, 0x74, 0x65, 0x70, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, + 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, + 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, + 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, + 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, + 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, + 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x73, 0x74, 0x65, + 0x70, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x74, 0x65, 0x70, 0x7d, 0x22, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, + 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x7d, + 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, - 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, - 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, - 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, - 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, - 0x20, 0x73, 0x74, 0x65, 0x70, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x74, 0x65, - 0x70, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, - 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, - 0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, + 0x3d, 0x20, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, 0x6d, + 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, 0x6d, + 0x65, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, + 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, + 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, + 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, + 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, + 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, + 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, + 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, + 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, + 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x49, 0x6e, 0x74, 0x46, 0x69, - 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, - 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, - 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, - 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, - 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, - 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, - 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, - 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, - 0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, - 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, - 0x6c, 0x79, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, - 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, - 0x6c, 0x65, 0x64, 0x3e, 0x55, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, - 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, + 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, + 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, + 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x52, 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, - 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, - 0x75, 0x6c, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, - 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x75, 0x74, 0x6f, - 0x73, 0x61, 0x76, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x20, 0x6f, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x72, 0x79, 0x20, 0x63, - 0x68, 0x61, 0x6e, 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, - 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x24, 0x7b, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, - 0x74, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, - 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, - 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, - 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x74, 0x79, - 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, - 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, - 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, + 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, + 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x55, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, + 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x3d, 0x20, + 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, + 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, + 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, + 0x6f, 0x6e, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3e, + 0x55, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, + 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, + 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, + 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x75, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, + 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x61, 0x6c, 0x6c, + 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x3c, + 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, + 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, + 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x6f, 0x6e, + 0x20, 0x65, 0x76, 0x65, 0x72, 0x79, 0x20, 0x63, 0x68, 0x61, 0x6e, 0x67, + 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x75, + 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5d, 0x29, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, - 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, - 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, - 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x3e, 0x55, - 0x73, 0x65, 0x72, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, - 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, - 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, - 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, - 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, - 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, + 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x7d, + 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, + 0x74, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3c, 0x2f, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, + 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, + 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, + 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, + 0x74, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, + 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, + 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, + 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, + 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, + 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, + 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x3e, 0x55, 0x73, 0x65, 0x72, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, + 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x20, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x75, 0x73, 0x65, + 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, + 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, + 0x62, 0x6f, 0x74, 0x22, 0x3e, 0x42, 0x6f, 0x74, 0x20, 0x6e, 0x61, 0x6d, + 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, + 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, + 0x22, 0x63, 0x68, 0x61, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x22, + 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, + 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, + 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, + 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, - 0x6f, 0x72, 0x3d, 0x22, 0x62, 0x6f, 0x74, 0x22, 0x3e, 0x42, 0x6f, 0x74, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, - 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, - 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x72, 0x22, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x63, 0x68, - 0x61, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, - 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, - 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, + 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, + 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, + 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, - 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, - 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, - 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x43, 0x68, 0x61, 0x74, 0x20, 0x68, + 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x43, 0x68, - 0x61, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, - 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, + 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, + 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x68, 0x69, 0x73, 0x74, + 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, + 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, + 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, - 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, - 0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, - 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x31, + 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, + 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x47, 0x72, - 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, - 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, - 0x72, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x67, 0x72, 0x61, - 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, - 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x55, 0x73, 0x65, 0x20, 0x67, - 0x62, 0x6e, 0x66, 0x20, 0x6f, 0x72, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, - 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2b, 0x63, 0x6f, 0x6e, 0x76, 0x65, - 0x72, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, - 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x22, 0x20, - 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, - 0x70, 0x72, 0x6f, 0x70, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x22, 0x20, - 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, - 0x22, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x70, - 0x31, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x32, 0x2c, 0x70, 0x72, 0x6f, 0x70, - 0x33, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, - 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, - 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, - 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x7d, 0x20, 0x2f, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, - 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, - 0x22, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, - 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, - 0x7d, 0x3e, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x4a, 0x53, - 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x3c, 0x2f, 0x62, - 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, - 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, - 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, - 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, - 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x22, 0x2c, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x65, - 0x64, 0x69, 0x63, 0x74, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, - 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x65, 0x6d, 0x70, 0x65, 0x72, - 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, - 0x20, 0x31, 0x2e, 0x35, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, - 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, - 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, - 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x7d, 0x29, 0x7d, 0x0a, + 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, + 0x72, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x65, - 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, - 0x74, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x65, 0x22, 0x2c, - 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, 0x6d, - 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, - 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, + 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, + 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, + 0x72, 0x3d, 0x22, 0x55, 0x73, 0x65, 0x20, 0x67, 0x62, 0x6e, 0x66, 0x20, + 0x6f, 0x72, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, + 0x6d, 0x61, 0x2b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x22, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, + 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, + 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, + 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, + 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x70, + 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x6f, 0x72, 0x64, + 0x65, 0x72, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x31, 0x2c, 0x70, 0x72, + 0x6f, 0x70, 0x32, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x33, 0x22, 0x20, 0x6f, + 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, + 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, + 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, + 0x4f, 0x72, 0x64, 0x65, 0x72, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, + 0x3d, 0x22, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x22, 0x20, 0x6f, 0x6e, + 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x76, + 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, + 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x3e, 0x43, 0x6f, + 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, + 0x63, 0x68, 0x65, 0x6d, 0x61, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, + 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, + 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, + 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, + 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, + 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, + 0x20, 0x22, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x69, 0x6f, 0x6e, + 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, + 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, 0x5f, 0x70, 0x72, 0x65, + 0x64, 0x69, 0x63, 0x74, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, + 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, + 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, + 0x20, 0x22, 0x54, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, + 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x35, + 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, + 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x65, + 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, - 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, - 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x43, 0x6f, 0x6e, 0x73, 0x69, - 0x64, 0x65, 0x72, 0x20, 0x4e, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, - 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, - 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, - 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, - 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, - 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, - 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x7d, 0x29, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, - 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x4b, - 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, - 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x2c, 0x20, 0x6d, 0x69, - 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x22, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x7d, - 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, - 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, - 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, - 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, - 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, - 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, - 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, + 0x74, 0x75, 0x72, 0x65, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, + 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x65, 0x6e, 0x61, 0x6c, 0x69, + 0x7a, 0x65, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x20, 0x73, 0x65, + 0x71, 0x75, 0x65, 0x6e, 0x63, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, + 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69, - 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, - 0x3e, 0x4d, 0x6f, 0x72, 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, - 0x73, 0x3c, 0x2f, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, - 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, - 0x46, 0x53, 0x2d, 0x5a, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, - 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, - 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x66, - 0x73, 0x5f, 0x7a, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, - 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, + 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, + 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, + 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3a, 0x20, 0x22, 0x43, 0x6f, 0x6e, 0x73, 0x69, 0x64, 0x65, 0x72, 0x20, + 0x4e, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x66, 0x6f, 0x72, + 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x22, 0x2c, 0x20, + 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, + 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, + 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, + 0x74, 0x5f, 0x6e, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x7d, 0x29, 0x7d, 0x0a, 0x20, + 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, + 0x74, 0x5f, 0x6e, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, + 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x4b, 0x20, 0x73, 0x61, 0x6d, + 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, + 0x20, 0x31, 0x30, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x2d, + 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, + 0x70, 0x5f, 0x6b, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x7d, 0x29, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, + 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, + 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, + 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, + 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, + 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x22, 0x2c, + 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, + 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, + 0x5f, 0x70, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, - 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50, 0x22, 0x2c, 0x20, 0x6d, - 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, - 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, - 0x20, 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x22, + 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x4d, 0x6f, 0x72, + 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x3c, 0x2f, 0x73, + 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, + 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, + 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, + 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x46, 0x53, 0x2d, 0x5a, + 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, + 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, - 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x7d, 0x29, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, - 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, - 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, - 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, - 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x70, 0x72, 0x65, - 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, - 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, - 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, - 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x46, 0x72, 0x65, 0x71, - 0x75, 0x65, 0x6e, 0x63, 0x79, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, - 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, - 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x66, 0x72, 0x65, 0x71, 0x75, - 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, - 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, - 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x66, - 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, - 0x61, 0x6c, 0x74, 0x79, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, - 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x72, 0x20, - 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, - 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x68, 0x72, 0x65, - 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, - 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, - 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, - 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x30, 0x22, 0x20, 0x63, 0x68, 0x65, - 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, - 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x30, 0x7d, 0x20, 0x6f, - 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, - 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x6e, 0x6f, 0x20, 0x4d, 0x69, 0x72, 0x6f, - 0x73, 0x74, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, - 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, - 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x31, 0x22, 0x20, 0x63, 0x68, - 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, - 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x31, 0x7d, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, - 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, - 0x61, 0x74, 0x20, 0x76, 0x31, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, - 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x32, 0x22, 0x20, 0x63, - 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, - 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x32, 0x7d, - 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, - 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, - 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, - 0x74, 0x61, 0x74, 0x20, 0x76, 0x32, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, - 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, - 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x74, 0x61, 0x75, - 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2e, 0x30, - 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, - 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x73, 0x74, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x66, + 0x73, 0x5f, 0x7a, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, + 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x79, 0x70, 0x69, 0x63, + 0x61, 0x6c, 0x20, 0x50, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, + 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x79, + 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, - 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, + 0x6c, 0x5f, 0x70, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, + 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x73, 0x65, + 0x6e, 0x63, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, + 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, + 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, + 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, + 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, + 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x73, + 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, + 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, + 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x3a, 0x20, 0x22, 0x46, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, + 0x79, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, + 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, + 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3a, 0x20, 0x22, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, + 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, + 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x66, 0x72, 0x65, 0x71, 0x75, + 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, + 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x72, 0x20, 0x2f, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, + 0x73, 0x73, 0x3d, 0x22, 0x74, 0x68, 0x72, 0x65, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, - 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c, - 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, - 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, - 0x5f, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, - 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, - 0x65, 0x74, 0x61, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, - 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69, - 0x6c, 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, 0x61, - 0x6e, 0x73, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x20, - 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x61, - 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, 0x20, - 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x6d, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x74, 0x65, 0x78, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, - 0x2f, 0x26, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, 0x3b, - 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, - 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, 0x67, - 0x2c, 0x20, 0x27, 0x26, 0x6c, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, - 0x63, 0x65, 0x28, 0x2f, 0x3e, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x67, - 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, - 0x23, 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, 0x24, - 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, 0x24, - 0x31, 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, - 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, - 0x5c, 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, - 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, - 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, - 0x2f, 0x5f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, 0x67, - 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, - 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, - 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, - 0x3f, 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, - 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, - 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, - 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, - 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, - 0x2f, 0x60, 0x60, 0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, 0x5c, - 0x73, 0x5c, 0x53, 0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, 0x67, - 0x2c, 0x20, 0x27, 0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, 0x64, - 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x3c, - 0x2f, 0x70, 0x72, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, - 0x28, 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, 0x2c, - 0x20, 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, - 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, - 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, - 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, 0x6e, - 0x67, 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, - 0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, 0x7b, - 0x20, 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, 0x20, - 0x7d, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, - 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20, - 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x21, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, 0x3e, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, - 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, - 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, - 0x65, 0x72, 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, 0x6d, 0x73, 0x2e, - 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, 0x7d, 0x6d, 0x73, - 0x20, 0x70, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x2c, 0x20, - 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, - 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x73, 0x65, 0x63, - 0x6f, 0x6e, 0x64, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, - 0x32, 0x29, 0x7d, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x70, - 0x65, 0x72, 0x20, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x41, 0x70, 0x70, 0x28, 0x70, 0x72, - 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, - 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, - 0x61, 0x69, 0x6e, 0x65, 0x72, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, 0x61, 0x64, 0x65, - 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, - 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x68, 0x65, - 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6d, 0x61, 0x69, 0x6e, 0x20, 0x69, - 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, - 0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3f, 0x20, - 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3a, 0x20, 0x43, 0x6f, - 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x7d, 0x20, 0x2f, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, - 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, 0x69, 0x74, 0x65, - 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, - 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74, - 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x3c, 0x24, 0x7b, 0x4d, 0x6f, - 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, - 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, + 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, + 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, + 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3d, 0x22, 0x30, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, + 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, + 0x20, 0x3d, 0x3d, 0x20, 0x30, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, + 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, + 0x20, 0x6e, 0x6f, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, + 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, + 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, + 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, + 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3d, 0x22, 0x31, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, + 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, + 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x31, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, + 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, + 0x31, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, + 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, + 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x3d, 0x22, 0x32, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, + 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, + 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, 0x32, 0x7d, 0x20, 0x6f, 0x6e, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, + 0x2f, 0x3e, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, + 0x76, 0x32, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, + 0x73, 0x74, 0x61, 0x74, 0x20, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x6d, + 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, + 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, + 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, + 0x61, 0x75, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, + 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, + 0x61, 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, + 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, + 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x7d, + 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, + 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x50, 0x6f, 0x77, 0x65, 0x72, 0x65, 0x64, - 0x20, 0x62, 0x79, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, - 0x22, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, - 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, 0x67, 0x65, 0x72, - 0x67, 0x61, 0x6e, 0x6f, 0x76, 0x2f, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, - 0x63, 0x70, 0x70, 0x22, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, - 0x70, 0x70, 0x3c, 0x2f, 0x61, 0x3e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x3c, - 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, - 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x22, - 0x3e, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x3c, 0x2f, 0x61, 0x3e, - 0x2e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, - 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, + 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, + 0x6c, 0x64, 0x28, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, + 0x53, 0x68, 0x6f, 0x77, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, + 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x31, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, + 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, 0x5f, 0x70, + 0x72, 0x6f, 0x62, 0x73, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x7d, 0x29, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x72, + 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, + 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x72, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, + 0x6f, 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, 0x28, 0x31, + 0x20, 0x2d, 0x20, 0x70, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x20, 0x3d, 0x20, + 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x31, + 0x39, 0x32, 0x20, 0x2a, 0x20, 0x70, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x60, 0x72, + 0x67, 0x62, 0x61, 0x28, 0x24, 0x7b, 0x72, 0x7d, 0x2c, 0x24, 0x7b, 0x67, + 0x7d, 0x2c, 0x30, 0x2c, 0x30, 0x2e, 0x33, 0x29, 0x60, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, + 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, + 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x6d, 0x73, 0x67, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x21, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, + 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, + 0x65, 0x73, 0x20, 0x7c, 0x7c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x30, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, + 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, + 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, + 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x31, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2f, 0x20, 0x4e, 0x6f, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, + 0x62, 0x79, 0x74, 0x65, 0x20, 0x70, 0x61, 0x69, 0x72, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, + 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, + 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, + 0x5b, 0x30, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, + 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, 0x74, 0x68, 0x28, 0x27, + 0x62, 0x79, 0x74, 0x65, 0x3a, 0x20, 0x5c, 0x5c, 0x27, 0x29, 0x29, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, + 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, + 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x62, 0x20, 0x3d, 0x3e, + 0x20, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x3a, + 0x20, 0x70, 0x72, 0x6f, 0x62, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, + 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, + 0x69, 0x65, 0x73, 0x3a, 0x20, 0x5b, 0x70, 0x72, 0x6f, 0x62, 0x5d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, + 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3d, 0x24, + 0x7b, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, 0x61, 0x7d, 0x20, + 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x73, + 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7d, 0x20, + 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, + 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, + 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6f, 0x75, 0x6e, + 0x64, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, 0x66, 0x69, + 0x6e, 0x64, 0x28, 0x70, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x2e, 0x74, 0x6f, + 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x6d, 0x73, + 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x66, 0x6f, + 0x75, 0x6e, 0x64, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, + 0x6c, 0x6f, 0x72, 0x28, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2e, 0x70, 0x72, + 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, 0x73, + 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, + 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, + 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, + 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, 0x72, 0x6f, + 0x62, 0x2d, 0x73, 0x65, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x70, 0x72, + 0x6f, 0x62, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x28, 0x70, 0x2c, 0x20, + 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, + 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x69, 0x74, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x60, 0x70, 0x72, + 0x6f, 0x62, 0x3a, 0x20, 0x24, 0x7b, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, + 0x7d, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, + 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, + 0x20, 0x27, 0x30, 0x2e, 0x33, 0x65, 0x6d, 0x27, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, + 0x6f, 0x75, 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, + 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3f, 0x20, 0x70, + 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, + 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, + 0x24, 0x7b, 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x7d, + 0x3a, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, + 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x70, + 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x20, 0x2a, 0x20, 0x31, 0x30, 0x30, 0x29, + 0x7d, 0x25, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x7d, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, + 0x7b, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, + 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, + 0x72, 0x20, 0x7d, 0x7d, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, + 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3d, 0x24, 0x7b, 0x70, + 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, + 0x65, 0x6e, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, + 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x29, 0x20, 0x3f, 0x20, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x60, + 0x20, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70, - 0x70, 0x29, 0x2c, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, - 0x2e, 0x62, 0x6f, 0x64, 0x79, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f, - 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65, - 0x61, 0x64, 0x3e, 0x0a, 0x0a, 0x3c, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, - 0x3c, 0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, 0x68, - 0x74, 0x6d, 0x6c, 0x3e, 0x0a + 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, 0x61, 0x6e, + 0x73, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x20, 0x72, + 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x61, 0x72, + 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, 0x20, 0x28, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x6d, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x74, 0x65, 0x78, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, + 0x26, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, 0x3b, 0x27, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, + 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, 0x67, 0x2c, + 0x20, 0x27, 0x26, 0x6c, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x28, 0x2f, 0x3e, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x67, 0x74, + 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x23, + 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f, + 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, 0x24, 0x31, + 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, + 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, + 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, + 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, + 0x5f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, 0x67, 0x2c, + 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, + 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, + 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, + 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, + 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x2f, + 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, + 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, + 0x60, 0x60, 0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, 0x5c, 0x73, + 0x5c, 0x53, 0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, 0x67, 0x2c, + 0x20, 0x27, 0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, 0x64, 0x65, + 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x3c, 0x2f, + 0x70, 0x72, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, + 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, 0x2c, 0x20, + 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, + 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, + 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x62, + 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, 0x6e, 0x67, + 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, + 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, 0x7b, 0x20, + 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, 0x20, 0x7d, + 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, + 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20, 0x28, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, + 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, 0x3e, 0x60, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, + 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, + 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, + 0x72, 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, 0x6d, 0x73, 0x2e, 0x74, + 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, 0x7d, 0x6d, 0x73, 0x20, + 0x70, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x2c, 0x20, 0x24, + 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, + 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x73, 0x65, 0x63, 0x6f, + 0x6e, 0x64, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x32, + 0x29, 0x7d, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x70, 0x65, + 0x72, 0x20, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, + 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, + 0x72, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, + 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x20, 0x3d, + 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x66, + 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, + 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, + 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x27, + 0x30, 0x70, 0x78, 0x27, 0x2c, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, + 0x27, 0x30, 0x70, 0x78, 0x27, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x62, 0x75, + 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, + 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x20, 0x3d, + 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, + 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, + 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, + 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x63, 0x74, 0x20, 0x3d, 0x20, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, + 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x42, 0x6f, 0x75, + 0x6e, 0x64, 0x69, 0x6e, 0x67, 0x43, 0x6c, 0x69, 0x65, 0x6e, 0x74, 0x52, + 0x65, 0x63, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, + 0x2e, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x2b, 0x20, 0x77, 0x69, + 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x59, + 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, + 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, 0x2e, 0x6c, 0x65, 0x66, 0x74, + 0x20, 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, + 0x72, 0x6f, 0x6c, 0x6c, 0x58, 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x21, 0x69, 0x73, + 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x61, + 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, + 0x73, 0x69, 0x64, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, + 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x6f, 0x70, 0x6f, + 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, + 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x21, 0x70, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, + 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, + 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, + 0x20, 0x26, 0x26, 0x20, 0x21, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, + 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, + 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, + 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, + 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x61, 0x64, + 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, + 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, + 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, + 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, + 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, + 0x65, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, + 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, + 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, + 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x5d, 0x29, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, + 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x7d, 0x20, 0x72, 0x65, 0x66, + 0x3d, 0x24, 0x7b, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, + 0x7d, 0x20, 0x6f, 0x6e, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, + 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, + 0x72, 0x7d, 0x3e, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, + 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x20, 0x26, 0x26, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, + 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x20, 0x69, 0x6e, 0x74, + 0x6f, 0x3d, 0x22, 0x23, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x66, 0x3d, + 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, + 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, + 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, + 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x2c, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6c, 0x65, + 0x66, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x70, 0x6f, 0x70, + 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, + 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x53, 0x6f, 0x75, 0x72, 0x63, + 0x65, 0x3a, 0x20, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x20, 0x28, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, + 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, + 0x2f, 0x64, 0x65, 0x76, 0x65, 0x6c, 0x6f, 0x70, 0x69, 0x74, 0x2f, 0x70, + 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, + 0x2f, 0x62, 0x6c, 0x6f, 0x62, 0x2f, 0x6d, 0x61, 0x73, 0x74, 0x65, 0x72, + 0x2f, 0x73, 0x72, 0x63, 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, + 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x2e, 0x6a, 0x73, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x2a, 0x20, 0x52, 0x65, 0x64, 0x69, 0x72, + 0x65, 0x63, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, + 0x67, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x65, 0x73, 0x63, 0x65, 0x6e, 0x64, + 0x61, 0x6e, 0x74, 0x73, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x74, 0x68, + 0x65, 0x20, 0x67, 0x69, 0x76, 0x65, 0x6e, 0x20, 0x43, 0x53, 0x53, 0x20, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x20, 0x2a, 0x2f, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, + 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, + 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x55, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, + 0x6c, 0x65, 0x74, 0x20, 0x69, 0x20, 0x69, 0x6e, 0x20, 0x70, 0x72, 0x6f, + 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x5b, 0x69, 0x5d, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x73, 0x65, 0x74, + 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, + 0x74, 0x44, 0x69, 0x64, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, + 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x20, 0x3d, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, + 0x61, 0x79, 0x65, 0x72, 0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, 0x74, 0x68, + 0x69, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, + 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, + 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, + 0x72, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, + 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x66, + 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, + 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, + 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x76, 0x65, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x6e, 0x6f, + 0x64, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x79, 0x70, + 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x20, 0x3f, 0x20, + 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, + 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x6e, + 0x6f, 0x64, 0x65, 0x29, 0x20, 0x3a, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, + 0x79, 0x65, 0x72, 0x28, 0x73, 0x68, 0x6f, 0x77, 0x20, 0x3d, 0x20, 0x74, + 0x72, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x29, 0x20, + 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x6c, 0x65, 0x61, + 0x6e, 0x20, 0x75, 0x70, 0x20, 0x6f, 0x6c, 0x64, 0x20, 0x6e, 0x6f, 0x64, + 0x65, 0x20, 0x69, 0x66, 0x20, 0x6d, 0x6f, 0x76, 0x69, 0x6e, 0x67, 0x20, + 0x62, 0x61, 0x73, 0x65, 0x73, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x21, + 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, + 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, + 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, + 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x26, 0x26, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x2f, 0x3e, 0x60, + 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x3d, + 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, + 0x64, 0x65, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, + 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, + 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, + 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, + 0x73, 0x68, 0x6f, 0x77, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, + 0x72, 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, + 0x78, 0x79, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, + 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, + 0x65, 0x72, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, + 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x68, 0x69, 0x67, 0x68, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x74, 0x68, + 0x61, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x73, 0x20, 0x69, + 0x74, 0x73, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x63, 0x68, 0x69, + 0x6c, 0x64, 0x20, 0x69, 0x66, 0x20, 0x69, 0x74, 0x20, 0x65, 0x78, 0x69, + 0x73, 0x74, 0x73, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, + 0x75, 0x73, 0x65, 0x64, 0x20, 0x61, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, + 0x6e, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x20, 0x72, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x78, + 0x79, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, + 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, + 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, + 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x67, 0x65, 0x74, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x43, + 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x7b, 0x20, 0x63, 0x68, 0x69, + 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7d, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, + 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, + 0x41, 0x70, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, + 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, + 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6d, 0x61, 0x69, + 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74, 0x53, + 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3a, + 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x7d, + 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, + 0x69, 0x74, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x4d, 0x65, 0x73, + 0x73, 0x61, 0x67, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x7d, 0x20, 0x2f, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, + 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x3c, 0x24, + 0x7b, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, + 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x7d, 0x20, 0x2f, 0x3e, + 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, 0x50, 0x6f, 0x77, 0x65, + 0x72, 0x65, 0x64, 0x20, 0x62, 0x79, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, + 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, + 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, + 0x67, 0x65, 0x72, 0x67, 0x61, 0x6e, 0x6f, 0x76, 0x2f, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x22, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, + 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x61, 0x3e, 0x20, 0x61, 0x6e, + 0x64, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, + 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x67, 0x6d, 0x6c, 0x2e, + 0x61, 0x69, 0x22, 0x3e, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x3c, + 0x2f, 0x61, 0x3e, 0x2e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, + 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, + 0x28, 0x41, 0x70, 0x70, 0x29, 0x2c, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, + 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, 0x72, 0x79, 0x53, 0x65, 0x6c, + 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x27, 0x23, 0x63, 0x6f, 0x6e, 0x74, + 0x61, 0x69, 0x6e, 0x65, 0x72, 0x27, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x3c, 0x2f, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3e, 0x0a, 0x3c, 0x2f, + 0x68, 0x65, 0x61, 0x64, 0x3e, 0x0a, 0x0a, 0x3c, 0x62, 0x6f, 0x64, 0x79, + 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, + 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x22, 0x3e, + 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, + 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, + 0x22, 0x3e, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x3c, 0x2f, 0x62, + 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, 0x68, 0x74, 0x6d, 0x6c, + 0x3e, 0x0a }; -unsigned int index_html_len = 20333; +unsigned int index_html_len = 27218; diff --git a/examples/server/public/index.html b/examples/server/public/index.html index 5eedb0b28..959a9b9a6 100644 --- a/examples/server/public/index.html +++ b/examples/server/public/index.html @@ -102,6 +102,17 @@ padding: 0.5em; } + .prob-set { + padding: 0.3em; + border-bottom: 1px solid #ccc; + } + + .popover-content { + position: absolute; + background-color: white; + padding: 0.2em; + box-shadow: 0 0 10px rgba(0, 0, 0, 0.1); + } textarea { padding: 5px; @@ -133,11 +144,17 @@ font-size: 80%; color: #888; } + + @media (prefers-color-scheme: dark) { + .popover-content { + background-color: black; + } + } +
    +
    diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 1e6d10c1d..025b385cc 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -124,8 +124,9 @@ static void server_log(const char *level, const char *function, int line, static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token) { std::string out = token == -1 ? "" : llama_token_to_str(ctx, token); - // if first bit is 1, meaning it's a partial character - if (out.size() > 0 && (out[0] & 0x80) == 0x80) + // if the size is 1 and first bit is 1, meaning it's a partial character + // (size > 1 meaning it's already a known token) + if (out.size() == 1 && (out[0] & 0x80) == 0x80) { std::stringstream ss; ss << std::hex << (out[0] & 0xff); @@ -1321,59 +1322,86 @@ int main(int argc, char **argv) while (llama.has_next_token) { const completion_token_output token_with_probs = llama.doCompletion(); - const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok); - if (llama.multibyte_pending > 0) { + if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) { continue; } + const std::string token_text = llama_token_to_str(llama.ctx, token_with_probs.tok); size_t pos = std::min(sent_count, llama.generated_text.size()); const std::string str_test = llama.generated_text.substr(pos); + bool is_stop_full = false; size_t stop_pos = llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL); if (stop_pos != std::string::npos) { + is_stop_full = true; llama.generated_text.erase( llama.generated_text.begin() + pos + stop_pos, llama.generated_text.end()); pos = std::min(sent_count, llama.generated_text.size()); } else { + is_stop_full = false; stop_pos = llama.findStoppingStrings(str_test, token_text.size(), STOP_PARTIAL); } - const std::string to_send = llama.generated_text.substr(pos, stop_pos); - sent_count += to_send.size(); + if ( + stop_pos == std::string::npos || + // Send rest of the text if we are at the end of the generation + (!llama.has_next_token && !is_stop_full && stop_pos > 0) + ) { + const std::string to_send = llama.generated_text.substr(pos, std::string::npos); - std::vector probs_output = {}; + sent_count += to_send.size(); - if (llama.params.n_probs > 0) { - const std::vector to_send_toks = llama_tokenize(llama.ctx, to_send, false); - size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size()); - size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size()); - if (probs_pos < probs_stop_pos) { - probs_output = std::vector(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos); + std::vector probs_output = {}; + + if (llama.params.n_probs > 0) { + const std::vector to_send_toks = llama_tokenize(llama.ctx, to_send, false); + size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size()); + size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size()); + if (probs_pos < probs_stop_pos) { + probs_output = std::vector(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos); + } + sent_token_probs_index = probs_stop_pos; + } + + const json data = format_partial_response(llama, to_send, probs_output); + + const std::string str = + "data: " + + data.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + + if (!sink.write(str.data(), str.size())) { + LOG_VERBOSE("stream closed", {}); + llama_print_timings(llama.ctx); + return false; } - sent_token_probs_index = probs_stop_pos; } - const json data = llama.has_next_token - ? format_partial_response(llama, to_send, probs_output) - // Generation is done, send extra information. - : format_final_response(llama, to_send, llama.generated_token_probs); + if (!llama.has_next_token) { + // Generation is done, send extra information. + const json data = format_final_response(llama, "", llama.generated_token_probs); - const std::string str = - "data: " + - data.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; + const std::string str = + "data: " + + data.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; - LOG_VERBOSE("data stream", { - { "to_send", str } - }); + LOG_VERBOSE("data stream", { + { "to_send", str } + }); - if (!sink.write(str.data(), str.size())) { - LOG_VERBOSE("stream closed", {}); - llama_print_timings(llama.ctx); - return false; + if (!sink.write(str.data(), str.size())) { + LOG_VERBOSE("stream closed", {}); + llama_print_timings(llama.ctx); + return false; + } } } From 12e2e33a977af73e75885eeee91c5575a77f4e5f Mon Sep 17 00:00:00 2001 From: slaren Date: Fri, 25 Aug 2023 14:08:53 +0200 Subject: [PATCH 040/147] convert.py : export rope freq_base when converting CodeLlama from an HF model (#2773) --- convert.py | 34 ++++++++++++++++++---------------- 1 file changed, 18 insertions(+), 16 deletions(-) diff --git a/convert.py b/convert.py index 10276bf63..e58ea46e0 100755 --- a/convert.py +++ b/convert.py @@ -160,13 +160,14 @@ class Params: def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params': config = json.load(open(config_path)) - n_vocab = config["vocab_size"] - n_embd = config["hidden_size"] - n_layer = config["num_hidden_layers"] - n_ff = config["intermediate_size"] - n_head = config["num_attention_heads"] - n_head_kv = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head - f_norm_eps = config["rms_norm_eps"] + n_vocab = config["vocab_size"] + n_embd = config["hidden_size"] + n_layer = config["num_hidden_layers"] + n_ff = config["intermediate_size"] + n_head = config["num_attention_heads"] + n_head_kv = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head + f_norm_eps = config["rms_norm_eps"] + f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None n_mult = Params.find_n_mult(n_ff, n_embd) @@ -179,15 +180,16 @@ class Params: "Suggestion: provide 'config.json' of the model in the same directory containing model files.") return Params( - n_vocab = n_vocab, - n_embd = n_embd, - n_mult = n_mult, - n_layer = n_layer, - n_ctx = n_ctx, - n_ff = n_ff, - n_head = n_head, - n_head_kv = n_head_kv, - f_norm_eps = f_norm_eps, + n_vocab = n_vocab, + n_embd = n_embd, + n_mult = n_mult, + n_layer = n_layer, + n_ctx = n_ctx, + n_ff = n_ff, + n_head = n_head, + n_head_kv = n_head_kv, + f_norm_eps = f_norm_eps, + f_rope_freq_base = f_rope_freq_base, ) # LLaMA v2 70B params.json From 154725c5436808e5c519685d0279e850596dbe62 Mon Sep 17 00:00:00 2001 From: slaren Date: Fri, 25 Aug 2023 15:16:19 +0200 Subject: [PATCH 041/147] llama-bench : add model sizes (#2771) * llama-bench : add model sizes * more compact markdown output * back to GiB * adjust column sizes --- examples/llama-bench/llama-bench.cpp | 60 ++++++++++++++++++++++++---- llama.cpp | 18 ++++++++- llama.h | 6 ++- 3 files changed, 74 insertions(+), 10 deletions(-) diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 7a2811584..d0fe6d90d 100755 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -441,6 +441,8 @@ struct test { static const std::string gpu_info; std::string model_filename; std::string model_type; + uint64_t model_size; + uint64_t model_n_params; int n_batch; int n_threads; bool f32_kv; @@ -457,8 +459,10 @@ struct test { test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) { model_filename = inst.model; char buf[128]; - llama_model_type(lmodel, buf, sizeof(buf)); + llama_model_desc(lmodel, buf, sizeof(buf)); model_type = buf; + model_size = llama_model_size(lmodel); + model_n_params = llama_model_n_params(lmodel); n_batch = inst.n_batch; n_threads = inst.n_threads; f32_kv = inst.f32_kv; @@ -524,7 +528,7 @@ struct test { "build_commit", "build_number", "cuda", "opencl", "metal", "gpu_blas", "blas", "cpu_info", "gpu_info", - "model_filename", "model_type", + "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_threads", "f16_kv", "n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split", "n_prompt", "n_gen", "test_time", @@ -538,6 +542,7 @@ struct test { static field_type get_field_type(const std::string & field) { if (field == "build_number" || field == "n_batch" || field == "n_threads" || + field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" || field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "avg_ns" || field == "stddev_ns") { @@ -573,7 +578,7 @@ struct test { build_commit, std::to_string(build_number), std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas), cpu_info, gpu_info, - model_filename, model_type, + model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params), std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv), std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str, std::to_string(n_prompt), std::to_string(n_gen), test_time, @@ -709,8 +714,15 @@ struct markdown_printer : public printer { return -30; } if (field == "t/s") { - return 15; + return 16; } + if (field == "size" || field == "params") { + return 10; + } + if (field == "n_gpu_layers") { + return 3; + } + int width = std::max((int)field.length(), 10); if (test::get_field_type(field) == test::STRING) { @@ -719,9 +731,28 @@ struct markdown_printer : public printer { return width; } + static std::string get_field_display_name(const std::string & field) { + if (field == "n_gpu_layers") { + return "ngl"; + } + if (field == "n_threads") { + return "threads"; + } + if (field == "mul_mat_q") { + return "mmq"; + } + if (field == "tensor_split") { + return "ts"; + } + return field; + } + void print_header(const cmd_params & params) override { // select fields to print - fields = { "model", "backend" }; + fields.push_back("model"); + fields.push_back("size"); + fields.push_back("params"); + fields.push_back("backend"); bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS"; if (!is_cpu_backend) { fields.push_back("n_gpu_layers"); @@ -752,7 +783,7 @@ struct markdown_printer : public printer { fprintf(fout, "|"); for (const auto & field : fields) { - fprintf(fout, " %*s |", get_field_width(field), field.c_str()); + fprintf(fout, " %*s |", get_field_width(field), get_field_display_name(field).c_str()); } fprintf(fout, "\n"); fprintf(fout, "|"); @@ -769,12 +800,26 @@ struct markdown_printer : public printer { fprintf(fout, "|"); for (const auto & field : fields) { std::string value; + char buf[128]; if (field == "model") { value = t.model_type; + } else if (field == "size") { + if (t.model_size < 1024*1024*1024) { + snprintf(buf, sizeof(buf), "%.2f MiB", t.model_size / 1024.0 / 1024.0); + } else { + snprintf(buf, sizeof(buf), "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0); + } + value = buf; + } else if (field == "params") { + if (t.model_n_params < 1000*1000*1000) { + snprintf(buf, sizeof(buf), "%.2f M", t.model_n_params / 1e6); + } else { + snprintf(buf, sizeof(buf), "%.2f B", t.model_n_params / 1e9); + } + value = buf; } else if (field == "backend") { value = test::get_backend(); } else if (field == "test") { - char buf[128]; if (t.n_prompt > 0 && t.n_gen == 0) { snprintf(buf, sizeof(buf), "pp %d", t.n_prompt); } else if (t.n_gen > 0 && t.n_prompt == 0) { @@ -785,7 +830,6 @@ struct markdown_printer : public printer { } value = buf; } else if (field == "t/s") { - char buf[128]; snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts()); value = buf; } else if (vmap.find(field) != vmap.end()) { diff --git a/llama.cpp b/llama.cpp index d12b6d1cb..4529ac822 100644 --- a/llama.cpp +++ b/llama.cpp @@ -5297,13 +5297,29 @@ int llama_model_n_embd(const struct llama_model * model) { return model->hparams.n_embd; } -int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size) { +int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) { return snprintf(buf, buf_size, "%s %s %s", model->name.c_str(), llama_model_type_name(model->type), llama_model_ftype_name(model->ftype).c_str()); } +uint64_t llama_model_size(const struct llama_model * model) { + uint64_t size = 0; + for (const auto & it : model->tensors_by_name) { + size += ggml_nbytes(it.second); + } + return size; +} + +uint64_t llama_model_n_params(const struct llama_model * model) { + uint64_t nparams = 0; + for (const auto & it : model->tensors_by_name) { + nparams += ggml_nelements(it.second); + } + return nparams; +} + int llama_model_quantize( const char * fname_inp, const char * fname_out, diff --git a/llama.h b/llama.h index 2bcf94e0f..d47468172 100644 --- a/llama.h +++ b/llama.h @@ -254,7 +254,11 @@ extern "C" { LLAMA_API int llama_model_n_embd (const struct llama_model * model); // Get a string describing the model type - LLAMA_API int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size); + LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); + // Returns the total size of all the tensors in the model in bytes + LLAMA_API uint64_t llama_model_size(const struct llama_model * model); + // Returns the total number of parameters in the model + LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); // Returns 0 on success LLAMA_API int llama_model_quantize( From 28b2c996ca0ab90a5669946084f13443ec98e241 Mon Sep 17 00:00:00 2001 From: Nigel Bosch Date: Fri, 25 Aug 2023 09:41:52 -0500 Subject: [PATCH 042/147] convert.py : Get rope scale from HuggingFace models (#2772) * Get rope scale from HF models * Save rope scale only for linear scaling * Rewrite for clarity --- convert.py | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/convert.py b/convert.py index e58ea46e0..4f3e92798 100755 --- a/convert.py +++ b/convert.py @@ -105,6 +105,7 @@ class Params: f_norm_eps: float f_rope_freq_base: Optional[float] = None + f_rope_scale: Optional[float] = None ftype: Optional[GGMLFileType] = None @@ -169,6 +170,11 @@ class Params: f_norm_eps = config["rms_norm_eps"] f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None + if "rope_scaling" in config and config["rope_scaling"].get("type") == "linear": + f_rope_scale = config["rope_scaling"].get("factor") + else: + f_rope_scale = None + n_mult = Params.find_n_mult(n_ff, n_embd) if "max_sequence_length" in config: @@ -190,6 +196,7 @@ class Params: n_head_kv = n_head_kv, f_norm_eps = f_norm_eps, f_rope_freq_base = f_rope_freq_base, + f_rope_scale = f_rope_scale, ) # LLaMA v2 70B params.json @@ -773,6 +780,9 @@ class OutputFile: if params.f_rope_freq_base: self.gguf.add_rope_freq_base(params.f_rope_freq_base) + if params.f_rope_scale: + self.gguf.add_rope_scale_linear(params.f_rope_scale) + if params.ftype: self.gguf.add_file_type(params.ftype) From c82742ac9cd96fd34aa961978805c1d8a361d589 Mon Sep 17 00:00:00 2001 From: Matt Pulver Date: Fri, 25 Aug 2023 11:18:48 -0400 Subject: [PATCH 043/147] llama : add llama_beam_search() (#2267) * Add llama_beam_search(). * Add '// Beam search' heading to llama.{h,cpp} after llama_grammar_accept_token(). * Add space around * pointers and & references. * Add spaces around comparison and assignment operators. * Prefer west const. * Use llama_ prefix for structs in global namespace. * Delete obsolete comment from an earlier revision. * Change eos to eob in llama_beam and llama_beam_view structs. --- common/common.h | 1 + examples/CMakeLists.txt | 1 + examples/beam_search/CMakeLists.txt | 8 + examples/beam_search/beam_search.cpp | 188 ++++++++++++++++++++ examples/server/server.cpp | 90 ++++++++-- llama.cpp | 251 +++++++++++++++++++++++++++ llama.h | 37 ++++ 7 files changed, 563 insertions(+), 13 deletions(-) create mode 100644 examples/beam_search/CMakeLists.txt create mode 100644 examples/beam_search/beam_search.cpp diff --git a/common/common.h b/common/common.h index 17d271e67..ce61265f8 100644 --- a/common/common.h +++ b/common/common.h @@ -28,6 +28,7 @@ struct gpt_params { int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors float tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens. + int32_t n_beams = 0; // if non-zero then use beam search of given width. float rope_freq_base = 10000.0f; // RoPE base frequency float rope_freq_scale = 1.0f; // RoPE frequency scaling factor diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index d2176c910..94b785224 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -25,6 +25,7 @@ else() add_subdirectory(simple) add_subdirectory(embd-input) add_subdirectory(llama-bench) + add_subdirectory(beam_search) if (LLAMA_METAL) add_subdirectory(metal) endif() diff --git a/examples/beam_search/CMakeLists.txt b/examples/beam_search/CMakeLists.txt new file mode 100644 index 000000000..b29e01092 --- /dev/null +++ b/examples/beam_search/CMakeLists.txt @@ -0,0 +1,8 @@ +set(TARGET beam_search) +add_executable(${TARGET} beam_search.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) +if(TARGET BUILD_INFO) + add_dependencies(${TARGET} BUILD_INFO) +endif() diff --git a/examples/beam_search/beam_search.cpp b/examples/beam_search/beam_search.cpp new file mode 100644 index 000000000..1c04fabc2 --- /dev/null +++ b/examples/beam_search/beam_search.cpp @@ -0,0 +1,188 @@ +#ifndef _GNU_SOURCE +#define _GNU_SOURCE +#endif + +#include "common.h" +#include "llama.h" +#include "build-info.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) +#include +#include +#elif defined (_WIN32) +#define WIN32_LEAN_AND_MEAN +#define NOMINMAX +#include +#include +#endif + +// Used for debugging to print out beam tokens. +struct ostream_beam_view { + llama_context * ctx; + llama_beam_view beam_view; +}; +std::ostream& operator<<(std::ostream& os, const ostream_beam_view & obv) { + os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens("; + for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) { + os << llama_token_to_str(obv.ctx, obv.beam_view.tokens[i]); + } + return os << ')'; +} + +// Put here anything you want back in beam_search_callback(). +struct beam_search_callback_data { + llama_context * ctx; + std::vector response; +}; + +// In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same. +// For example, eob can be flagged due to maximum token length, stop words, etc. +bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, const size_t n_tokens) { + return n_tokens && tokens[n_tokens-1] == llama_token_eos(callback_data.ctx); +} + +// Function matching type llama_beam_search_callback_fn_t. +// Custom callback example is called each time the beams lengths increase: +// * Show progress by printing ',' following by number of convergent beam tokens if any. +// * When all beams converge to a common prefix, they are made available in beams_state.beams[0]. +// This is also called when the stop condition is met. +// Collect tokens into std::vector response which is pointed to by callback_data. +void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_state) { + auto& callback_data = *static_cast(callback_data_ptr); + // Mark beams as EOS as needed. + for (size_t i = 0 ; i < beams_state.n_beams ; ++i) { + llama_beam_view& beam_view = beams_state.beam_views[i]; + if (!beam_view.eob && is_at_eob(callback_data, beam_view.tokens, beam_view.n_tokens)) { + beam_view.eob = true; + } + } + printf(","); // Show progress + if (const size_t n = beams_state.common_prefix_length) { + callback_data.response.resize(callback_data.response.size() + n); + assert(0u < beams_state.n_beams); + const llama_token * tokens = beams_state.beam_views[0].tokens; + std::copy(tokens, tokens + n, callback_data.response.end() - n); + printf("%lu", n); + } + fflush(stdout); +#if 1 // DEBUG: print current beams for this iteration + std::cout << "\n\nCurrent beams (last_call=" << beams_state.last_call << "):\n"; + for (size_t i = 0 ; i < beams_state.n_beams ; ++i) { + std::cout << "beams["< 3 ) + { + params.prompt = argv[3]; + } + + if ( params.prompt.empty() ) + { + params.prompt = "### Request:\nHow many countries are there?\n\n### Response:\n"; + } + + //--------------------------------- + // Init LLM : + //--------------------------------- + + llama_backend_init(params.numa); + + llama_model * model; + llama_context * ctx; + + std::tie(model, ctx) = llama_init_from_gpt_params( params ); + + if ( model == NULL ) + { + fprintf( stderr , "%s: error: unable to load model\n" , __func__ ); + return 1; + } + + //--------------------------------- + // Tokenize the prompt : + //--------------------------------- + + std::vector tokens_list = llama_tokenize(ctx, params.prompt, true); + + const size_t max_context_size = llama_n_ctx( ctx ); + const size_t max_tokens_list_size = max_context_size - 4 ; + + if (tokens_list.size() > max_tokens_list_size) + { + fprintf( stderr , "%s: error: prompt too long (%lu tokens, max %lu)\n" , + __func__ , tokens_list.size() , max_tokens_list_size ); + return 1; + } + + fprintf( stderr, "\n\n" ); + + // Print the tokens from the prompt : + + for( auto id : tokens_list ) + { + std::cout << llama_token_to_str(ctx, id); + } + std::cout << std::flush; + + int n_past = llama_get_kv_cache_token_count(ctx); + if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past, params.n_threads)) + { + fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ ); + return 1; + } + n_past += tokens_list.size(); + + beam_search_callback_data callback_data{ctx, {}}; + size_t const beam_width = static_cast(params.n_beams); + int const n_predict = 256; + llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict, params.n_threads); + + std::cout << "\n\n"; + for (llama_token const token_id : callback_data.response) { + std::cout << llama_token_to_str(ctx,token_id); + } + std::cout << std::endl; + + llama_free( ctx ); + llama_free_model( model ); + + llama_backend_free(); + + return 0; +} diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 025b385cc..3300553f9 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1209,6 +1209,62 @@ static void log_server_request(const Request &req, const Response &res) }); } +bool is_at_eob(llama_server_context & server_context, const llama_token * tokens, const size_t n_tokens) { + return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx); +} + +// Function matching type llama_beam_search_callback_fn_t. +// Custom callback example is called each time the beams lengths increase: +// * Show progress by printing ',' following by number of convergent beam tokens if any. +// * When all beams converge to a common prefix, they are made available in beams_state.beams[0]. +// This is also called when the stop condition is met. +// Collect tokens into std::vector response which is pointed to by callback_data. +void beam_search_callback(void * callback_data, llama_beams_state beams_state) { + auto & llama = *static_cast(callback_data); + // Mark beams as EOS as needed. + for (size_t i = 0 ; i < beams_state.n_beams ; ++i) { + llama_beam_view& beam_view = beams_state.beam_views[i]; + if (!beam_view.eob && is_at_eob(llama, beam_view.tokens, beam_view.n_tokens)) { + beam_view.eob = true; + } + } + printf(","); // Show progress + if (const size_t n = beams_state.common_prefix_length) { + llama.generated_token_probs.resize(llama.generated_token_probs.size() + n); + assert(0u < beams_state.n_beams); + const llama_token * tokens = beams_state.beam_views[0].tokens; + const auto map = [](llama_token tok) { return completion_token_output{{},tok}; }; + std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map); + printf("%lu", n); + } + fflush(stdout); +#if 0 // DEBUG: print current beams for this iteration + std::cout << "\n\nCurrent beams:\n"; + for (size_t i=0 ; i < beams_state.n_beams ; ++i) { + std::cout << "beams["<t_sample_us += ggml_time_us() - t_start_sample_us; } +// +// Beam search +// + +struct llama_beam { + std::vector tokens; + float p; // Cumulative beam probability (renormalized relative to all beams) + bool eob; // Initialize end-of-beam to false. Callback sets this to true. + // Sort beams by probability. In case of ties, prefer beams at eob. + bool operator<(const llama_beam & rhs) const { + return std::make_pair(p, eob) < std::make_pair(rhs.p, rhs.eob); + } + // Shift off first n tokens and discard them. + void shift_tokens(const size_t n) { + if (n) { + std::copy(tokens.begin() + n, tokens.end(), tokens.begin()); + tokens.resize(tokens.size() - n); + } + } + llama_beam_view view() const { return {tokens.data(), tokens.size(), p, eob}; } +}; + +// A struct for calculating logit-related info. +struct llama_logit_info { + const float * const logits; + const int n_vocab; + const float max_l; + const float normalizer; + struct sum_exp { + float max_l; + float operator()(float sum, float l) const { return sum + std::exp(l - max_l); } + }; + llama_logit_info(llama_context * ctx) + : logits(llama_get_logits(ctx)) + , n_vocab(llama_n_vocab(ctx)) + , max_l(*std::max_element(logits, logits + n_vocab)) + , normalizer(1.0f / std::accumulate(logits, logits + n_vocab, 0.0f, sum_exp{max_l})) + { } + llama_token_data get_token_data(const llama_token token_id) const { + constexpr auto p = std::numeric_limits::quiet_NaN(); // never used + return {token_id, logits[token_id], p}; + } + // Return top k token_data by logit. + std::vector top_k(size_t k) { + std::vector min_heap; // min-heap by logit + const llama_token k_min = std::min(static_cast(k), n_vocab); + min_heap.reserve(k_min); + for (llama_token token_id = 0 ; token_id < k_min ; ++token_id) { + min_heap.push_back(get_token_data(token_id)); + } + auto comp = [](const llama_token_data & a, const llama_token_data & b) { return a.logit > b.logit; }; + std::make_heap(min_heap.begin(), min_heap.end(), comp); + for (llama_token token_id = k_min ; token_id < n_vocab ; ++token_id) { + if (min_heap.front().logit < logits[token_id]) { + std::pop_heap(min_heap.begin(), min_heap.end(), comp); + min_heap.back().id = token_id; + min_heap.back().logit = logits[token_id]; + std::push_heap(min_heap.begin(), min_heap.end(), comp); + } + } + return min_heap; + } + float probability_from_logit(float logit) { + return normalizer * std::exp(logit - max_l); + } +}; + +struct llama_beam_search_data { + llama_context * ctx; + size_t n_beams; + int n_past; + int n_predict; + int n_threads; + std::vector beams; + std::vector next_beams; + + // Re-calculated on each loop iteration + size_t common_prefix_length; + + // Used to communicate to/from callback on beams state. + std::vector beam_views; + + llama_beam_search_data(llama_context * ctx, size_t n_beams, int n_past, int n_predict, int n_threads) + : ctx(ctx) + , n_beams(n_beams) + , n_past(n_past) + , n_predict(n_predict) + , n_threads(n_threads) + , beam_views(n_beams) { + beams.reserve(n_beams); + next_beams.reserve(n_beams); + } + + // Collapse beams to a single beam given by index. + void collapse_beams(const size_t beam_idx) { + if (0u < beam_idx) { + std::swap(beams[0], beams[beam_idx]); + } + beams.resize(1); + } + + // Min-heaps are used to efficiently collect the top-k elements (k=n_beams). + // The repetative patterns below reflect the 2 stages of heaps: + // * Gather elements until the vector is full, then call std::make_heap() on it. + // * If the heap is full and a new element is found that should be included, pop the + // least element to the back(), replace it with the new, then push it into the heap. + void fill_next_beams_by_top_probabilities(llama_beam & beam) { + // Min-heaps use a greater-than comparator. + const auto comp = [](const llama_beam & a, const llama_beam & b) { return a.p > b.p; }; + if (beam.eob) { + // beam is at end-of-sentence, so just copy it to next_beams if its probability is high enough. + if (next_beams.size() < n_beams) { + next_beams.push_back(std::move(beam)); + if (next_beams.size() == n_beams) { + std::make_heap(next_beams.begin(), next_beams.end(), comp); + } + } else if (next_beams.front().p < beam.p) { + std::pop_heap(next_beams.begin(), next_beams.end(), comp); + next_beams.back() = std::move(beam); + std::push_heap(next_beams.begin(), next_beams.end(), comp); + } + } else { + // beam is not at end-of-sentence, so branch with next top_k tokens. + if (!beam.tokens.empty()) { + llama_eval(ctx, beam.tokens.data(), beam.tokens.size(), n_past, n_threads); + } + llama_logit_info logit_info(ctx); + std::vector next_tokens = logit_info.top_k(n_beams); + size_t i=0; + if (next_beams.size() < n_beams) { + for (; next_beams.size() < n_beams ; ++i) { + llama_beam next_beam = beam; + next_beam.tokens.push_back(next_tokens[i].id); + next_beam.p *= logit_info.probability_from_logit(next_tokens[i].logit); + next_beams.push_back(std::move(next_beam)); + } + std::make_heap(next_beams.begin(), next_beams.end(), comp); + } else { + for (; next_beams.front().p == 0.0f ; ++i) { + std::pop_heap(next_beams.begin(), next_beams.end(), comp); + next_beams.back() = beam; + next_beams.back().tokens.push_back(next_tokens[i].id); + next_beams.back().p *= logit_info.probability_from_logit(next_tokens[i].logit); + std::push_heap(next_beams.begin(), next_beams.end(), comp); + } + } + for (; i < n_beams ; ++i) { + const float next_p = beam.p * logit_info.probability_from_logit(next_tokens[i].logit); + if (next_beams.front().p < next_p) { + std::pop_heap(next_beams.begin(), next_beams.end(), comp); + next_beams.back() = beam; + next_beams.back().tokens.push_back(next_tokens[i].id); + next_beams.back().p = next_p; + std::push_heap(next_beams.begin(), next_beams.end(), comp); + } + } + } + } + + // Find common_prefix_length based on beams. + // Requires beams is not empty. + size_t find_common_prefix_length() { + size_t common_prefix_length = beams[0].tokens.size(); + for (size_t i = 1 ; i < beams.size() ; ++i) { + common_prefix_length = std::min(common_prefix_length, beams[i].tokens.size()); + for (size_t j = 0 ; j < common_prefix_length ; ++j) { + if (beams[0].tokens[j] != beams[i].tokens[j]) { + common_prefix_length = j; + break; + } + } + } + return common_prefix_length; + } + + // Construct beams_state to send back to caller via the callback function. + // Side effect: set common_prefix_length = find_common_prefix_length(); + llama_beams_state get_beams_state(const bool last_call) { + for (size_t i = 0 ; i < beams.size() ; ++i) { + beam_views[i] = beams[i].view(); + } + common_prefix_length = find_common_prefix_length(); + return {beam_views.data(), beams.size(), common_prefix_length, last_call}; + } + + // Loop: + // * while i < n_predict, AND + // * any of the beams have not yet reached end-of-beam (eob), AND + // * the highest probability beam(s) (plural in case of ties) are not at end-of-sentence + // (since all other beam probabilities can only decrease) + void loop(const llama_beam_search_callback_fn_t callback, void * const callback_data) { + beams.push_back({{}, 1.0f, false}); // Start with one empty beam w/ probability = 1.0 and !eob. + const auto not_eob = [](const llama_beam & beam) { return !beam.eob; }; + for (int i = 0 ; i < n_predict && std::any_of(beams.begin(),beams.end(),not_eob) && + !beams[top_beam_index()].eob ; ++i) { + callback(callback_data, get_beams_state(false)); // Sets common_prefix_length + update_beams_from_beam_views(); // Update values (p,eob) that callback may have changed. + if (common_prefix_length) { + llama_eval(ctx, beams[0].tokens.data(), common_prefix_length, n_past, n_threads); + n_past += common_prefix_length; + } + // Zero-out next_beam probabilities to place them last in following min-heap. + std::for_each(next_beams.begin(), next_beams.end(), [](llama_beam & beam) { beam.p = 0.0f; }); + for (llama_beam & beam : beams) { + beam.shift_tokens(common_prefix_length); + fill_next_beams_by_top_probabilities(beam); + } + // next_beams become the beams of next/final iteration. Swap them to re-use memory. + beams.swap(next_beams); + renormalize_beam_probabilities(beams); + } + collapse_beams(top_beam_index()); + callback(callback_data, get_beams_state(true)); + } + + // As beams grow, the cumulative probabilities decrease. + // Renormalize them to avoid floating point underflow. + static void renormalize_beam_probabilities(std::vector & beams) { + const auto sum_p = [](float sum, llama_beam & beam) { return sum + beam.p; }; + const float inv_sum = 1.0f / std::accumulate(beams.begin(), beams.end(), 0.0f, sum_p); + std::for_each(beams.begin(), beams.end(), [=](llama_beam & beam) { beam.p *= inv_sum; }); + } + + // Assumes beams is non-empty. Uses llama_beam::operator<() for ordering. + size_t top_beam_index() { + return std::max_element(beams.begin(), beams.end()) - beams.begin(); + } + + // Copy (p,eob) for each beam which may have been changed by the callback. + void update_beams_from_beam_views() { + for (size_t i = 0 ; i < beams.size() ; ++i) { + beams[i].p = beam_views[i].p; + beams[i].eob = beam_views[i].eob; + } + } +}; + +void llama_beam_search(llama_context * ctx, + llama_beam_search_callback_fn_t callback, void * callback_data, + size_t n_beams, int n_past, int n_predict, int n_threads) { + assert(ctx); + const int64_t t_start_sample_us = ggml_time_us(); + + llama_beam_search_data beam_search_data(ctx, n_beams, n_past, n_predict, n_threads); + + beam_search_data.loop(callback, callback_data); + + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + ctx->n_sample++; +} + // // quantization // diff --git a/llama.h b/llama.h index d47468172..86737200f 100644 --- a/llama.h +++ b/llama.h @@ -469,6 +469,43 @@ extern "C" { /// @details Accepts the sampled token into the grammar LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token); + // + // Beam search + // + + struct llama_beam_view { + const llama_token * tokens; + size_t n_tokens; + float p; // Cumulative beam probability (renormalized relative to all beams) + bool eob; // Callback should set this to true when a beam is at end-of-beam. + }; + + // Passed to beam_search_callback function. + // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams + // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks. + // These pointers are valid only during the synchronous callback, so should not be saved. + struct llama_beams_state { + llama_beam_view * beam_views; + size_t n_beams; // Number of elements in beam_views[]. + size_t common_prefix_length; // Current max length of prefix tokens shared by all beams. + bool last_call; // True iff this is the last callback invocation. + }; + + // Type of pointer to the beam_search_callback function. + // void* callback_data is any custom data passed to llama_beam_search, that is subsequently + // passed back to beam_search_callback. This avoids having to use global variables in the callback. + typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, llama_beams_state); + + /// @details Deterministically returns entire sentence constructed by a beam search. + /// @param ctx Pointer to the llama_context. + /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state. + /// @param callback_data A pointer that is simply passed back to callback. + /// @param n_beams Number of beams to use. + /// @param n_past Number of tokens already evaluated. + /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier. + /// @param n_threads Number of threads as passed to llama_eval(). + LLAMA_API void llama_beam_search(struct llama_context * ctx, llama_beam_search_callback_fn_t callback, void * callback_data, size_t n_beams, int n_past, int n_predict, int n_threads); + // Performance information LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx); LLAMA_API void llama_print_timings(struct llama_context * ctx); From d046dcee081118c9071bbc63dacdb359a58c467a Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Fri, 25 Aug 2023 19:05:02 +0300 Subject: [PATCH 044/147] Faster perplexity computation (#2786) Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 67 +++++++++++++++++++++++++----- 1 file changed, 56 insertions(+), 11 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index a7bd9db2a..18635932b 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -6,6 +6,8 @@ #include #include #include +#include +#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data @@ -27,6 +29,40 @@ std::vector softmax(const std::vector& logits) { return probs; } +float log_softmax(int n_vocab, const float * logits, int tok) { + float max_logit = logits[0]; + for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]); + double sum_exp = 0.0; + for (int i = 0; i < n_vocab; ++i) sum_exp += expf(logits[i] - max_logit); + return logits[tok] - max_logit - log(sum_exp); +} + +void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector& workers, + double& nll, double& nll2) { + + std::mutex mutex; + int counter = 0; + auto compute = [&mutex, &counter, &nll, &nll2, n_vocab, logits, tokens, n_token] () { + double local_nll = 0, local_nll2 = 0; + while (true) { + std::unique_lock lock(mutex); + int i = counter++; + if (i >= n_token) { + nll += local_nll; nll2 += local_nll2; + break; + } + lock.unlock(); + double v = -log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]); + local_nll += v; + local_nll2 += v*v; + } + }; + for (auto& w : workers) w = std::thread(compute); + compute(); + for (auto& w : workers) w.join(); + +} + void perplexity_v2(llama_context * ctx, const gpt_params & params) { // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` @@ -166,9 +202,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) { int count = 0; double nll = 0.0; + double nll2 = 0.0; fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch); + std::vector workers(std::thread::hardware_concurrency() - 1); + for (int i = 0; i < n_chunk; ++i) { const int start = i * params.n_ctx; const int end = start + params.n_ctx; @@ -228,26 +267,32 @@ void perplexity(llama_context * ctx, const gpt_params & params) { // Example, we have a context window of 512, we will compute perplexity for each of the // last 256 tokens. Then, we split the input up into context window size chunks to // process the entire prompt. - for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) { - // Calculate probability of next token, given the previous ones. - const std::vector tok_logits( - logits.begin() + (j + 0) * n_vocab, - logits.begin() + (j + 1) * n_vocab); + const int first = std::min(512, params.n_ctx/2); + process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first, workers, nll, nll2); + count += params.n_ctx - first - 1; - const float prob = softmax(tok_logits)[tokens[start + j + 1]]; - - nll += -std::log(prob); - ++count; - } // perplexity is e^(average negative log-likelihood) if (params.ppl_output_type == 0) { printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); } else { - printf("%8d %.4lf\n", i*params.n_ctx, std::exp(nll / count)); + double av = nll/count; + double av2 = nll2/count - av*av; + if (av2 > 0) av2 = sqrt(av2/(count-1)); + printf("%8d %.4lf %4lf %4lf\n", i*params.n_ctx, std::exp(nll / count), av, av2); } fflush(stdout); } printf("\n"); + nll2 /= count; + nll /= count; + nll2 -= nll * nll; + if (nll2 > 0) { + nll2 = sqrt(nll2/(count-1)); + double ppl = exp(nll); + printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl); + } else { + printf("Unexpected negative standard deviation of log(prob)\n"); + } } std::vector hellaswag_evaluate_tokens(llama_context * ctx, const std::vector& tokens, int n_past, int n_batch, From 232caf3c1581a6cb023571780ff41dc2d66d1ca0 Mon Sep 17 00:00:00 2001 From: Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com> Date: Fri, 25 Aug 2023 09:17:15 -0700 Subject: [PATCH 045/147] llama : fix struct decl (#2790) --- llama.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.h b/llama.h index 86737200f..b77dd7735 100644 --- a/llama.h +++ b/llama.h @@ -485,7 +485,7 @@ extern "C" { // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks. // These pointers are valid only during the synchronous callback, so should not be saved. struct llama_beams_state { - llama_beam_view * beam_views; + struct llama_beam_view * beam_views; size_t n_beams; // Number of elements in beam_views[]. size_t common_prefix_length; // Current max length of prefix tokens shared by all beams. bool last_call; // True iff this is the last callback invocation. From bae5c5f679e043371bc2b4dffff8d4964d6cb953 Mon Sep 17 00:00:00 2001 From: lon <114724657+longregen@users.noreply.github.com> Date: Sat, 26 Aug 2023 10:07:43 +0200 Subject: [PATCH 046/147] examples : skip unnecessary external lib in server README.md how-to (#2804) --- examples/server/README.md | 23 ++++++++++------------- 1 file changed, 10 insertions(+), 13 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index 77997f98d..7105e9020 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -77,34 +77,31 @@ You need to have [Node.js](https://nodejs.org/en) installed. ```bash mkdir llama-client cd llama-client -npm init -npm install axios ``` Create a index.js file and put inside this: ```javascript -const axios = require("axios"); - const prompt = `Building a website can be done in 10 simple steps:`; async function Test() { - let result = await axios.post("http://127.0.0.1:8080/completion", { - prompt, - n_predict: 512, - }); - - // the response is received until completion finish - console.log(result.data.content); + let response = await fetch("http://127.0.0.1:8080/completion", { + method: 'POST', + body: JSON.stringify({ + prompt, + n_predict: 512, + }) + }) + console.log((await response.json()).content) } -Test(); +Test() ``` And run it: ```bash -node . +node index.js ``` ## API Endpoints From 2ba83c8685177faea3399db9564f9c52df75c366 Mon Sep 17 00:00:00 2001 From: klosax <131523366+klosax@users.noreply.github.com> Date: Sat, 26 Aug 2023 13:45:53 +0200 Subject: [PATCH 047/147] Fix spm whitespaces (#2806) * llama.cpp : fix spm whitespace escaping + clean up * main.cpp : spm - add whitespace in front of prompt * test-tokenizer-0.cpp : spm - add whitespace in front of prompt --- examples/main/main.cpp | 17 ++++++++++---- llama.cpp | 48 +++++++++++--------------------------- tests/test-tokenizer-0.cpp | 3 ++- 3 files changed, 27 insertions(+), 41 deletions(-) diff --git a/examples/main/main.cpp b/examples/main/main.cpp index cb8747c2b..4665b82fe 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -189,12 +189,19 @@ int main(int argc, char ** argv) { } } - const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; + // Add BOS if SPM tokenizer + const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; // tokenize the prompt std::vector embd_inp; + + if (llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM) { + // Add a space in front of the first character to match OG llama tokenizer behavior + params.prompt.insert(0, 1, ' '); + } + if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) { - embd_inp = ::llama_tokenize(ctx, params.prompt, is_spm); + embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos); } else { embd_inp = session_tokens; } @@ -210,9 +217,9 @@ int main(int argc, char ** argv) { int original_prompt_len = 0; if (ctx_guidance) { params.cfg_negative_prompt.insert(0, 1, ' '); - guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, is_spm); + guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos); - std::vector original_inp = ::llama_tokenize(ctx, params.prompt, is_spm); + std::vector original_inp = ::llama_tokenize(ctx, params.prompt, add_bos); original_prompt_len = original_inp.size(); guidance_offset = (int)guidance_inp.size() - original_prompt_len; } @@ -259,7 +266,7 @@ int main(int argc, char ** argv) { } // prefix & suffix for instruct mode - const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", is_spm); + const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos); const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false); // in instruct mode, we inject a prefix and a suffix to each input by the user diff --git a/llama.cpp b/llama.cpp index 7d8b9a0ac..b0a3b5768 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1635,7 +1635,7 @@ static void llm_load_hparams( } // TODO: This should probably be in llama.h -static std::vector llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos, bool escape); +static std::vector llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos); static void llm_load_vocab( llama_model_loader & ml, @@ -1737,7 +1737,7 @@ static void llm_load_vocab( } // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n' - vocab.linefeed_id = llama_tokenize_internal(vocab, "\n", false, false)[0]; + vocab.linefeed_id = llama_tokenize_internal(vocab, "\n", false)[0]; // special tokens GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID)); @@ -3027,14 +3027,8 @@ static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) { } static std::string llama_escape_whitespace(const std::string& text) { - std::string result = "\xe2\x96\x81"; - for (size_t offs = 0; offs < text.length(); ++offs) { - if (text[offs] == ' ') { - result += "\xe2\x96\x81"; - } else { - result += text[offs]; - } - } + std::string result = text; + replace_all(result, " ", "\xe2\x96\x81"); return result; } @@ -3219,7 +3213,7 @@ struct llm_bigram_bpe { }; struct llm_tokenizer_bpe { - llm_tokenizer_bpe(const llama_vocab & vocab, bool g2ws): vocab(vocab) { flag_g2ws = g2ws; } + llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {} void tokenize(const std::string & text, std::vector & output) { int final_prev_index = -1; @@ -3371,8 +3365,6 @@ private: return words; } - bool flag_g2ws = false; - const llama_vocab & vocab; std::vector symbols; @@ -3381,39 +3373,26 @@ private: llm_bigram_bpe::queue work_queue; }; -static std::vector llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos, bool escape) { +static std::vector llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos) { std::vector output; if (raw_text.empty()) { return output; } + if (bos && vocab.special_bos_id != -1) { + output.push_back(vocab.special_bos_id); + } + switch (vocab.type) { case LLAMA_VOCAB_TYPE_SPM: { llm_tokenizer_spm tokenizer(vocab); - - if (bos) { - output.push_back(vocab.special_bos_id); - } - - std::string text; - if (escape) { - text = llama_escape_whitespace(raw_text); - } else { - text = raw_text; - } - - tokenizer.tokenize(text, output); + tokenizer.tokenize(llama_escape_whitespace(raw_text), output); } break; case LLAMA_VOCAB_TYPE_BPE: { - llm_tokenizer_bpe tokenizer(vocab, escape); - - if (bos && vocab.special_bos_id != -1) { - output.push_back(vocab.special_bos_id); - } - + llm_tokenizer_bpe tokenizer(vocab); tokenizer.tokenize(raw_text, output); } break; }; @@ -6095,8 +6074,7 @@ int llama_tokenize_with_model( llama_token * tokens, int n_max_tokens, bool add_bos) { - auto escape = llama_vocab_get_type(model->vocab) == LLAMA_VOCAB_TYPE_SPM; - auto res = llama_tokenize_internal(model->vocab, text, add_bos, escape); + auto res = llama_tokenize_internal(model->vocab, text, add_bos); if (n_max_tokens < (int) res.size()) { LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp index f3ee851a3..7e9ac9188 100644 --- a/tests/test-tokenizer-0.cpp +++ b/tests/test-tokenizer-0.cpp @@ -100,7 +100,8 @@ int main(int argc, char **argv) { bool success = true; for (const auto & test_kv : k_tests()) { - std::vector res = llama_tokenize(ctx, test_kv.first, true); + // Add a space in front of the first character to match OG llama tokenizer behavior + std::vector res = llama_tokenize(ctx, " " + test_kv.first, true); fprintf(stderr, "%s : '%s' tokenized to '%s'\n", __func__, test_kv.first.c_str(), unescape_whitespace(ctx, res).c_str()); From a2ca4e9de9da45ed0bb1c34935d5ec80cebc22d5 Mon Sep 17 00:00:00 2001 From: Nigel Bosch Date: Sat, 26 Aug 2023 07:11:17 -0500 Subject: [PATCH 048/147] Handle null rope scaling value (#2793) --- convert.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/convert.py b/convert.py index 4f3e92798..d44e5a8c4 100755 --- a/convert.py +++ b/convert.py @@ -170,7 +170,8 @@ class Params: f_norm_eps = config["rms_norm_eps"] f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None - if "rope_scaling" in config and config["rope_scaling"].get("type") == "linear": + rope_scaling = config.get("rope_scaling") + if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear": f_rope_scale = config["rope_scaling"].get("factor") else: f_rope_scale = None From f305bad11e10ad09e396faed2e37f4f845f5d566 Mon Sep 17 00:00:00 2001 From: Volodymyr Vitvitskyi <72226+signalpillar@users.noreply.github.com> Date: Sat, 26 Aug 2023 14:25:39 +0100 Subject: [PATCH 049/147] flake : build llama.cpp on Intel with nix (#2795) Problem ------- `nix build` fails with missing `Accelerate.h`. Changes ------- - Fix build of the llama.cpp with nix for Intel: add the same SDK frameworks as for ARM - Add `quantize` app to the output of nix flake - Extend nix devShell with llama-python so we can use convertScript Testing ------- Testing the steps with nix: 1. `nix build` Get the model and then 2. `nix develop` and then `python convert.py models/llama-2-7b.ggmlv3.q4_0.bin` 3. `nix run llama.cpp#quantize -- open_llama_7b/ggml-model-f16.gguf ./models/ggml-model-q4_0.bin 2` 4. `nix run llama.cpp#llama -- -m models/ggml-model-q4_0.bin -p "What is nix?" -n 400 --temp 0.8 -e -t 8` Co-authored-by: Volodymyr Vitvitskyi --- flake.nix | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/flake.nix b/flake.nix index 616b90252..d454cedc3 100644 --- a/flake.nix +++ b/flake.nix @@ -21,6 +21,12 @@ CoreGraphics CoreVideo ] + else if isDarwin then + with pkgs.darwin.apple_sdk.frameworks; [ + Accelerate + CoreGraphics + CoreVideo + ] else with pkgs; [ openblas ] ); @@ -80,8 +86,13 @@ type = "app"; program = "${self.packages.${system}.default}/bin/llama"; }; + apps.quantize = { + type = "app"; + program = "${self.packages.${system}.default}/bin/quantize"; + }; apps.default = self.apps.${system}.llama; devShells.default = pkgs.mkShell { + buildInputs = [ llama-python ]; packages = nativeBuildInputs ++ osSpecific; }; }); From 771551a793c9976ed9cdfe7b8c69536af32af9f9 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sat, 26 Aug 2023 16:48:53 +0300 Subject: [PATCH 050/147] Fix HellaSwag (#2805) Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 20 +++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 18635932b..fd89852d6 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -351,6 +351,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) { fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count); const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; + fprintf(stderr, "================================= is_spm = %d\n", is_spm); // This is needed as usual for LLaMA models const bool add_bos = is_spm; @@ -406,6 +407,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) { double acc = 0.0f; const int n_vocab = llama_n_vocab(ctx); + std::vector> ending_tokens(4); + std::vector tok_logits(n_vocab); for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) { @@ -413,11 +416,21 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) { std::vector context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos); size_t context_size = context_embd.size(); + for (int i = 0; i < 4; ++i) { + ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[i], add_bos); + for (int k = 0; k < int(context_size); ++k) { + if (ending_tokens[i][k] != context_embd[k]) { + fprintf(stderr, "Oops: ending %d of task %d differs from context at position %d\n",i,int(task_idx),k); + break; + } + } + } + // Do the 1st ending // In this case we include the context when evaluating - auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos); + //auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos); + auto query_embd = ending_tokens[0]; auto query_size = query_embd.size(); - //printf("First query: %d\n",(int)query_size); // Stop if query wont fit the ctx window if (query_size > (size_t)params.n_ctx) { @@ -462,7 +475,8 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) { for (size_t ending_idx = 1; ending_idx < 4; ending_idx++) { // Tokenize the query - query_embd = ::llama_tokenize(ctx, hs_data[task_idx].ending[ending_idx], false); + query_embd.resize(ending_tokens[ending_idx].size() - context_size); + std::memcpy(query_embd.data(), ending_tokens[ending_idx].data() + context_size, query_embd.size()*sizeof(int)); query_size = query_embd.size(); // Stop if query wont fit the ctx window From 7592375403a0bd0456d5ec2cdf8350e591f04fb0 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sat, 26 Aug 2023 17:27:49 +0300 Subject: [PATCH 051/147] Better perplexity for 2- and 3-bit quantization for LLaMA-v2-70B (#2807) * Better perplexity for 2- and 3-bit quantization for the 70B model * PR comment --------- Co-authored-by: Iwan Kawrakow --- llama.cpp | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/llama.cpp b/llama.cpp index b0a3b5768..52fcaceff 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4653,6 +4653,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s std::unique_ptr ml(new llama_model_loader(fname_inp, /*use_mmap*/ false)); + llama_model model; + llm_load_arch(*ml, model); + llm_load_hparams(*ml, model, 0, 0, 0); + const size_t align = GGUF_DEFAULT_ALIGNMENT; struct gguf_context * ctx_out = gguf_init_empty(); @@ -4678,6 +4682,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s ++n_feed_forward_w2; } } + if (n_attention_wv != n_feed_forward_w2 || (uint32_t)n_attention_wv != model.hparams.n_layer) { + LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n", + __func__, n_attention_wv, n_feed_forward_w2, model.hparams.n_layer); + } int i_attention_wv = 0; int i_feed_forward_w2 = 0; @@ -4769,6 +4777,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) && (i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K; + if (model.type == MODEL_70B) { + // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is + // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with + // nearly negligible increase in model size by quantizing this tensor with more bits: + if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K; + } ++i_attention_wv; } else if (name.find("ffn_down.weight") != std::string::npos) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; From 04f4b1eb10f3e25750ca3e530265ce2841730e6b Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 26 Aug 2023 17:37:35 +0300 Subject: [PATCH 052/147] k-quants : remove unnecessary tensor shape restrictions (#2811) --- llama.cpp | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/llama.cpp b/llama.cpp index 52fcaceff..59105db1c 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4762,8 +4762,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { int nx = tensor->ne[0]; - int ny = tensor->ne[1]; - if (nx % QK_K == 0 && ny % QK_K == 0) { + if (nx % QK_K == 0) { new_type = GGML_TYPE_Q6_K; } } else if (name.find("attn_v.weight") != std::string::npos) { @@ -4812,8 +4811,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; - if (nx % QK_K != 0 || ny % QK_K != 0) { - LLAMA_LOG_INFO("\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K); + if (nx % QK_K != 0) { + LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K); convert_incompatible_tensor = true; } } From 50526f37eba0b28336700890242ff282b949cd83 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Sat, 26 Aug 2023 12:53:52 -0400 Subject: [PATCH 053/147] llama : use std::abs in llama_sample_tail_free (#2800) Plain 'abs' casts the input to int. --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 59105db1c..2b88485a8 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3887,7 +3887,7 @@ void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * // Calculate absolute value of second derivatives for (size_t i = 0; i < second_derivatives.size(); ++i) { - second_derivatives[i] = abs(second_derivatives[i]); + second_derivatives[i] = std::abs(second_derivatives[i]); } // Normalize the second derivatives From 72f895c923ba98b8f2af294440206f35915c0501 Mon Sep 17 00:00:00 2001 From: "Dr. Tom Murphy VII Ph.D" <499244+tom7@users.noreply.github.com> Date: Sat, 26 Aug 2023 14:12:56 -0400 Subject: [PATCH 054/147] main : fix bug (penalize_nl=false doesn't work) + suppress warning on mingw (#1528) * Fix bug in main.cpp where penalize_nl=false has no effect. It modifies the underlying logits array, but at this point we are already working on the candidates copy. * Suppress redefinition warning for NOMINMAX on mingw. In my installation, this macro is already defined by /usr/lib/gcc/x86_64-w64-mingw32/11/include/c++/x86_64-w64-mingw32/bits/os_defines.h:45. * main : fix indentation * main : pass ctx to llama_token_nl() --------- Co-authored-by: Georgi Gerganov --- examples/main/main.cpp | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 4665b82fe..11d7a7e4f 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -604,7 +604,12 @@ int main(int argc, char ** argv) { last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, last_n_repeat, alpha_frequency, alpha_presence); if (!penalize_nl) { - logits[llama_token_nl(ctx)] = nl_logit; + for (size_t idx = 0; idx < candidates_p.size; idx++) { + if (candidates_p.data[idx].id == llama_token_nl(ctx)) { + candidates_p.data[idx].logit = nl_logit; + break; + } + } } if (grammar != NULL) { From 741ca7dd1cec0a0349494742b9083d6ef4cd73c5 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Sat, 26 Aug 2023 14:17:51 -0400 Subject: [PATCH 055/147] llama : move #includes out of _GNU_SOURCE conditional (#2817) --- llama.cpp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/llama.cpp b/llama.cpp index 2b88485a8..62889b3ed 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1,9 +1,6 @@ // Defines fileno on msys: #ifndef _GNU_SOURCE #define _GNU_SOURCE -#include -#include -#include #endif #include "llama.h" @@ -62,6 +59,9 @@ #include #include #include +#include +#include +#include #include #include #include From 61d1a2895eeca55e0c8b7018492f6ab9c90cff78 Mon Sep 17 00:00:00 2001 From: Tungsten842 Date: Sat, 26 Aug 2023 20:19:44 +0200 Subject: [PATCH 056/147] flake.nix : add rocm support and cleanup (#2808) --- flake.lock | 12 ++++++------ flake.nix | 43 +++++++++++++++++++++++-------------------- 2 files changed, 29 insertions(+), 26 deletions(-) diff --git a/flake.lock b/flake.lock index 33164e096..a7777d05d 100644 --- a/flake.lock +++ b/flake.lock @@ -5,11 +5,11 @@ "systems": "systems" }, "locked": { - "lastModified": 1685518550, - "narHash": "sha256-o2d0KcvaXzTrPRIo0kOLV0/QXHhDQ5DTi+OxcjO8xqY=", + "lastModified": 1692799911, + "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=", "owner": "numtide", "repo": "flake-utils", - "rev": "a1720a10a6cfe8234c0e93907ffe81be440f4cef", + "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44", "type": "github" }, "original": { @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1685931219, - "narHash": "sha256-8EWeOZ6LKQfgAjB/USffUSELPRjw88A+xTcXnOUvO5M=", + "lastModified": 1692913444, + "narHash": "sha256-1SvMQm2DwofNxXVtNWWtIcTh7GctEVrS/Xel/mdc6iY=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "7409480d5c8584a1a83c422530419efe4afb0d19", + "rev": "18324978d632ffc55ef1d928e81630c620f4f447", "type": "github" }, "original": { diff --git a/flake.nix b/flake.nix index d454cedc3..02095411e 100644 --- a/flake.nix +++ b/flake.nix @@ -6,6 +6,9 @@ outputs = { self, nixpkgs, flake-utils }: flake-utils.lib.eachDefaultSystem (system: let + name = "llama.cpp"; + src = ./.; + meta.mainProgram = "llama"; inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin; buildInputs = with pkgs; [ openmpi ]; osSpecific = with pkgs; buildInputs ++ @@ -31,7 +34,7 @@ with pkgs; [ openblas ] ); pkgs = import nixpkgs { inherit system; }; - nativeBuildInputs = with pkgs; [ cmake pkgconfig ]; + nativeBuildInputs = with pkgs; [ cmake ninja pkgconfig ]; llama-python = pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]); postPatch = '' @@ -44,35 +47,35 @@ mv $out/bin/server $out/bin/llama-server ''; cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ]; - in { + in + { packages.default = pkgs.stdenv.mkDerivation { - name = "llama.cpp"; - src = ./.; - postPatch = postPatch; - nativeBuildInputs = nativeBuildInputs; - buildInputs = osSpecific; + inherit name src meta postPatch nativeBuildInputs buildInputs postInstall; cmakeFlags = cmakeFlags ++ (if isAarch64 && isDarwin then [ - "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1" - "-DLLAMA_METAL=ON" - ] else [ - "-DLLAMA_BLAS=ON" - "-DLLAMA_BLAS_VENDOR=OpenBLAS" + "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1" + "-DLLAMA_METAL=ON" + ] else [ + "-DLLAMA_BLAS=ON" + "-DLLAMA_BLAS_VENDOR=OpenBLAS" ]); - postInstall = postInstall; - meta.mainProgram = "llama"; }; packages.opencl = pkgs.stdenv.mkDerivation { - name = "llama.cpp"; - src = ./.; - postPatch = postPatch; - nativeBuildInputs = nativeBuildInputs; + inherit name src meta postPatch nativeBuildInputs postInstall; buildInputs = with pkgs; buildInputs ++ [ clblast ]; cmakeFlags = cmakeFlags ++ [ "-DLLAMA_CLBLAST=ON" ]; - postInstall = postInstall; - meta.mainProgram = "llama"; + }; + packages.rocm = pkgs.stdenv.mkDerivation { + inherit name src meta postPatch nativeBuildInputs postInstall; + buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ]; + cmakeFlags = cmakeFlags ++ [ + "-DLLAMA_HIPBLAS=1" + "-DCMAKE_C_COMPILER=hipcc" + "-DCMAKE_CXX_COMPILER=hipcc" + "-DCMAKE_POSITION_INDEPENDENT_CODE=ON" + ]; }; apps.llama-server = { type = "app"; From c7d92e6dfec3f54849f3a0ba373054d29f321ea2 Mon Sep 17 00:00:00 2001 From: Tim Miller Date: Sun, 27 Aug 2023 03:27:07 +0900 Subject: [PATCH 057/147] llama : use Unicode Escape Sequence to replace encoded characters (#2814) The use of special characters within source files can break compiling on some computers with different region and language settings. Using Unicode escape sequences should allow for the code to be compiled on all setups without needing to change your computers settings or switch regions. --- llama.cpp | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/llama.cpp b/llama.cpp index 62889b3ed..05c54c213 100644 --- a/llama.cpp +++ b/llama.cpp @@ -955,10 +955,10 @@ struct llama_vocab { id linefeed_id = 13; int find_bpe_rank(std::string token_left, std::string token_right) const { - replace_all(token_left, " ", "Ġ"); - replace_all(token_left, "\n", "Ċ"); - replace_all(token_right, " ", "Ġ"); - replace_all(token_right, "\n", "Ċ"); + replace_all(token_left, " ", "\u0120"); + replace_all(token_left, "\n", "\u010A"); + replace_all(token_right, " ", "\u0120"); + replace_all(token_right, "\n", "\u010A"); auto it = bpe_ranks.find(std::make_pair(token_left, token_right)); if (it == bpe_ranks.end()) { From 730d9c681e339b76407659344e5a2cd50af7d7d5 Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Sat, 26 Aug 2023 14:13:36 -0600 Subject: [PATCH 058/147] convert.py : advanced option (#2753) * Allow convert.py to convert to q8_0 Fix issue with bounded_parallel_map and greedy consuming iterator Display elapsed time during conversion * Add --concurrency option Minor improvements to help text Clean up bounded_parallel_map function a bit * Massive speed improvement thanks to Cebtenzzre * Refactor types --- convert.py | 206 ++++++++++++++++++++++++++++++++++------------------- 1 file changed, 133 insertions(+), 73 deletions(-) diff --git a/convert.py b/convert.py index d44e5a8c4..a15e6ccd2 100755 --- a/convert.py +++ b/convert.py @@ -3,6 +3,7 @@ import gguf import argparse import concurrent.futures +from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor import copy import enum import faulthandler @@ -17,13 +18,14 @@ import re import signal import struct import sys +import time import zipfile import numpy as np from abc import ABCMeta, abstractmethod from dataclasses import dataclass from pathlib import Path -from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Iterable, List, Literal, Optional, Sequence, Tuple, TypeVar, Union) +from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, TypeVar, Union) from sentencepiece import SentencePieceProcessor # type: ignore if TYPE_CHECKING: @@ -37,30 +39,70 @@ NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]' ARCH=gguf.MODEL_ARCH.LLAMA NAMES=gguf.MODEL_TENSOR_NAMES[ARCH] +DEFAULT_CONCURRENCY = 8 # # data types # @dataclass(frozen=True) -class UnquantizedDataType: +class DataType: name: str + dtype: 'np.dtype[Any]' + valid_conversions: List[str] -DT_F16 = UnquantizedDataType('F16') -DT_F32 = UnquantizedDataType('F32') -DT_I32 = UnquantizedDataType('I32') -DT_BF16 = UnquantizedDataType('BF16') + def elements_to_bytes(self, n_elements: int) -> int: + return n_elements * self.dtype.itemsize -DataType = Union[UnquantizedDataType] +@dataclass(frozen=True) +class UnquantizedDataType(DataType): + pass -DATA_TYPE_TO_NUMPY: Dict[DataType, 'np.dtype[Any]'] = { - DT_BF16: np.dtype(np.uint16), - DT_F16: np.dtype(np.float16), - DT_F32: np.dtype(np.float32), - DT_I32: np.dtype(np.int32), -} +DT_F16 = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0']) +DT_F32 = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0']) +DT_I32 = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = []) +DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0']) -NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = \ - {dtype: data_type for (data_type, dtype) in DATA_TYPE_TO_NUMPY.items()} +@dataclass(frozen=True) +class QuantizedDataType(DataType): + block_size: int + quantized_dtype: 'np.dtype[Any]' + ggml_type: gguf.GGMLQuantizationType + + def quantize(self, arr: NDArray) -> NDArray: + raise NotImplementedError(f'Quantization for {self.name} not implemented') + + def elements_to_bytes(self, n_elements: int) -> int: + assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}' + return self.quantized_dtype.itemsize * (n_elements // self.block_size) + +@dataclass(frozen=True) +class Q8_0QuantizedDataType(QuantizedDataType): + # Mini Q8_0 quantization in Python! + def quantize(self, arr: NDArray) -> NDArray: + assert arr.size % self.block_size == 0 and arr.size != 0, f'Bad array size {arr.size}' + assert arr.dtype == np.float32, f'Bad array type {arr.dtype}' + n_blocks = arr.size // self.block_size + blocks = arr.reshape((n_blocks, self.block_size)) + # Much faster implementation of block quantization contributed by @Cebtenzzre + def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[Tuple[Any, Any]]: + d = abs(blocks).max(axis = 1) / np.float32(127) + with np.errstate(divide = 'ignore'): + qs = (blocks / d[:, None]).round() + qs[d == 0] = 0 + yield from zip(d, qs) + return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype) + +DT_Q8_0 = Q8_0QuantizedDataType('Q8_0', + dtype = np.dtype(np.float32), valid_conversions = [], + ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32, + quantized_dtype = np.dtype([('d', ' DataType: - if len(tensor.shape) == 1: - # 1D tensors are always F32. - return DT_F32 - elif self == GGMLFileType.AllF32: - return DT_F32 - elif self == GGMLFileType.MostlyF16: - return DT_F16 - else: + dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self) + if dt is None: raise ValueError(self) + # 1D tensors are always F32. + return dt if len(tensor.shape) > 1 else DT_F32 +GGML_FILE_TYPE_TO_DATA_TYPE: Dict[GGMLFileType, DataType] = { + GGMLFileType.AllF32 : DT_F32, + GGMLFileType.MostlyF16 : DT_F16, + GGMLFileType.MostlyQ8_0: DT_Q8_0, +} # # hparams loading @@ -415,7 +459,7 @@ class UnquantizedTensor(Tensor): self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype] def astype(self, data_type: DataType) -> Tensor: - dtype = DATA_TYPE_TO_NUMPY[data_type] + dtype = data_type.dtype if self.data_type == DT_BF16: self.ndarray = bf16_to_fp32(self.ndarray) return UnquantizedTensor(self.ndarray.astype(dtype)) @@ -454,22 +498,6 @@ def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, conv GGMLCompatibleTensor = Union[UnquantizedTensor] -class DeferredPermutedTensor(Tensor): - def __init__(self, base: Tensor, n_head: int, n_head_kv: int) -> None: - self.base = base - self.n_head = n_head - self.data_type = self.base.data_type - - def astype(self, data_type: DataType) -> Tensor: - return self.base.astype(data_type).permute(self.n_head, self.n_head_kv) - - def to_ggml(self) -> GGMLCompatibleTensor: - return self.base.to_ggml().permute(self.n_head, self.n_head_kv) - - def permute(self, n_head: int, n_head_kv: int) -> Tensor: - raise Exception("shouldn't permute twice") - - @dataclass class LazyTensor: _load: Callable[[], Tensor] @@ -479,7 +507,9 @@ class LazyTensor: def load(self) -> Tensor: ret = self._load() - assert ret.data_type == self.data_type, (self.data_type, ret.data_type, self.description) + # Should be okay if it maps to the same numpy type? + assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \ + (self.data_type, ret.data_type, self.description) return ret def astype(self, data_type: DataType) -> 'LazyTensor': @@ -490,8 +520,8 @@ class LazyTensor: return LazyTensor(load, self.shape, data_type, f'convert({data_type}) {self.description}') def validate_conversion_to(self, data_type: DataType) -> None: - if data_type == self.data_type: - return + if data_type != self.data_type and data_type.name not in self.data_type.valid_conversions: + raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.') LazyModel = Dict[str, LazyTensor] @@ -617,9 +647,7 @@ class LazyUnpickler(pickle.Unpickler): info = self.zip_file.getinfo(filename) def load(offset: int, elm_count: int) -> NDArray: - dtype = DATA_TYPE_TO_NUMPY.get(data_type) - if dtype is None: - raise Exception("tensor stored in unsupported format") + dtype = data_type.dtype fp = self.zip_file.open(info) fp.seek(offset * dtype.itemsize) size = elm_count * dtype.itemsize @@ -683,7 +711,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus: def convert(info: Dict[str, Any]) -> LazyTensor: data_type = SAFETENSORS_DATA_TYPES[info['dtype']] - numpy_dtype = DATA_TYPE_TO_NUMPY[data_type] + numpy_dtype = data_type.dtype shape: List[int] = info['shape'] begin, end = info['data_offsets'] assert 0 <= begin <= end <= len(byte_buf) @@ -723,23 +751,35 @@ def lazy_load_file(path: Path) -> ModelPlus: In = TypeVar('In') Out = TypeVar('Out') -def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int) -> Iterable[Out]: +def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, factory: Callable = ThreadPoolExecutor) -> Iterable[Out]: '''Parallel map, but with backpressure. If the caller doesn't call `next` fast enough, this will stop calling `func` at some point rather than letting results pile up in memory. Specifically, there is a max of one output value buffered per thread.''' - with concurrent.futures.ThreadPoolExecutor() as executor: + if concurrency < 2: + yield from map(func, iterable) + # Not reached. + iterable = iter(iterable) + with factory(max_workers = max_workers) as executor: futures: List[concurrent.futures.Future[Out]] = [] - items_rev = list(iterable)[::-1] - for i in range(min(concurrency, len(items_rev))): - futures.append(executor.submit(func, items_rev.pop())) + done = False + for _ in range(concurrency): + try: + futures.append(executor.submit(func, next(iterable))) + except StopIteration: + done = True + break + while futures: result = futures.pop(0).result() - if items_rev: - futures.append(executor.submit(func, items_rev.pop())) + while not done and len(futures) < concurrency: + try: + futures.append(executor.submit(func, next(iterable))) + except StopIteration: + done = True + break yield result - def check_vocab_size(params: Params, vocab: Vocab) -> None: if params.n_vocab != vocab.vocab_size: assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab) @@ -804,12 +844,11 @@ class OutputFile: self.gguf.add_token_types(toktypes) def add_tensor_info(self, name: str, tensor: LazyTensor) -> None: - n_elements = 1 - for dim in tensor.shape: - n_elements *= dim - data_type = DATA_TYPE_TO_NUMPY[tensor.data_type] - data_nbytes = n_elements * data_type.itemsize - self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes) + n_elements = int(np.prod(tensor.shape)) + raw_dtype = getattr(tensor.data_type, 'ggml_type', None) + data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype + data_nbytes = tensor.data_type.elements_to_bytes(n_elements) + self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype = raw_dtype) def write_meta(self) -> None: self.gguf.write_header_to_file() @@ -835,7 +874,20 @@ class OutputFile: of.close() @staticmethod - def write_all(fname_out: Path, params: Params, model: LazyModel, vocab: Vocab) -> None: + def do_item(item: Tuple[str, LazyTensor]) -> Tuple[DataType, NDArray]: + name, lazy_tensor = item + tensor = lazy_tensor.load().to_ggml() + return (lazy_tensor.data_type, tensor.ndarray) + + @staticmethod + def maybe_do_quantize(item: Tuple[DataType, NDArray]) -> NDArray: + dt, arr = item + if not isinstance(dt, QuantizedDataType): + return arr + return dt.quantize(arr) + + @staticmethod + def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, concurrency: int = DEFAULT_CONCURRENCY) -> None: check_vocab_size(params, vocab) of = OutputFile(fname_out) @@ -851,16 +903,19 @@ class OutputFile: of.write_meta() of.write_tensor_info() - def do_item(item: Tuple[str, LazyTensor]) -> NDArray: - name, lazy_tensor = item - return lazy_tensor.load().to_ggml().ndarray - # tensor data - ndarrays = bounded_parallel_map(do_item, model.items(), concurrency=8) + ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency) + if ftype == GGMLFileType.MostlyQ8_0: + ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, factory = ProcessPoolExecutor) + else: + ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner) + + start = time.time() for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)): + elapsed = time.time() - start size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) padi = len(str(len(model))) - print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type}") + print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}") of.gguf.write_tensor_data(ndarray) of.close() @@ -872,6 +927,8 @@ def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFi return GGMLFileType.AllF32 if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)): return GGMLFileType.MostlyF16 + if output_type_str == "q8_0": + return GGMLFileType.MostlyQ8_0 name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()} @@ -918,7 +975,7 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel: print(f"skipping tensor {name_new}") continue else: - print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type} | {lazy_tensor.shape}") + print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}") out[name_new] = lazy_tensor return out @@ -1023,6 +1080,7 @@ def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path: namestr = { GGMLFileType.AllF32: "f32", GGMLFileType.MostlyF16: "f16", + GGMLFileType.MostlyQ8_0:"q8_0", }[file_type] ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf" if ret in model_paths: @@ -1046,12 +1104,13 @@ def main(args_in: Optional[List[str]] = None) -> None: parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outtype", choices=["f32", "f16"], help="output format (default: based on input)") + parser.add_argument("--outtype", choices=["f32", "f16", "q8_0"], help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)") parser.add_argument("--vocabtype", choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm") parser.add_argument("--ctx", type=int, help="model training context (default: based on input)") + parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY) args = parser.parse_args(args_in) if args.dump_single: @@ -1073,6 +1132,7 @@ def main(args_in: Optional[List[str]] = None) -> None: params.ftype = { "f32": GGMLFileType.AllF32, "f16": GGMLFileType.MostlyF16, + "q8_0": GGMLFileType.MostlyQ8_0, }[args.outtype] print(f"params = {params}") @@ -1104,7 +1164,7 @@ def main(args_in: Optional[List[str]] = None) -> None: params.ftype = ftype print(f"Writing {outfile}, format {ftype}") - OutputFile.write_all(outfile, params, model, vocab) + OutputFile.write_all(outfile, ftype, params, model, vocab, concurrency = args.concurrency) print(f"Wrote {outfile}") From c1ac54b77aaba10d029084d152be786102010eb2 Mon Sep 17 00:00:00 2001 From: Bruce MacDonald Date: Sat, 26 Aug 2023 16:11:45 -0700 Subject: [PATCH 059/147] server : add `/detokenize` endpoint (#2802) * Add a /detokenize endpoint to the example server * remove trailing white-space --- examples/server/README.md | 6 ++++++ examples/server/server.cpp | 21 +++++++++++++++++++++ 2 files changed, 27 insertions(+) diff --git a/examples/server/README.md b/examples/server/README.md index 7105e9020..517608046 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -164,6 +164,12 @@ node index.js Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`. +- **POST** `/detokenize`: Convert tokens to text. + + *Options:* + + `tokens`: Set the tokens to detokenize. + - **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does. *Options:* diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 3300553f9..a4b4d6418 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1104,6 +1104,12 @@ static json format_tokenizer_response(const std::vector &tokens) {"tokens", tokens}}; } +static json format_detokenized_response(std::string content) +{ + return json{ + {"content", content}}; +} + template static T json_value(const json &body, const std::string &key, const T &default_value) { @@ -1501,6 +1507,21 @@ int main(int argc, char **argv) const json data = format_tokenizer_response(tokens); return res.set_content(data.dump(), "application/json"); }); + svr.Post("/detokenize", [&llama](const Request &req, Response &res) + { + auto lock = llama.lock(); + + const json body = json::parse(req.body); + std::string content; + if (body.count("tokens") != 0) + { + const std::vector tokens = body["tokens"]; + content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend()); + } + + const json data = format_detokenized_response(content); + return res.set_content(data.dump(), "application/json"); }); + svr.Post("/embedding", [&llama](const Request &req, Response &res) { auto lock = llama.lock(); From 789c8c945a2814e1487e18e68823d9926e3b1454 Mon Sep 17 00:00:00 2001 From: slaren Date: Sun, 27 Aug 2023 09:03:27 +0200 Subject: [PATCH 060/147] ci : add LoRA test to CI (#2650) * ci : add lora test ggml-ci * move lora summary to the top, add lora logs ggml-ci * ci : decrease CPU ppl runs to 2 to avoide 20 min timeout ggml-ci * add 7b lora test use 1 thread for CUDA generation tests ggml-ci * add test with q8_0 (cpu only) ggml-ci --------- Co-authored-by: Georgi Gerganov --- ci/run.sh | 140 +++++++++++++++++++++++++++++++++++++++++++++--------- 1 file changed, 118 insertions(+), 22 deletions(-) diff --git a/ci/run.sh b/ci/run.sh index e1486e7c1..942b2e00c 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -196,17 +196,17 @@ function gg_run_open_llama_3b_v2 { (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log - (time ./bin/perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log - (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log - (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log - (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log - (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log - (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log - (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log - (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log - (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log - (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log + (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log + (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log + (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log + (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log + (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log + (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log + (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log + (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log function check_ppl { qnt="$1" @@ -233,6 +233,48 @@ function gg_run_open_llama_3b_v2 { check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log + # lora + function compare_ppl { + qnt="$1" + ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1) + ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1) + + if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then + printf ' - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2" + return 20 + fi + + printf ' - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2" + return 0 + } + + path_lora="../models-mnt/open-llama/3B-v2/lora" + path_shakespeare="../models-mnt/shakespeare" + + shakespeare="${path_shakespeare}/shakespeare.txt" + lora_shakespeare="${path_lora}/ggml-adapter-model.bin" + + gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json + gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin + gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt + + python3 ../convert-lora-to-ggml.py ${path_lora} + + # f16 + (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log + (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log + compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log + + # q8_0 + (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log + (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log + compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log + + # q8_0 + f16 lora-base + (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log + compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log + + set +e } @@ -242,6 +284,7 @@ function gg_sum_open_llama_3b_v2 { gg_printf 'OpenLLaMA 3B-v2:\n' gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)" gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)" + gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)" gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)" gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)" gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)" @@ -253,6 +296,11 @@ function gg_sum_open_llama_3b_v2 { gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)" gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)" gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)" + gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)" + gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)" + gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)" + gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)" + gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)" } # open_llama_7b_v2 @@ -310,17 +358,17 @@ function gg_run_open_llama_7b_v2 { ./bin/quantize ${model_f16} ${model_q5_k} q5_k ./bin/quantize ${model_f16} ${model_q6_k} q6_k - (time ./bin/main --model ${model_f16} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/main --model ${model_q8_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log - (time ./bin/main --model ${model_q4_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log - (time ./bin/main --model ${model_q4_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log - (time ./bin/main --model ${model_q5_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log - (time ./bin/main --model ${model_q5_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log - (time ./bin/main --model ${model_q2_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log - (time ./bin/main --model ${model_q3_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log - (time ./bin/main --model ${model_q4_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log - (time ./bin/main --model ${model_q5_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log - (time ./bin/main --model ${model_q6_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/main --model ${model_f16} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log + (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log + (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log + (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log + (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log + (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log + (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log + (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log + (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log (time ./bin/perplexity --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log @@ -359,6 +407,48 @@ function gg_run_open_llama_7b_v2 { check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log + # lora + function compare_ppl { + qnt="$1" + ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1) + ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1) + + if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then + printf ' - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2" + return 20 + fi + + printf ' - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2" + return 0 + } + + path_lora="../models-mnt/open-llama/7B-v2/lora" + path_shakespeare="../models-mnt/shakespeare" + + shakespeare="${path_shakespeare}/shakespeare.txt" + lora_shakespeare="${path_lora}/ggml-adapter-model.bin" + + gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json + gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin + gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt + + python3 ../convert-lora-to-ggml.py ${path_lora} + + # f16 + (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log + (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log + compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log + + # currently not supported by the CUDA backend + # q8_0 + #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log + #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log + #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log + + # q8_0 + f16 lora-base + #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log + #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log + set +e } @@ -368,6 +458,7 @@ function gg_sum_open_llama_7b_v2 { gg_printf 'OpenLLaMA 7B-v2:\n' gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)" gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)" + gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)" gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)" gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)" gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)" @@ -379,6 +470,11 @@ function gg_sum_open_llama_7b_v2 { gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)" gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)" gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)" + gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)" + gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)" + #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)" + #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)" + #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)" } ## main From 1591e2e590762011b43b10a9b6e04f13f98f2aa5 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Przemys=C5=82aw=20Pawe=C5=82czyk?= Date: Sun, 27 Aug 2023 10:10:25 +0200 Subject: [PATCH 061/147] ggml : detect SSSE3 (#2825) * ggml : add ggml_cpu_has_ssse3 * llama : show SSSE3 in system info --- ggml.c | 8 ++++++++ ggml.h | 1 + llama.cpp | 1 + 3 files changed, 10 insertions(+) diff --git a/ggml.c b/ggml.c index 8cb5c404f..394fb459f 100644 --- a/ggml.c +++ b/ggml.c @@ -20516,6 +20516,14 @@ int ggml_cpu_has_sse3(void) { #endif } +int ggml_cpu_has_ssse3(void) { +#if defined(__SSSE3__) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_vsx(void) { #if defined(__POWER9_VECTOR__) return 1; diff --git a/ggml.h b/ggml.h index 421c0df60..b418153ba 100644 --- a/ggml.h +++ b/ggml.h @@ -1944,6 +1944,7 @@ extern "C" { GGML_API int ggml_cpu_has_clblast (void); GGML_API int ggml_cpu_has_gpublas (void); GGML_API int ggml_cpu_has_sse3 (void); + GGML_API int ggml_cpu_has_ssse3 (void); GGML_API int ggml_cpu_has_vsx (void); // diff --git a/llama.cpp b/llama.cpp index 05c54c213..e956c0163 100644 --- a/llama.cpp +++ b/llama.cpp @@ -6194,6 +6194,7 @@ const char * llama_print_system_info(void) { s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | "; s += "BLAS = " + std::to_string(ggml_cpu_has_blas()) + " | "; s += "SSE3 = " + std::to_string(ggml_cpu_has_sse3()) + " | "; + s += "SSSE3 = " + std::to_string(ggml_cpu_has_ssse3()) + " | "; s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | "; return s.c_str(); From edd4c1481708fcd788b0e423268304fd26e2b125 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 27 Aug 2023 14:19:19 +0300 Subject: [PATCH 062/147] llama : more tokenizer fixes (#2810) * tests : write a Python tokenizer test (wip) * llama : prefix input text for tokenization with whitespace * llama : distinguish pieces from decoded text + fix detokenization * common : add comments * examples : no longer manually add leading space when tokenizing * tests : use Python to generate tokenizer tests for C++ * tests : add option to tokenize text files ggml-ci * tests : add test-tokenizer-1.py * llama.cpp : fix LF token * hellaswag : move the concat space for clarity * tests : add falcon tests (py + cpp, currently do not pass Unicode) ggml-ci * common : temporary separate llama_detokenize calls for SPM and BPE --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> --- common/common.cpp | 39 +++- common/common.h | 22 ++- examples/beam_search/beam_search.cpp | 6 +- examples/embd-input/embd-input-lib.cpp | 2 +- examples/embedding/embedding.cpp | 5 +- examples/main/main.cpp | 20 +- examples/perplexity/perplexity.cpp | 4 +- examples/save-load-state/save-load-state.cpp | 4 +- examples/server/server.cpp | 16 +- examples/simple/simple.cpp | 4 +- .../train-text-from-scratch.cpp | 4 +- llama.cpp | 60 +++--- llama.h | 10 +- tests/CMakeLists.txt | 6 +- tests/test-tokenizer-0-falcon.cpp | 178 +++++++++++++++++ tests/test-tokenizer-0-falcon.py | 83 ++++++++ tests/test-tokenizer-0-llama.cpp | 182 ++++++++++++++++++ tests/test-tokenizer-0-llama.py | 95 +++++++++ tests/test-tokenizer-0.cpp | 141 -------------- tests/test-tokenizer-1.cpp | 14 +- 20 files changed, 671 insertions(+), 224 deletions(-) create mode 100644 tests/test-tokenizer-0-falcon.cpp create mode 100644 tests/test-tokenizer-0-falcon.py create mode 100644 tests/test-tokenizer-0-llama.cpp create mode 100644 tests/test-tokenizer-0-llama.py delete mode 100644 tests/test-tokenizer-0.cpp diff --git a/common/common.cpp b/common/common.cpp index ff19ec4e5..0d91a6a35 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -733,12 +733,12 @@ std::vector llama_tokenize( return result; } -std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) { +std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) { std::vector result(8, 0); - const int n_tokens = llama_token_to_str(ctx, token, result.data(), result.size()); + const int n_tokens = llama_token_to_piece(ctx, token, result.data(), result.size()); if (n_tokens < 0) { result.resize(-n_tokens); - int check = llama_token_to_str(ctx, token, result.data(), result.size()); + int check = llama_token_to_piece(ctx, token, result.data(), result.size()); GGML_ASSERT(check == -n_tokens); } else { result.resize(n_tokens); @@ -746,3 +746,36 @@ std::string llama_token_to_str(const struct llama_context * ctx, llama_token tok return std::string(result.data(), result.size()); } + +std::string llama_detokenize_spm(llama_context * ctx, const std::vector & tokens) { + const llama_token bos_id = llama_token_bos(ctx); + + std::string piece; + std::string result; + + for (size_t i = 0; i < tokens.size(); ++i) { + piece = llama_token_to_piece(ctx, tokens[i]); + + // remove the leading space of the first non-BOS token + if (((tokens[0] == bos_id && i == 1) || (tokens[0] != bos_id && i == 0)) && piece[0] == ' ') { + piece = piece.substr(1); + } + + result += piece; + } + + return result; +} + +std::string llama_detokenize_bpe(llama_context * ctx, const std::vector & tokens) { + std::string piece; + std::string result; + + for (size_t i = 0; i < tokens.size(); ++i) { + piece = llama_token_to_piece(ctx, tokens[i]); + + result += piece; + } + + return result; +} diff --git a/common/common.h b/common/common.h index ce61265f8..97fda2be7 100644 --- a/common/common.h +++ b/common/common.h @@ -116,11 +116,31 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param // Vocab utils // +// tokenizes a string into a vector of tokens +// should work similar to Python's `tokenizer.encode` std::vector llama_tokenize( struct llama_context * ctx, const std::string & text, bool add_bos); -std::string llama_token_to_str( +// tokenizes a token into a piece +// should work similar to Python's `tokenizer.id_to_piece` +std::string llama_token_to_piece( const struct llama_context * ctx, llama_token token); + +// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function +// that takes into account the tokenizer type and decides how to handle the leading space +// +// detokenizes a vector of tokens into a string +// should work similar to Python's `tokenizer.decode` +// removes the leading space from the first non-BOS token +std::string llama_detokenize_spm( + llama_context * ctx, + const std::vector & tokens); + +// detokenizes a vector of tokens into a string +// should work similar to Python's `tokenizer.decode` +std::string llama_detokenize_bpe( + llama_context * ctx, + const std::vector & tokens); diff --git a/examples/beam_search/beam_search.cpp b/examples/beam_search/beam_search.cpp index 1c04fabc2..42c7c7254 100644 --- a/examples/beam_search/beam_search.cpp +++ b/examples/beam_search/beam_search.cpp @@ -35,7 +35,7 @@ struct ostream_beam_view { std::ostream& operator<<(std::ostream& os, const ostream_beam_view & obv) { os << "p(" << obv.beam_view.p << ") eob(" << std::boolalpha << obv.beam_view.eob << ") tokens("; for (size_t i = 0 ; i < obv.beam_view.n_tokens ; ++i) { - os << llama_token_to_str(obv.ctx, obv.beam_view.tokens[i]); + os << llama_token_to_piece(obv.ctx, obv.beam_view.tokens[i]); } return os << ')'; } @@ -156,7 +156,7 @@ int main(int argc, char ** argv) for( auto id : tokens_list ) { - std::cout << llama_token_to_str(ctx, id); + std::cout << llama_token_to_piece(ctx, id); } std::cout << std::flush; @@ -175,7 +175,7 @@ int main(int argc, char ** argv) std::cout << "\n\n"; for (llama_token const token_id : callback_data.response) { - std::cout << llama_token_to_str(ctx,token_id); + std::cout << llama_token_to_piece(ctx,token_id); } std::cout << std::endl; diff --git a/examples/embd-input/embd-input-lib.cpp b/examples/embd-input/embd-input-lib.cpp index 8a6ad882e..036bdb398 100644 --- a/examples/embd-input/embd-input-lib.cpp +++ b/examples/embd-input/embd-input-lib.cpp @@ -214,7 +214,7 @@ const char * sampling(struct MyModel * mymodel) { if (id == llama_token_eos(ctx)) { ret = ""; } else { - ret = llama_token_to_str(ctx, id); + ret = llama_token_to_piece(ctx, id); } eval_id(mymodel, id); return ret.c_str(); diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 38395c75b..93d583b5c 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -56,9 +56,6 @@ int main(int argc, char ** argv) { int n_past = 0; - // Add a space in front of the first character to match OG llama tokenizer behavior - params.prompt.insert(0, 1, ' '); - // tokenize the prompt auto embd_inp = ::llama_tokenize(ctx, params.prompt, true); @@ -67,7 +64,7 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str()); fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); for (int i = 0; i < (int) embd_inp.size(); i++) { - fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]).c_str()); + fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str()); } fprintf(stderr, "\n"); } diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 11d7a7e4f..3ce57f436 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -195,11 +195,6 @@ int main(int argc, char ** argv) { // tokenize the prompt std::vector embd_inp; - if (llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM) { - // Add a space in front of the first character to match OG llama tokenizer behavior - params.prompt.insert(0, 1, ' '); - } - if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) { embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos); } else { @@ -216,7 +211,6 @@ int main(int argc, char ** argv) { int guidance_offset = 0; int original_prompt_len = 0; if (ctx_guidance) { - params.cfg_negative_prompt.insert(0, 1, ' '); guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos); std::vector original_inp = ::llama_tokenize(ctx, params.prompt, add_bos); @@ -285,7 +279,7 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str()); fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); for (int i = 0; i < (int) embd_inp.size(); i++) { - fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_str(ctx, embd_inp[i]).c_str()); + fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str()); } if (ctx_guidance) { @@ -293,14 +287,14 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str()); fprintf(stderr, "%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size()); for (int i = 0; i < (int) guidance_inp.size(); i++) { - fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_str(ctx, guidance_inp[i]).c_str()); + fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str()); } } if (params.n_keep > 0) { fprintf(stderr, "%s: static prompt based on n_keep: '", __func__); for (int i = 0; i < params.n_keep; i++) { - fprintf(stderr, "%s", llama_token_to_str(ctx, embd_inp[i]).c_str()); + fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_inp[i]).c_str()); } fprintf(stderr, "'\n"); } @@ -456,7 +450,7 @@ int main(int argc, char ** argv) { //printf("\n---\n"); //printf("resetting: '"); //for (int i = 0; i < (int) embd.size(); i++) { - // printf("%s", llama_token_to_str(ctx, embd[i])); + // printf("%s", llama_token_to_piece(ctx, embd[i])); //} //printf("'\n"); //printf("\n---\n"); @@ -509,7 +503,7 @@ int main(int argc, char ** argv) { input_size = embd_guidance.size(); //fprintf(stderr, "\n---------------------\n"); //for (int i = 0; i < (int) embd_guidance.size(); i++) { - //fprintf(stderr, "%s", llama_token_to_str(ctx, embd_guidance[i])); + //fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_guidance[i])); //} //fprintf(stderr, "\n---------------------\n"); } else { @@ -673,7 +667,7 @@ int main(int argc, char ** argv) { // display text if (input_echo) { for (auto id : embd) { - printf("%s", llama_token_to_str(ctx, id).c_str()); + printf("%s", llama_token_to_piece(ctx, id).c_str()); } fflush(stdout); } @@ -689,7 +683,7 @@ int main(int argc, char ** argv) { if (params.antiprompt.size()) { std::string last_output; for (auto id : last_n_tokens) { - last_output += llama_token_to_str(ctx, id); + last_output += llama_token_to_piece(ctx, id); } is_antiprompt = false; diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index fd89852d6..b596d0626 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -392,7 +392,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) { hs_data[i].context = prompt_lines[idx*6]; hs_data[i].gold_ending_idx = std::stoi( prompt_lines[idx*6+1] ); for (size_t j=0; j < 4; j++) { - hs_data[i].ending[j] = " " + prompt_lines[idx*6+2+j]; + hs_data[i].ending[j] = prompt_lines[idx*6+2+j]; } // Delete the selected random example from the prompt @@ -417,7 +417,7 @@ void hellaswag_score(llama_context * ctx, const gpt_params & params) { size_t context_size = context_embd.size(); for (int i = 0; i < 4; ++i) { - ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[i], add_bos); + ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + " " + hs_data[task_idx].ending[i], add_bos); for (int k = 0; k < int(context_size); ++k) { if (ending_tokens[i][k] != context_embd[k]) { fprintf(stderr, "Oops: ending %d of task %d differs from context at position %d\n",i,int(task_idx),k); diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index 3db61b754..573bc4ef9 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -87,7 +87,7 @@ int main(int argc, char ** argv) { } llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; auto next_token = llama_sample_token(ctx, &candidates_p); - auto next_token_str = llama_token_to_str(ctx, next_token); + auto next_token_str = llama_token_to_piece(ctx, next_token); last_n_tokens_data.push_back(next_token); printf("%s", next_token_str.c_str()); @@ -147,7 +147,7 @@ int main(int argc, char ** argv) { } llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; auto next_token = llama_sample_token(ctx2, &candidates_p); - auto next_token_str = llama_token_to_str(ctx2, next_token); + auto next_token_str = llama_token_to_piece(ctx2, next_token); last_n_tokens_data.push_back(next_token); printf("%s", next_token_str.c_str()); diff --git a/examples/server/server.cpp b/examples/server/server.cpp index a4b4d6418..89a3311f5 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -94,7 +94,7 @@ static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end) std::string ret; for (; begin != end; ++begin) { - ret += llama_token_to_str(ctx, *begin); + ret += llama_token_to_piece(ctx, *begin); } return ret; } @@ -123,7 +123,7 @@ static void server_log(const char *level, const char *function, int line, // format incomplete utf-8 multibyte character for output static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token) { - std::string out = token == -1 ? "" : llama_token_to_str(ctx, token); + std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token); // if the size is 1 and first bit is 1, meaning it's a partial character // (size > 1 meaning it's already a known token) if (out.size() == 1 && (out[0] & 0x80) == 0x80) @@ -286,7 +286,6 @@ struct llama_server_context std::vector p; if (first) { - s.insert(0, 1, ' '); // add a space if it's the first p = ::llama_tokenize(ctx, s, add_bos); first = false; } @@ -309,7 +308,6 @@ struct llama_server_context else { auto s = json_prompt.template get(); - s.insert(0, 1, ' '); // always add a first space prompt_tokens = ::llama_tokenize(ctx, s, add_bos); } @@ -566,7 +564,7 @@ struct llama_server_context if (!embd.empty() && embd.back() == llama_token_eos(ctx)) { - // stopping_word = llama_token_to_str(ctx, embd.back()); + // stopping_word = llama_token_to_piece(ctx, embd.back()); has_next_token = false; stopped_eos = true; LOG_VERBOSE("eos token found", {}); @@ -613,7 +611,7 @@ struct llama_server_context { const completion_token_output token_with_probs = nextToken(); - const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(ctx, token_with_probs.tok); + const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok); generated_text += token_text; if (params.n_probs > 0) @@ -1254,7 +1252,7 @@ void beam_search_callback(void * callback_data, llama_beams_state beams_state) { struct token_translator { llama_context * ctx; - std::string operator()(llama_token tok) const { return llama_token_to_str(ctx, tok); } + std::string operator()(llama_token tok) const { return llama_token_to_piece(ctx, tok); } std::string operator()(completion_token_output cto) const { return (*this)(cto.tok); } }; @@ -1364,7 +1362,7 @@ int main(int argc, char **argv) while (llama.has_next_token) { const completion_token_output token_with_probs = llama.doCompletion(); - const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_str(llama.ctx, token_with_probs.tok); + const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(llama.ctx, token_with_probs.tok); stop_pos = llama.findStoppingStrings(llama.generated_text, token_text.size(), STOP_FULL); @@ -1395,7 +1393,7 @@ int main(int argc, char **argv) if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) { continue; } - const std::string token_text = llama_token_to_str(llama.ctx, token_with_probs.tok); + const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok); size_t pos = std::min(sent_count, llama.generated_text.size()); diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp index 132f7fbf9..4ee85faca 100644 --- a/examples/simple/simple.cpp +++ b/examples/simple/simple.cpp @@ -63,7 +63,7 @@ int main(int argc, char ** argv) { fprintf(stderr, "\n\n"); for (auto id : tokens_list) { - fprintf(stderr, "%s", llama_token_to_str(ctx, id).c_str()); + fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str()); } fflush(stderr); @@ -112,7 +112,7 @@ int main(int argc, char ** argv) { } // print the new token : - printf("%s", llama_token_to_str(ctx, new_token_id).c_str()); + printf("%s", llama_token_to_piece(ctx, new_token_id).c_str()); fflush(stdout); // push this new token for next evaluation diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp index 79b117df7..12d153417 100644 --- a/examples/train-text-from-scratch/train-text-from-scratch.cpp +++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp @@ -1964,7 +1964,7 @@ void print_matrix(struct ggml_tensor * probs) { void print_token(struct llama_context * ctx, llama_token token) { - printf("%s", llama_token_to_str(ctx, token).c_str()); + printf("%s", llama_token_to_piece(ctx, token).c_str()); } void print_tokens(struct llama_context* ctx, struct ggml_tensor * tokens) { @@ -2202,7 +2202,7 @@ int tokenize_file(struct llama_context * lctx, const char * filename, std::vecto const char * in = buf.data(); const char * end = buf.data() + buf.size(); for (int i = 0; i < (int) out.size(); ++i) { - std::string s = llama_token_to_str(lctx, out[i]); + std::string s = llama_token_to_piece(lctx, out[i]); int len = s.length(); if (in >= end) { printf("%s: unexpected end of original text.\n", __func__); diff --git a/llama.cpp b/llama.cpp index e956c0163..2a8af4ee9 100644 --- a/llama.cpp +++ b/llama.cpp @@ -796,12 +796,12 @@ static void llama_nop(struct ggml_tensor * tensor) { // don't offload by default (void) tensor; } -static std::string llama_token_to_text(const struct llama_context * ctx, llama_token token) { +static std::string llama_token_to_str(const struct llama_context * ctx, llama_token token) { std::vector result(8, 0); - const int n_tokens = llama_token_to_str(ctx, token, result.data(), result.size()); + const int n_tokens = llama_token_to_piece(ctx, token, result.data(), result.size()); if (n_tokens < 0) { result.resize(-n_tokens); - int check = llama_token_to_str(ctx, token, result.data(), result.size()); + int check = llama_token_to_piece(ctx, token, result.data(), result.size()); GGML_ASSERT(check == -n_tokens); } else { result.resize(n_tokens); @@ -1635,7 +1635,8 @@ static void llm_load_hparams( } // TODO: This should probably be in llama.h -static std::vector llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos); +static std::vector llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos); +static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch); static void llm_load_vocab( llama_model_loader & ml, @@ -1737,7 +1738,11 @@ static void llm_load_vocab( } // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n' - vocab.linefeed_id = llama_tokenize_internal(vocab, "\n", false)[0]; + if (vocab.type == LLAMA_VOCAB_TYPE_SPM) { + vocab.linefeed_id = llama_byte_to_token(vocab, '\n'); + } else { + vocab.linefeed_id = llama_tokenize_internal(vocab, "\n", false)[0]; + } // special tokens GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID)); @@ -3026,10 +3031,8 @@ static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) { return vocab.token_to_id.at(buf); } -static std::string llama_escape_whitespace(const std::string& text) { - std::string result = text; - replace_all(result, " ", "\xe2\x96\x81"); - return result; +static void llama_escape_whitespace(std::string & text) { + replace_all(text, " ", "\xe2\x96\x81"); } static void llama_unescape_whitespace(std::string & word) { @@ -3373,22 +3376,31 @@ private: llm_bigram_bpe::queue work_queue; }; -static std::vector llama_tokenize_internal(const llama_vocab & vocab, const std::string & raw_text, bool bos) { +static std::vector llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos) { std::vector output; - if (raw_text.empty()) { - return output; - } + // OG tokenizer behavior: + // + // tokenizer.encode('', add_bos=True) returns [1] + // tokenizer.encode('', add_bos=False) returns [] if (bos && vocab.special_bos_id != -1) { output.push_back(vocab.special_bos_id); } + if (raw_text.empty()) { + return output; + } + switch (vocab.type) { case LLAMA_VOCAB_TYPE_SPM: { + // without adding this leading whitespace, we do not get the same results as the original tokenizer + raw_text = " " + raw_text; + llm_tokenizer_spm tokenizer(vocab); - tokenizer.tokenize(llama_escape_whitespace(raw_text), output); + llama_escape_whitespace(raw_text); + tokenizer.tokenize(raw_text, output); } break; case LLAMA_VOCAB_TYPE_BPE: { @@ -4078,16 +4090,16 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c std::vector candidates_grammar; for (size_t i = 0; i < candidates->size; ++i) { - const llama_token id = candidates->data[i].id; - const std::string text = llama_token_to_text(ctx, id); + const llama_token id = candidates->data[i].id; + const std::string piece = llama_token_to_str(ctx, id); if (id == eos) { if (!allow_eos) { candidates->data[i].logit = -INFINITY; } - } else if (text.empty() || text[0] == 0) { + } else if (piece.empty() || piece[0] == 0) { candidates->data[i].logit = -INFINITY; } else { - candidates_decoded.push_back(decode_utf8(text.c_str(), grammar->partial_utf8)); + candidates_decoded.push_back(decode_utf8(piece.c_str(), grammar->partial_utf8)); candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second }); } } @@ -4291,10 +4303,10 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar GGML_ASSERT(false); } - const std::string text = llama_token_to_text(ctx, token); + const std::string piece = llama_token_to_str(ctx, token); // Note terminating 0 in decoded string - const auto decoded = decode_utf8(text.c_str(), grammar->partial_utf8); + const auto decoded = decode_utf8(piece.c_str(), grammar->partial_utf8); const auto & code_points = decoded.first; for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it); @@ -6101,12 +6113,12 @@ int llama_tokenize_with_model( return res.size(); } -int llama_token_to_str(const struct llama_context * ctx, llama_token token, char * buf, int length) { - return llama_token_to_str_with_model(&ctx->model, token, buf, length); +int llama_token_to_piece(const struct llama_context * ctx, llama_token token, char * buf, int length) { + return llama_token_to_piece_with_model(&ctx->model, token, buf, length); } -// does not write null-terminator to str -int llama_token_to_str_with_model(const struct llama_model * model, llama_token token, char * buf, int length) { +// does not write null-terminator to buf +int llama_token_to_piece_with_model(const struct llama_model * model, llama_token token, char * buf, int length) { if (0 <= token && token < llama_model_n_vocab(model)) { if (llama_is_normal_token(model->vocab, token)) { std::string result = model->vocab.id_to_token[token].text; diff --git a/llama.h b/llama.h index b77dd7735..b084fe23c 100644 --- a/llama.h +++ b/llama.h @@ -381,15 +381,17 @@ extern "C" { int n_max_tokens, bool add_bos); - // Token Id -> String. Uses the vocabulary in the provided context - // Does not write null terminator to the buffer - LLAMA_API int llama_token_to_str( + // Token Id -> Piece. + // Uses the vocabulary in the provided context. + // Does not write null terminator to the buffer. + // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. + LLAMA_API int llama_token_to_piece( const struct llama_context * ctx, llama_token token, char * buf, int length); - LLAMA_API int llama_token_to_str_with_model( + LLAMA_API int llama_token_to_piece_with_model( const struct llama_model * model, llama_token token, char * buf, diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 2afaf86b1..ca1f39d31 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -25,8 +25,10 @@ endfunction() llama_build_and_test_executable(test-quantize-fns.cpp) llama_build_and_test_executable(test-quantize-perf.cpp) llama_build_and_test_executable(test-sampling.cpp) -llama_build_executable(test-tokenizer-0.cpp) -llama_test_executable (test-tokenizer-0.llama test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) +llama_build_executable(test-tokenizer-0-llama.cpp) +llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf) +llama_build_executable(test-tokenizer-0-falcon.cpp) +#llama_test_executable (test-tokenizer-0-falcon test-tokenizer-0-falcon.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) llama_build_executable(test-tokenizer-1.cpp) # test-tokenizer-1 requires a BPE vocab. re-enable when we have one. #llama_test_executable (test-tokenizer-1.llama test-tokenizer-1.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) diff --git a/tests/test-tokenizer-0-falcon.cpp b/tests/test-tokenizer-0-falcon.cpp new file mode 100644 index 000000000..836fb8ad2 --- /dev/null +++ b/tests/test-tokenizer-0-falcon.cpp @@ -0,0 +1,178 @@ +#include "llama.h" +#include "common.h" + +#include +#include +#include +#include +#include + +// generate using test-tokenizer-0-falcon.py +static const std::map> & k_tests() { + static std::map> _k_tests = { + { "" , { }, }, + { " " , { 204, }, }, + { " " , { 258, }, }, + { " " , { 466, }, }, + { "\t" , { 192, }, }, + { "\n" , { 193, }, }, + { "\t\n" , { 19125, }, }, + { "Hello world" , { 9856, 1079, }, }, + { " Hello world" , { 23090, 1079, }, }, + { "Hello World" , { 9856, 2889, }, }, + { " Hello World" , { 23090, 2889, }, }, + { " Hello World!" , { 23090, 2889, 12, }, }, + { "Hello, world!" , { 9856, 23, 1079, 12, }, }, + { " Hello, world!" , { 23090, 23, 1079, 12, }, }, + { " this is 🦙.cpp" , { 414, 304, 3346, 111, 231, 25, 29247, }, }, + { "w048 7tuijk dsdfhu" , { 98, 55866, 204, 34, 16682, 7149, 36190, 6869, 11481, }, }, + { "нещо на Български" , { 150, 133, 6207, 151, 215, 150, 134, 5052, 133, 6279, 5052, 223, 151, 216, 49679, 123, 53110, 47043, 7795, }, }, + { "កាន់តែពិសេសអាចខលចេញ" , { 38154, 206, 38154, 126, 38154, 225, 167, 237, 217, 38154, 221, 167, 237, 208, 38154, 228, 38154, 127, 38154, 237, 167, 237, 207, 38154, 237, 38154, 107, 38154, 126, 38154, 211, 38154, 207, 38154, 233, 38154, 211, 167, 237, 207, 38154, 215, }, }, + { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", { 2571, 232, 206, 204, 19, 11003, 20, 8196, 126, 283, 219, 48778, 116, 13392, 204, 19, 51831, 732, 63209, 1741, 7955, 522, 20, 22438, 211, 204, 19, 7927, 53360, 325, 504, 701, 946, 10930, 20, }, }, + { "Hello" , { 9856, }, }, + { " Hello" , { 23090, }, }, + { " Hello" , { 204, 23090, }, }, + { " Hello" , { 258, 23090, }, }, + { " Hello" , { 466, 23090, }, }, + { " Hello\n Hello" , { 466, 23090, 742, 23090, }, }, + }; + + return _k_tests; +} + +int main(int argc, char **argv) { + if (argc < 2) { + fprintf(stderr, "Usage: %s vocab-file [text-file]\n", argv[0]); + return 1; + } + + const std::string fname = argv[1]; + + std::string fname_text; + if (argc > 2) { + fname_text = argv[2]; + } + + fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str()); + + llama_model * model; + llama_context * ctx; + + llama_backend_init(false); + + // load the vocab + { + auto lparams = llama_context_default_params(); + + lparams.vocab_only = true; + + model = llama_load_model_from_file(fname.c_str(), lparams); + + if (model == NULL) { + fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); + return 1; + } + + ctx = llama_new_context_with_model(model, lparams); + + if (ctx == NULL) { + fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); + llama_free_model(model); + return 1; + } + } + + if (llama_vocab_type(ctx) != LLAMA_VOCAB_TYPE_BPE) { + fprintf(stderr, "%s : error: vocab type is not SPM\n", __func__); + llama_free_model(model); + llama_free(ctx); + return 2; + } + + bool success = true; + + for (const auto & test_kv : k_tests()) { + const std::vector res = llama_tokenize(ctx, test_kv.first, false); + + printf("\n"); + printf("src: '%s'\n", test_kv.first.c_str()); + printf("res: '%s'\n", llama_detokenize_bpe(ctx, res).c_str()); + printf("tok: "); + for (const auto & tok : res) { + printf("%d ", tok); + } + printf("\n"); + + bool correct = res.size() == test_kv.second.size(); + + for (int i = 0; i < (int) res.size() && correct; ++i) { + if (test_kv.second[i] != res[i]) { + correct = false; + } + } + + if (!correct) { + fprintf(stderr, "%s : failed test: '%s'\n", __func__, test_kv.first.c_str()); + fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__, + llama_detokenize_bpe(ctx, res).c_str(), + llama_detokenize_bpe(ctx, test_kv.second).c_str()); + fprintf(stderr, "%s : expected tokens: ", __func__); + for (const auto & t : test_kv.second) { + fprintf(stderr, "%6d, ", t); + } + fprintf(stderr, "\n"); + fprintf(stderr, "%s : got tokens: ", __func__); + for (const auto & t : res) { + fprintf(stderr, "%6d, ", t); + } + fprintf(stderr, "\n"); + + success = false; + } + } + + if (!fname_text.empty()) { + fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str()); + + std::string text; + { + std::ifstream ifs(fname_text); + if (!ifs) { + fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_text.c_str()); + return 1; + } + text = std::string(std::istreambuf_iterator(ifs), std::istreambuf_iterator()); + } + + fprintf(stderr, "%s : text size: %zu\n", __func__, text.size()); + + const std::vector res = llama_tokenize(ctx, text, true); + + fprintf(stderr, "%s : tokens: %zu\n", __func__, res.size()); + + { + const std::string fname_out = fname_text + ".tokcpp"; + + std::ofstream ofs(fname_out); + if (!ofs) { + fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_out.c_str()); + return 1; + } + + for (const auto & tok : res) { + ofs << tok << " "; + } + + ofs << "\n"; + } + + fprintf(stderr, "%s : tokens written to '%s'\n", __func__, (fname_text + ".tokcpp").c_str()); + } + + llama_free_model(model); + llama_free(ctx); + + llama_backend_free(); + + return success ? 0 : 3; +} diff --git a/tests/test-tokenizer-0-falcon.py b/tests/test-tokenizer-0-falcon.py new file mode 100644 index 000000000..9c8c1c7d1 --- /dev/null +++ b/tests/test-tokenizer-0-falcon.py @@ -0,0 +1,83 @@ +# tests with BPE tokenizer + +import os +import sys +import argparse + +from transformers import AutoTokenizer + +parser = argparse.ArgumentParser() +parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file") +parser.add_argument("--fname-tok", help="path to a text file to tokenize") +args = parser.parse_args() + +dir_tokenizer = args.dir_tokenizer + +tokenizer = AutoTokenizer.from_pretrained(dir_tokenizer) + +tests = [ + "", + " ", + " ", + " ", + "\t", + "\n", + "\t\n", + "Hello world", + " Hello world", + "Hello World", + " Hello World", + " Hello World!", + "Hello, world!", + " Hello, world!", + " this is 🦙.cpp", + "w048 7tuijk dsdfhu", + "нещо на Български", + "កាន់តែពិសេសអាចខលចេញ", + "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", + "Hello", + " Hello", + " Hello", + " Hello", + " Hello", + " Hello\n Hello", + ] + +for text in tests: + print('text: ', text) + print(tokenizer.encode(text)) + print(tokenizer.decode(tokenizer.encode(text))) + +print("\n\ntests for C++:\n") +for text in tests: + res = tokenizer.encode(text) + + k = text.replace('\n', '\\n') + k = k.replace('\t', '\\t') + k = '"' + k + '"' + print("{ %-24s, { " % k, end='') + for x in res: + print("%7d," % x, end='') + print(" }, },") + +print(tokenizer.encode('hello')) +print(tokenizer.encode('world')) +print(tokenizer.encode(' world')) +print(tokenizer.encode('hello world')) + +fname_tok = args.fname_tok +if fname_tok: + print('tokenizing file: ', fname_tok) + fname_out = fname_tok + '.tok' + with open(fname_tok, 'r') as f: + lines = f.readlines() + s = ''.join(lines) + res = tokenizer.encode(s) + # write to file + with open(fname_out, 'w') as f: + for x in res: + f.write(str(x) + ' ') + f.write('\n') + print('len(res): ', len(res)) + print('len(lines): ', len(lines)) + print('results written to: ', fname_out) diff --git a/tests/test-tokenizer-0-llama.cpp b/tests/test-tokenizer-0-llama.cpp new file mode 100644 index 000000000..8630742c6 --- /dev/null +++ b/tests/test-tokenizer-0-llama.cpp @@ -0,0 +1,182 @@ +#include "llama.h" +#include "common.h" + +#include +#include +#include +#include +#include + +// generate using test-tokenizer-0-llama.py +static const std::map> & k_tests() { + static std::map> _k_tests = { + { "" , { }, }, + { " " , { 259, }, }, + { " " , { 1678, }, }, + { " " , { 268, }, }, + { "\t" , { 29871, 12, }, }, + { "\n" , { 29871, 13, }, }, + { "\t\n" , { 29871, 12, 13, }, }, + { "Hello world" , { 15043, 3186, }, }, + { " Hello world" , { 29871, 15043, 3186, }, }, + { "Hello World" , { 15043, 2787, }, }, + { " Hello World" , { 29871, 15043, 2787, }, }, + { " Hello World!" , { 29871, 15043, 2787, 29991, }, }, + { "Hello, world!" , { 15043, 29892, 3186, 29991, }, }, + { " Hello, world!" , { 29871, 15043, 29892, 3186, 29991, }, }, + { " this is 🦙.cpp" , { 29871, 445, 338, 29871, 243, 162, 169, 156, 29889, 8223, }, }, + { "w048 7tuijk dsdfhu" , { 281, 29900, 29946, 29947, 29871, 29955, 9161, 13535, 18031, 2176, 6905, }, }, + { "нещо на Български" , { 1538, 4851, 665, 1386, 29713, 1305, }, }, + { "កាន់តែពិសេសអាចខលចេញ" , { 29871, 31849, 31324, 31934, 228, 162, 142, 228, 161, 146, 228, 162, 133, 228, 161, 153, 228, 161, 186, 31708, 228, 162, 132, 31708, 228, 161, 165, 31324, 228, 161, 136, 228, 161, 132, 228, 161, 158, 228, 161, 136, 228, 162, 132, 228, 161, 140, }, }, + { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", { 29871, 243, 162, 157, 131, 313, 8945, 29897, 29871, 243, 162, 155, 185, 30722, 243, 162, 143, 174, 30598, 313, 20787, 953, 3848, 275, 16125, 630, 29897, 29871, 31681, 313, 6194, 953, 29877, 2397, 393, 756, 967, 1914, 5993, 29897, }, }, + { "Hello" , { 15043, }, }, + { " Hello" , { 29871, 15043, }, }, + { " Hello" , { 259, 15043, }, }, + { " Hello" , { 1678, 15043, }, }, + { " Hello" , { 268, 15043, }, }, + { " Hello\n Hello" , { 268, 15043, 13, 1678, 15043, }, }, + }; + + return _k_tests; +} + +int main(int argc, char **argv) { + if (argc < 2) { + fprintf(stderr, "Usage: %s vocab-file [text-file]\n", argv[0]); + return 1; + } + + const std::string fname = argv[1]; + + std::string fname_text; + if (argc > 2) { + fname_text = argv[2]; + } + + fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str()); + + llama_model * model; + llama_context * ctx; + + llama_backend_init(false); + + // load the vocab + { + auto lparams = llama_context_default_params(); + + lparams.vocab_only = true; + + model = llama_load_model_from_file(fname.c_str(), lparams); + + if (model == NULL) { + fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); + return 1; + } + + ctx = llama_new_context_with_model(model, lparams); + + if (ctx == NULL) { + fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); + llama_free_model(model); + return 1; + } + } + + if (llama_vocab_type(ctx) != LLAMA_VOCAB_TYPE_SPM) { + fprintf(stderr, "%s : error: vocab type is not SPM\n", __func__); + llama_free_model(model); + llama_free(ctx); + return 2; + } + + bool success = true; + + for (const auto & test_kv : k_tests()) { + const std::vector res_bos = llama_tokenize(ctx, test_kv.first, true); + const std::vector res_nobos = llama_tokenize(ctx, test_kv.first, false); + + printf("\n"); + printf("src: '%s'\n", test_kv.first.c_str()); + printf("res: '%s'\n", llama_detokenize_spm(ctx, res_bos).c_str()); + printf("tok: "); + for (const auto & tok : res_bos) { + printf("%d ", tok); + } + printf("\n"); + + bool correct = res_nobos.size() == test_kv.second.size() && res_bos.size() == res_nobos.size() + 1 && res_bos[0] == 1; + + for (int i = 0; i < (int) res_nobos.size() && correct; ++i) { + if (test_kv.second[i] != res_bos[i + 1]) { + correct = false; + } + if (test_kv.second[i] != res_nobos[i]) { + correct = false; + } + } + + if (!correct) { + fprintf(stderr, "%s : failed test: '%s'\n", __func__, test_kv.first.c_str()); + fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__, + llama_detokenize_spm(ctx, res_nobos).c_str(), + llama_detokenize_spm(ctx, test_kv.second).c_str()); + fprintf(stderr, "%s : expected tokens: ", __func__); + for (const auto & t : test_kv.second) { + fprintf(stderr, "%6d, ", t); + } + fprintf(stderr, "\n"); + fprintf(stderr, "%s : got tokens: ", __func__); + for (const auto & t : res_nobos) { + fprintf(stderr, "%6d, ", t); + } + fprintf(stderr, "\n"); + + success = false; + } + } + + if (!fname_text.empty()) { + fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str()); + + std::string text; + { + std::ifstream ifs(fname_text); + if (!ifs) { + fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_text.c_str()); + return 1; + } + text = std::string(std::istreambuf_iterator(ifs), std::istreambuf_iterator()); + } + + fprintf(stderr, "%s : text size: %zu\n", __func__, text.size()); + + const std::vector res = llama_tokenize(ctx, text, true); + + fprintf(stderr, "%s : tokens: %zu\n", __func__, res.size()); + + { + const std::string fname_out = fname_text + ".tokcpp"; + + std::ofstream ofs(fname_out); + if (!ofs) { + fprintf(stderr, "%s : error: could not open file '%s'\n", __func__, fname_out.c_str()); + return 1; + } + + for (const auto & tok : res) { + ofs << tok << " "; + } + + ofs << "\n"; + } + + fprintf(stderr, "%s : tokens written to '%s'\n", __func__, (fname_text + ".tokcpp").c_str()); + } + + llama_free_model(model); + llama_free(ctx); + + llama_backend_free(); + + return success ? 0 : 3; +} diff --git a/tests/test-tokenizer-0-llama.py b/tests/test-tokenizer-0-llama.py new file mode 100644 index 000000000..bc164ee29 --- /dev/null +++ b/tests/test-tokenizer-0-llama.py @@ -0,0 +1,95 @@ +# tests with SPM tokenizer + +import os +import sys +import argparse + +from sentencepiece import SentencePieceProcessor + +parser = argparse.ArgumentParser() +parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file") +parser.add_argument("--fname-tok", help="path to a text file to tokenize") +args = parser.parse_args() + +dir_tokenizer = args.dir_tokenizer + +tokenizer = SentencePieceProcessor(dir_tokenizer + '/tokenizer.model') + +tests = [ + "", + " ", + " ", + " ", + "\t", + "\n", + "\t\n", + "Hello world", + " Hello world", + "Hello World", + " Hello World", + " Hello World!", + "Hello, world!", + " Hello, world!", + " this is 🦙.cpp", + "w048 7tuijk dsdfhu", + "нещо на Български", + "កាន់តែពិសេសអាចខលចេញ", + "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", + "Hello", + " Hello", + " Hello", + " Hello", + " Hello", + " Hello\n Hello", + ] + + +for text in tests: + print('text: ', text) + print('\nwith bos:') + print(tokenizer.encode(text, add_bos=True)) + print(tokenizer.decode(tokenizer.encode(text, add_bos=True))) + print('\nwithout bos:') + print(tokenizer.encode(text, add_bos=False)) + print(tokenizer.decode(tokenizer.encode(text, add_bos=False))) + +print("'" + tokenizer.id_to_piece(15043) + "'") # '_Hello' +print("'" + tokenizer.id_to_piece(29871) + "'") # '_' +print("'" + tokenizer.decode([15043]) + "'") # 'Hello' +print("'" + tokenizer.decode([15043, 15043]) + "'") # 'Hello Hello' +print("'" + tokenizer.decode([29871, 15043]) + "'") # ' Hello' +print("'" + tokenizer.decode([29871, 15043, 29871, 15043]) + "'") # ' Hello Hello' + +print("\n\ntests for C++:\n") +for text in tests: + res = tokenizer.encode(text, add_bos=False) + + k = text.replace('\n', '\\n') + k = k.replace('\t', '\\t') + k = '"' + k + '"' + print("{ %-24s, { " % k, end='') + for x in res: + print("%7d," % x, end='') + print(" }, },") + +print(tokenizer.encode('hello')) +print(tokenizer.encode('world')) +print(tokenizer.encode(' world')) +print(tokenizer.encode('hello world')) + +fname_tok = args.fname_tok +if fname_tok: + print('tokenizing file: ', fname_tok) + fname_out = fname_tok + '.tok' + with open(fname_tok, 'r') as f: + lines = f.readlines() + s = ''.join(lines) + res = tokenizer.encode(s, add_bos=True) + # write to file + with open(fname_out, 'w') as f: + for x in res: + f.write(str(x) + ' ') + f.write('\n') + print('len(res): ', len(res)) + print('len(lines): ', len(lines)) + print('results written to: ', fname_out) diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp deleted file mode 100644 index 7e9ac9188..000000000 --- a/tests/test-tokenizer-0.cpp +++ /dev/null @@ -1,141 +0,0 @@ -#include "llama.h" -#include "common.h" - -#include -#include -#include -#include - -static std::string unescape_whitespace(llama_context* ctx, const std::vector& tokens) { - std::string result; - for (size_t i = 0; i < tokens.size(); ++i) { - result += llama_token_to_str(ctx, tokens[i]); - } - return result; -} - -static const std::map> & k_tests() { - static std::map> _k_tests = { - { " ", {1, 259, }, }, - { " ", { 1, 1678, }, }, - { " ", { 1, 268, }, }, - { "\t", { 1, 29871, 12, }, }, - { "\n", { 1, 29871, 13, }, }, - { "\t\n", { 1, 29871, 12, 13, }, }, - { "Hello world", { 1, 15043, 3186, }, }, - { " Hello world", { 1, 29871, 15043, 3186, }, }, - { "Hello World", { 1, 15043, 2787, }, }, - { " Hello World", { 1, 29871, 15043, 2787, }, }, - { " Hello World!", { 1, 29871, 15043, 2787, 29991, }, }, - { " this is 🦙.cpp", { 1, 29871, 445, 338, 29871, 243, 162, 169, 156, 29889, 8223, }, }, - { "w048 7tuijk dsdfhu", { 1, 281, 29900, 29946, 29947, 29871, 29955, 9161, 13535, 18031, 2176, 6905, }, }, - { "нещо на Български", { 1, 1538, 4851, 665, 1386, 29713, 1305, }, }, - { "កាន់តែពិសេសអាចខលចេញ", { 1, 29871, 31849, 31324, 31934, 228, 162, 142, 228, 161, - 146, 228, 162, 133, 228, 161, 153, 228, 161, 186, - 31708, 228, 162, 132, 31708, 228, 161, 165, 31324, 228, - 161, 136, 228, 161, 132, 228, 161, 158, 228, 161, - 136, 228, 162, 132, 228, 161, 140, }, }, - { "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)", - { 1, 29871, 243, 162, 157, 131, 313, 8945, 29897, 29871, - 243, 162, 155, 185, 30722, 243, 162, 143, 174, 30598, - 313, 20787, 953, 3848, 275, 16125, 630, 29897, 29871, 31681, - 313, 6194, 953, 29877, 2397, 393, 756, 967, 1914, 5993, 29897, }, }, - { "Hello", { 1, 15043 }, }, - { " Hello", { 1, 29871, 15043 }, }, - { " Hello", { 1, 259, 15043 }, }, - { " Hello", { 1, 1678, 15043 }, }, - { " Hello", { 1, 268, 15043 }, }, - { " Hello\n Hello", { 1, 268, 15043, 13, 1678, 15043 }, }, - }; - - return _k_tests; -} - -int main(int argc, char **argv) { - if (argc < 2) { - fprintf(stderr, "Usage: %s \n", argv[0]); - return 1; - } - - const std::string fname = argv[1]; - - fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str()); - - llama_model * model; - llama_context * ctx; - - llama_backend_init(false); - - // load the vocab - { - auto lparams = llama_context_default_params(); - - lparams.vocab_only = true; - - model = llama_load_model_from_file(fname.c_str(), lparams); - - if (model == NULL) { - fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); - return 1; - } - - ctx = llama_new_context_with_model(model, lparams); - - if (ctx == NULL) { - fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str()); - llama_free_model(model); - return 1; - } - } - - const int n_vocab = llama_n_vocab(ctx); - - if (n_vocab != 32000) { - fprintf(stderr, "%s : expected 32000 tokens, got %d\n", __func__, n_vocab); - llama_free_model(model); - llama_free(ctx); - return 2; - } - - bool success = true; - - for (const auto & test_kv : k_tests()) { - // Add a space in front of the first character to match OG llama tokenizer behavior - std::vector res = llama_tokenize(ctx, " " + test_kv.first, true); - fprintf(stderr, "%s : '%s' tokenized to '%s'\n", - __func__, test_kv.first.c_str(), unescape_whitespace(ctx, res).c_str()); - - bool correct = res.size() == test_kv.second.size(); - - for (int i = 0; i < (int) res.size() && correct; ++i) { - if (res[i] != test_kv.second[i]) { - correct = false; - } - } - - if (!correct) { - fprintf(stderr, "%s : failed test: '%s'\n", __func__, test_kv.first.c_str()); - fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__, - unescape_whitespace(ctx, res).c_str(), unescape_whitespace(ctx, test_kv.second).c_str()); - fprintf(stderr, "%s : expected tokens: ", __func__); - for (const auto & t : test_kv.second) { - fprintf(stderr, "%6d, ", t); - } - fprintf(stderr, "\n"); - fprintf(stderr, "%s : got tokens: ", __func__); - for (const auto & t : res) { - fprintf(stderr, "%6d, ", t); - } - fprintf(stderr, "\n"); - - success = false; - } - } - - llama_free_model(model); - llama_free(ctx); - - llama_backend_free(); - - return success ? 0 : 3; -} diff --git a/tests/test-tokenizer-1.cpp b/tests/test-tokenizer-1.cpp index bd607d12b..ce4f2898c 100644 --- a/tests/test-tokenizer-1.cpp +++ b/tests/test-tokenizer-1.cpp @@ -22,14 +22,6 @@ static std::string escape_whitespace(const std::string& text) { return result; } -static std::string unescape_whitespace(llama_context * ctx, const std::vector & tokens) { - std::string result; - for (size_t i = 0; i < tokens.size(); ++i) { - result += llama_token_to_str(ctx, tokens[i]); - } - return result; -} - int main(int argc, char **argv) { if (argc < 2) { fprintf(stderr, "Usage: %s \n", argv[0]); @@ -72,13 +64,13 @@ int main(int argc, char **argv) { const int n_vocab = llama_n_vocab(ctx); for (int i = 0; i < n_vocab; ++i) { - std::string forward = llama_token_to_str(ctx, i); + std::string forward = llama_token_to_piece(ctx, i); std::vector tokens = llama_tokenize(ctx, forward, false); if (tokens.size() == 1) { if (i != tokens[0]) { - std::string backward = llama_token_to_str(ctx, tokens[0]); + std::string backward = llama_token_to_piece(ctx, tokens[0]); fprintf(stderr, "%s : error: token %d is string %s but bpe returns token %d %s\n", - __func__, i, llama_token_to_str(ctx, i).c_str(), tokens[0], backward.c_str()); + __func__, i, llama_token_to_piece(ctx, i).c_str(), tokens[0], backward.c_str()); return 2; } } From d0cee0d36d5be95a0d9088b674dbb27354107221 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 27 Aug 2023 14:19:54 +0300 Subject: [PATCH 063/147] gguf : add 64-bit support (GGUF v2) (#2821) * gguf : bump version to 2 * gguf : add support for 64-bit (no backwards comp yet) * gguf : v1 backwards comp * gguf.py : bump GGUF version * gguf.py : uint64_t on all lengths, sizes and counts, enums still uint32_t * gguf.py : string lengths uint32_t * gguf : update all counts to 64-bit * gguf.py : string len uint64_t and n_dims uint32_t * gguf : fix typo * llama.cpp : print gguf version --------- Co-authored-by: klosax <131523366+klosax@users.noreply.github.com> --- examples/gguf/gguf.cpp | 3 + ggml.c | 137 +++++++++++++++++++++++++++++++++++------ ggml.h | 11 +++- gguf-py/gguf/gguf.py | 36 ++++++++--- llama.cpp | 4 +- 5 files changed, 164 insertions(+), 27 deletions(-) diff --git a/examples/gguf/gguf.cpp b/examples/gguf/gguf.cpp index dee00df87..cda517bde 100644 --- a/examples/gguf/gguf.cpp +++ b/examples/gguf/gguf.cpp @@ -30,6 +30,9 @@ bool gguf_ex_write(const std::string & fname) { gguf_set_val_u32 (ctx, "some.parameter.uint32", 0x12345678); gguf_set_val_i32 (ctx, "some.parameter.int32", -0x12345679); gguf_set_val_f32 (ctx, "some.parameter.float32", 0.123456789f); + gguf_set_val_u64 (ctx, "some.parameter.uint64", 0x123456789abcdef0ull); + gguf_set_val_i64 (ctx, "some.parameter.int64", -0x123456789abcdef1ll); + gguf_set_val_f64 (ctx, "some.parameter.float64", 0.1234567890123456789); gguf_set_val_bool(ctx, "some.parameter.bool", true); gguf_set_val_str (ctx, "some.parameter.string", "hello world"); diff --git a/ggml.c b/ggml.c index 394fb459f..855d519bf 100644 --- a/ggml.c +++ b/ggml.c @@ -19394,7 +19394,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i //////////////////////////////////////////////////////////////////////////////// struct gguf_str { - uint32_t n; + uint64_t n; // GGUFv2 char * data; }; @@ -19408,9 +19408,12 @@ static const size_t GGUF_TYPE_SIZE[GGUF_TYPE_COUNT] = { [GGUF_TYPE_FLOAT32] = sizeof(float), [GGUF_TYPE_BOOL] = sizeof(bool), [GGUF_TYPE_STRING] = sizeof(struct gguf_str), + [GGUF_TYPE_UINT64] = sizeof(uint64_t), + [GGUF_TYPE_INT64] = sizeof(int64_t), + [GGUF_TYPE_FLOAT64] = sizeof(double), [GGUF_TYPE_ARRAY] = 0, // undefined }; -static_assert(GGUF_TYPE_COUNT == 10, "GGUF_TYPE_COUNT != 10"); +static_assert(GGUF_TYPE_COUNT == 13, "GGUF_TYPE_COUNT != 13"); static const char * GGUF_TYPE_NAME[GGUF_TYPE_COUNT] = { [GGUF_TYPE_UINT8] = "u8", @@ -19423,8 +19426,11 @@ static const char * GGUF_TYPE_NAME[GGUF_TYPE_COUNT] = { [GGUF_TYPE_BOOL] = "bool", [GGUF_TYPE_STRING] = "str", [GGUF_TYPE_ARRAY] = "arr", + [GGUF_TYPE_UINT64] = "u64", + [GGUF_TYPE_INT64] = "i64", + [GGUF_TYPE_FLOAT64] = "f64", }; -static_assert(GGUF_TYPE_COUNT == 10, "GGUF_TYPE_COUNT != 10"); +static_assert(GGUF_TYPE_COUNT == 13, "GGUF_TYPE_COUNT != 13"); union gguf_value { uint8_t uint8; @@ -19434,6 +19440,9 @@ union gguf_value { uint32_t uint32; int32_t int32; float float32; + uint64_t uint64; + int64_t int64; + double float64; bool bool_; struct gguf_str str; @@ -19441,7 +19450,7 @@ union gguf_value { struct { enum gguf_type type; - uint32_t n; + uint64_t n; // GGUFv2 void * data; } arr; }; @@ -19449,8 +19458,6 @@ union gguf_value { struct gguf_kv { struct gguf_str key; - uint32_t n_bytes; // TODO: is this actually needed? - enum gguf_type type; union gguf_value value; }; @@ -19458,15 +19465,15 @@ struct gguf_kv { struct gguf_header { uint32_t magic; uint32_t version; - uint32_t n_tensors; - uint32_t n_kv; + uint64_t n_tensors; // GGUFv2 + uint64_t n_kv; // GGUFv2 }; struct gguf_tensor_info { struct gguf_str name; uint32_t n_dims; - uint32_t ne[GGML_MAX_DIMS]; + uint64_t ne[GGML_MAX_DIMS]; enum ggml_type type; @@ -19497,19 +19504,32 @@ static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset) return n == size; } -static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) { +// NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 +static bool gguf_fread_str_cur(FILE * file, struct gguf_str * p, size_t * offset) { p->n = 0; p->data = NULL; bool ok = true; - // TODO: how to avoid mallocs for strings? ok = ok && gguf_fread_el(file, &p->n, sizeof(p->n), offset); p->data = calloc(p->n + 1, 1); ok = ok && gguf_fread_el(file, p->data, p->n, offset); return ok; } +static bool gguf_fread_str_v1(FILE * file, struct gguf_str * p, size_t * offset) { + p->n = 0; + p->data = NULL; + + bool ok = true; + + uint32_t n = 0; + ok = ok && gguf_fread_el(file, &n, sizeof(n), offset); p->data = calloc(n + 1, 1); p->n = n; + ok = ok && gguf_fread_el(file, p->data, p->n, offset); + + return ok; +} + struct gguf_context * gguf_init_empty(void) { struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context)); @@ -19565,8 +19585,21 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ctx->data = NULL; ok = ok && gguf_fread_el(file, &ctx->header.version, sizeof(ctx->header.version), &offset); - ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset); - ok = ok && gguf_fread_el(file, &ctx->header.n_kv, sizeof(ctx->header.n_kv), &offset); + + if (ctx->header.version == 1) { + // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 + uint32_t n_tensors = 0; + uint32_t n_kv = 0; + + ok = ok && gguf_fread_el(file, &n_tensors, sizeof(n_tensors), &offset); + ok = ok && gguf_fread_el(file, &n_kv, sizeof(n_kv), &offset); + + ctx->header.n_tensors = n_tensors; + ctx->header.n_kv = n_kv; + } else { + ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset); + ok = ok && gguf_fread_el(file, &ctx->header.n_kv, sizeof(ctx->header.n_kv), &offset); + } if (!ok) { fprintf(stderr, "%s: failed to read header\n", __func__); @@ -19576,6 +19609,12 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p } } + // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 + bool (* gguf_fread_str)(FILE *, struct gguf_str *, size_t *) = gguf_fread_str_cur; + if (ctx->header.version == 1) { + gguf_fread_str = gguf_fread_str_v1; + } + // read the kv pairs { ctx->kv = GGML_ALIGNED_MALLOC(ctx->header.n_kv * sizeof(struct gguf_kv)); @@ -19585,9 +19624,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p //fprintf(stderr, "%s: reading kv %d\n", __func__, i); - ok = ok && gguf_fread_str(file, &kv->key, &offset); - //ok = ok && gguf_fread_el (file, &kv->n_bytes, sizeof(kv->n_bytes), &offset); - ok = ok && gguf_fread_el (file, &kv->type, sizeof(kv->type), &offset); + ok = ok && gguf_fread_str(file, &kv->key, &offset); + ok = ok && gguf_fread_el (file, &kv->type, sizeof(kv->type), &offset); //fprintf(stderr, "%s: reading kv with key %s\n", __func__, kv->key.data); @@ -19599,12 +19637,23 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p case GGUF_TYPE_UINT32: ok = ok && gguf_fread_el (file, &kv->value.uint32, sizeof(kv->value.uint32), &offset); break; case GGUF_TYPE_INT32: ok = ok && gguf_fread_el (file, &kv->value.int32, sizeof(kv->value.int32), &offset); break; case GGUF_TYPE_FLOAT32: ok = ok && gguf_fread_el (file, &kv->value.float32, sizeof(kv->value.float32), &offset); break; + case GGUF_TYPE_UINT64: ok = ok && gguf_fread_el (file, &kv->value.uint64, sizeof(kv->value.uint64), &offset); break; + case GGUF_TYPE_INT64: ok = ok && gguf_fread_el (file, &kv->value.int64, sizeof(kv->value.int64), &offset); break; + case GGUF_TYPE_FLOAT64: ok = ok && gguf_fread_el (file, &kv->value.float64, sizeof(kv->value.float64), &offset); break; case GGUF_TYPE_BOOL: ok = ok && gguf_fread_el (file, &kv->value.bool_, sizeof(kv->value.bool_), &offset); break; case GGUF_TYPE_STRING: ok = ok && gguf_fread_str(file, &kv->value.str, &offset); break; case GGUF_TYPE_ARRAY: { ok = ok && gguf_fread_el(file, &kv->value.arr.type, sizeof(kv->value.arr.type), &offset); - ok = ok && gguf_fread_el(file, &kv->value.arr.n, sizeof(kv->value.arr.n), &offset); + + if (ctx->header.version == 1) { + // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 + uint32_t n = 0; + ok = ok && gguf_fread_el(file, &n, sizeof(n), &offset); + kv->value.arr.n = n; + } else { + ok = ok && gguf_fread_el(file, &kv->value.arr.n, sizeof(kv->value.arr.n), &offset); + } switch (kv->value.arr.type) { case GGUF_TYPE_UINT8: @@ -19614,6 +19663,9 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p case GGUF_TYPE_UINT32: case GGUF_TYPE_INT32: case GGUF_TYPE_FLOAT32: + case GGUF_TYPE_UINT64: + case GGUF_TYPE_INT64: + case GGUF_TYPE_FLOAT64: case GGUF_TYPE_BOOL: { kv->value.arr.data = malloc(kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type]); @@ -19660,7 +19712,14 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ok = ok && gguf_fread_str(file, &info->name, &offset); ok = ok && gguf_fread_el (file, &info->n_dims, sizeof(info->n_dims), &offset); for (uint32_t j = 0; j < info->n_dims; ++j) { - ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset); + if (ctx->header.version == 1) { + // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 + uint32_t t = 0; + ok = ok && gguf_fread_el(file, &t, sizeof(t), &offset); + info->ne[j] = t; + } else { + ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset); + } } ok = ok && gguf_fread_el (file, &info->type, sizeof(info->type), &offset); ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset), &offset); @@ -19954,6 +20013,18 @@ float gguf_get_val_f32(struct gguf_context * ctx, int i) { return ctx->kv[i].value.float32; } +uint64_t gguf_get_val_u64(struct gguf_context * ctx, int i) { + return ctx->kv[i].value.uint64; +} + +int64_t gguf_get_val_i64(struct gguf_context * ctx, int i) { + return ctx->kv[i].value.int64; +} + +double gguf_get_val_f64(struct gguf_context * ctx, int i) { + return ctx->kv[i].value.float64; +} + bool gguf_get_val_bool(struct gguf_context * ctx, int i) { return ctx->kv[i].value.bool_; } @@ -20056,6 +20127,27 @@ void gguf_set_val_f32(struct gguf_context * ctx, const char * key, float val) { ctx->kv[idx].value.float32 = val; } +void gguf_set_val_u64(struct gguf_context * ctx, const char * key, uint64_t val) { + const int idx = gguf_get_or_add_key(ctx, key); + + ctx->kv[idx].type = GGUF_TYPE_UINT64; + ctx->kv[idx].value.uint64 = val; +} + +void gguf_set_val_i64(struct gguf_context * ctx, const char * key, int64_t val) { + const int idx = gguf_get_or_add_key(ctx, key); + + ctx->kv[idx].type = GGUF_TYPE_INT64; + ctx->kv[idx].value.int64 = val; +} + +void gguf_set_val_f64(struct gguf_context * ctx, const char * key, double val) { + const int idx = gguf_get_or_add_key(ctx, key); + + ctx->kv[idx].type = GGUF_TYPE_FLOAT64; + ctx->kv[idx].value.float64 = val; +} + void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val) { const int idx = gguf_get_or_add_key(ctx, key); @@ -20106,6 +20198,9 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) { case GGUF_TYPE_UINT32: gguf_set_val_u32 (ctx, src->kv[i].key.data, src->kv[i].value.uint32); break; case GGUF_TYPE_INT32: gguf_set_val_i32 (ctx, src->kv[i].key.data, src->kv[i].value.int32); break; case GGUF_TYPE_FLOAT32: gguf_set_val_f32 (ctx, src->kv[i].key.data, src->kv[i].value.float32); break; + case GGUF_TYPE_UINT64: gguf_set_val_u64 (ctx, src->kv[i].key.data, src->kv[i].value.uint64); break; + case GGUF_TYPE_INT64: gguf_set_val_i64 (ctx, src->kv[i].key.data, src->kv[i].value.int64); break; + case GGUF_TYPE_FLOAT64: gguf_set_val_f64 (ctx, src->kv[i].key.data, src->kv[i].value.float64); break; case GGUF_TYPE_BOOL: gguf_set_val_bool(ctx, src->kv[i].key.data, src->kv[i].value.bool_); break; case GGUF_TYPE_STRING: gguf_set_val_str (ctx, src->kv[i].key.data, src->kv[i].value.str.data); break; case GGUF_TYPE_ARRAY: @@ -20267,6 +20362,9 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf, case GGUF_TYPE_UINT32: gguf_bwrite_el (buf, &kv->value.uint32, sizeof(kv->value.uint32) ); break; case GGUF_TYPE_INT32: gguf_bwrite_el (buf, &kv->value.int32, sizeof(kv->value.int32) ); break; case GGUF_TYPE_FLOAT32: gguf_bwrite_el (buf, &kv->value.float32, sizeof(kv->value.float32)); break; + case GGUF_TYPE_UINT64: gguf_bwrite_el (buf, &kv->value.uint64, sizeof(kv->value.uint64) ); break; + case GGUF_TYPE_INT64: gguf_bwrite_el (buf, &kv->value.int64, sizeof(kv->value.int64) ); break; + case GGUF_TYPE_FLOAT64: gguf_bwrite_el (buf, &kv->value.float64, sizeof(kv->value.float64)); break; case GGUF_TYPE_BOOL: gguf_bwrite_el (buf, &kv->value.bool_, sizeof(kv->value.bool_) ); break; case GGUF_TYPE_STRING: gguf_bwrite_str(buf, &kv->value.str ); break; case GGUF_TYPE_ARRAY: @@ -20282,6 +20380,9 @@ static void gguf_write_to_buf(struct gguf_context * ctx, struct gguf_buf * buf, case GGUF_TYPE_UINT32: case GGUF_TYPE_INT32: case GGUF_TYPE_FLOAT32: + case GGUF_TYPE_UINT64: + case GGUF_TYPE_INT64: + case GGUF_TYPE_FLOAT64: case GGUF_TYPE_BOOL: { gguf_bwrite_el(buf, kv->value.arr.data, kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type]); diff --git a/ggml.h b/ggml.h index b418153ba..792ca6e42 100644 --- a/ggml.h +++ b/ggml.h @@ -216,7 +216,7 @@ #define GGML_EXIT_ABORTED 1 #define GGUF_MAGIC 0x46554747 // "GGUF" -#define GGUF_VERSION 1 +#define GGUF_VERSION 2 #define GGUF_DEFAULT_ALIGNMENT 32 @@ -1827,6 +1827,9 @@ extern "C" { GGUF_TYPE_BOOL = 7, GGUF_TYPE_STRING = 8, GGUF_TYPE_ARRAY = 9, + GGUF_TYPE_UINT64 = 10, + GGUF_TYPE_INT64 = 11, + GGUF_TYPE_FLOAT64 = 12, GGUF_TYPE_COUNT, // marks the end of the enum }; @@ -1867,6 +1870,9 @@ extern "C" { GGML_API uint32_t gguf_get_val_u32 (struct gguf_context * ctx, int i); GGML_API int32_t gguf_get_val_i32 (struct gguf_context * ctx, int i); GGML_API float gguf_get_val_f32 (struct gguf_context * ctx, int i); + GGML_API uint64_t gguf_get_val_u64 (struct gguf_context * ctx, int i); + GGML_API int64_t gguf_get_val_i64 (struct gguf_context * ctx, int i); + GGML_API double gguf_get_val_f64 (struct gguf_context * ctx, int i); GGML_API bool gguf_get_val_bool(struct gguf_context * ctx, int i); GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i); GGML_API int gguf_get_arr_n (struct gguf_context * ctx, int i); @@ -1886,6 +1892,9 @@ extern "C" { GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val); GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t val); GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float val); + GGML_API void gguf_set_val_u64 (struct gguf_context * ctx, const char * key, uint64_t val); + GGML_API void gguf_set_val_i64 (struct gguf_context * ctx, const char * key, int64_t val); + GGML_API void gguf_set_val_f64 (struct gguf_context * ctx, const char * key, double val); GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool val); GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val); GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n); diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py index f4db7001b..838a2c0f8 100644 --- a/gguf-py/gguf/gguf.py +++ b/gguf-py/gguf/gguf.py @@ -13,7 +13,7 @@ from typing import Any, IO, List, Optional # GGUF_MAGIC = 0x46554747 -GGUF_VERSION = 1 +GGUF_VERSION = 2 GGUF_DEFAULT_ALIGNMENT = 32 # general @@ -365,6 +365,9 @@ class GGUFValueType(IntEnum): BOOL = 7 STRING = 8 ARRAY = 9 + UINT64 = 10 + INT64 = 11 + FLOAT64 = 12 @staticmethod def get_type(val): @@ -378,6 +381,7 @@ class GGUFValueType(IntEnum): return GGUFValueType.BOOL elif isinstance(val, int): return GGUFValueType.INT32 + # TODO: need help with 64-bit types in Python else: print("Unknown type: "+str(type(val))) sys.exit() @@ -400,8 +404,8 @@ class GGUFWriter: def write_header_to_file(self): self.fout.write(struct.pack(" Date: Sun, 27 Aug 2023 14:44:35 +0300 Subject: [PATCH 064/147] readme : update hot topics --- README.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/README.md b/README.md index 95471fdbb..f15a583b0 100644 --- a/README.md +++ b/README.md @@ -11,6 +11,10 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ ### Hot topics +- ## IMPORTANT: Tokenizer fixes and API change (developers and projects using `llama.cpp` built-in tokenization must read): https://github.com/ggerganov/llama.cpp/pull/2810 + +- ## GGUFv2 adds support for 64-bit sizes + backwards compatible: https://github.com/ggerganov/llama.cpp/pull/2821 + - Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717 - A new file format has been introduced: [GGUF](https://github.com/ggerganov/llama.cpp/pull/2398) From a6d1189fdd4c1ab4ba23f9d777f8950901dcffb2 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sun, 27 Aug 2023 15:19:59 +0300 Subject: [PATCH 065/147] k_quants tuning for Falcon-7b (#2816) * Make ggml-cuda.cu build with QK_K = 64 Using LLAMA_CUDA_FORCE_DMMV = ON and -nommq it runs and produces a meaningful result. * k_quants tuning for Falcon-7b --------- Co-authored-by: Iwan Kawrakow --- ggml-cuda.cu | 25 +++++++++++++++++-------- llama.cpp | 43 ++++++++++++++++++++++++++++++++++--------- 2 files changed, 51 insertions(+), 17 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 83d53c13c..d83aefc9a 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -306,11 +306,11 @@ typedef struct { #define QI4_K (QK_K / (4*QR4_K)) #ifdef GGML_QKK_64 typedef struct { - half d[2]; // super-block scales/mins + half dm[2]; // super-block scales/mins uint8_t scales[2]; // 4-bit block scales/mins uint8_t qs[QK_K/2]; // 4--bit quants } block_q4_K; -static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding"); +static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding"); #else typedef struct { half2 dm; // super-block scale for quantized scales/mins @@ -737,8 +737,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, float const int tid = threadIdx.x; const uint8_t * q = x[i].qs; float * y = yy + i*QK_K; - const float d = (float)x[i].d[0]; - const float m = (float)x[i].d[1]; + const float d = (float)x[i].dm[0]; + const float m = (float)x[i].dm[1]; y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4); y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4); #endif @@ -1155,8 +1155,8 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const uint16_t * a = (const uint16_t *)x[i].scales; aux16[0] = a[0] & 0x0f0f; aux16[1] = (a[0] >> 4) & 0x0f0f; - const float d = (float)x[i].d[0]; - const float m = (float)x[i].d[1]; + const float d = (float)x[i].dm[0]; + const float m = (float)x[i].dm[1]; float sum = 0.f; for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2]) @@ -2845,8 +2845,8 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1( aux16[0] = a[0] & 0x0f0f; aux16[1] = (a[0] >> 4) & 0x0f0f; - const float dall = bq4_K->d[0]; - const float dmin = bq4_K->d[1]; + const float dall = bq4_K->dm[0]; + const float dmin = bq4_K->dm[1]; const float d8_1 = __low2float(bq8_1[0].ds); const float d8_2 = __low2float(bq8_1[1].ds); @@ -2929,7 +2929,11 @@ template static __device__ __forceinlin const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd; +#if QK_K == 256 x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm; +#else + x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]}; +#endif } #pragma unroll @@ -3119,7 +3123,9 @@ template static __device__ __forceinlin const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd; +#if QK_K == 256 x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm; +#endif } #pragma unroll @@ -4709,6 +4715,8 @@ static void ggml_mul_mat_q3_K_q8_1_cuda( const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) { +#if QK_K == 256 + int id; CUDA_CHECK(cudaGetDevice(&id)); const int compute_capability = g_compute_capabilities[id]; @@ -4740,6 +4748,7 @@ static void ggml_mul_mat_q3_K_q8_1_cuda( mul_mat_q3_K<<>> (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); } +#endif } static void ggml_mul_mat_q4_K_q8_1_cuda( diff --git a/llama.cpp b/llama.cpp index df103a6e5..e9868f5d0 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4776,7 +4776,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { int nx = tensor->ne[0]; - if (nx % QK_K == 0) { + if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) { + new_type = GGML_TYPE_Q8_0; + } + else if (new_type != GGML_TYPE_Q8_0) { new_type = GGML_TYPE_Q6_K; } } else if (name.find("attn_v.weight") != std::string::npos) { @@ -4800,17 +4803,39 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } else if (name.find("ffn_down.weight") != std::string::npos) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { - new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; + new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K + : model.arch != LLM_ARCH_FALCON || use_more_bits(i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K + : GGML_TYPE_Q3_K; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) { + new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) { + if (model.arch == LLM_ARCH_FALCON) { + new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K : + use_more_bits(i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; + } else { + if (use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; + } + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && i_feed_forward_w2 < 4) { + new_type = GGML_TYPE_Q5_K; } - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; - else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && - use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < 4) new_type = GGML_TYPE_Q5_K; ++i_feed_forward_w2; } else if (name.find("attn_output.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; + if (model.arch != LLM_ARCH_FALCON) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; + } else { + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K; + } + } + else if (name.find("attn_qkv.weight") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K; } else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; From 25423e9185b7c2a1881ed8f85cc752a12370be9d Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 27 Aug 2023 15:24:40 +0300 Subject: [PATCH 066/147] scripts : helper convert script --- scripts/convert-gg.sh | 26 ++++++++++++++++++++++++++ scripts/qnt-all.sh | 2 ++ scripts/run-all-perf.sh | 2 ++ scripts/run-all-ppl.sh | 2 ++ 4 files changed, 32 insertions(+) create mode 100755 scripts/convert-gg.sh diff --git a/scripts/convert-gg.sh b/scripts/convert-gg.sh new file mode 100755 index 000000000..01fda16fd --- /dev/null +++ b/scripts/convert-gg.sh @@ -0,0 +1,26 @@ +#!/bin/bash + +set -e + +# LLaMA v1 +python3 convert.py ../llama1/7B --outfile models/llama-7b/ggml-model-f16.gguf --outtype f16 +python3 convert.py ../llama1/13B --outfile models/llama-13b/ggml-model-f16.gguf --outtype f16 +python3 convert.py ../llama1/30B --outfile models/llama-30b/ggml-model-f16.gguf --outtype f16 +python3 convert.py ../llama1/65B --outfile models/llama-65b/ggml-model-f16.gguf --outtype f16 + +# LLaMA v2 +python3 convert.py ../llama2/llama-2-7b --outfile models/llama-7b-v2/ggml-model-f16.gguf --outtype f16 +python3 convert.py ../llama2/llama-2-13b --outfile models/llama-13b-v2/ggml-model-f16.gguf --outtype f16 +python3 convert.py ../llama2/llama-2-70b --outfile models/llama-70b-v2/ggml-model-f16.gguf --outtype f16 + +# Code Llama +python3 convert.py ../codellama/CodeLlama-7b/ --outfile models/codellama-7b/ggml-model-f16.gguf --outtype f16 +python3 convert.py ../codellama/CodeLlama-13b/ --outfile models/codellama-13b/ggml-model-f16.gguf --outtype f16 +python3 convert.py ../codellama/CodeLlama-34b/ --outfile models/codellama-34b/ggml-model-f16.gguf --outtype f16 + +# Falcon +python3 convert-falcon-hf-to-gguf.py ../falcon/falcon-7b 1 +mv -v ../falcon/falcon-7b/ggml-model-f16.gguf models/falcon-7b/ggml-model-f16.gguf + +python3 convert-falcon-hf-to-gguf.py ../falcon/falcon-40b 1 +mv -v ../falcon/falcon-40b/ggml-model-f16.gguf models/falcon-40b/ggml-model-f16.gguf diff --git a/scripts/qnt-all.sh b/scripts/qnt-all.sh index 1b3d07da5..1376e4194 100755 --- a/scripts/qnt-all.sh +++ b/scripts/qnt-all.sh @@ -20,6 +20,8 @@ fi model="$1" out="../tmp/results-${model}" +set -e + mkdir -p ${out} for q in ${qnt[@]}; do diff --git a/scripts/run-all-perf.sh b/scripts/run-all-perf.sh index 91a6d853f..7391e3dd5 100755 --- a/scripts/run-all-perf.sh +++ b/scripts/run-all-perf.sh @@ -20,6 +20,8 @@ fi model="$1" out="../tmp/results-${model}" +set -e + mkdir -p ${out} mstr="" diff --git a/scripts/run-all-ppl.sh b/scripts/run-all-ppl.sh index 366d0866c..f643ca3ae 100755 --- a/scripts/run-all-ppl.sh +++ b/scripts/run-all-ppl.sh @@ -17,6 +17,8 @@ if [ ! -z "$3" ]; then args="$3" fi +set -e + model="$1" out="../tmp/results-${model}" From da7455d0467b5f5cc2e45d0dcffaf098df13db63 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 27 Aug 2023 15:52:34 +0300 Subject: [PATCH 067/147] readme : fix headings --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index f15a583b0..bf3eb0b76 100644 --- a/README.md +++ b/README.md @@ -11,9 +11,9 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ ### Hot topics -- ## IMPORTANT: Tokenizer fixes and API change (developers and projects using `llama.cpp` built-in tokenization must read): https://github.com/ggerganov/llama.cpp/pull/2810 +- #### IMPORTANT: Tokenizer fixes and API change (developers and projects using `llama.cpp` built-in tokenization must read): https://github.com/ggerganov/llama.cpp/pull/2810 -- ## GGUFv2 adds support for 64-bit sizes + backwards compatible: https://github.com/ggerganov/llama.cpp/pull/2821 +- GGUFv2 adds support for 64-bit sizes + backwards compatible: https://github.com/ggerganov/llama.cpp/pull/2821 - Added support for Falcon models: https://github.com/ggerganov/llama.cpp/pull/2717 From eaa13a48ff4136f01c1cdb79cacd61b67ec53095 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 27 Aug 2023 16:40:48 +0300 Subject: [PATCH 068/147] falcon : fix CUDA inference by making K and Q contiguous (#2830) * falcon : fix CUDA inference by making K and Q contiguous ggml-ci * cuda : add assert to guard from non-cont ropes --- ggml-cuda.cu | 2 ++ llama.cpp | 10 ++++++---- 2 files changed, 8 insertions(+), 4 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index d83aefc9a..d76a25dc2 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -6337,9 +6337,11 @@ void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_ASSERT(src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); + GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented const int mode = ((int32_t *) dst->op_params)[2]; const bool is_glm = mode & 4; + ggml_cuda_op(src0, src1, dst, ggml_cuda_op_rope, true, !is_glm); // flatten support not implemented for glm } diff --git a/llama.cpp b/llama.cpp index e9868f5d0..0d12d9cca 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2642,18 +2642,20 @@ static struct ggml_cgraph * llm_build_falcon( const size_t wsize = ggml_type_size(cur->type); - struct ggml_tensor * tmpq = ggml_view_3d( + // TODO: these 2 ggml_conts are technically not needed, but we add them until CUDA support for + // non-contiguous views is added for the rope operator + struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_3d( ctx0, cur, n_embd_head, n_head, N, wsize * n_embd_head, wsize * n_embd_head * (n_head + 2 * n_head_kv), - 0); + 0)); offload_func_kq(tmpq); - struct ggml_tensor * tmpk = ggml_view_3d( + struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_3d( ctx0, cur, n_embd_head, n_head_kv, N, wsize * n_embd_head, wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * n_head); + wsize * n_embd_head * n_head)); offload_func_kq(tmpk); struct ggml_tensor * tmpv = ggml_view_3d( From 463173a6c0ff353055eb90665794884c888c790f Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sun, 27 Aug 2023 16:50:33 +0300 Subject: [PATCH 069/147] llama : speedup tokenization (#2831) * Speedup tokenization On current master it takes ~3.2 seconds to tokenize Wikitext. With this change it becomes ~525 ms. * Fixit: it was missing the piece after the last found occurence --------- Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 4 ++++ llama.cpp | 15 ++++++++++----- 2 files changed, 14 insertions(+), 5 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index b596d0626..ebafa0c29 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -190,10 +190,14 @@ void perplexity(llama_context * ctx, const gpt_params & params) { const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; const bool add_bos = is_spm; + auto tim1 = std::chrono::high_resolution_clock::now(); fprintf(stderr, "%s: tokenizing the input ..\n", __func__); auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos); + auto tim2 = std::chrono::high_resolution_clock::now(); + fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast(tim2-tim1).count()); + const int n_chunk_max = tokens.size() / params.n_ctx; const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max); diff --git a/llama.cpp b/llama.cpp index 0d12d9cca..0bb8fcd6e 100644 --- a/llama.cpp +++ b/llama.cpp @@ -114,12 +114,17 @@ static size_t utf8_len(char src) { } void replace_all(std::string & s, const std::string & search, const std::string & replace) { - for (size_t pos = 0; ; pos += replace.length()) { - pos = s.find(search, pos); - if (pos == std::string::npos) break; - s.erase(pos, search.length()); - s.insert(pos, replace); + std::string result; + for (size_t pos = 0; ; pos += search.length()) { + auto new_pos = s.find(search, pos); + if (new_pos == std::string::npos) { + result += s.substr(pos, s.size() - pos); + break; + } + result += s.substr(pos, new_pos - pos) + replace; + pos = new_pos; } + s = std::move(result); } static void zeros(std::ofstream & file, size_t n) { From 230d46c723edf5999752e4cb67fd94edb19ef9c7 Mon Sep 17 00:00:00 2001 From: Olivier Chafik Date: Sun, 27 Aug 2023 15:13:31 +0100 Subject: [PATCH 070/147] examples : update llama2.c converter to read vocab and write models in GGUF format (#2751) * llama2.c: direct gguf output (WIP) * Simplify vector building logic * llama2.c gguf conversion: fix token types in converter * llama2.c: support copying vocab from a llama gguf model file * llama2.c: update default path for vocab model + readme * llama2.c: use defines for gguf keys * llama2.c: escape whitespaces w/ U+2581 in vocab converter the llama.cpp way * llama2.c converter: cleanups + take n_ff from config --- examples/convert-llama2c-to-ggml/README.md | 8 +- .../convert-llama2c-to-ggml.cpp | 340 +++++++++++------- 2 files changed, 220 insertions(+), 128 deletions(-) diff --git a/examples/convert-llama2c-to-ggml/README.md b/examples/convert-llama2c-to-ggml/README.md index fd561fcbc..0f37d295b 100644 --- a/examples/convert-llama2c-to-ggml/README.md +++ b/examples/convert-llama2c-to-ggml/README.md @@ -12,18 +12,14 @@ usage: ./convert-llama2c-to-ggml [options] options: -h, --help show this help message and exit - --copy-vocab-from-model FNAME model path from which to copy vocab (default 'tokenizer.bin') + --copy-vocab-from-model FNAME path of gguf llama model or llama2.c vocabulary from which to copy vocab (default 'models/7B/ggml-model-f16.gguf') --llama2c-model FNAME [REQUIRED] model path from which to load Karpathy's llama2.c model --llama2c-output-model FNAME model path to save the converted llama2.c model (default ak_llama_model.bin') ``` An example command using a model from [karpathy/tinyllamas](https://huggingface.co/karpathy/tinyllamas) is as follows: -`$ ./convert-llama2c-to-ggml --copy-vocab-from-model ../llama2.c/tokenizer.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.ggmlv3.bin` - -For now the generated model is in the legacy GGJTv3 format, so you need to convert it to gguf manually: - -`$ python ./convert-llama-ggmlv3-to-gguf.py --eps 1e-5 --input stories42M.ggmlv3.bin --output stories42M.gguf.bin` +`$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin` Now you can use the model with a command like: diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index f8a58dc7a..51d90ea6a 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -10,9 +10,48 @@ #include #include #include +#include #include #include +// GGUF keys & tensor names. + +#define KV_GENERAL_ARCHITECTURE "general.architecture" +#define KV_GENERAL_NAME "general.name" + +#define KV_TOKENIZER_MODEL "tokenizer.ggml.model" +#define KV_TOKENIZER_LIST "tokenizer.ggml.tokens" +#define KV_TOKENIZER_TOKEN_TYPE "tokenizer.ggml.token_type" +#define KV_TOKENIZER_SCORES "tokenizer.ggml.scores" +#define KV_TOKENIZER_BOS_ID "tokenizer.ggml.bos_token_id" +#define KV_TOKENIZER_EOS_ID "tokenizer.ggml.eos_token_id" +#define KV_TOKENIZER_UNK_ID "tokenizer.ggml.unknown_token_id" +#define KV_TOKENIZER_SEP_ID "tokenizer.ggml.seperator_token_id" +#define KV_TOKENIZER_PAD_ID "tokenizer.ggml.padding_token_id" +#define KV_TOKENIZER_HF_JSON "tokenizer.huggingface.json" + +#define KV_CONTEXT_LENGTH "llama.context_length" +#define KV_EMBEDDING_LENGTH "llama.embedding_length" +#define KV_BLOCK_COUNT "llama.block_count" +#define KV_FEED_FORWARD_LENGTH "llama.feed_forward_length" +#define KV_ATTENTION_HEAD_COUNT "llama.attention.head_count" +#define KV_ATTENTION_HEAD_COUNT_KV "llama.attention.head_count_kv" +#define KV_ATTENTION_LAYERNORM_RMS_EPS "llama.attention.layer_norm_rms_epsilon" +#define KV_ROPE_DIMENSION_COUNT "llama.rope.dimension_count" + +#define TN_TOKEN_EMBD "token_embd.weight" +#define TN_OUTPUT_NORM "output_norm.weight" +#define TN_OUTPUT "output.weight" +#define TN_ATTN_NORM "blk.%d.attn_norm.weight" +#define TN_ATTN_Q "blk.%d.attn_q.weight" +#define TN_ATTN_K "blk.%d.attn_k.weight" +#define TN_ATTN_V "blk.%d.attn_v.weight" +#define TN_ATTN_OUTPUT "blk.%d.attn_output.weight" +#define TN_FFN_NORM "blk.%d.ffn_norm.weight" +#define TN_FFN_GATE "blk.%d.ffn_gate.weight" +#define TN_FFN_DOWN "blk.%d.ffn_down.weight" +#define TN_FFN_UP "blk.%d.ffn_up.weight" + #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data #endif @@ -20,6 +59,11 @@ #define LLAMA_FILE_MAGIC_GGJT 0x67676a74u // 'ggjt' #define LLAMA_FILE_VERSION_GGJT_V3 3 +#define TOKENIZER_NAME "llama" +#define UNKNOWN_TOKEN_ID 0 +#define BOS_TOKEN_ID 1 +#define EOS_TOKEN_ID 2 + //////////////////////////////////////// llama2.c model structs and functions to load models, alloc memory etc. typedef struct { int dim; // transformer dimension @@ -183,6 +227,7 @@ struct my_llama_hparams { uint32_t n_vocab = 32000; uint32_t n_ctx = 512; // this is provided as user input? uint32_t n_embd = 4096; + uint32_t n_ff = 11008; uint32_t n_mult = 4; uint32_t n_head = 32; uint32_t n_layer = 32; @@ -214,6 +259,8 @@ struct my_llama_layer { struct my_llama_model { struct ggml_context * ctx = NULL; + std::string name; + my_llama_hparams hparams; struct ggml_tensor * tok_embeddings; @@ -276,18 +323,13 @@ struct train_params { int mem_compute1_gb; }; -uint32_t get_n_ff(const struct my_llama_hparams* hparams) { - const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult; - return n_ff; -} - void print_params(struct my_llama_hparams * params) { printf("%s: n_vocab: %d\n", __func__, params->n_vocab); printf("%s: n_ctx: %d\n", __func__, params->n_ctx); printf("%s: n_embd: %d\n", __func__, params->n_embd); printf("%s: n_mult: %d\n", __func__, params->n_mult); printf("%s: n_head: %d\n", __func__, params->n_head); - printf("%s: n_ff: %d\n", __func__, get_n_ff(params)); + printf("%s: n_ff: %d\n", __func__, params->n_ff); printf("%s: n_layer: %d\n", __func__, params->n_layer); printf("%s: n_rot: %d\n", __func__, params->n_rot); } @@ -299,7 +341,7 @@ void init_model(struct my_llama_model * model) { const uint32_t n_layer = hparams.n_layer; const uint32_t n_vocab = hparams.n_vocab; - const uint32_t n_ff = get_n_ff(&hparams); + const uint32_t n_ff = hparams.n_ff; struct ggml_context * ctx = model->ctx; model->train_its = 0; @@ -481,21 +523,6 @@ struct llama_file { return std::string(chars.data(), len); } - void write_raw(const void * ptr, size_t size) { - if (size == 0) { - return; - } - errno = 0; - size_t ret = std::fwrite(ptr, size, 1, fp); - if (ret != 1) { - throw std::runtime_error(format("write error: %s", strerror(errno))); - } - } - - void write_u32(std::uint32_t val) { - write_raw(&val, sizeof(val)); - } - ~llama_file() { if (fp) { std::fclose(fp); @@ -503,30 +530,6 @@ struct llama_file { } }; -void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) { - if (tensor == NULL) { - file->write_u32(0); - file->write_u32(0); - file->write_u32(GGML_TYPE_F32); - file->seek((0-file->tell()) & 31, SEEK_CUR); - return; - } - const char * name = ggml_get_name(tensor); - uint32_t name_len = strlen(name); - uint32_t nd = tensor->n_dims; - uint32_t ne[4] = { (uint32_t)tensor->ne[0], - (uint32_t)tensor->ne[1], - (uint32_t)tensor->ne[2], - (uint32_t)tensor->ne[3] }; - file->write_u32(nd); - file->write_u32(name_len); - file->write_u32(tensor->type); - file->write_raw(ne, sizeof(ne[0]) * nd); - file->write_raw(name, name_len); - file->seek((0-file->tell()) & 31, SEEK_CUR); - file->write_raw(tensor->data, ggml_nbytes(tensor)); -} - bool is_ggml_file(const char *filename) { llama_file file(filename, "rb"); if (file.size < 4) { @@ -536,48 +539,96 @@ bool is_ggml_file(const char *filename) { return magic == GGUF_MAGIC; } +static std::string llama_escape_whitespaces(const std::string& text) { + std::ostringstream out; + for (char c : text) { + if (c == ' ') out << "\xe2\x96\x81"; + else out << c; + } + return out.str(); +} + void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) { -#pragma message("TODO: implement reading vocabulary using gguf") -// // heuristic to infer whether vocab is from ggml or from llama2.c vocabulary -// if (is_ggml_file(filename)) { -// -// struct llama_context_params llama_params = llama_context_default_params(); -// llama_params.vocab_only = true; -// -// struct llama_model * lmodel = llama_load_model_from_file(filename, llama_params); -// struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params); -// -// const int n_vocab = llama_n_vocab(lctx); -// vocab->id_to_token.resize(n_vocab); -// for (int i=0; iid_to_token[i].text = llama_token_get_text(lctx, i); -// vocab->id_to_token[i].score = llama_token_get_score(lctx, i); -// vocab->id_to_token[i].type = llama_token_get_type(lctx, i); -// vocab->token_to_id.emplace(vocab->id_to_token[i].text, i); -// } -// llama_free(lctx); -// llama_free_model(lmodel); -// } else - { // assume llama2.c vocabulary - printf("Assuming llama2.c vocabulary since %s is not a ggml file\n", filename); + if (is_ggml_file(filename)) { + struct ggml_context * ctx_data = NULL; + + struct gguf_init_params params = { + /*.no_alloc = */ false, + /*.ctx = */ &ctx_data, + }; + + struct gguf_context * ctx = gguf_init_from_file(filename, params); + GGML_ASSERT(ctx != NULL); + + const int model_idx = gguf_find_key(ctx, KV_TOKENIZER_MODEL); + GGML_ASSERT(model_idx >= 0); + std::string tokenizer_name = gguf_get_val_str(ctx, model_idx); + GGML_ASSERT(tokenizer_name == TOKENIZER_NAME); + + const int token_idx = gguf_find_key(ctx, KV_TOKENIZER_LIST); + GGML_ASSERT(token_idx >= 0); + + const int score_idx = gguf_find_key(ctx, KV_TOKENIZER_SCORES); + GGML_ASSERT(score_idx >= 0); + const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx); + + const int toktype_idx = gguf_find_key(ctx, KV_TOKENIZER_TOKEN_TYPE); + GGML_ASSERT(toktype_idx >= 0); + const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx); + + const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx); + + vocab->id_to_token.resize(n_vocab); + + for (uint32_t i = 0; i < n_vocab; i++) { + std::string word = gguf_get_arr_str(ctx, token_idx, i); + + vocab->token_to_id[word] = i; + + auto & token_data = vocab->id_to_token[i]; + token_data.text = std::move(word); + token_data.score = scores[i]; + token_data.type = (llama_token_type) toktypes[i]; + } + ggml_free(ctx_data); + gguf_free(ctx); + } else { + // assume llama2.c vocabulary + printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename); llama_file file(filename, "rb"); const int n_vocab = config->vocab_size; /* uint32_t max_token_length = */ file.read_u32(); // unused vocab->id_to_token.resize(n_vocab); - for (int i=0; i single byte tokens. - char byte_val; - if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) { - char cstr[2] = { byte_val, 0 }; - text = cstr; + + unsigned char byte_val; + llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL; + if (id == UNKNOWN_TOKEN_ID) { + text = ""; + type = LLAMA_TOKEN_TYPE_UNKNOWN; + } else if (id == BOS_TOKEN_ID) { + text = ""; + type = LLAMA_TOKEN_TYPE_CONTROL; + } else if (id == EOS_TOKEN_ID) { + text = ""; + type = LLAMA_TOKEN_TYPE_CONTROL; + } else if (text.empty()) { + type = LLAMA_TOKEN_TYPE_CONTROL; + } else if (sscanf(text.c_str(), "<0x%02hhX>", &byte_val) == 1) { + // Text of byte tokens is already in the expected format. + type = LLAMA_TOKEN_TYPE_BYTE; + } else { + type = LLAMA_TOKEN_TYPE_NORMAL; } - vocab->id_to_token[i].text = text; - vocab->id_to_token[i].score = score; - vocab->id_to_token[i].type = LLAMA_TOKEN_TYPE_UNDEFINED; - vocab->token_to_id.emplace(text, i); + text = llama_escape_whitespaces(text); + + vocab->id_to_token[id].text = text; + vocab->id_to_token[id].score = score; + vocab->id_to_token[id].type = type; + vocab->token_to_id.emplace(text, id); } } } @@ -619,33 +670,6 @@ void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * kar } void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) { - struct llama_file file(filename, "wb"); - if (file.fp == NULL) { - return; - } - -#pragma message("TODO: implement file saving using gguf") - // write_magic - file.write_u32(LLAMA_FILE_MAGIC_GGJT); // magic - file.write_u32(LLAMA_FILE_VERSION_GGJT_V3); // version - // write_hparams - file.write_u32(model->hparams.n_vocab); - file.write_u32(model->hparams.n_embd); - file.write_u32(model->hparams.n_mult); - file.write_u32(model->hparams.n_head); - file.write_u32(model->hparams.n_layer); - file.write_u32(model->hparams.n_rot); - file.write_u32(LLAMA_FTYPE_ALL_F32); - - // write_vocab - for now we are just writing the existing BPE voc. assuming karpathy's vocabulary is the same. idk. - uint32_t n_vocab = model->hparams.n_vocab; - for (uint32_t i = 0; i < n_vocab; i++) { - const auto & token_data = vocab->id_to_token.at(i); - file.write_u32((uint32_t) token_data.text.size()); - file.write_raw(token_data.text.data(), token_data.text.size()); - file.write_raw(&token_data.score, sizeof(token_data.score)); - } - // stuff AK weights into GG weights one by one. // w->token_embedding_table -> model->tok_embeddings // float* -> struct ggml_tensor @@ -658,8 +682,7 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod // for rms-att-weight int row_length = model->hparams.n_embd; const auto & hparams = model->hparams; - //int n_ff = model->hparams.n_embd; - int n_ff = get_n_ff(&hparams); + int n_ff = model->hparams.n_ff; for (uint32_t i = 0; i < model->hparams.n_layer; ++i){ auto & layer = model->layers[i]; @@ -677,28 +700,91 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod stuff_karpathy_weights_into_gg(layer.w2 , &w->w2[i*n_ff*row_length]); stuff_karpathy_weights_into_gg(layer.w3 , &w->w3[i*row_length*n_ff]); } + + struct gguf_context * ctx = gguf_init_empty(); + + std::vector tokens; + std::vector scores; + std::vector token_types; + for (const llama_vocab::token_data & token_data : vocab->id_to_token) { + tokens.push_back(token_data.text.c_str()); + scores.push_back(token_data.score); + token_types.push_back(token_data.type); + } + gguf_set_arr_str(ctx, KV_TOKENIZER_LIST, tokens.data(), tokens.size()); + gguf_set_arr_data(ctx, KV_TOKENIZER_SCORES, GGUF_TYPE_FLOAT32, scores.data(), scores.size()); + gguf_set_arr_data(ctx, KV_TOKENIZER_TOKEN_TYPE, GGUF_TYPE_INT32, token_types.data(), token_types.size()); + + gguf_set_val_str(ctx, KV_TOKENIZER_MODEL, TOKENIZER_NAME); + + gguf_set_val_str(ctx, KV_GENERAL_ARCHITECTURE, "llama"); + gguf_set_val_str(ctx, KV_GENERAL_NAME, "llama"); + + // special tokens + gguf_set_val_u32(ctx, KV_TOKENIZER_UNK_ID, UNKNOWN_TOKEN_ID); + gguf_set_val_u32(ctx, KV_TOKENIZER_BOS_ID, BOS_TOKEN_ID); + gguf_set_val_u32(ctx, KV_TOKENIZER_EOS_ID, EOS_TOKEN_ID); + gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1); + gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1); + + gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx); + gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd); + gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff); + gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head); + // n_head_kv is optional, default to n_head + // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...); + gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer); + gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot); + gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f); + // write tensors - write_tensor(&file, model->tok_embeddings); - write_tensor(&file, model->norm); - write_tensor(&file, model->output); // ? + ggml_set_name(model->tok_embeddings, TN_TOKEN_EMBD); + gguf_add_tensor(ctx, model->tok_embeddings); + + ggml_set_name(model->norm, TN_OUTPUT_NORM); + gguf_add_tensor(ctx, model->norm); + + ggml_set_name(model->output, TN_OUTPUT); + gguf_add_tensor(ctx, model->output); + for (uint32_t i = 0; i < model->hparams.n_layer; ++i) { auto & layer = model->layers[i]; - write_tensor(&file, layer.attention_norm); - write_tensor(&file, layer.wq); - write_tensor(&file, layer.wk); - write_tensor(&file, layer.wv); - write_tensor(&file, layer.wo); - write_tensor(&file, layer.ffn_norm); - write_tensor(&file, layer.w1); - write_tensor(&file, layer.w2); - write_tensor(&file, layer.w3); + ggml_format_name(layer.wq, TN_ATTN_Q, i); + gguf_add_tensor(ctx, layer.wq); + + ggml_format_name(layer.wk, TN_ATTN_K, i); + gguf_add_tensor(ctx, layer.wk); + + ggml_format_name(layer.wv, TN_ATTN_V, i); + gguf_add_tensor(ctx, layer.wv); + + ggml_format_name(layer.wo, TN_ATTN_OUTPUT, i); + gguf_add_tensor(ctx, layer.wo); + + ggml_format_name(layer.attention_norm, TN_ATTN_NORM, i); + gguf_add_tensor(ctx, layer.attention_norm); + + ggml_format_name(layer.w1, TN_FFN_GATE, i); + gguf_add_tensor(ctx, layer.w1); + + ggml_format_name(layer.w2, TN_FFN_DOWN, i); + gguf_add_tensor(ctx, layer.w2); + + ggml_format_name(layer.w3, TN_FFN_UP, i); + gguf_add_tensor(ctx, layer.w3); + + ggml_format_name(layer.ffn_norm, TN_FFN_NORM, i); + gguf_add_tensor(ctx, layer.ffn_norm); } + + gguf_write_to_file(ctx, filename, false); + gguf_free(ctx); } struct train_params get_default_train_params() { struct train_params params; - params.fn_vocab_model = "tokenizer.bin"; + params.fn_vocab_model = "models/7B/ggml-model-f16.gguf"; params.fn_llama2c_output_model = "ak_llama_model.bin"; params.fn_train_data = "shakespeare.txt"; params.fn_checkpoint_in = "checkpoint.bin"; @@ -751,7 +837,7 @@ void print_usage(int /*argc*/, char ** argv, const struct train_params * params) fprintf(stderr, "\n"); fprintf(stderr, "options:\n"); fprintf(stderr, " -h, --help show this help message and exit\n"); - fprintf(stderr, " --copy-vocab-from-model FNAME llama2.c vocabulary or ggmlv3 model path from which to copy vocab (default '%s')\n", params->fn_vocab_model); + fprintf(stderr, " --copy-vocab-from-model FNAME path of gguf llama model or llama2.c vocabulary from which to copy vocab (default '%s')\n", params->fn_vocab_model); fprintf(stderr, " --llama2c-model FNAME [REQUIRED] model path from which to load Karpathy's llama2.c model\n"); fprintf(stderr, " --llama2c-output-model FNAME model path to save the converted llama2.c model (default %s')\n", params->fn_llama2c_output_model); fprintf(stderr, "\n"); @@ -812,6 +898,14 @@ bool params_parse(int argc, char ** argv, struct train_params * params) { return true; } +std::string basename(const std::string &path) { + size_t pos = path.find_last_of("/"); + if (pos == std::string::npos) { + return path; + } + return path.substr(pos + 1); +} + int main(int argc, char ** argv) { struct train_params params = get_default_train_params(); if (!params_parse(argc, argv, ¶ms)) { @@ -840,6 +934,7 @@ int main(int argc, char ** argv) { model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx); model.hparams.n_ctx = params.n_ctx; model.hparams.n_embd = config.dim; //params.n_embd; + model.hparams.n_ff = config.hidden_dim; model.hparams.n_mult = 32;//params.n_mult; model.hparams.n_head = config.n_heads; //params.n_head; model.hparams.n_layer = config.n_layers; //params.n_layer; @@ -853,6 +948,7 @@ int main(int argc, char ** argv) { model.ctx = ggml_init(lcparams); init_model(&model); + model.name = basename(params.fn_llama2c_model); save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model); printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model); From c10704d01e21e3dbe4d6ca1026ebff85349dd239 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 27 Aug 2023 18:55:41 +0300 Subject: [PATCH 071/147] llama : fix MPI threads (close #2827) --- llama.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 0bb8fcd6e..72d2d1de0 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2845,7 +2845,6 @@ static bool llama_eval_internal( GGML_ASSERT(n_tokens > 0); GGML_ASSERT(n_past >= 0); - GGML_ASSERT(n_threads > 0); // TODO: keep the values of n_batch and n_ctx // GGML_ASSERT(n_tokens <= n_batch); // GGML_ASSERT(n_past + n_tokens <= n_ctx); @@ -2856,6 +2855,8 @@ static bool llama_eval_internal( ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads); #endif + GGML_ASSERT(n_threads > 0); + const int N = n_tokens; const auto & model = lctx.model; From 103cfafc774f6feb3172b5d4d39681c965b17eba Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 27 Aug 2023 21:50:22 +0300 Subject: [PATCH 072/147] gguf : fix strings to not be null-terminated (#2839) * gguf : fix strings to not be null-terminated ggml-ci * gguf : fix gguf_add_tensor name --- ggml.c | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/ggml.c b/ggml.c index 855d519bf..767c19ae2 100644 --- a/ggml.c +++ b/ggml.c @@ -19524,8 +19524,8 @@ static bool gguf_fread_str_v1(FILE * file, struct gguf_str * p, size_t * offset) bool ok = true; uint32_t n = 0; - ok = ok && gguf_fread_el(file, &n, sizeof(n), offset); p->data = calloc(n + 1, 1); p->n = n; - ok = ok && gguf_fread_el(file, p->data, p->n, offset); + ok = ok && gguf_fread_el(file, &n, sizeof(n), offset); p->data = calloc(n + 1, 1); p->n = n; + ok = ok && gguf_fread_el(file, p->data, p->n, offset); return ok; } @@ -20071,7 +20071,7 @@ static int gguf_get_or_add_key(struct gguf_context * ctx, const char * key) { const int n_kv = gguf_get_n_kv(ctx); ctx->kv = realloc(ctx->kv, (n_kv + 1) * sizeof(struct gguf_kv)); - ctx->kv[n_kv].key.n = strlen(key) + 1; + ctx->kv[n_kv].key.n = strlen(key); ctx->kv[n_kv].key.data = strdup(key); ctx->header.n_kv++; @@ -20159,7 +20159,7 @@ void gguf_set_val_str(struct gguf_context * ctx, const char * key, const char * const int idx = gguf_get_or_add_key(ctx, key); ctx->kv[idx].type = GGUF_TYPE_STRING; - ctx->kv[idx].value.str.n = strlen(val) + 1; + ctx->kv[idx].value.str.n = strlen(val); ctx->kv[idx].value.str.data = strdup(val); } @@ -20182,7 +20182,7 @@ void gguf_set_arr_str(struct gguf_context * ctx, const char * key, const char ** ctx->kv[idx].value.arr.data = malloc(n*sizeof(struct gguf_str)); for (int i = 0; i < n; i++) { struct gguf_str * str = &((struct gguf_str *)ctx->kv[idx].value.arr.data)[i]; - str->n = strlen(data[i]) + 1; + str->n = strlen(data[i]); str->data = strdup(data[i]); } } @@ -20229,7 +20229,7 @@ void gguf_add_tensor( const int idx = ctx->header.n_tensors; ctx->infos = realloc(ctx->infos, (idx + 1)*sizeof(struct gguf_tensor_info)); - ctx->infos[idx].name.n = strlen(tensor->name) + 1; + ctx->infos[idx].name.n = strlen(tensor->name); ctx->infos[idx].name.data = strdup(tensor->name); for (int i = 0; i < GGML_MAX_DIMS; ++i) { From 3e8ff47af620a31e0810c58a41e4b089145982ef Mon Sep 17 00:00:00 2001 From: JohnnyB Date: Mon, 28 Aug 2023 07:31:24 +0100 Subject: [PATCH 073/147] devops : added systemd units and set versioning to use date. (#2835) * Corrections and systemd units * Missing dependency clblast --- ....srpm.spec => llama-cpp-clblast.srpm.spec} | 44 ++++++++++++++---- ...s.srpm.spec => llama-cpp-cublas.srpm.spec} | 26 ++++++++++- .devops/llama-cpp.srpm.spec | 45 +++++++++++++++---- 3 files changed, 96 insertions(+), 19 deletions(-) rename .devops/{lamma-cpp-clblast.srpm.spec => llama-cpp-clblast.srpm.spec} (56%) rename .devops/{lamma-cpp-cublas.srpm.spec => llama-cpp-cublas.srpm.spec} (71%) diff --git a/.devops/lamma-cpp-clblast.srpm.spec b/.devops/llama-cpp-clblast.srpm.spec similarity index 56% rename from .devops/lamma-cpp-clblast.srpm.spec rename to .devops/llama-cpp-clblast.srpm.spec index 739c68281..076f29695 100644 --- a/.devops/lamma-cpp-clblast.srpm.spec +++ b/.devops/llama-cpp-clblast.srpm.spec @@ -13,12 +13,13 @@ # It is up to the user to install the correct vendor-specific support. Name: llama.cpp-clblast -Version: master +Version: %( date "+%%Y%%m%%d" ) Release: 1%{?dist} -Summary: OpenCL Inference of LLaMA model in pure C/C++ +Summary: OpenCL Inference of LLaMA model in C/C++ License: MIT Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz -BuildRequires: coreutils make gcc-c++ git mesa-libOpenCL-devel +BuildRequires: coreutils make gcc-c++ git mesa-libOpenCL-devel clblast-devel +Requires: clblast URL: https://github.com/ggerganov/llama.cpp %define debug_package %{nil} @@ -35,18 +36,43 @@ make -j LLAMA_CLBLAST=1 %install mkdir -p %{buildroot}%{_bindir}/ -cp -p main %{buildroot}%{_bindir}/llamacppclblast -cp -p server %{buildroot}%{_bindir}/llamacppclblastserver -cp -p simple %{buildroot}%{_bindir}/llamacppclblastsimple +cp -p main %{buildroot}%{_bindir}/llamaclblast +cp -p server %{buildroot}%{_bindir}/llamaclblastserver +cp -p simple %{buildroot}%{_bindir}/llamaclblastsimple + +mkdir -p %{buildroot}/usr/lib/systemd/system +%{__cat} < %{buildroot}/usr/lib/systemd/system/llamaclblast.service +[Unit] +Description=Llama.cpp server, CPU only (no GPU support in this build). +After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target + +[Service] +Type=simple +EnvironmentFile=/etc/sysconfig/llama +ExecStart=/usr/bin/llamaclblastserver $LLAMA_ARGS +ExecReload=/bin/kill -s HUP $MAINPID +Restart=never + +[Install] +WantedBy=default.target +EOF + +mkdir -p %{buildroot}/etc/sysconfig +%{__cat} < %{buildroot}/etc/sysconfig/llama +LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin" +EOF %clean rm -rf %{buildroot} rm -rf %{_builddir}/* %files -%{_bindir}/llamacppclblast -%{_bindir}/llamacppclblastserver -%{_bindir}/llamacppclblastsimple +%{_bindir}/llamaclblast +%{_bindir}/llamaclblastserver +%{_bindir}/llamaclblastsimple +/usr/lib/systemd/system/llamaclblast.service +%config /etc/sysconfig/llama + %pre diff --git a/.devops/lamma-cpp-cublas.srpm.spec b/.devops/llama-cpp-cublas.srpm.spec similarity index 71% rename from .devops/lamma-cpp-cublas.srpm.spec rename to .devops/llama-cpp-cublas.srpm.spec index 75d32fbe7..f847ebb1e 100644 --- a/.devops/lamma-cpp-cublas.srpm.spec +++ b/.devops/llama-cpp-cublas.srpm.spec @@ -13,7 +13,7 @@ # It is up to the user to install the correct vendor-specific support. Name: llama.cpp-cublas -Version: master +Version: %( date "+%%Y%%m%%d" ) Release: 1%{?dist} Summary: CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL) License: MIT @@ -40,6 +40,28 @@ cp -p main %{buildroot}%{_bindir}/llamacppcublas cp -p server %{buildroot}%{_bindir}/llamacppcublasserver cp -p simple %{buildroot}%{_bindir}/llamacppcublassimple +mkdir -p %{buildroot}/usr/lib/systemd/system +%{__cat} < %{buildroot}/usr/lib/systemd/system/llamacublas.service +[Unit] +Description=Llama.cpp server, CPU only (no GPU support in this build). +After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target + +[Service] +Type=simple +EnvironmentFile=/etc/sysconfig/llama +ExecStart=/usr/bin/llamacppcublasserver $LLAMA_ARGS +ExecReload=/bin/kill -s HUP $MAINPID +Restart=never + +[Install] +WantedBy=default.target +EOF + +mkdir -p %{buildroot}/etc/sysconfig +%{__cat} < %{buildroot}/etc/sysconfig/llama +LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin" +EOF + %clean rm -rf %{buildroot} rm -rf %{_builddir}/* @@ -48,6 +70,8 @@ rm -rf %{_builddir}/* %{_bindir}/llamacppcublas %{_bindir}/llamacppcublasserver %{_bindir}/llamacppcublassimple +/usr/lib/systemd/system/llamacublas.service +%config /etc/sysconfig/llama %pre diff --git a/.devops/llama-cpp.srpm.spec b/.devops/llama-cpp.srpm.spec index c65251a5a..446213d69 100644 --- a/.devops/llama-cpp.srpm.spec +++ b/.devops/llama-cpp.srpm.spec @@ -6,6 +6,7 @@ # Notes for llama.cpp: # 1. Tags are currently based on hash - which will not sort asciibetically. # We need to declare standard versioning if people want to sort latest releases. +# In the meantime, YYYYMMDD format will be used. # 2. Builds for CUDA/OpenCL support are separate, with different depenedencies. # 3. NVidia's developer repo must be enabled with nvcc, cublas, clblas, etc installed. # Example: https://developer.download.nvidia.com/compute/cuda/repos/fedora37/x86_64/cuda-fedora37.repo @@ -13,12 +14,13 @@ # It is up to the user to install the correct vendor-specific support. Name: llama.cpp -Version: master +Version: %( date "+%%Y%%m%%d" ) Release: 1%{?dist} Summary: CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL) License: MIT Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz -BuildRequires: coreutils make gcc-c++ git +BuildRequires: coreutils make gcc-c++ git libstdc++-devel +Requires: libstdc++ URL: https://github.com/ggerganov/llama.cpp %define debug_package %{nil} @@ -26,27 +28,52 @@ URL: https://github.com/ggerganov/llama.cpp %description CPU inference for Meta's Lllama2 models using default options. +Models are not included in this package and must be downloaded separately. %prep -%autosetup +%setup -n llama.cpp-master %build make -j %install mkdir -p %{buildroot}%{_bindir}/ -cp -p main %{buildroot}%{_bindir}/llamacpp -cp -p server %{buildroot}%{_bindir}/llamacppserver -cp -p simple %{buildroot}%{_bindir}/llamacppsimple +cp -p main %{buildroot}%{_bindir}/llama +cp -p server %{buildroot}%{_bindir}/llamaserver +cp -p simple %{buildroot}%{_bindir}/llamasimple + +mkdir -p %{buildroot}/usr/lib/systemd/system +%{__cat} < %{buildroot}/usr/lib/systemd/system/llama.service +[Unit] +Description=Llama.cpp server, CPU only (no GPU support in this build). +After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target + +[Service] +Type=simple +EnvironmentFile=/etc/sysconfig/llama +ExecStart=/usr/bin/llamaserver $LLAMA_ARGS +ExecReload=/bin/kill -s HUP $MAINPID +Restart=never + +[Install] +WantedBy=default.target +EOF + +mkdir -p %{buildroot}/etc/sysconfig +%{__cat} < %{buildroot}/etc/sysconfig/llama +LLAMA_ARGS="-m /opt/llama2/ggml-model-f32.bin" +EOF %clean rm -rf %{buildroot} rm -rf %{_builddir}/* %files -%{_bindir}/llamacpp -%{_bindir}/llamacppserver -%{_bindir}/llamacppsimple +%{_bindir}/llama +%{_bindir}/llamaserver +%{_bindir}/llamasimple +/usr/lib/systemd/system/llama.service +%config /etc/sysconfig/llama %pre From ebcee207b6058b7f695bb5c203ad87b1066a9790 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Mon, 28 Aug 2023 02:32:25 -0400 Subject: [PATCH 074/147] quantize : make output filename optional again (#2823) * quantize : make output filename optional again * quantize : fix path parsing on Windows suggested by @slaren --- examples/quantize/quantize.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index d172f645a..df9a214fc 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -100,7 +100,7 @@ int main(int argc, char ** argv) { } } - if (argc - arg_idx < 3) { + if (argc - arg_idx < 2) { usage(argv[0]); } @@ -114,7 +114,7 @@ int main(int argc, char ** argv) { std::string ftype_str; if (try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) { std::string fpath; - const size_t pos = fname_inp.find_last_of('/'); + const size_t pos = fname_inp.find_last_of("/\\"); if (pos != std::string::npos) { fpath = fname_inp.substr(0, pos + 1); } From f55538c3ccba9b926846ef862fa830cea08c433e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 28 Aug 2023 10:59:08 +0300 Subject: [PATCH 075/147] metal : fix memory leak (#2762) * metal : fix memory leak * metal : fix encoders memory leak * metal : clean up more memory resources * metal : fix more leaks * metal : reuse dispatch queue + autoreleasepool * metal : reuse array for command buffers and encoders * ggml : assert for odd number of blocks on ARM 15M tinyllama is an example --- ggml-metal.h | 1 + ggml-metal.m | 100 +++++++++++++++++++++++++++++++++++++++++---------- ggml.c | 11 +++--- 3 files changed, 88 insertions(+), 24 deletions(-) diff --git a/ggml-metal.h b/ggml-metal.h index 00202b787..fca28d37e 100644 --- a/ggml-metal.h +++ b/ggml-metal.h @@ -24,6 +24,7 @@ // max memory buffers that can be mapped to the device #define GGML_METAL_MAX_BUFFERS 16 +#define GGML_METAL_MAX_COMMAND_BUFFERS 32 struct ggml_tensor; struct ggml_cgraph; diff --git a/ggml-metal.m b/ggml-metal.m index 06eb3872e..ad2ee8cf5 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -33,12 +33,15 @@ struct ggml_metal_buffer { struct ggml_metal_context { int n_cb; - float * logits; - id device; id queue; id library; + id command_buffers [GGML_METAL_MAX_COMMAND_BUFFERS]; + id command_encoders[GGML_METAL_MAX_COMMAND_BUFFERS]; + + dispatch_queue_t d_queue; + int n_buffers; struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS]; @@ -114,12 +117,13 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context)); - ctx->n_cb = n_cb; + ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS); ctx->device = MTLCreateSystemDefaultDevice(); ctx->queue = [ctx->device newCommandQueue]; ctx->n_buffers = 0; ctx->concur_list_len = 0; + ctx->d_queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT); #if 0 // compile from source string and show compile log @@ -239,9 +243,67 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { void ggml_metal_free(struct ggml_metal_context * ctx) { fprintf(stderr, "%s: deallocating\n", __func__); +#define GGML_METAL_DEL_KERNEL(name) \ + [ctx->function_##name release]; \ + [ctx->pipeline_##name release]; + + GGML_METAL_DEL_KERNEL(add); + GGML_METAL_DEL_KERNEL(add_row); + GGML_METAL_DEL_KERNEL(mul); + GGML_METAL_DEL_KERNEL(mul_row); + GGML_METAL_DEL_KERNEL(scale); + GGML_METAL_DEL_KERNEL(silu); + GGML_METAL_DEL_KERNEL(relu); + GGML_METAL_DEL_KERNEL(gelu); + GGML_METAL_DEL_KERNEL(soft_max); + GGML_METAL_DEL_KERNEL(diag_mask_inf); + GGML_METAL_DEL_KERNEL(get_rows_f16); + GGML_METAL_DEL_KERNEL(get_rows_q4_0); + GGML_METAL_DEL_KERNEL(get_rows_q4_1); + GGML_METAL_DEL_KERNEL(get_rows_q8_0); + GGML_METAL_DEL_KERNEL(get_rows_q2_K); + GGML_METAL_DEL_KERNEL(get_rows_q3_K); + GGML_METAL_DEL_KERNEL(get_rows_q4_K); + GGML_METAL_DEL_KERNEL(get_rows_q5_K); + GGML_METAL_DEL_KERNEL(get_rows_q6_K); + GGML_METAL_DEL_KERNEL(rms_norm); + GGML_METAL_DEL_KERNEL(norm); + GGML_METAL_DEL_KERNEL(mul_mat_f16_f32); + GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32); + GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32); + GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32); + GGML_METAL_DEL_KERNEL(mul_mat_q2_K_f32); + GGML_METAL_DEL_KERNEL(mul_mat_q3_K_f32); + GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32); + GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32); + GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32); + GGML_METAL_DEL_KERNEL(mul_mm_f16_f32); + GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32); + GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32); + GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32); + GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32); + GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32); + GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32); + GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32); + GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32); + GGML_METAL_DEL_KERNEL(rope); + GGML_METAL_DEL_KERNEL(alibi_f32); + GGML_METAL_DEL_KERNEL(cpy_f32_f16); + GGML_METAL_DEL_KERNEL(cpy_f32_f32); + GGML_METAL_DEL_KERNEL(cpy_f16_f16); + +#undef GGML_METAL_DEL_KERNEL + for (int i = 0; i < ctx->n_buffers; ++i) { [ctx->buffers[i].metal release]; } + + [ctx->library release]; + [ctx->queue release]; + [ctx->device release]; + + dispatch_release(ctx->d_queue); + free(ctx); } @@ -261,7 +323,7 @@ void ggml_metal_host_free(void * data) { } void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) { - ctx->n_cb = n_cb; + ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS); } int ggml_metal_if_optimized(struct ggml_metal_context * ctx) { @@ -507,6 +569,8 @@ void ggml_metal_graph_compute( struct ggml_cgraph * gf) { metal_printf("%s: evaluating graph\n", __func__); + @autoreleasepool { + // if there is ctx->concur_list, dispatch concurrently // else fallback to serial dispatch MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor; @@ -521,29 +585,25 @@ void ggml_metal_graph_compute( const int n_cb = ctx->n_cb; - NSMutableArray * command_buffers = [NSMutableArray arrayWithCapacity:n_cb]; - for (int i = 0; i < n_cb; ++i) { - command_buffers[i] = [ctx->queue commandBuffer]; + ctx->command_buffers[i] = [ctx->queue commandBuffer]; // enqueue the command buffers in order to specify their execution order - [command_buffers[i] enqueue]; - } + [ctx->command_buffers[i] enqueue]; - // TODO: is this the best way to start threads? - dispatch_queue_t queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT); + ctx->command_encoders[i] = [ctx->command_buffers[i] computeCommandEncoderWithDescriptor: edesc]; + } for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) { const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb; - dispatch_async(queue, ^{ + dispatch_async(ctx->d_queue, ^{ size_t offs_src0 = 0; size_t offs_src1 = 0; size_t offs_dst = 0; - id command_buffer = command_buffers[cb_idx]; - - id encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc]; + id command_buffer = ctx->command_buffers[cb_idx]; + id encoder = ctx->command_encoders[cb_idx]; const int node_start = (cb_idx + 0) * n_nodes_per_cb; const int node_end = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes); @@ -1117,17 +1177,19 @@ void ggml_metal_graph_compute( } // wait for all threads to finish - dispatch_barrier_sync(queue, ^{}); - - [command_buffers[n_cb - 1] waitUntilCompleted]; + dispatch_barrier_sync(ctx->d_queue, ^{}); // check status of command buffers // needed to detect if the device ran out-of-memory for example (#1881) for (int i = 0; i < n_cb; i++) { - MTLCommandBufferStatus status = (MTLCommandBufferStatus) [command_buffers[i] status]; + [ctx->command_buffers[i] waitUntilCompleted]; + + MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status]; if (status != MTLCommandBufferStatusCompleted) { fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status); GGML_ASSERT(false); } } + + } } diff --git a/ggml.c b/ggml.c index 767c19ae2..54f426bc0 100644 --- a/ggml.c +++ b/ggml.c @@ -2436,7 +2436,6 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * const int nb = n / qk; assert(n % qk == 0); - assert(nb % 2 == 0); const block_q4_0 * restrict x = vx; const block_q8_0 * restrict y = vy; @@ -2445,6 +2444,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); + GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb for (int i = 0; i < nb; i += 2) { const block_q4_0 * restrict x0 = &x[i + 0]; const block_q4_0 * restrict x1 = &x[i + 1]; @@ -2623,6 +2623,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * } // Main loop + GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb for (int i = 2; i < nb; i+=2) { _mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0); _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0); @@ -2706,7 +2707,6 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const int nb = n / qk; assert(n % qk == 0); - assert(nb % 2 == 0); const block_q4_1 * restrict x = vx; const block_q8_1 * restrict y = vy; @@ -2718,6 +2718,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * float summs = 0; + GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb for (int i = 0; i < nb; i += 2) { const block_q4_1 * restrict x0 = &x[i + 0]; const block_q4_1 * restrict x1 = &x[i + 1]; @@ -2832,7 +2833,6 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * const int nb = n / qk; assert(n % qk == 0); - assert(nb % 2 == 0); assert(qk == QK5_0); const block_q5_0 * restrict x = vx; @@ -2848,6 +2848,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * uint64_t tmp0[4]; uint64_t tmp1[4]; + GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb for (int i = 0; i < nb; i += 2) { const block_q5_0 * restrict x0 = &x[i]; const block_q5_0 * restrict x1 = &x[i + 1]; @@ -3072,7 +3073,6 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * const int nb = n / qk; assert(n % qk == 0); - assert(nb % 2 == 0); assert(qk == QK5_1); const block_q5_1 * restrict x = vx; @@ -3091,6 +3091,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * uint64_t tmp0[4]; uint64_t tmp1[4]; + GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb for (int i = 0; i < nb; i += 2) { const block_q5_1 * restrict x0 = &x[i]; const block_q5_1 * restrict x1 = &x[i + 1]; @@ -3328,7 +3329,6 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * const int nb = n / qk; assert(n % qk == 0); - assert(nb % 2 == 0); const block_q8_0 * restrict x = vx; const block_q8_0 * restrict y = vy; @@ -3337,6 +3337,7 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); + GGML_ASSERT(nb % 2 == 0); // TODO: handle odd nb for (int i = 0; i < nb; i += 2) { const block_q8_0 * restrict x0 = &x[i + 0]; const block_q8_0 * restrict x1 = &x[i + 1]; From dd0dc366dab10e8df28d3924e7f313b5c695e908 Mon Sep 17 00:00:00 2001 From: igarnier Date: Mon, 28 Aug 2023 10:19:59 +0200 Subject: [PATCH 076/147] llama.h : add missing struct keyword for C compat in callback type (#2847) --- llama.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.h b/llama.h index b084fe23c..7bb681d61 100644 --- a/llama.h +++ b/llama.h @@ -496,7 +496,7 @@ extern "C" { // Type of pointer to the beam_search_callback function. // void* callback_data is any custom data passed to llama_beam_search, that is subsequently // passed back to beam_search_callback. This avoids having to use global variables in the callback. - typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, llama_beams_state); + typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state); /// @details Deterministically returns entire sentence constructed by a beam search. /// @param ctx Pointer to the llama_context. From 92b1bbd2ec43c82ec0530ba3c8758846c5790c75 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Mon, 28 Aug 2023 13:23:55 +0200 Subject: [PATCH 077/147] CUDA: fix RoPE asserts, block sizes (#2833) --- ggml-cuda.cu | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index d76a25dc2..5fd625630 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -4908,8 +4908,8 @@ static void scale_f32_cuda(const float * x, float * dst, const float scale, cons static void rope_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p0, const float p_delta, const int p_delta_rows, const float theta_scale, cudaStream_t stream) { - GGML_ASSERT(nrows % 2 == 0); // GG: is this assert really needed? I don't see why - const dim3 block_dims(1, 2*CUDA_ROPE_BLOCK_SIZE, 1); + GGML_ASSERT(ncols % 2 == 0); + const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); const dim3 block_nums(nrows, num_blocks_x, 1); rope_f32<<>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale); @@ -4917,7 +4917,8 @@ static void rope_f32_cuda(const float * x, float * dst, const int ncols, const i static void rope_neox_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p0, const float p_delta, const int p_delta_rows, const float theta_scale, cudaStream_t stream) { - const dim3 block_dims(1, 2*CUDA_ROPE_BLOCK_SIZE, 1); + GGML_ASSERT(ncols % 2 == 0); + const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); const dim3 block_nums(nrows, num_blocks_x, 1); rope_neox_f32<<>>(x, dst, ncols, p0, p_delta, p_delta_rows, theta_scale); From 35feac6560387cf0484371af3d9b12bff678e0b9 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Mon, 28 Aug 2023 14:24:53 +0300 Subject: [PATCH 078/147] ggml : sync (mem align to header + conv_transpose_2d fixes + ggml_alloc) (#2852) * ggml : sync (mem align to header + conv_transpose_2d fixes) ggml-ci * ggml-alloc : minor fix * ggml-alloc : sync more fixes --- ggml-alloc.c | 6 +++--- ggml.c | 22 ++++++++-------------- ggml.h | 18 +++++++++++++----- 3 files changed, 24 insertions(+), 22 deletions(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index 1ef011654..140e9a2a7 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -268,7 +268,7 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) /*.parse_seq = */ {0}, /*.parse_seq_len = */ 0, #ifdef GGML_ALLOCATOR_DEBUG - /*.allocated_tensors = */ = {0}, + /*.allocated_tensors = */ {0}, #endif }; @@ -297,7 +297,7 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) { /*.parse_seq = */ {0}, /*.parse_seq_len = */ 0, #ifdef GGML_ALLOCATOR_DEBUG - /*.allocated_tensors = */ = {0}, + /*.allocated_tensors = */ {0}, #endif }; @@ -556,7 +556,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n( struct ggml_tensor * view_src = get_view_source(parent); struct hash_node * view_src_hn = hash_get(ht, view_src); view_src_hn->n_views -= 1; - AT_PRINTF("view_src %s\n", view_src->name); + AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views); if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) { ggml_allocator_free_tensor(alloc, view_src); } diff --git a/ggml.c b/ggml.c index 54f426bc0..dadb30757 100644 --- a/ggml.c +++ b/ggml.c @@ -157,12 +157,6 @@ typedef void * thread_ret_t; //#define GGML_SOFT_MAX_ACCELERATE #endif -#if UINTPTR_MAX == 0xFFFFFFFF - #define GGML_MEM_ALIGN 4 -#else - #define GGML_MEM_ALIGN 16 -#endif - // // logging // @@ -7098,11 +7092,13 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0( }; struct ggml_tensor* result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + + ggml_set_op_params_i32(result, 0, stride); + result->op = GGML_OP_CONV_TRANSPOSE_2D; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; - result->src[2] = ggml_new_i32(ctx, stride); return result; } @@ -13498,7 +13494,6 @@ static void ggml_compute_forward_conv_transpose_2d( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - const struct ggml_tensor * opt0, struct ggml_tensor * dst) { GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -13558,7 +13553,7 @@ static void ggml_compute_forward_conv_transpose_2d( return; } - const int32_t stride = ((const int32_t*)(opt0->data))[0]; + const int32_t stride = ggml_get_op_params_i32(dst, 0); // total patches in dst const int np = ne2; @@ -13571,7 +13566,7 @@ static void ggml_compute_forward_conv_transpose_2d( const int ip1 = MIN(ip0 + dp, np); ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - ggml_fp16_t * const wdata_src = (ggml_fp16_t *) params->wdata + nk; + ggml_fp16_t * const wdata_src = wdata + nk; for (int i2 = ip0; i2 < ip1; i2++) { // Cout float * dst_data = (float *)((char *) dst->data + i2*nb2); @@ -13583,9 +13578,8 @@ static void ggml_compute_forward_conv_transpose_2d( for (int i00 = 0; i00 < ne00; i00++) { float v = 0; ggml_vec_dot_f16(ne03, &v, - (ggml_fp16_t *) wdata_src + i1n, - (ggml_fp16_t *) wdata_kernel + i01*ne00*ne03 + i00*ne03); - + wdata_src + i1n, + wdata_kernel + i01*ne00*ne03 + i00*ne03); dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v; } } @@ -15732,7 +15726,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm } break; case GGML_OP_CONV_TRANSPOSE_2D: { - ggml_compute_forward_conv_transpose_2d(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor); + ggml_compute_forward_conv_transpose_2d(params, tensor->src[0], tensor->src[1], tensor); } break; case GGML_OP_POOL_1D: { diff --git a/ggml.h b/ggml.h index 792ca6e42..4ef3d5253 100644 --- a/ggml.h +++ b/ggml.h @@ -130,13 +130,16 @@ // The data of the tensor is accessed via the "data" pointer. For example: // // { -// struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2, 3); +// const int nx = 2; +// const int ny = 3; // -// // a[2, 1] = 1.0f; -// *(float *) ((char *) a->data + 2*a->nb[1] + 1*a->nb[0]) = 1.0f; +// struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, nx, ny); // -// // a[0, 2] = 2.0f; -// *(float *) ((char *) a->data + 0*a->nb[1] + 2*a->nb[0]) = 2.0f; +// for (int y = 0; y < ny; y++) { +// for (int x = 0; x < nx; x++) { +// *(float *) ((char *) a->data + y*a->nb[1] + x*a->nb[0]) = x + y; +// } +// } // // ... // } @@ -211,6 +214,11 @@ #define GGML_MAX_OP_PARAMS 32 #define GGML_DEFAULT_N_THREADS 4 +#if UINTPTR_MAX == 0xFFFFFFFF + #define GGML_MEM_ALIGN 4 +#else + #define GGML_MEM_ALIGN 16 +#endif #define GGML_EXIT_SUCCESS 0 #define GGML_EXIT_ABORTED 1 From 3af6b86301ddfb11bb68e91dfc030b611b0d8426 Mon Sep 17 00:00:00 2001 From: Ronny Brendel Date: Mon, 28 Aug 2023 14:51:08 +0200 Subject: [PATCH 079/147] ggml : tiny ggml_vec_dot_q4_K_q8_K AVX2 improvement (#2819) --- k_quants.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/k_quants.c b/k_quants.c index 82bf81697..3a9b1dafd 100644 --- a/k_quants.c +++ b/k_quants.c @@ -2694,13 +2694,13 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; __m256i p16l = _mm256_maddubs_epi16(q4l, q8l); p16l = _mm256_madd_epi16(scale_l, p16l); - sumi = _mm256_add_epi32(sumi, p16l); const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; __m256i p16h = _mm256_maddubs_epi16(q4h, q8h); p16h = _mm256_madd_epi16(scale_h, p16h); - sumi = _mm256_add_epi32(sumi, p16h); + const __m256i sumj = _mm256_add_epi32(p16l, p16h); + sumi = _mm256_add_epi32(sumi, sumj); } __m256 vd = _mm256_set1_ps(d); From be475f60af1a54e8de81466ccc907d080cf6df1a Mon Sep 17 00:00:00 2001 From: grahameth <96447521+grahameth@users.noreply.github.com> Date: Mon, 28 Aug 2023 17:38:12 +0200 Subject: [PATCH 080/147] llama.cpp : fix wrong vsnprintf call in MS compiler (#2856) Co-authored-by: grahameth <-> --- llama.cpp | 4 ---- 1 file changed, 4 deletions(-) diff --git a/llama.cpp b/llama.cpp index 72d2d1de0..da8ff64d0 100644 --- a/llama.cpp +++ b/llama.cpp @@ -6257,10 +6257,6 @@ void llama_log_set(llama_log_callback log_callback, void * user_data) { g_state.log_callback_user_data = user_data; } -#if defined(_MSC_VER) && !defined(vsnprintf) -#define vsnprintf _vsnprintf -#endif - static void llama_log_internal_v(llama_log_level level, const char * format, va_list args) { va_list args_copy; va_copy(args_copy, args); From 75fafcbcccc280a5b3883bc76d0a2dabf474d094 Mon Sep 17 00:00:00 2001 From: alonfaraj Date: Mon, 28 Aug 2023 18:38:35 +0300 Subject: [PATCH 081/147] make : fix tests build (#2855) * makefile: - fix test name - add missing tests build * editorconfig : fixes --------- Co-authored-by: Georgi Gerganov --- .gitignore | 5 ++++- Makefile | 14 ++++++++++---- 2 files changed, 14 insertions(+), 5 deletions(-) diff --git a/.gitignore b/.gitignore index e5faab774..7a3f3fff4 100644 --- a/.gitignore +++ b/.gitignore @@ -63,10 +63,13 @@ poetry.toml # Test binaries tests/test-grammar-parser +tests/test-llama-grammar tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling -tests/test-tokenizer-0 +tests/test-tokenizer-0-llama +tests/test-tokenizer-0-falcon +tests/test-tokenizer-1 diff --git a/Makefile b/Makefile index a3400e491..e60821dd5 100644 --- a/Makefile +++ b/Makefile @@ -2,7 +2,7 @@ BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple server embd-input-test gguf llama-bench # Binaries only useful for tests -TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0 +TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1 default: $(BUILD_TARGETS) @@ -442,10 +442,10 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o llama.o common.o $(OBJS) +tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o common.o grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) -tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o common.o $(OBJS) +tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o common.o $(OBJS) @@ -466,5 +466,11 @@ tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) -tests/test-tokenizer-0: tests/test-tokenizer-0.cpp build-info.h ggml.o llama.o common.o $(OBJS) +tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + +tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + +tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) From 6b73ef120114beb5664ea94aab48d07ed248ee52 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Mon, 28 Aug 2023 17:59:39 +0200 Subject: [PATCH 082/147] YAML result logging + preset script (#2657) --- common/common.cpp | 331 +++++++++++++++++++++++++++-- common/common.h | 18 ++ examples/main/main.cpp | 78 ++++++- examples/perplexity/perplexity.cpp | 141 +++++++++--- examples/server/server.cpp | 2 +- llama.cpp | 29 +++ llama.h | 3 + run_with_preset.py | 140 ++++++++++++ 8 files changed, 700 insertions(+), 42 deletions(-) create mode 100755 run_with_preset.py diff --git a/common/common.cpp b/common/common.cpp index 0d91a6a35..4a0d43c13 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1,15 +1,20 @@ #include "common.h" +#include "build-info.h" +#include "llama.h" -#include -#include -#include -#include -#include -#include #include -#include -#include +#include +#include +#include +#include +#include +#include +#include #include +#include +#include +#include +#include #if defined(__APPLE__) && defined(__MACH__) #include @@ -19,11 +24,14 @@ #if defined(_WIN32) #define WIN32_LEAN_AND_MEAN #define NOMINMAX +#include +#include #include #include #include #else #include +#include #include #endif @@ -93,7 +101,6 @@ void process_escapes(std::string& input) { bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { bool invalid_param = false; - bool escape_prompt = false; std::string arg; gpt_params default_params; const std::string arg_prefix = "--"; @@ -125,8 +132,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.prompt = argv[i]; - } else if (arg == "-e") { - escape_prompt = true; + } else if (arg == "-e" || arg == "--escape") { + params.escape = true; } else if (arg == "--prompt-cache") { if (++i >= argc) { invalid_param = true; @@ -415,6 +422,16 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.antiprompt.push_back(argv[i]); + } else if (arg == "-ld" || arg == "--logdir") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.logdir = argv[i]; + + if (params.logdir.back() != DIRECTORY_SEPARATOR) { + params.logdir += DIRECTORY_SEPARATOR; + } } else if (arg == "--perplexity") { params.perplexity = true; } else if (arg == "--ppl-stride") { @@ -520,7 +537,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { exit(1); } - if (escape_prompt) { + if (params.escape) { process_escapes(params.prompt); process_escapes(params.input_prefix); process_escapes(params.input_suffix); @@ -546,7 +563,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stdout, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads); fprintf(stdout, " -p PROMPT, --prompt PROMPT\n"); fprintf(stdout, " prompt to start generation with (default: empty)\n"); - fprintf(stdout, " -e process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n"); + fprintf(stdout, " -e, --escape process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n"); fprintf(stdout, " --prompt-cache FNAME file to cache prompt state for faster startup (default: none)\n"); fprintf(stdout, " --prompt-cache-all if specified, saves user input and generations to cache as well.\n"); fprintf(stdout, " not supported with --interactive or other interactive options\n"); @@ -627,6 +644,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stdout, " --lora-base FNAME optional model to use as a base for the layers modified by the LoRA adapter\n"); fprintf(stdout, " -m FNAME, --model FNAME\n"); fprintf(stdout, " model path (default: %s)\n", params.model.c_str()); + fprintf(stdout, " -ld LOGDIR, --logdir LOGDIR\n"); + fprintf(stdout, " path under which to save YAML logs (no logging if unset)\n"); fprintf(stdout, "\n"); } @@ -779,3 +798,289 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector> converter; + std::wstring wpath = converter.from_bytes(path); + + // if the path already exists, check whether it's a directory + const DWORD attributes = GetFileAttributesW(wpath.c_str()); + if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) { + return true; + } + + size_t pos_slash = 0; + + // process path from front to back, procedurally creating directories + while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) { + const std::wstring subpath = wpath.substr(0, pos_slash); + const wchar_t * test = subpath.c_str(); + + const bool success = CreateDirectoryW(test, NULL); + if (!success) { + const DWORD error = GetLastError(); + + // if the path already exists, ensure that it's a directory + if (error == ERROR_ALREADY_EXISTS) { + const DWORD attributes = GetFileAttributesW(subpath.c_str()); + if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) { + return false; + } + } else { + return false; + } + } + + pos_slash += 1; + } + + return true; +#else + // if the path already exists, check whether it's a directory + struct stat info; + if (stat(path.c_str(), &info) == 0) { + return S_ISDIR(info.st_mode); + } + + size_t pos_slash = 1; // skip leading slashes for directory creation + + // process path from front to back, procedurally creating directories + while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) { + const std::string subpath = path.substr(0, pos_slash); + struct stat info; + + // if the path already exists, ensure that it's a directory + if (stat(subpath.c_str(), &info) == 0) { + if (!S_ISDIR(info.st_mode)) { + return false; + } + } else { + // create parent directories + const int ret = mkdir(subpath.c_str(), 0755); + if (ret != 0) { + return false; + } + } + + pos_slash += 1; + } + + return true; +#endif // _WIN32 +} + +void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector & data) { + if (data.empty()) { + fprintf(stream, "%s:\n", prop_name); + return; + } + + fprintf(stream, "%s: [", prop_name); + for (size_t i = 0; i < data.size() - 1; ++i) { + fprintf(stream, "%e, ", data[i]); + } + fprintf(stream, "%e]\n", data.back()); +} + +void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector & data) { + if (data.empty()) { + fprintf(stream, "%s:\n", prop_name); + return; + } + + fprintf(stream, "%s: [", prop_name); + for (size_t i = 0; i < data.size() - 1; ++i) { + fprintf(stream, "%d, ", data[i]); + } + fprintf(stream, "%d]\n", data.back()); +} + +void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data) { + std::string data_str(data == NULL ? "" : data); + + if (data_str.empty()) { + fprintf(stream, "%s:\n", prop_name); + return; + } + + size_t pos_start = 0; + size_t pos_found = 0; + + if (!data_str.empty() && (std::isspace(data_str[0]) || std::isspace(data_str.back()))) { + data_str = std::regex_replace(data_str, std::regex("\n"), "\\n"); + data_str = std::regex_replace(data_str, std::regex("\""), "\\\""); + data_str = "\"" + data_str + "\""; + fprintf(stream, "%s: %s\n", prop_name, data_str.c_str()); + return; + } + + if (data_str.find('\n') == std::string::npos) { + fprintf(stream, "%s: %s\n", prop_name, data_str.c_str()); + return; + } + + fprintf(stream, "%s: |\n", prop_name); + while ((pos_found = data_str.find('\n', pos_start)) != std::string::npos) { + fprintf(stream, " %s\n", data_str.substr(pos_start, pos_found-pos_start).c_str()); + pos_start = pos_found + 1; + } +} + +std::string get_sortable_timestamp() { + using clock = std::chrono::system_clock; + + const clock::time_point current_time = clock::now(); + const time_t as_time_t = clock::to_time_t(current_time); + char timestamp_no_ns[100]; + std::strftime(timestamp_no_ns, 100, "%Y_%m_%d-%H_%M_%S", std::localtime(&as_time_t)); + + const int64_t ns = std::chrono::duration_cast( + current_time.time_since_epoch() % 1000000000).count(); + char timestamp_ns[10]; + snprintf(timestamp_ns, 11, "%09ld", ns); + + return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns); +} + +void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx, + const std::string & timestamp, const std::vector & prompt_tokens, const char * model_desc) { + fprintf(stream, "build_commit: %s\n", BUILD_COMMIT); + fprintf(stream, "build_number: %d\n", BUILD_NUMBER); + fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false"); + fprintf(stream, "cpu_has_avx: %s\n", ggml_cpu_has_avx() ? "true" : "false"); + fprintf(stream, "cpu_has_avx2: %s\n", ggml_cpu_has_avx2() ? "true" : "false"); + fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false"); + fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false"); + fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false"); + fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false"); + fprintf(stream, "cpu_has_cublas: %s\n", ggml_cpu_has_cublas() ? "true" : "false"); + fprintf(stream, "cpu_has_clblast: %s\n", ggml_cpu_has_clblast() ? "true" : "false"); + fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false"); + fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false"); + fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false"); + fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false"); + fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false"); + fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false"); + fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false"); + fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false"); + fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false"); + +#ifdef NDEBUG + fprintf(stream, "debug: false\n"); +#else + fprintf(stream, "debug: true\n"); +#endif // NDEBUG + + fprintf(stream, "model_desc: %s\n", model_desc); + fprintf(stream, "n_vocab: %d # output size of the final layer, 32001 for some models\n", llama_n_vocab(lctx)); + +#ifdef __OPTIMIZE__ + fprintf(stream, "optimize: true\n"); +#else + fprintf(stream, "optimize: false\n"); +#endif // __OPTIMIZE__ + + fprintf(stream, "time: %s\n", timestamp.c_str()); + + fprintf(stream, "\n"); + fprintf(stream, "###############\n"); + fprintf(stream, "# User Inputs #\n"); + fprintf(stream, "###############\n"); + fprintf(stream, "\n"); + + fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str()); + fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch); + dump_string_yaml_multiline(stream, "cfg_negative_prompt", params.cfg_negative_prompt.c_str()); + fprintf(stream, "cfg_scale: %f # default: 1.0\n", params.cfg_scale); + fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks); + fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false"); + fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx); + fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false"); + fprintf(stream, "export: %s # default: false\n", params.export_cgraph ? "true" : "false"); + fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n"); + fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", params.frequency_penalty); + dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str()); + fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n"); + fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false"); + fprintf(stream, "hellaswag_tasks: %ld # default: 400\n", params.hellaswag_tasks); + + const auto logit_bias_eos = params.logit_bias.find(llama_token_eos(lctx)); + const bool ignore_eos = logit_bias_eos != params.logit_bias.end() && logit_bias_eos->second == -INFINITY; + fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false"); + + dump_string_yaml_multiline(stream, "in_prefix", params.input_prefix.c_str()); + fprintf(stream, "in_prefix_bos: %s # default: false\n", params.input_prefix_bos ? "true" : "false"); + dump_string_yaml_multiline(stream, "in_suffix", params.input_prefix.c_str()); + fprintf(stream, "instruct: %s # default: false\n", params.instruct ? "true" : "false"); + fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false"); + fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false"); + fprintf(stream, "keep: %d # default: 0\n", params.n_keep); + fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str()); + + fprintf(stream, "logit_bias:\n"); + for (std::pair lb : params.logit_bias) { + if (ignore_eos && lb.first == logit_bias_eos->first) { + continue; + } + fprintf(stream, " %d: %f", lb.first, lb.second); + } + + fprintf(stream, "lora: %s\n", params.lora_adapter.c_str()); + fprintf(stream, "lora_base: %s\n", params.lora_base.c_str()); + fprintf(stream, "low_vram: %s # default: false\n", params.low_vram ? "true" : "false"); + fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu); + fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false"); + fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", params.mirostat); + fprintf(stream, "mirostat_ent: %f # default: 5.0\n", params.mirostat_tau); + fprintf(stream, "mirostat_lr: %f # default: 0.1\n", params.mirostat_eta); + fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false"); + fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str()); + fprintf(stream, "mtest: %s # default: false\n", params.mem_test ? "true" : "false"); + fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false"); + fprintf(stream, "n_gpu_layers: %d # default: 0\n", params.n_gpu_layers); + fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict); + fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", params.n_probs); + fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false"); + fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false"); + fprintf(stream, "no_penalize_nl: %s # default: false\n", !params.penalize_nl ? "true" : "false"); + fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false"); + fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type); + fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride); + fprintf(stream, "presence_penalty: %f # default: 0.0\n", params.presence_penalty); + dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str()); + fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str()); + fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false"); + fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false"); + dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens); + fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false"); + fprintf(stream, "repeat_penalty: %f # default: 1.1\n", params.repeat_penalty); + + fprintf(stream, "reverse_prompt:\n"); + for (std::string ap : params.antiprompt) { + size_t pos = 0; + while ((pos = ap.find('\n', pos)) != std::string::npos) { + ap.replace(pos, 1, "\\n"); + pos += 1; + } + + fprintf(stream, " - %s\n", ap.c_str()); + } + + fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base); + fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale); + fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed); + fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false"); + fprintf(stream, "temp: %f # default: 0.8\n", params.temp); + + const std::vector tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES); + dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector); + + fprintf(stream, "tfs: %f # default: 1.0\n", params.tfs_z); + fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency()); + fprintf(stream, "top_k: %d # default: 40\n", params.top_k); + fprintf(stream, "top_p: %f # default: 0.95\n", params.top_p); + fprintf(stream, "typical_p: %f # default: 1.0\n", params.typical_p); + fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false"); +} diff --git a/common/common.h b/common/common.h index 97fda2be7..c15373144 100644 --- a/common/common.h +++ b/common/common.h @@ -11,6 +11,12 @@ #include #include +#ifdef _WIN32 +#define DIRECTORY_SEPARATOR '\\' +#else +#define DIRECTORY_SEPARATOR '/' +#endif // _WIN32 + // // CLI argument parsing // @@ -61,6 +67,7 @@ struct gpt_params { std::string input_suffix = ""; // string to suffix user inputs with std::string grammar = ""; // optional BNF-like grammar to constrain sampling std::vector antiprompt; // string upon seeing which more user input is prompted + std::string logdir = ""; // directory in which to save YAML log files std::string lora_adapter = ""; // lora adapter path std::string lora_base = ""; // base model path for the lora adapter @@ -82,6 +89,7 @@ struct gpt_params { bool prompt_cache_ro = false; // open the prompt cache read-only and do not update it bool embedding = false; // get only sentence embedding + bool escape = false; // escape "\n", "\r", "\t", "\'", "\"", and "\\" bool interactive_first = false; // wait for user input immediately bool multiline_input = false; // reverse the usage of `\` bool simple_io = false; // improves compatibility with subprocesses and limited consoles @@ -144,3 +152,13 @@ std::string llama_detokenize_spm( std::string llama_detokenize_bpe( llama_context * ctx, const std::vector & tokens); + +bool create_directory_with_parents(const std::string & path); +void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector & data); +void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector & data); +void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data); +std::string get_sortable_timestamp(); + +void dump_non_result_info_yaml( + FILE * stream, const gpt_params & params, const llama_context * lctx, + const std::string & timestamp, const std::vector & prompt_tokens, const char * model_desc); diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 3ce57f436..89cc4f602 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -17,6 +17,7 @@ #include #include #include +#include #include #include @@ -36,9 +37,57 @@ #pragma warning(disable: 4244 4267) // possible loss of data #endif -static llama_context ** g_ctx; +static llama_context ** g_ctx; +static llama_model ** g_model; +static gpt_params * g_params; +static std::vector * g_input_tokens; +static std::ostringstream * g_output_ss; +static std::vector * g_output_tokens; static bool is_interacting = false; +void write_logfile( + const llama_context * ctx, const gpt_params & params, const llama_model * model, + const std::vector input_tokens, const std::string output, const std::vector output_tokens) { + + if (params.logdir.empty()) { + return; + } + + const std::string timestamp = get_sortable_timestamp(); + + const bool success = create_directory_with_parents(params.logdir); + if (!success) { + fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n", + __func__, params.logdir.c_str()); + return; + } + + const std::string logfile_path = params.logdir + timestamp + ".yml"; + FILE * logfile = fopen(logfile_path.c_str(), "w"); + + if (logfile == NULL) { + fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str()); + return; + } + + fprintf(logfile, "binary: main\n"); + char model_desc[128]; + llama_model_desc(model, model_desc, sizeof(model_desc)); + dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc); + + fprintf(logfile, "\n"); + fprintf(logfile, "######################\n"); + fprintf(logfile, "# Generation Results #\n"); + fprintf(logfile, "######################\n"); + fprintf(logfile, "\n"); + + dump_string_yaml_multiline(logfile, "output", output.c_str()); + dump_vector_int_yaml(logfile, "output_tokens", output_tokens); + + llama_dump_timing_info_yaml(logfile, ctx); + fclose(logfile); +} + #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) void sigint_handler(int signo) { if (signo == SIGINT) { @@ -48,6 +97,7 @@ void sigint_handler(int signo) { console::cleanup(); printf("\n"); llama_print_timings(*g_ctx); + write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens); _exit(130); } } @@ -56,6 +106,7 @@ void sigint_handler(int signo) { int main(int argc, char ** argv) { gpt_params params; + g_params = ¶ms; if (gpt_params_parse(argc, argv, params) == false) { return 1; @@ -116,6 +167,7 @@ int main(int argc, char ** argv) { llama_model * model; llama_context * ctx; llama_context * ctx_guidance = NULL; + g_model = &model; g_ctx = &ctx; // load the model and apply lora adapter, if any @@ -397,6 +449,10 @@ int main(int argc, char ** argv) { int n_session_consumed = 0; int n_past_guidance = 0; + std::vector input_tokens; g_input_tokens = &input_tokens; + std::vector output_tokens; g_output_tokens = &output_tokens; + std::ostringstream output_ss; g_output_ss = &output_ss; + // the first thing we will do is to output the prompt, so set color accordingly console::set_display(console::prompt); @@ -667,7 +723,15 @@ int main(int argc, char ** argv) { // display text if (input_echo) { for (auto id : embd) { - printf("%s", llama_token_to_piece(ctx, id).c_str()); + const std::string token_str = llama_token_to_piece(ctx, id); + printf("%s", token_str.c_str()); + + if (embd.size() > 1) { + input_tokens.push_back(id); + } else { + output_tokens.push_back(id); + output_ss << token_str; + } } fflush(stdout); } @@ -761,6 +825,8 @@ int main(int argc, char ** argv) { printf("%s", params.input_suffix.c_str()); } + const size_t original_size = embd_inp.size(); + // instruct mode: insert instruction prefix if (params.instruct && !is_antiprompt) { n_consumed = embd_inp.size(); @@ -775,6 +841,12 @@ int main(int argc, char ** argv) { embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end()); } + for (size_t i = original_size; i < embd_inp.size(); ++i) { + const llama_token token = embd_inp[i]; + output_tokens.push_back(token); + output_ss << llama_token_to_piece(ctx, token); + } + n_remain -= line_inp.size(); } @@ -817,6 +889,8 @@ int main(int argc, char ** argv) { } llama_print_timings(ctx); + write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens); + if (ctx_guidance) { llama_free(ctx_guidance); } llama_free(ctx); llama_free_model(model); diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index ebafa0c29..aeb774c5f 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -3,16 +3,79 @@ #include "build-info.h" #include +#include +#include #include #include -#include #include #include +#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data #endif +struct results_perplexity { + std::vector tokens; + double ppl_value; + std::vector logits; + std::vector probs; +}; + +struct results_log_softmax { + double log_softmax; + float logit; + float prob; +}; + +void write_logfile(const llama_context * ctx, const gpt_params & params, + const llama_model * model, const struct results_perplexity & results) { + + if (params.logdir.empty()) { + return; + } + + if (params.hellaswag) { + fprintf(stderr, "%s: warning: logging results is not implemented for HellaSwag. No files will be written.\n", __func__); + return; + } + + const std::string timestamp = get_sortable_timestamp(); + + const bool success = create_directory_with_parents(params.logdir); + if (!success) { + fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n", + __func__, params.logdir.c_str()); + return; + } + + const std::string logfile_path = params.logdir + timestamp + ".yml"; + FILE * logfile = fopen(logfile_path.c_str(), "w"); + + if (logfile == NULL) { + fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str()); + return; + } + + fprintf(logfile, "binary: main\n"); + char model_desc[128]; + llama_model_desc(model, model_desc, sizeof(model_desc)); + dump_non_result_info_yaml(logfile, params, ctx, timestamp, results.tokens, model_desc); + + fprintf(logfile, "\n"); + fprintf(logfile, "######################\n"); + fprintf(logfile, "# Perplexity Results #\n"); + fprintf(logfile, "######################\n"); + fprintf(logfile, "\n"); + + dump_vector_float_yaml(logfile, "logits", results.logits); + fprintf(logfile, "ppl_value: %f\n", results.ppl_value); + dump_vector_float_yaml(logfile, "probs", results.probs); + + llama_dump_timing_info_yaml(logfile, ctx); + fclose(logfile); +} + std::vector softmax(const std::vector& logits) { std::vector probs(logits.size()); float max_logit = logits[0]; @@ -29,20 +92,20 @@ std::vector softmax(const std::vector& logits) { return probs; } -float log_softmax(int n_vocab, const float * logits, int tok) { +results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) { float max_logit = logits[0]; for (int i = 1; i < n_vocab; ++i) max_logit = std::max(max_logit, logits[i]); double sum_exp = 0.0; for (int i = 0; i < n_vocab; ++i) sum_exp += expf(logits[i] - max_logit); - return logits[tok] - max_logit - log(sum_exp); + return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp}; } -void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector& workers, - double& nll, double& nll2) { +void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector & workers, + double & nll, double & nll2, float * logit_history, float * prob_history) { std::mutex mutex; int counter = 0; - auto compute = [&mutex, &counter, &nll, &nll2, n_vocab, logits, tokens, n_token] () { + auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () { double local_nll = 0, local_nll2 = 0; while (true) { std::unique_lock lock(mutex); @@ -52,34 +115,43 @@ void process_logits(int n_vocab, const float * logits, const int * tokens, int n break; } lock.unlock(); - double v = -log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]); + const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]); + const double v = -results.log_softmax; local_nll += v; local_nll2 += v*v; + + logit_history[i] = results.logit; + prob_history[i] = results.prob; } }; - for (auto& w : workers) w = std::thread(compute); + for (auto & w : workers) w = std::thread(compute); compute(); - for (auto& w : workers) w.join(); + for (auto & w : workers) w.join(); } -void perplexity_v2(llama_context * ctx, const gpt_params & params) { +results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) { // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` // Output: `perplexity: 13.5106 [114/114]` // BOS tokens will be added for each chunk before eval - if (params.ppl_stride <= 0) { - fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride); - return; - } - const bool is_spm = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; const bool add_bos = is_spm; fprintf(stderr, "%s: tokenizing the input ..\n", __func__); - auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos); + std::vector tokens = ::llama_tokenize(ctx, params.prompt, add_bos); + std::vector logit_history; + std::vector prob_history; + + logit_history.resize(tokens.size()); + prob_history.resize(tokens.size()); + + if (params.ppl_stride <= 0) { + fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride); + return {tokens, -1, logit_history, prob_history}; + } const int calc_chunk = params.n_ctx; @@ -88,7 +160,7 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) { if (int(tokens.size()) <= calc_chunk) { fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__, tokens.size(), params.n_ctx, params.ppl_stride); - return; + return {tokens, -1, logit_history, prob_history}; } const int n_chunk_max = (tokens.size() - calc_chunk + params.ppl_stride - 1) / params.ppl_stride; @@ -120,7 +192,7 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) { //fprintf(stderr, " Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch); if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) { //fprintf(stderr, "%s : failed to eval\n", __func__); - return; + return {tokens, -1, logit_history, prob_history}; } // save original token and restore it after eval @@ -161,6 +233,8 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) { logits.begin() + (j + 1) * n_vocab); const float prob = softmax(tok_logits)[tokens[start + j + 1]]; + logit_history[start + j + 1] = tok_logits[tokens[start + j + 1]]; + prob_history[start + j + 1] = prob; nll += -std::log(prob); ++count; @@ -174,12 +248,14 @@ void perplexity_v2(llama_context * ctx, const gpt_params & params) { fflush(stdout); } printf("\n"); + + return {tokens, std::exp(nll / count), logit_history, prob_history}; } -void perplexity(llama_context * ctx, const gpt_params & params) { +results_perplexity perplexity(llama_context * ctx, const gpt_params & params) { + if (params.ppl_stride > 0) { - perplexity_v2(ctx, params); - return; + return perplexity_v2(ctx, params); } // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research @@ -193,11 +269,17 @@ void perplexity(llama_context * ctx, const gpt_params & params) { auto tim1 = std::chrono::high_resolution_clock::now(); fprintf(stderr, "%s: tokenizing the input ..\n", __func__); - auto tokens = ::llama_tokenize(ctx, params.prompt, add_bos); + std::vector tokens = ::llama_tokenize(ctx, params.prompt, add_bos); auto tim2 = std::chrono::high_resolution_clock::now(); fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast(tim2-tim1).count()); + std::vector logit_history; + logit_history.resize(tokens.size()); + + std::vector prob_history; + prob_history.resize(tokens.size()); + const int n_chunk_max = tokens.size() / params.n_ctx; const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max); @@ -236,7 +318,7 @@ void perplexity(llama_context * ctx, const gpt_params & params) { if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * n_batch, params.n_threads)) { fprintf(stderr, "%s : failed to eval\n", __func__); - return; + return {tokens, -1, logit_history, prob_history}; } // restore the original token in case it was set to BOS @@ -272,7 +354,8 @@ void perplexity(llama_context * ctx, const gpt_params & params) { // last 256 tokens. Then, we split the input up into context window size chunks to // process the entire prompt. const int first = std::min(512, params.n_ctx/2); - process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first, workers, nll, nll2); + process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, params.n_ctx - 1 - first, + workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); count += params.n_ctx - first - 1; // perplexity is e^(average negative log-likelihood) @@ -287,16 +370,19 @@ void perplexity(llama_context * ctx, const gpt_params & params) { fflush(stdout); } printf("\n"); + nll2 /= count; nll /= count; + const double ppl = exp(nll); nll2 -= nll * nll; if (nll2 > 0) { nll2 = sqrt(nll2/(count-1)); - double ppl = exp(nll); printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl); } else { printf("Unexpected negative standard deviation of log(prob)\n"); } + + return {tokens, ppl, logit_history, prob_history}; } std::vector hellaswag_evaluate_tokens(llama_context * ctx, const std::vector& tokens, int n_past, int n_batch, @@ -604,13 +690,16 @@ int main(int argc, char ** argv) { params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info()); } + struct results_perplexity results; if (params.hellaswag) { hellaswag_score(ctx, params); } else { - perplexity(ctx, params); + results = perplexity(ctx, params); } llama_print_timings(ctx); + write_logfile(ctx, params, model, results); + llama_free(ctx); llama_free_model(model); diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 89a3311f5..b485a5ead 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -719,7 +719,7 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, fprintf(stdout, " -ts SPLIT --tensor-split SPLIT\n"); fprintf(stdout, " how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n"); fprintf(stdout, " -mg i, --main-gpu i the GPU to use for scratch and small tensors\n"); - fprintf(stdout, " -lv, --low-vram don't allocate VRAM scratch buffer\n"); + fprintf(stdout, " -lv, --low-vram don't allocate VRAM scratch buffer\n"); fprintf(stdout, " -nommq, --no-mul-mat-q\n"); fprintf(stdout, " use cuBLAS instead of custom mul_mat_q CUDA kernels.\n"); fprintf(stdout, " Not recommended since this is both slower and uses more VRAM.\n"); diff --git a/llama.cpp b/llama.cpp index da8ff64d0..11697ee65 100644 --- a/llama.cpp +++ b/llama.cpp @@ -6247,6 +6247,35 @@ const char * llama_print_system_info(void) { return s.c_str(); } +void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) { + + fprintf(stream, "\n"); + fprintf(stream, "###########\n"); + fprintf(stream, "# Timings #\n"); + fprintf(stream, "###########\n"); + fprintf(stream, "\n"); + + fprintf(stream, "mst_eval: %.2f # ms / token during generation\n", + 1.0e-3 * ctx->t_eval_us / ctx->n_eval); + fprintf(stream, "mst_p_eval: %.2f # ms / token during prompt processing\n", + 1.0e-3 * ctx->t_p_eval_us / ctx->n_p_eval); + fprintf(stream, "mst_sample: %.2f # ms / token during sampling\n", + 1.0e-3 * ctx->t_sample_us / ctx->n_sample); + fprintf(stream, "n_eval: %d # number of tokens generated (excluding the first one)\n", ctx->n_eval); + fprintf(stream, "n_p_eval: %d # number of tokens processed in batches at the beginning\n", ctx->n_p_eval); + fprintf(stream, "n_sample: %d # number of sampled tokens\n", ctx->n_sample); + fprintf(stream, "t_eval_us: %ld # total microseconds spent generating tokens\n", ctx->t_eval_us); + fprintf(stream, "t_load_us: %ld # total microseconds spent loading the model\n", ctx->t_load_us); + fprintf(stream, "t_p_eval_us: %ld # total microseconds spent prompt processing\n", ctx->t_p_eval_us); + fprintf(stream, "t_sample_us: %ld # total microseconds spent sampling\n", ctx->t_sample_us); + fprintf(stream, "ts_eval: %.2f # tokens / second during generation\n", + 1.0e6 * ctx->n_eval / ctx->t_eval_us); + fprintf(stream, "ts_p_eval: %.2f # tokens / second during prompt processing\n", + 1.0e6 * ctx->n_p_eval / ctx->t_p_eval_us); + fprintf(stream, "ts_sample: %.2f # tokens / second during sampling\n", + 1.0e6 * ctx->n_sample / ctx->t_sample_us); +} + // For internal test use const std::vector>& llama_internal_get_tensor_map(struct llama_context * ctx) { return ctx->model.tensors_by_name; diff --git a/llama.h b/llama.h index 7bb681d61..b38d3be20 100644 --- a/llama.h +++ b/llama.h @@ -10,6 +10,7 @@ #endif // GGML_USE_CUBLAS #include #include +#include #include #ifdef LLAMA_SHARED @@ -520,6 +521,8 @@ extern "C" { // If this is not called, or NULL is supplied, everything is output on stderr. LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data); + LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx); + #ifdef __cplusplus } #endif diff --git a/run_with_preset.py b/run_with_preset.py new file mode 100755 index 000000000..8f90f52a9 --- /dev/null +++ b/run_with_preset.py @@ -0,0 +1,140 @@ +#!/usr/bin/env python3 + +import argparse +import os +import subprocess +import sys + +import yaml + +CLI_ARGS_MAIN_PERPLEXITY = [ + "batch-size", "cfg-negative-prompt", "cfg-scale", "chunks", "color", "ctx-size", "escape", + "export", "file", "frequency-penalty", "grammar", "grammar-file", "hellaswag", + "hellaswag-tasks", "ignore-eos", "in-prefix", "in-prefix-bos", "in-suffix", "instruct", + "interactive", "interactive-first", "keep", "logdir", "logit-bias", "lora", "lora-base", + "low-vram", "main-gpu", "memory-f32", "mirostat", "mirostat-ent", "mirostat-lr", "mlock", + "model", "mtest", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q", + "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt", + "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "random-prompt", "repeat-last-n", + "repeat-penalty", "reverse-prompt", "rope-freq-base", "rope-freq-scale", "rope-scale", "seed", + "simple-io", "tensor-split", "threads", "temp", "tfs", "top-k", "top-p", "typical", + "verbose-prompt" +] + +CLI_ARGS_LLAMA_BENCH = [ + "batch-size", "memory-f32", "low-vram", "model", "mul-mat-q", "n-gen", "n-gpu-layers", + "n-prompt", "output", "repetitions", "tensor-split", "threads", "verbose" +] + +CLI_ARGS_SERVER = [ + "alias", "batch-size", "ctx-size", "embedding", "host", "memory-f32", "lora", "lora-base", + "low-vram", "main-gpu", "mlock", "model", "n-gpu-layers", "n-probs", "no-mmap", "no-mul-mat-q", + "numa", "path", "port", "rope-freq-base", "timeout", "rope-freq-scale", "tensor-split", + "threads", "verbose" +] + +description = """Run llama.cpp binaries with presets from YAML file(s). +To specify which binary should be run, specify the "binary" property (main, perplexity, llama-bench, and server are supported). +To get a preset file template, run a llama.cpp binary with the "--logdir" CLI argument. + +Formatting considerations: +- The YAML property names are the same as the CLI argument names of the corresponding binary. +- Properties must use the long name of their corresponding llama.cpp CLI arguments. +- Like the llama.cpp binaries the property names do not differentiate between hyphens and underscores. +- Flags must be defined as ": true" to be effective. +- To define the logit_bias property, the expected format is ": " in the "logit_bias" namespace. +- To define multiple "reverse_prompt" properties simultaneously the expected format is a list of strings. +- To define a tensor split, pass a list of floats. +""" +usage = "run_with_preset.py [-h] [yaml_files ...] [-- ...]" +epilog = (" -- specify additional CLI ars to be passed to the binary (override all preset files). " + "Unknown args will be ignored.") + +parser = argparse.ArgumentParser( + description=description, usage=usage, epilog=epilog, formatter_class=argparse.RawTextHelpFormatter) +parser.add_argument("-bin", "--binary", help="The binary to run.") +parser.add_argument("yaml_files", nargs="*", + help="Arbitrary number of YAML files from which to read preset values. " + "If two files specify the same values the later one will be used.") + +known_args, unknown_args = parser.parse_known_args() + +if not known_args.yaml_files and not unknown_args: + parser.print_help() + sys.exit(0) + +props = dict() + +for yaml_file in known_args.yaml_files: + with open(yaml_file, "r") as f: + props.update(yaml.load(f, yaml.SafeLoader)) + +props = {prop.replace("_", "-"): val for prop, val in props.items()} + +binary = props.pop("binary", "main") +if known_args.binary: + binary = known_args.binary + +if os.path.exists(f"./{binary}"): + binary = f"./{binary}" + +if binary.lower().endswith("main") or binary.lower().endswith("perplexity"): + cli_args = CLI_ARGS_MAIN_PERPLEXITY +elif binary.lower().endswith("llama-bench"): + cli_args = CLI_ARGS_LLAMA_BENCH +elif binary.lower().endswith("server"): + cli_args = CLI_ARGS_SERVER +else: + print(f"Unknown binary: {binary}") + sys.exit(1) + +command_list = [binary] + +for cli_arg in cli_args: + value = props.pop(cli_arg, None) + + if not value or value == -1: + continue + + if cli_arg == "logit-bias": + for token, bias in value.items(): + command_list.append("--logit-bias") + command_list.append(f"{token}{bias:+}") + continue + + if cli_arg == "reverse-prompt" and not isinstance(value, str): + for rp in value: + command_list.append("--reverse-prompt") + command_list.append(str(rp)) + continue + + command_list.append(f"--{cli_arg}") + + if cli_arg == "tensor-split": + command_list.append(",".join([str(v) for v in value])) + continue + + value = str(value) + + if value != "True": + command_list.append(str(value)) + +num_unused = len(props) +if num_unused > 10: + print(f"The preset file contained a total of {num_unused} unused properties.") +elif num_unused > 0: + print("The preset file contained the following unused properties:") + for prop, value in props.items(): + print(f" {prop}: {value}") + +command_list += unknown_args + +sp = subprocess.Popen(command_list) + +while sp.returncode is None: + try: + sp.wait() + except KeyboardInterrupt: + pass + +sys.exit(sp.returncode) From 43033b7bb4858da4f591715b3babdf906c9b7cbc Mon Sep 17 00:00:00 2001 From: slaren Date: Mon, 28 Aug 2023 19:19:18 +0200 Subject: [PATCH 083/147] llama-bench : set locale to utf8 (#2832) --- examples/llama-bench/llama-bench.cpp | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index d0fe6d90d..bf3a487ab 100755 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -3,6 +3,9 @@ #include #include #include +#include +#include +#include #include #include #include @@ -10,7 +13,6 @@ #include #include #include -#include #include #include @@ -916,6 +918,9 @@ static void llama_null_log_callback(enum llama_log_level level, const char * tex } int main(int argc, char ** argv) { + // try to set locale for unicode characters in markdown + setlocale(LC_CTYPE, ".UTF-8"); + #if !defined(NDEBUG) fprintf(stderr, "warning: asserts enabled, performance may be affected\n"); #endif From 44c117f41ee01c5ac8fb86bba041f08d8b87b46d Mon Sep 17 00:00:00 2001 From: xaedes Date: Mon, 28 Aug 2023 21:51:47 +0200 Subject: [PATCH 084/147] train : mem usage and other improvements (#2439) * fix track_max_mem in forward_batch_wo_cache_flash_attn_train * remove unnecessary Adam(W) optimizer tensors. reduces optimizer memory overhead from 7*modelsize to 2*modelsize. additionally allows to optimize models with more than 2^31 parameters by replacing int with int64_t. bumps training checkpoint file version, but old checkpoints can still be read. new version with less tensors is saved. * add gradient clipping to AdamW * Fix reset of unused g->nodes and g->grads to NULL * implement gradient checkpointing for training reduces memory overhead from O(n_layer) to O(sqrt(n_layer)) as explained in readme of https://github.com/cybertronai/gradient-checkpointing * remove unused compute buffer 3 * add and use function ggml_build_backward_expand to avoid stack overflows with large maximum number of nodes GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); * change AdamW decay parameter to work like the torch AdamW decay parameter It is now relative to Adam learning rate `alpha*sched`. Before that it was relative to `sched` only. `alpha` being the maximum learning rate and `sched` being a scaling parameter in [0..1] * change default AdamW weight decay parameter used in training to 0.1 as used in nanoGPT * change default AdamW weight decay parameter defined in ggml to 0.0, making Adam default instead of AdamW btw: the default weight decay parameter for torch.optim.AdamW is 0.01 * bug fixes for cross entropy loss ggml_cross_entropy_loss: sums where not correctly added in workload of each thread ggml_cross_entropy_loss_back: simplify backward process, reducing numerical issues guard usage of exp f16 lookup in cross entropy by #define GGML_CROSS_ENTROPY_EXP_FP16 cross entropy loss is only used once during training, but it is quite sensitive to numerical errors introduced by exp-f16-lookup. so exp-f16-lookup for cross entropy loss is disabled by default, trading better gradients for very slightly worse runtime performance. * fix test-grad0 for cross_entropy_loss the second argument to cross_entropy_loss must sum up to 1 for each row * fix test-grad0 for soft_max dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) * improve finite differences of test-grad0 by using double instead of float * change cross_entropy_loss to output average over all rows this helps keeping the loss and gradients in a sane range * improve gradient checkpointing sqrt(n_layers) is only the best checkpoint step when mem size of checkpoints and mem size of layers are equal. since layers require more memory than the single-tensor-checkpoint we use, the optimal values are compute different: ``` given: n, u, v objective: minimize(a*u+b*v) where a*b=n, a>0, b>0 b=n/a minimize(a*u+v*n/a) diff(a*u+v*n/a, a) = u - (v*n/a)/a diff(a*u+v*n/a, a) == 0 u - (v*n/a)/a == 0 u == v*n/(a*a) u*a*a = v*n a*a = v*n/u a = sqrt(n*v/u) ``` this change results in more checkpoints, requiring less layers to store between checkpoints, overall improving memory usage. * disable gradient checkpointing debug output * llama : fix rope usage in train-text-from-scratch after ChatGLM change * add more training parameters: --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. --adam-min-alpha N Adam minimum learning rate alpha, usually 0.1 * alpha * replace memcpy with reshape operation so that the graph is not cut at the input this makes it possible to store other values into the input tensor and then simply recompute the graph without rebuilding it * remove unused function argument from get_example_targets_batch * measure and print total training time * add optimization callback to ggml_opt_resume_g this callback is called before each iteration with custom data and pointer to learning schedule parameter (only used in Adam(W)). can be used for dynamic learning schedule and setting input data for batches before each iteration * use optimization callback in training allows dynamic learning schedule and different batch data for each iteration without relying on low n_iter and high n_examples parameters reduces runtime by avoiding restart of optimization function and improves training convergence by providing a different batch for each iteration * add minimum number of tensor dimensions to apply weight decay (default 2) this allows to not apply weight decay to bias parameters * rename training parameter cos-decay-alpha to cos-decay-min and clarify that adam-min-alpha also applies to warmup * fix increase of model.train_samples and model.train_tokens now that each optimizer iteration gets its own batch we need to multiply by number of opt iterations * change sampling parameters for prediction after training to defaults of common.h and clarify what is context for prediction and what are generated tokens * tighten abs error bounds for cross_entropy_loss in test-grad0 * add conditional compilation of using F16 exp in flash attention uncomment `// #define GGML_FLASH_ATTN_EXP_FP16` to enable usage of f16 exp in flash attention * tighten abs error bounds for flash_attn in test-grad0 * tighten abs error bounds for sqrt in test-grad0 * remove out-commented vectorized code of opt_adam the vectorized code might be bit faster for low number of parameters, but it had a big memory usage overhead * ggml : update ggml_rms_norm_back with configurable eps * llama training : fix ggml_rms_norm_back calls to pass configurable eps * remove trailing whitespace * add train function using automatic gradient checkpointing backward pass and allocator * in train function replace add_inplace by regular add because using add_inplace seems to result in different gradients * don't use allocate hash_map on context because the context has no_alloc=True when using memory allocator resulting in NULL data pointers * correctly clone reshape and permute operations by also cloning tensor->nb values * fix variable name and add missing type cast * terminate recursive tensor cloning when reaching tensor without src tensors * correctly clone view tensors by setting data pointers without this the checkpointing would only work when being used together with memory allocator * fix variable names * swap arguments to commutative ops to be the same as in `forward_batch_wo_cache_flash_attn` * add input tensors as checkpoints so that recursive tensor cloning of gradient checkpointing terminates on input tensors * fix variable name and add missing boolean negation * make sure some tensors are not reallocated by inserting new temporary nodes depending on them: output and parameter gradient tensors need to be available at the end of the graph execution parameter gradient tensors also need to be available before the graph execution because they are set to zero before each optimizer iteration checkpoint tensors are allocated all together to reduce memory allocator fragmentation afterwards, in addition to the temporary nodes, we also need to reset the temporary leafs * fix ASSERT to work with zero layers * add training options whether to use allocator and/or unified training function * integrate unified training function which may use memory allocator the unified training function also supports arguments whether to use flash attention and/or gradient checkpointing * format name of cloned tensors with " (clone)" suffix * set names for tensors in unified train function for easier debugging * allocate graph on context using ggml_new_graph * remove handwritten training functions * remove unused training parameters "use_scratch" and "use_unified" * remove trailing whitespace * remove unused train params: mem_compute1_gb & mem_compute2_gb mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented) * remove unused forward_batch function * add debug asserts in ggml_allocr_alloc to some common pitfalls when using this function directly * only use ggml_allocr_alloc when tensor has NULL data and is no view * fix test when to create temporary backward graph temporary backward graph is only necessary when using checkpointing * fix memory "leak" in optimizers each iteration a new cplan with new memory for work data was allocated. now cplan creation only happens at the start of optimization, with each iteration reusing the cplan and its work data. * reverse order of for loop in ggml_build_backward_expand to save memory when using gradient checkpointing and allocator with this loop order gradient checkpointing with allocator on 16 layer model saves 13% memory; 2 layer memory it saves 2% memory. the computation results are the same * add missing lctx argument to get_example_targets_batch * implement llama model file saving using gguf checkpoint loading and saving disabled, to be replaced by loading and saving via gguf * implement loading/saving of checkpointing files using GGUF * bug fixes * add checkpoint file version for future compatibility * update readme with gguf filenames * save & load opt->just_initialized value * add first draft for checkpoint conversion script * add gguf arch and ftype * save opt parameter counter as uint64 * add gguf key and tensor names for optimizer and training * add layer_norm_rms_eps to checkpoint convert script * use same GGUF_GET_KEY macro as in llama.cpp * use norm_rms_eps, and rope parameters and command line options to set them * fix memory corruption bug in gguf ctx->kv and ctx->infos was reallocated using not-aligned realloc, but freed with aligned free. to fix this a GGML_ALIGNED_REALLOC was added, but there is no posix_memalign_realloc function. so on non-windows and non-mingw32 platforms we fall back to aligned malloc, followed by copying and freeing the old data. * add gguf example cmake file * bug fixes in tokenize_file * bug fixes in load_llama_model_gguf * bug fix: init model when no checkpoint was loaded * bug fix in read_tensor_by_name * bug fix in load_opt_context_gguf * avoid printing lots of spaced on the unusual case that loss gets nan * set name of tensors with empty name from what was read from gguf * remove trailing whitespace * print data checksums before saving and after loading to verify correctness * bug fixes for convert-train-checkpoint-to-gguf * temporarily add code to write old checkpoint files used to verify that old checkpoint files are correctly converted to gguf * bug fixes for convert-train-checkpoint-to-gguf.py loading checkpoints with opt_version=0 * remove code used to verify correctness of checkpoint file conversion * remove trailing whitespace * remove prediction related code use main for prediction, it is better optimized * update train-text-from-scratch README.md * fix non-windows GGML_ALIGNED_REALLOC * add missing blank line at end of file * remove GGML_ALIGNED_REALLOC and use normal malloc/realloc/free for gguf ctx->kv & ctx->infos * train : fix compile warnings --------- Co-authored-by: Georgi Gerganov --- common/common.cpp | 5 +- .../convert-llama2c-to-ggml.cpp | 1 - examples/gguf/CMakeLists.txt | 5 + examples/train-text-from-scratch/README.md | 14 +- .../convert-train-checkpoint-to-gguf.py | 492 +++ .../train-text-from-scratch.cpp | 3502 ++++++----------- ggml-alloc.c | 4 + ggml.c | 335 +- ggml.h | 29 +- llama.cpp | 9 +- tests/test-grad0.cpp | 52 +- 11 files changed, 1940 insertions(+), 2508 deletions(-) create mode 100644 examples/gguf/CMakeLists.txt create mode 100644 examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py diff --git a/common/common.cpp b/common/common.cpp index 4a0d43c13..90fe2e84e 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -15,6 +15,7 @@ #include #include #include +#include #if defined(__APPLE__) && defined(__MACH__) #include @@ -938,8 +939,8 @@ std::string get_sortable_timestamp() { const int64_t ns = std::chrono::duration_cast( current_time.time_since_epoch() % 1000000000).count(); - char timestamp_ns[10]; - snprintf(timestamp_ns, 11, "%09ld", ns); + char timestamp_ns[11]; + snprintf(timestamp_ns, 11, "%09" PRId64, ns); return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns); } diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index 51d90ea6a..e9e070b1f 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -681,7 +681,6 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod // for rms-att-weight int row_length = model->hparams.n_embd; - const auto & hparams = model->hparams; int n_ff = model->hparams.n_ff; for (uint32_t i = 0; i < model->hparams.n_layer; ++i){ diff --git a/examples/gguf/CMakeLists.txt b/examples/gguf/CMakeLists.txt new file mode 100644 index 000000000..7d1806af3 --- /dev/null +++ b/examples/gguf/CMakeLists.txt @@ -0,0 +1,5 @@ +set(TARGET gguf) +add_executable(${TARGET} gguf.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/train-text-from-scratch/README.md b/examples/train-text-from-scratch/README.md index 726ec47c0..f4ffcd987 100644 --- a/examples/train-text-from-scratch/README.md +++ b/examples/train-text-from-scratch/README.md @@ -8,15 +8,15 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s # train ./bin/train-text-from-scratch \ - --vocab-model ../models/ggml-vocab.bin \ + --vocab-model ../models/ggml-vocab-llama.gguf \ --ctx 64 --embd 256 --head 8 --layer 16 \ - --checkpoint-in chk-shakespeare-256x16.bin \ - --checkpoint-out chk-shakespeare-256x16.bin \ - --model-out ggml-shakespeare-256x16-f32.bin \ + --checkpoint-in chk-shakespeare-256x16.gguf \ + --checkpoint-out chk-shakespeare-256x16.gguf \ + --model-out ggml-shakespeare-256x16-f32.gguf \ --train-data "shakespeare.txt" \ - -t 6 -b 16 -n 32 --seed 1 --adam-iter 16 \ - --print-details-interval 0 --predict 16 --use-flash + -t 6 -b 16 --seed 1 --adam-iter 256 \ + --no-checkpointing # predict -./bin/main -m ggml-shakespeare-256x16-f32.bin +./bin/main -m ggml-shakespeare-256x16-f32.gguf ``` diff --git a/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py new file mode 100644 index 000000000..01b3ee92a --- /dev/null +++ b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py @@ -0,0 +1,492 @@ +#!/usr/bin/env python3 +# train-text-from-scratch checkpoint --> gguf conversion + +import argparse +import gguf +import os +import struct +import sys +import numpy as np +from pathlib import Path + +# gguf constants +LLM_KV_OPTIMIZER_TYPE = "optimizer.type" +LLM_KV_OPTIMIZER_TYPE_ADAM = "adam" +LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs" +LLM_KV_OPTIMIZER_FILE_VERSION = "optimizer.file_version" +LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT = "optimizer.convergence_past_count" +LLM_KV_OPTIMIZER_PARAMETER_COUNT = "optimizer.parameter_count" +LLM_KV_OPTIMIZER_ITERATION_COUNT = "optimizer.iteration_count" +LLM_KV_OPTIMIZER_JUST_INITIALIZED = "optimizer.just_initialized" +LLM_KV_OPTIMIZER_ADAM_BEST_LOSS = "optimizer.adam.best_loss" +LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS = "optimizer.adam.previous_loss" +LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT = "optimizer.adam.no_improvement_count" +LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count" +LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS = "optimizer.lbfgs.best_loss" +LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP = "optimizer.lbfgs.line_search_step" +LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J = "optimizer.lbfgs.line_search_j" +LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K = "optimizer.lbfgs.line_search_k" +LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END = "optimizer.lbfgs.line_search_end" +LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count" + +LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS = "optimizer.adam.first_moments" +LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS = "optimizer.adam.second_moments" +LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values" + +LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS = "optimizer.lbfgs.current_parameters" +LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters" +LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS = "optimizer.lbfgs.current_gradients" +LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS = "optimizer.lbfgs.previous_gradients" +LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION = "optimizer.lbfgs.search_direction" +LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES = "optimizer.lbfgs.past_loss_values" +LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA = "optimizer.lbfgs.memory_alpha" +LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS = "optimizer.lbfgs.memory_ys" +LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S = "optimizer.lbfgs.memory_s" +LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y = "optimizer.lbfgs.memory_y" + +LLM_KV_TRAINING_FILE_VERSION = "training.file_version" +LLM_KV_TRAINING_ITERATION_COUNT = "training.iteration_count" +LLM_KV_TRAINING_SAMPLE_COUNT = "training.sample_count" +LLM_KV_TRAINING_TOKEN_COUNT = "training.token_count" + +class Tensor: + def __init__(self, dtype='f', ne=None): + if ne is None: + ne = [] + self.dtype = dtype + self.ne = ne + self.nbytes = 0 + if self.dtype == 'f': + if len(self.ne) == 0: + self.nbytes = 0 + else: + self.nbytes = int(np.product(self.ne)) * 4 + else: + raise ValueError(f"Unhandled data type '{self.dtype}'") + + def load(self, data, offset): + nd = struct.unpack(' 0 else []) + + self.lbfgs_x = Tensor('f', [self.nx]) + self.lbfgs_xp = Tensor('f', [self.nx]) + self.lbfgs_g = Tensor('f', [self.nx]) + self.lbfgs_gp = Tensor('f', [self.nx]) + self.lbfgs_d = Tensor('f', [self.nx]) + self.lbfgs_pf = Tensor('f', [self.past] if self.past > 0 else []) + self.lbfgs_lmal = Tensor('f', [self.lbfgs_m]) + self.lbfgs_lmys = Tensor('f', [self.lbfgs_m]) + self.lbfgs_lms = Tensor('f', [self.nx, self.lbfgs_m]) + self.lbfgs_lmy = Tensor('f', [self.nx, self.lbfgs_m]) + + if self.type == 0: + # these tensors are stored, but we don't need their data + x = Tensor('f', [self.nx]) + g = Tensor('f', [self.nx]) + g2 = Tensor('f', [self.nx]) + mh = Tensor('f', [self.nx]) + vh = Tensor('f', [self.nx]) + + offset = x.load(data, offset) + offset = g.load(data, offset) + offset = g2.load(data, offset) + offset = self.adam_m.load(data, offset) + offset = self.adam_v.load(data, offset) + offset = mh.load(data, offset) + offset = vh.load(data, offset) + offset = self.adam_pf.load(data, offset) + + self.adam_fx_best = struct.unpack(' 0 else []) + + self.lbfgs_x = Tensor('f', [self.nx]) + self.lbfgs_xp = Tensor('f', [self.nx]) + self.lbfgs_g = Tensor('f', [self.nx]) + self.lbfgs_gp = Tensor('f', [self.nx]) + self.lbfgs_d = Tensor('f', [self.nx]) + self.lbfgs_pf = Tensor('f', [self.past] if self.past > 0 else []) + self.lbfgs_lmal = Tensor('f', [self.lbfgs_m]) + self.lbfgs_lmys = Tensor('f', [self.lbfgs_m]) + self.lbfgs_lms = Tensor('f', [self.nx, self.lbfgs_m]) + self.lbfgs_lmy = Tensor('f', [self.nx, self.lbfgs_m]) + + # forgot to save type in version 1: + # guess self.type from number of remaining bytes + size_type_0 = 12 + sum([t.max_storage_size() for t in + [self.adam_m, self.adam_v] + +([self.adam_pf] if (self.past > 0) else [])]) + size_type_1 = 24 + sum([t.max_storage_size() for t in + [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g, + self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf, + self.lbfgs_lmal, self.lbfgs_lmys, + self.lbfgs_lms, self.lbfgs_lmy] + +([self.lbfgs_pf] if (self.past > 0) else [])]) + # due to alignment padding the size might not by exact + # but the difference in size for both types is significant, + # so we can just use whichever is closest + remaining = len(data) - offset + if abs(remaining - size_type_0) < abs(remaining - size_type_1): + self.type = 0 + else: + self.type = 1 + + if self.type == 0: + offset = self.adam_m.load(data, offset) + offset = self.adam_v.load(data, offset) + offset = self.adam_pf.load(data,offset) + + self.adam_fx_best = struct.unpack(' 0: + self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES) + + elif self.type == 1: + gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS) + gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m) + gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best) + gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step) + gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j) + gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k) + gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end) + gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement) + + self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS) + self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS) + self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS) + self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS) + self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION) + if self.past > 0: + self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES) + self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA) + self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS) + self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S) + self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y) + else: + raise ValueError('Unknown optimizer type') + +class ModelParams: + def __init__(self): + pass + + def load(self, data, offset): + self.n_vocab = struct.unpack(' @@ -17,8 +18,6 @@ #pragma warning(disable: 4244 4267) // possible loss of data #endif -static const float rms_norm_eps = 1e-5f; - struct random_normal_distribution { std::mt19937 gen; std::normal_distribution rd; @@ -63,17 +62,6 @@ float frand_uniform(struct random_uniform_distribution * rnd) { return rnd->rd(rnd->gen); } -void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * graph, int n_threads) { - struct ggml_cplan plan = ggml_graph_plan(graph, n_threads); - - if (plan.work_size > 0) { - buf.resize(plan.work_size); - plan.work_data = buf.data(); - } - - ggml_graph_compute(graph, &plan); -} - struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) { float scale = 1.0f; // xavier switch (tensor->n_dims) { @@ -167,29 +155,20 @@ struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struc return tensor; } -struct llama_vocab { - using id = int32_t; - using token = std::string; - using ttype = llama_token_type; - - struct token_data { - token text; - float score; - ttype type; - }; - - std::unordered_map token_to_id; - std::vector id_to_token; -}; - struct my_llama_hparams { uint32_t n_vocab = 32000; - uint32_t n_ctx = 512; // this is provided as user input? + uint32_t n_ctx = 512; uint32_t n_embd = 4096; - uint32_t n_mult = 4; uint32_t n_head = 32; uint32_t n_layer = 32; uint32_t n_rot = 64; + uint32_t n_ff = 11008; + + // float f_norm_eps = 1e-5; // falcon + float f_norm_rms_eps = 1e-5; // llama + + float rope_freq_base = 10000.0f; + float rope_freq_scale = 1.0f; bool operator!=(const my_llama_hparams& other) const { return memcmp(this, &other, sizeof(my_llama_hparams)); @@ -215,17 +194,6 @@ struct my_llama_layer { struct ggml_tensor * w3; }; -struct my_llama_kv_cache { - struct ggml_context * ctx = NULL; - - struct ggml_tensor * k; - struct ggml_tensor * v; - - // llama_ctx_buffer buf; - - int n; // number of tokens currently in the cache -}; - struct my_llama_model { struct ggml_context * ctx = NULL; @@ -243,18 +211,91 @@ struct my_llama_model { uint32_t train_tokens = 0; }; -uint32_t get_n_ff(const struct my_llama_hparams* hparams) { - const uint32_t n_ff = ((2*(4*hparams->n_embd)/3 + hparams->n_mult - 1)/hparams->n_mult)*hparams->n_mult; - return n_ff; -} +// gguf constants +const char * LLM_KV_OPTIMIZER_TYPE = "optimizer.type"; +const char * LLM_KV_OPTIMIZER_TYPE_ADAM = "adam"; +const char * LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs"; +const char * LLM_KV_OPTIMIZER_FILE_VERSION = "optimizer.file_version"; +const char * LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT = "optimizer.convergence_past_count"; +const char * LLM_KV_OPTIMIZER_PARAMETER_COUNT = "optimizer.parameter_count"; +const char * LLM_KV_OPTIMIZER_ITERATION_COUNT = "optimizer.iteration_count"; +const char * LLM_KV_OPTIMIZER_JUST_INITIALIZED = "optimizer.just_initialized"; +const char * LLM_KV_OPTIMIZER_ADAM_BEST_LOSS = "optimizer.adam.best_loss"; +const char * LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS = "optimizer.adam.previous_loss"; +const char * LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT = "optimizer.adam.no_improvement_count"; +const char * LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count"; +const char * LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS = "optimizer.lbfgs.best_loss"; +const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP = "optimizer.lbfgs.line_search_step"; +const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J = "optimizer.lbfgs.line_search_j"; +const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K = "optimizer.lbfgs.line_search_k"; +const char * LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END = "optimizer.lbfgs.line_search_end"; +const char * LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count"; + +const char * LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS = "optimizer.adam.first_moments"; +const char * LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS = "optimizer.adam.second_moments"; +const char * LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values"; + +const char * LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS = "optimizer.lbfgs.current_parameters"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS = "optimizer.lbfgs.current_gradients"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS = "optimizer.lbfgs.previous_gradients"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION = "optimizer.lbfgs.search_direction"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES = "optimizer.lbfgs.past_loss_values"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA = "optimizer.lbfgs.memory_alpha"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS = "optimizer.lbfgs.memory_ys"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S = "optimizer.lbfgs.memory_s"; +const char * LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y = "optimizer.lbfgs.memory_y"; + +const char * LLM_KV_TRAINING_FILE_VERSION = "training.file_version"; +const char * LLM_KV_TRAINING_ITERATION_COUNT = "training.iteration_count"; +const char * LLM_KV_TRAINING_SAMPLE_COUNT = "training.sample_count"; +const char * LLM_KV_TRAINING_TOKEN_COUNT = "training.token_count"; + +// gguf constants (sync with gguf.py) + +const char * LLM_KV_GENERAL_ARCHITECTURE = "general.architecture"; +const char * LLM_KV_GENERAL_FILE_TYPE = "general.file_type"; + +const char * LLM_KV_CONTEXT_LENGTH = "%s.context_length"; +const char * LLM_KV_EMBEDDING_LENGTH = "%s.embedding_length"; +const char * LLM_KV_BLOCK_COUNT = "%s.block_count"; +const char * LLM_KV_FEED_FORWARD_LENGTH = "%s.feed_forward_length"; +const char * LLM_KV_ATTENTION_HEAD_COUNT = "%s.attention.head_count"; +const char * LLM_KV_ATTENTION_LAYERNORM_RMS_EPS = "%s.attention.layer_norm_rms_epsilon"; +const char * LLM_KV_ROPE_DIMENSION_COUNT = "%s.rope.dimension_count"; +const char * LLM_KV_ROPE_FREQ_BASE = "%s.rope.freq_base"; // TODO load in llama.cpp +const char * LLM_KV_ROPE_SCALE_LINEAR = "%s.rope.scale_linear"; + +const char * LLM_KV_TOKENIZER_MODEL = "tokenizer.ggml.model"; +const char * LLM_KV_TOKENIZER_LIST = "tokenizer.ggml.tokens"; +const char * LLM_KV_TOKENIZER_TOKEN_TYPE = "tokenizer.ggml.token_type"; +const char * LLM_KV_TOKENIZER_SCORES = "tokenizer.ggml.scores"; +const char * LLM_KV_TOKENIZER_MERGES = "tokenizer.ggml.merges"; +const char * LLM_KV_TOKENIZER_BOS_ID = "tokenizer.ggml.bos_token_id"; +const char * LLM_KV_TOKENIZER_EOS_ID = "tokenizer.ggml.eos_token_id"; +const char * LLM_KV_TOKENIZER_UNK_ID = "tokenizer.ggml.unknown_token_id"; +const char * LLM_KV_TOKENIZER_SEP_ID = "tokenizer.ggml.seperator_token_id"; +const char * LLM_KV_TOKENIZER_PAD_ID = "tokenizer.ggml.padding_token_id"; + +const char * LLM_TENSOR_TOKEN_EMBD = "token_embd"; +const char * LLM_TENSOR_OUTPUT_NORM = "output_norm"; +const char * LLM_TENSOR_OUTPUT = "output"; +const char * LLM_TENSOR_ATTN_NORM = "blk.%d.attn_norm"; +const char * LLM_TENSOR_ATTN_Q = "blk.%d.attn_q"; +const char * LLM_TENSOR_ATTN_K = "blk.%d.attn_k"; +const char * LLM_TENSOR_ATTN_V = "blk.%d.attn_v"; +const char * LLM_TENSOR_ATTN_OUT = "blk.%d.attn_output"; +const char * LLM_TENSOR_FFN_NORM = "blk.%d.ffn_norm"; +const char * LLM_TENSOR_FFN_GATE = "blk.%d.ffn_gate"; +const char * LLM_TENSOR_FFN_DOWN = "blk.%d.ffn_down"; +const char * LLM_TENSOR_FFN_UP = "blk.%d.ffn_up"; void print_params(struct my_llama_hparams * params) { printf("%s: n_vocab: %d\n", __func__, params->n_vocab); printf("%s: n_ctx: %d\n", __func__, params->n_ctx); printf("%s: n_embd: %d\n", __func__, params->n_embd); - printf("%s: n_mult: %d\n", __func__, params->n_mult); printf("%s: n_head: %d\n", __func__, params->n_head); - printf("%s: n_ff: %d\n", __func__, get_n_ff(params)); + printf("%s: n_ff: %d\n", __func__, params->n_ff); printf("%s: n_layer: %d\n", __func__, params->n_layer); printf("%s: n_rot: %d\n", __func__, params->n_rot); } @@ -265,8 +306,7 @@ void init_model(struct my_llama_model * model) { const uint32_t n_embd = hparams.n_embd; const uint32_t n_layer = hparams.n_layer; const uint32_t n_vocab = hparams.n_vocab; - - const uint32_t n_ff = get_n_ff(&hparams); + const uint32_t n_ff = hparams.n_ff; struct ggml_context * ctx = model->ctx; @@ -274,20 +314,31 @@ void init_model(struct my_llama_model * model) { model->train_samples = 0; model->train_tokens = 0; + std::vector tn_buf; + tn_buf.resize(GGML_MAX_NAME); + auto tn = [&tn_buf](const char * key) -> const char * { + snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key); + return tn_buf.data(); + }; + auto tni = [&tn_buf](const char * key, int bid) -> const char * { + snprintf(tn_buf.data(), tn_buf.size(), key, bid); + std::string s = tn_buf.data(); + snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str()); + return tn_buf.data(); + }; + model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab); model->norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); model->output = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab); - ggml_set_name(model->tok_embeddings, "tok_embeddings.weight"); - ggml_set_name(model->norm, "norm.weight"); - ggml_set_name(model->output, "output.weight"); + ggml_set_name(model->tok_embeddings, tn(LLM_TENSOR_TOKEN_EMBD)); + ggml_set_name(model->norm, tn(LLM_TENSOR_OUTPUT_NORM)); + ggml_set_name(model->output, tn(LLM_TENSOR_OUTPUT)); model->layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { auto & layer = model->layers[i]; - std::string layers_i = "layers." + std::to_string(i); - layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); @@ -301,18 +352,18 @@ void init_model(struct my_llama_model * model) { layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd); layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff); - ggml_set_name(layer.attention_norm, (layers_i + ".attention_norm.weight").c_str()); + ggml_set_name(layer.attention_norm, tni(LLM_TENSOR_ATTN_NORM, i)); - ggml_set_name(layer.wq, (layers_i + ".attention.wq.weight").c_str()); - ggml_set_name(layer.wk, (layers_i + ".attention.wk.weight").c_str()); - ggml_set_name(layer.wv, (layers_i + ".attention.wv.weight").c_str()); - ggml_set_name(layer.wo, (layers_i + ".attention.wo.weight").c_str()); + ggml_set_name(layer.wq, tni(LLM_TENSOR_ATTN_Q, i)); + ggml_set_name(layer.wk, tni(LLM_TENSOR_ATTN_K, i)); + ggml_set_name(layer.wv, tni(LLM_TENSOR_ATTN_V, i)); + ggml_set_name(layer.wo, tni(LLM_TENSOR_ATTN_OUT, i)); - ggml_set_name(layer.ffn_norm, (layers_i + ".ffn_norm.weight").c_str()); + ggml_set_name(layer.ffn_norm, tni(LLM_TENSOR_FFN_NORM, i)); - ggml_format_name(layer.w1, "%s.feed_forward.w1.weight", layers_i.c_str()); - ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str()); - ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str()); + ggml_set_name(layer.w1, tni(LLM_TENSOR_FFN_GATE, i)); + ggml_set_name(layer.w2, tni(LLM_TENSOR_FFN_DOWN, i)); + ggml_set_name(layer.w3, tni(LLM_TENSOR_FFN_UP, i)); } } @@ -371,267 +422,6 @@ void randomize_model(struct my_llama_model * model, int seed, float mean, float } } -bool init_kv_cache(struct my_llama_kv_cache* cache, struct my_llama_model * model, int n_batch) { - const auto & hparams = model->hparams; - - const uint32_t n_ctx = hparams.n_ctx; - const uint32_t n_embd = hparams.n_embd; - const uint32_t n_layer = hparams.n_layer; - - const int64_t n_mem = n_layer*n_ctx*n_batch; - const int64_t n_elements = n_embd*n_mem; - - // cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB); - - // struct ggml_init_params params; - // params.mem_size = cache.buf.size; - // params.mem_buffer = cache.buf.addr; - // params.no_alloc = false; - if (!cache->ctx) { - struct ggml_init_params params; - params.mem_size = 2u*n_elements*ggml_type_size(GGML_TYPE_F32) + 2u*1024*1024; - params.mem_buffer = NULL; - params.no_alloc = false; - - cache->ctx = ggml_init(params); - - if (!cache->ctx) { - fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__); - return false; - } - } - - cache->k = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements); - cache->v = ggml_new_tensor_1d(cache->ctx, GGML_TYPE_F32, n_elements); - - return true; -} - -struct ggml_tensor * forward( - struct my_llama_model * model, - struct my_llama_kv_cache * cache, - struct ggml_context * ctx0, - struct ggml_cgraph * gf, - struct ggml_tensor * tokens_input, - const int n_tokens, - const int n_past) { - - const int N = n_tokens; - - struct my_llama_kv_cache& kv_self = *cache; - const auto & hparams = model->hparams; - const int n_ctx = hparams.n_ctx; - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_head = hparams.n_head; - const int n_rot = hparams.n_rot; - - struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); - memcpy(tokens->data, tokens_input->data, N*ggml_element_size(tokens)); - - struct ggml_tensor * kc = kv_self.k; - struct ggml_tensor * vc = kv_self.v; - - // inpL shape [n_embd,N,1,1] - struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens); - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - struct ggml_tensor * cur; - - // lctx.use_buf(ctx0, 0); - - // norm - { - // cur shape [n_embd,N,1,1] - cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps); - - // cur = attention_norm*cur - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model->layers[il].attention_norm, cur), - cur); - } - - // self-attention - { - // compute Q and K and RoPE them - // wq shape [n_embd, n_embd, 1, 1] - // wk shape [n_embd, n_embd, 1, 1] - // Qcur shape [n_embd/n_head, n_head, N, 1] - // Kcur shape [n_embd/n_head, n_head, N, 1] - struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0); - struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0); - - // store key and value to memory - { - // compute the transposed [N, n_embd] V matrix - // wv shape [n_embd, n_embd, 1, 1] - // Vcur shape [n_embd, N, 1, 1] - struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wv, cur), n_embd, N))); - - // kv_self.k shape [n_embd * n_ctx * n_layer, 1] - // kv_self.v shape [n_embd * n_ctx * n_layer, 1] - // k shape [n_embd * N, 1] == kv_self.k[:,n_past:n_past+N,il,0] - // v shape [N, n_embd, 1, 1] == kv_self.v[:,n_past:n_past+N,il,0] - - /* { - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v)); - - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } //*/ - - kc = ggml_set_1d_inplace(ctx0, kc, ggml_reshape_1d(ctx0, Kcur, n_embd*N), (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); - vc = ggml_set_2d_inplace(ctx0, vc, Vcur, ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v)); - } - - // Qcur shape [n_embd/n_head, n_head, N, 1] - // Q shape [n_embd/n_head, N, n_head, 1] - struct ggml_tensor * Q = - ggml_permute(ctx0, - Qcur, - 0, 2, 1, 3); - - // kv_self.k shape [n_embd * n_ctx * n_layer, 1] - // K shape [n_embd/n_head, n_past + N, n_head, 1] - struct ggml_tensor * K = - ggml_permute(ctx0, - ggml_reshape_3d(ctx0, - ggml_view_1d(ctx0, kc, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kc)*n_embd), - n_embd/n_head, n_head, n_past + N), - 0, 2, 1, 3); - - // K * Q - // KQ shape [n_past + N, N, n_head, 1] - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - - // KQ_scaled = KQ / sqrt(n_embd/n_head) - // KQ_scaled shape [n_past + N, N, n_head, 1] - struct ggml_tensor * KQ_scaled = - ggml_scale(ctx0, - KQ, - ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head))); - - // KQ_masked = mask_past(KQ_scaled) - // KQ_masked shape [n_past + N, N, n_head, 1] - struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); - - // KQ = soft_max(KQ_masked) - // KQ_soft_max shape [n_past + N, N, n_head, 1] - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - - // split cached V into n_head heads - //// V shape [n_past + N, n_embd/n_head, n_head, 1] - // V shape [n_past + N, n_embd/n_head, n_head, 1] == kv_self.v[:,:(n_past+N),il,1] - struct ggml_tensor * V = - ggml_view_3d(ctx0, vc, - n_past + N, n_embd/n_head, n_head, - n_ctx*ggml_element_size(vc), - n_ctx*ggml_element_size(vc)*n_embd/n_head, - il*n_ctx*ggml_element_size(vc)*n_embd); - - // KQV shape [n_embd/n_head, N, n_head, 1] - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - // KQV_merged shape [n_embd/n_head, n_head, N, 1] - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - // KQV_merged shape - - // cur = KQV_merged.contiguous().view(n_embd, N) - // cur shape [n_embd,N,1,1] - cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N); - // cur = ggml_cpy(ctx0, - // KQV_merged, - // ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); - - // projection (no bias) - // cur shape [n_embd,N,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].wo, - cur); - } - - // lctx.use_buf(ctx0, 1); - - // inpFF shape [n_embd,N,1,1] - struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); - - // feed-forward network - { - // norm - { - // cur shape [n_embd,N,1,1] - cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps); - - // cur = ffn_norm*cur - // cur shape [n_embd,N,1,1] - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model->layers[il].ffn_norm, cur), - cur); - } - - // tmp shape [n_ff,N,1,1] - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model->layers[il].w3, - cur); - - // cur shape [n_ff,N,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].w1, - cur); - - // SILU activation - // cur shape [n_ff,N,1,1] - cur = ggml_silu(ctx0, cur); - - // cur shape [n_ff,N,1,1] - cur = ggml_mul(ctx0, cur, tmp); - - // cur shape [n_embd,N,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].w2, - cur); - } - - // cur shape [n_embd,N,1,1] - cur = ggml_add(ctx0, cur, inpFF); - - // input for next layer - // inpL shape [n_embd,N,1,1] - inpL = cur; - } - - // norm - { - - // inpL shape [n_embd,N,1,1] - inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps); - - // inpL = norm*inpL - // inpL shape [n_embd,N,1,1] - inpL = ggml_mul(ctx0, - ggml_repeat(ctx0, model->norm, inpL), - inpL); - - //embeddings = inpL; - } - - // lm_head - // inpL shape [n_vocab,N,1,1] - inpL = ggml_mul_mat(ctx0, model->output, inpL); - - // run the computation - ggml_build_forward_expand(gf, inpL); - - return inpL; -} - void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) { GGML_ASSERT(tensor->n_dims == 1); GGML_ASSERT(tensor->ne[0] == ne0); @@ -658,786 +448,222 @@ void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int6 GGML_ASSERT(tensor->ne[3] == ne3); } -struct ggml_tensor * forward_batch( - struct my_llama_model * model, - struct my_llama_kv_cache * cache, - struct ggml_context * ctx0, - struct ggml_cgraph * gf, - struct ggml_tensor * tokens_input, - const int n_tokens, - const int n_past, - const int n_batch) { - - const int N = n_tokens; - - struct my_llama_kv_cache& kv_self = *cache; - const auto & hparams = model->hparams; - const int n_ctx = hparams.n_ctx; - const int n_vocab = hparams.n_vocab; - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_head = hparams.n_head; - const int n_rot = hparams.n_rot; - const int n_ff = get_n_ff(&hparams); - - struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch); - memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch); - - struct ggml_tensor * kc = kv_self.k; - struct ggml_tensor * vc = kv_self.v; - - // inpL shape [n_embd,N*n_batch,1] - struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens); - assert_shape_2d(inpL, n_embd, N*n_batch); - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - struct ggml_tensor * cur; - - // lctx.use_buf(ctx0, 0); - - // norm - { - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps); - assert_shape_2d(cur, n_embd, N*n_batch); - - // cur = attention_norm*cur - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model->layers[il].attention_norm, cur), - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - // self-attention - { - // compute Q and K and RoPE them - // wq shape [n_embd, n_embd, 1, 1] - // wk shape [n_embd, n_embd, 1, 1] - // Qcur shape [n_embd/n_head, n_head, N, n_batch] - // Kcur shape [n_embd/n_head, n_head, N, n_batch] - struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0); - struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0); - assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch); - assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch); - - // store key and value to memory - { - // compute the transposed [N, n_embd] V matrix - // wv shape [n_embd, n_embd, 1, 1] - // Vcur shape [N, n_embd, n_batch, 1] - struct ggml_tensor * Vcur = ggml_cont(ctx0, - ggml_permute(ctx0, - ggml_reshape_3d(ctx0, - ggml_mul_mat(ctx0, - model->layers[il].wv, - cur), - n_embd, N, n_batch), - 1, 0, 2, 3)); - assert_shape_3d(Vcur, N, n_embd, n_batch); - - // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer] - // kv_self.v shape [n_ctx * n_embd * n_batch * n_layer] - // k shape [n_embd * N, n_batch] == kv_self.k[:,n_past:n_past+N,:,il] - // v shape [N, n_embd, n_batch, 1] == kv_self.v[:,n_past:n_past+N,:,il] - - /* { - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v)); - - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } //*/ - - kc = ggml_set_2d_inplace(ctx0, kc, - ggml_reshape_2d(ctx0, Kcur, n_embd*N, n_batch), - ggml_element_size(kc)*n_embd*n_ctx, - (ggml_element_size(kc)*n_embd)*(il*n_batch*n_ctx + n_past)); - vc = ggml_set_2d_inplace(ctx0, vc, - ggml_reshape_2d(ctx0, Vcur, N*n_embd, n_batch), - ggml_element_size(vc)*n_ctx*n_embd, - ggml_element_size(vc)*(n_past + il*n_embd*n_batch*n_ctx)); - - assert_shape_1d(kc, n_embd * n_ctx * n_batch * n_layer); - assert_shape_1d(vc, n_embd * n_ctx * n_batch * n_layer); - } - - // Qcur shape [n_embd/n_head, n_head, N, n_batch] - // Q shape [n_embd/n_head, N, n_head, n_batch] - struct ggml_tensor * Q = - ggml_permute(ctx0, - Qcur, - 0, 2, 1, 3); - assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch); - - // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer] - // K shape [n_embd/n_head, n_past + N, n_head, n_batch] - struct ggml_tensor * K = - ggml_permute(ctx0, - ggml_reshape_4d(ctx0, - ggml_view_3d(ctx0, - kc, - n_embd, - (n_past + N), - n_batch, - n_embd*ggml_element_size(kc), - n_ctx*n_embd*ggml_element_size(kc), - il*n_batch*n_ctx*n_embd*ggml_element_size(kc)), - n_embd/n_head, n_head, n_past + N, n_batch), - 0, 2, 1, 3); - assert_shape_4d(K, n_embd/n_head, n_past + N, n_head, n_batch); - - // K * Q - // KQ shape [n_past + N, N, n_head, n_batch] - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - assert_shape_4d(KQ, n_past + N, N, n_head, n_batch); - - // KQ_scaled = KQ / sqrt(n_embd/n_head) - // KQ_scaled shape [n_past + N, N, n_head, n_batch] - struct ggml_tensor * KQ_scaled = - ggml_scale_inplace(ctx0, - KQ, - ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head))); - assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch); - - // KQ_masked = mask_past(KQ_scaled) - // KQ_masked shape [n_past + N, N, n_head, n_batch] - struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); - assert_shape_4d(KQ_masked, n_past + N, N, n_head, n_batch); - - // KQ = soft_max(KQ_masked) - // KQ_soft_max shape [n_past + N, N, n_head, n_batch] - struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); - assert_shape_4d(KQ_soft_max, n_past + N, N, n_head, n_batch); - - // split cached V into n_head heads - // kv_self.v shape [n_ctx * n_embd * n_batch * n_layer] - // V shape [n_past + N, n_embd/n_head, n_head, n_batch] == kv_self.v[:(n_past+N),:,:,il] - struct ggml_tensor * V = - ggml_view_4d(ctx0, vc, - n_past + N, n_embd/n_head, n_head, n_batch, - ggml_element_size(vc)*n_ctx, - ggml_element_size(vc)*n_ctx*n_embd/n_head, - ggml_element_size(vc)*n_ctx*n_embd, - il*n_batch*n_ctx*n_embd*ggml_element_size(vc)); - assert_shape_4d(V, n_past + N, n_embd/n_head, n_head, n_batch); - - // KQV shape [n_embd/n_head, N, n_head, n_batch] - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - // KQV_merged shape [n_embd/n_head, n_head, N, n_batch] - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch); - // KQV_merged shape - - // cur = KQV_merged.contiguous().view(n_embd, N) - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch); - assert_shape_2d(cur, n_embd, N*n_batch); - // cur = ggml_cpy(ctx0, - // KQV_merged, - // ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); - - // projection (no bias) - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].wo, - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - // lctx.use_buf(ctx0, 1); - - // inpFF shape [n_embd,N*n_batch,1,1] - struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA); - assert_shape_2d(inpFF, n_embd, N*n_batch); - - // feed-forward network - { - // norm - { - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps); - assert_shape_2d(cur, n_embd, N*n_batch); - - // cur = ffn_norm*cur - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model->layers[il].ffn_norm, cur), - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - // tmp shape [n_ff,N*n_batch,1,1] - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model->layers[il].w3, - cur); - assert_shape_2d(tmp, n_ff, N*n_batch); - - // cur shape [n_ff,N*n_batch,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].w1, - cur); - assert_shape_2d(cur, n_ff, N*n_batch); - - // SILU activation - // cur shape [n_ff,N*n_batch,1,1] - cur = ggml_silu(ctx0, cur); - assert_shape_2d(cur, n_ff, N*n_batch); - - // cur shape [n_ff,N*n_batch,1,1] - cur = ggml_mul(ctx0, cur, tmp); - assert_shape_2d(cur, n_ff, N*n_batch); - - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].w2, - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_add_inplace(ctx0, cur, inpFF); - assert_shape_2d(cur, n_embd, N*n_batch); - - // input for next layer - // inpL shape [n_embd,N*n_batch,1,1] - inpL = cur; - assert_shape_2d(inpL, n_embd, N*n_batch); - } - - // norm - { - - // inpL shape [n_embd,N*n_batch,1,1] - inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps); - assert_shape_2d(inpL, n_embd, N*n_batch); - - // inpL = norm*inpL - // inpL shape [n_embd,N*n_batch,1,1] - inpL = ggml_mul(ctx0, - ggml_repeat(ctx0, model->norm, inpL), - inpL); - - assert_shape_2d(inpL, n_embd, N*n_batch); - - //embeddings = inpL; - } - - // lm_head - // inpL shape [n_vocab,N*n_batch,1,1] - inpL = ggml_mul_mat(ctx0, model->output, inpL); - assert_shape_2d(inpL, n_vocab, N*n_batch); - - { - // inpL shape [n_vocab,N,n_batch,1] - inpL = ggml_reshape_3d(ctx0, - inpL, - n_vocab, N, n_batch); - assert_shape_3d(inpL, n_vocab, N, n_batch); - } - - // run the computation - ggml_build_forward_expand(gf, inpL); - - return inpL; +static size_t hash(void * p) { + return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE; } -struct ggml_tensor * forward_batch_wo_cache( - struct my_llama_model * model, - struct ggml_context * ctx0, - struct ggml_cgraph * gf, - struct ggml_tensor * tokens_input, - const int n_tokens, - const int n_batch) { +static size_t hash_find(void * hash_table[], void * p) { + size_t h = hash(p); - const int n_past = 0; - const int N = n_tokens; - - const auto & hparams = model->hparams; - //const int n_ctx = hparams.n_ctx; - const int n_vocab = hparams.n_vocab; - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_head = hparams.n_head; - const int n_rot = hparams.n_rot; - const int n_ff = get_n_ff(&hparams); - - struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch); - memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch); - - // inpL shape [n_embd,N*n_batch,1] - struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens); - assert_shape_2d(inpL, n_embd, N*n_batch); - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - struct ggml_tensor * cur; - - // lctx.use_buf(ctx0, 0); - - // norm - { - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps); - assert_shape_2d(cur, n_embd, N*n_batch); - - // cur = attention_norm*cur - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model->layers[il].attention_norm, cur), - cur); - assert_shape_2d(cur, n_embd, N*n_batch); + // linear probing + size_t i = h; + while (hash_table[i] != NULL && hash_table[i] != p) { + i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE; + if (i == h) { + // visited all hash table entries -> not found + return GGML_GRAPH_HASHTABLE_SIZE; } - - // self-attention - { - // compute Q and K and RoPE them - // wq shape [n_embd, n_embd, 1, 1] - // wk shape [n_embd, n_embd, 1, 1] - // Qcur shape [n_embd/n_head, n_head, N, n_batch] - // Kcur shape [n_embd/n_head, n_head, N, n_batch] - struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0); - struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0); - assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch); - assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch); - - // Vcur shape [N, n_batch, n_embd/n_head, n_head] - struct ggml_tensor * Vcur = ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, cur, model->layers[il].wv), N, n_batch, n_embd/n_head, n_head); - assert_shape_4d(Vcur, N, n_batch, n_embd/n_head, n_head); - - // Qcur shape [n_embd/n_head, n_head, N, n_batch] - // Q shape [n_embd/n_head, N, n_head, n_batch] - struct ggml_tensor * Q = - ggml_permute(ctx0, - Qcur, - 0, 2, 1, 3); - assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch); - - // kv_self.k shape [n_embd * n_ctx * n_batch * n_layer] - // K shape [n_embd/n_head, N, n_head, n_batch] - struct ggml_tensor * K = - ggml_permute(ctx0, - Kcur, - 0, 2, 1, 3); - assert_shape_4d(K, n_embd/n_head, N, n_head, n_batch); - - // K * Q - // KQ shape [N, N, n_head, n_batch] - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - assert_shape_4d(KQ, N, N, n_head, n_batch); - - // KQ_scaled = KQ / sqrt(n_embd/n_head) - // KQ_scaled shape [N, N, n_head, n_batch] - struct ggml_tensor * KQ_scaled = - ggml_scale_inplace(ctx0, - KQ, - ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head))); - assert_shape_4d(KQ_scaled, N, N, n_head, n_batch); - - // KQ_masked = mask_past(KQ_scaled) - // KQ_masked shape [N, N, n_head, n_batch] - struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past); - assert_shape_4d(KQ_masked, N, N, n_head, n_batch); - - // KQ = soft_max(KQ_masked) - // KQ_soft_max shape [N, N, n_head, n_batch] - struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); - assert_shape_4d(KQ_soft_max, N, N, n_head, n_batch); - - // Vcur shape [N, n_batch, n_embd/n_head, n_head] - // V shape [N, n_embd/n_head, n_head, n_batch] - struct ggml_tensor * V = - ggml_permute(ctx0, - Vcur, - 0, 3, 1, 2); - assert_shape_4d(V, N, n_embd/n_head, n_head, n_batch); - - // KQV shape [n_embd/n_head, N, n_head, n_batch] - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - // KQV_merged shape [n_embd/n_head, n_head, N, n_batch] - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch); - // KQV_merged shape - - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch); - assert_shape_2d(cur, n_embd, N*n_batch); - - // projection (no bias) - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].wo, - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - // lctx.use_buf(ctx0, 1); - - // inpFF shape [n_embd,N*n_batch,1,1] - struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA); - assert_shape_2d(inpFF, n_embd, N*n_batch); - - // feed-forward network - { - // norm - { - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps); - assert_shape_2d(cur, n_embd, N*n_batch); - - // cur = ffn_norm*cur - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model->layers[il].ffn_norm, cur), - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - // tmp shape [n_ff,N*n_batch,1,1] - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model->layers[il].w3, - cur); - assert_shape_2d(tmp, n_ff, N*n_batch); - - // cur shape [n_ff,N*n_batch,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].w1, - cur); - assert_shape_2d(cur, n_ff, N*n_batch); - - // SILU activation - // cur shape [n_ff,N*n_batch,1,1] - cur = ggml_silu(ctx0, cur); - assert_shape_2d(cur, n_ff, N*n_batch); - - // cur shape [n_ff,N*n_batch,1,1] - cur = ggml_mul(ctx0, cur, tmp); - assert_shape_2d(cur, n_ff, N*n_batch); - - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_mul_mat(ctx0, - model->layers[il].w2, - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - // cur shape [n_embd,N*n_batch,1,1] - cur = ggml_add_inplace(ctx0, cur, inpFF); - assert_shape_2d(cur, n_embd, N*n_batch); - - // input for next layer - // inpL shape [n_embd,N*n_batch,1,1] - inpL = cur; - assert_shape_2d(inpL, n_embd, N*n_batch); } - - // norm - { - - // inpL shape [n_embd,N*n_batch,1,1] - inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps); - assert_shape_2d(inpL, n_embd, N*n_batch); - - // inpL = norm*inpL - // inpL shape [n_embd,N*n_batch,1,1] - inpL = ggml_mul(ctx0, - ggml_repeat(ctx0, model->norm, inpL), - inpL); - - assert_shape_2d(inpL, n_embd, N*n_batch); - - //embeddings = inpL; - } - - // lm_head - // inpL shape [n_vocab,N*n_batch,1,1] - inpL = ggml_mul_mat(ctx0, model->output, inpL); - assert_shape_2d(inpL, n_vocab, N*n_batch); - - { - // inpL shape [n_vocab,N,n_batch,1] - inpL = ggml_reshape_3d(ctx0, - inpL, - n_vocab, N, n_batch); - assert_shape_3d(inpL, n_vocab, N, n_batch); - } - - // run the computation - ggml_build_forward_expand(gf, inpL); - - return inpL; + return i; } -struct ggml_tensor * forward_batch_wo_cache_flash_attn( - struct my_llama_model * model, - struct ggml_context * ctx0, - struct ggml_cgraph * gf, - struct ggml_tensor * tokens_input, - const int n_tokens, - const int n_batch) { +static bool hash_insert(void * hash_table[], void * p) { + //size_t h = hash(p); + size_t i = hash_find(hash_table, p); - const int n_past = 0; - const int N = n_tokens; + GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full - const auto & hparams = model->hparams; - //const int n_ctx = hparams.n_ctx; - const int n_vocab = hparams.n_vocab; - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_head = hparams.n_head; - const int n_rot = hparams.n_rot; - const int n_ff = get_n_ff(&hparams); - - struct ggml_tensor * tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch); - memcpy(tokens->data, tokens_input->data, ggml_element_size(tokens)*N*n_batch); - - struct ggml_tensor * inpL = ggml_get_rows(ctx0, model->tok_embeddings, tokens); - assert_shape_2d(inpL, n_embd, N*n_batch); - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - struct ggml_tensor * cur; - - // norm - { - cur = ggml_rms_norm(ctx0, inpL, rms_norm_eps); - assert_shape_2d(cur, n_embd, N*n_batch); - - // cur = attention_norm*cur - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model->layers[il].attention_norm, cur), - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - // self-attention - { - // compute Q and K and RoPE them - // wq shape [n_embd, n_embd, 1, 1] - // wk shape [n_embd, n_embd, 1, 1] - struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wq, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0); - struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, model->layers[il].wk, cur), n_embd/n_head, n_head, N, n_batch), n_past, n_rot, 0, 0); - assert_shape_4d(Qcur, n_embd/n_head, n_head, N, n_batch); - assert_shape_4d(Kcur, n_embd/n_head, n_head, N, n_batch); - - struct ggml_tensor * Vcur = ggml_reshape_4d(ctx0, ggml_mul_mat(ctx0, cur, model->layers[il].wv), N, n_batch, n_embd/n_head, n_head); - assert_shape_4d(Vcur, N, n_batch, n_embd/n_head, n_head); - - struct ggml_tensor * Q = - ggml_permute(ctx0, - Qcur, - 0, 2, 1, 3); - assert_shape_4d(Q, n_embd/n_head, N, n_head, n_batch); - - struct ggml_tensor * K = - ggml_permute(ctx0, - Kcur, - 0, 2, 1, 3); - assert_shape_4d(K, n_embd/n_head, N, n_head, n_batch); - - struct ggml_tensor * V = - ggml_permute(ctx0, - Vcur, - 0, 3, 1, 2); - assert_shape_4d(V, N, n_embd/n_head, n_head, n_batch); - - bool masked = true; - struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, masked); - assert_shape_4d(KQV, n_embd/n_head, N, n_head, n_batch); - - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - assert_shape_4d(KQV_merged, n_embd/n_head, n_head, N, n_batch); - cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, KQV_merged), n_embd, N*n_batch); - assert_shape_2d(cur, n_embd, N*n_batch); - - // projection (no bias) - cur = ggml_mul_mat(ctx0, - model->layers[il].wo, - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - struct ggml_tensor * inpFF = ggml_add_inplace(ctx0, cur, inpSA); - assert_shape_2d(inpFF, n_embd, N*n_batch); - - // feed-forward network - { - // norm - { - cur = ggml_rms_norm(ctx0, inpFF, rms_norm_eps); - assert_shape_2d(cur, n_embd, N*n_batch); - - // cur = ffn_norm*cur - cur = ggml_mul(ctx0, - ggml_repeat(ctx0, model->layers[il].ffn_norm, cur), - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model->layers[il].w3, - cur); - assert_shape_2d(tmp, n_ff, N*n_batch); - - cur = ggml_mul_mat(ctx0, - model->layers[il].w1, - cur); - assert_shape_2d(cur, n_ff, N*n_batch); - - // SILU activation - cur = ggml_silu(ctx0, cur); - assert_shape_2d(cur, n_ff, N*n_batch); - - cur = ggml_mul(ctx0, cur, tmp); - assert_shape_2d(cur, n_ff, N*n_batch); - - cur = ggml_mul_mat(ctx0, - model->layers[il].w2, - cur); - assert_shape_2d(cur, n_embd, N*n_batch); - } - - cur = ggml_add_inplace(ctx0, cur, inpFF); - assert_shape_2d(cur, n_embd, N*n_batch); - - // input for next layer - inpL = cur; - assert_shape_2d(inpL, n_embd, N*n_batch); + if (hash_table[i] == p) { + return true; } - // norm - { - - inpL = ggml_rms_norm(ctx0, inpL, rms_norm_eps); - assert_shape_2d(inpL, n_embd, N*n_batch); - - // inpL = norm*inpL - inpL = ggml_mul(ctx0, - ggml_repeat(ctx0, model->norm, inpL), - inpL); - - assert_shape_2d(inpL, n_embd, N*n_batch); - } - - // lm_head - inpL = ggml_mul_mat(ctx0, model->output, inpL); - assert_shape_2d(inpL, n_vocab, N*n_batch); - - { - inpL = ggml_reshape_3d(ctx0, - inpL, - n_vocab, N, n_batch); - assert_shape_3d(inpL, n_vocab, N, n_batch); - } - - // run the computation - ggml_build_forward_expand(gf, inpL); - - return inpL; + // insert + GGML_ASSERT(hash_table[i] == NULL); + hash_table[i] = p; + return false; } -// expand the graph nodes without creating leafs. -struct ggml_tensor * expand(struct ggml_cgraph * g, struct ggml_tensor * t) { - // check if already visited - for (int i = 0; i < g->n_nodes; i++) { - if (g->nodes[i] == t) { - return t; - } - } - - for (int i = 0; i < g->n_leafs; i++) { - if (g->leafs[i] == t) { - return t; - } - } - - for (int i = 0; i < GGML_MAX_SRC; ++i) { - if (t->src[i]) { - expand(g, t->src[i]); - } - } - - GGML_ASSERT(g->n_nodes < GGML_MAX_NODES); - - if (strlen(t->name) == 0) { - snprintf(t->name, sizeof(t->name), "node_%d", g->n_nodes); - } - - g->nodes[g->n_nodes] = t; - g->grads[g->n_nodes] = t->grad; - g->n_nodes++; - return t; +static bool hash_contains(void * hash_table[], void * p) { + size_t i = hash_find(hash_table, p); + return (i < GGML_GRAPH_HASHTABLE_SIZE) && (hash_table[i] == p); } -void graph_set_leafs_grads(struct ggml_cgraph * g) { - // moves leaf nodes to g->leafs. - // i.e. g->n_nodes might change. - int n_nodes = 0; - for (int i = 0; i < g->n_nodes; ++i) { - struct ggml_tensor * node = g->nodes[i]; - const bool is_leaf = node->op == GGML_OP_NONE && node->grad == NULL; - if (is_leaf) { - GGML_ASSERT(g->n_leafs < GGML_MAX_NODES); +struct hash_map { + void * keys[GGML_GRAPH_HASHTABLE_SIZE]; + void * vals[GGML_GRAPH_HASHTABLE_SIZE]; +}; +//static const size_t HASH_MAP_SIZE = sizeof(struct hash_map); - if (strlen(node->name) == 0) { - snprintf(node->name, sizeof(node->name), "leaf_%d", g->n_leafs); - } - - g->leafs[g->n_leafs] = node; - g->n_leafs++; - } else { - GGML_ASSERT(n_nodes < GGML_MAX_NODES); - - if (strlen(node->name) == 0) { - snprintf(node->name, sizeof(node->name), "node_%d", n_nodes); - } - - g->nodes[n_nodes] = node; - g->grads[n_nodes] = node->grad; - n_nodes++; - } +struct hash_map * new_hash_map() { + struct hash_map * result = new struct hash_map; + for (int i=0; ikeys[i] = NULL; + result->vals[i] = NULL; } - for (int i=n_nodes; i < g->n_nodes; ++i) { - g->nodes[n_nodes] = NULL; - g->grads[n_nodes] = NULL; - } - g->n_nodes = n_nodes; + return result; +}; + +void free_hash_map(struct hash_map * map) { + delete map; } -struct ggml_tensor * forward_batch_wo_cache_flash_attn_train( - struct my_llama_model * model, - struct ggml_context * ctx0, +static bool ggml_is_view(struct ggml_tensor * t) { + return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE || + t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY; +} + +static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) { + switch (t->op) { + case GGML_OP_PERMUTE: + case GGML_OP_RESHAPE: + case GGML_OP_TRANSPOSE: + case GGML_OP_VIEW: + return t->src[0]; + case GGML_OP_CPY: + return t->src[1]; + default: + return NULL; + } +} + +static struct ggml_tensor * get_view_source(struct ggml_tensor * t) { + struct ggml_tensor * parent = t; + do { + parent = get_view_parent(parent); + } while (ggml_is_view(parent)); + return parent; +} + +struct ggml_tensor * ggml_recompute_graph_node( + struct ggml_context * ctx, + struct ggml_cgraph * graph, + struct hash_map * replacements, + struct ggml_tensor * node) { + + if (node == NULL) { + return NULL; + } + + if (node->is_param) { + return node; + } + + if (!hash_contains(graph->visited_hash_table, node)) { + return node; + } + + int count_children = 0; + for (int k = 0; k < GGML_MAX_SRC; ++k) { + if (node->src[k]) { + ++count_children; + } + } + + if (count_children == 0) { + return node; + } + + size_t i = hash_find(replacements->keys, node); + GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full + if (replacements->keys[i] == node) { + return (struct ggml_tensor *) replacements->vals[i]; + } + + struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, node->n_dims, node->ne); + + // insert clone into replacements + GGML_ASSERT(replacements->keys[i] == NULL); // assert that we don't overwrite + replacements->keys[i] = node; + replacements->vals[i] = clone; + + clone->op = node->op; + clone->grad = node->grad; + clone->is_param = node->is_param; + clone->extra = node->extra; + for (int k = 0; k < GGML_MAX_DIMS; ++k) { + clone->nb[k] = node->nb[k]; + } + for (int k = 0; k < GGML_MAX_SRC; ++k) { + clone->src[k] = ggml_recompute_graph_node(ctx, graph, replacements, node->src[k]); + } + if (ggml_is_view(clone)) { + struct ggml_tensor * source = get_view_source(clone); + GGML_ASSERT(source != NULL); + clone->data = source->data; + } + + GGML_ASSERT(sizeof(node->op_params) == sizeof(int32_t) * (GGML_MAX_OP_PARAMS / sizeof(int32_t))); + GGML_ASSERT(sizeof(node->name) == GGML_MAX_NAME); + memcpy(clone->op_params, node->op_params, sizeof(node->op_params)); + ggml_format_name(clone, "%s (clone)", ggml_get_name(node)); + + return clone; +}; + +void ggml_build_backward_gradient_checkpointing( + struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, + struct ggml_cgraph * gb_tmp, + struct ggml_tensor * * checkpoints, + int n_checkpoints) { + *gb_tmp = *gf; + ggml_build_backward_expand(ctx, gf, gb_tmp, true); + + if (n_checkpoints <= 0) { + *gb = *gb_tmp; + return; + } + + struct hash_map * replacements = new_hash_map(); + + // insert checkpoints in replacements + for (int i = 0; i < n_checkpoints; ++i) { + size_t k = hash_find(replacements->keys, checkpoints[i]); + GGML_ASSERT(k < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full + GGML_ASSERT(replacements->keys[k] == NULL); // assert that we don't overwrite + replacements->keys[k] = checkpoints[i]; + replacements->vals[k] = checkpoints[i]; + } + + *gb = *gf; + // rewrite gb_tmp->nodes[gf->n_nodes:gb_tmp->n_nodes], + // replacing references to gb_tmp->nodes[0:gf->n_nodes] ( == gf->nodes[0:gf->n_nodes]), + // by recomputing them from checkpoints + for (int i = gf->n_nodes; in_nodes; ++i) { + struct ggml_tensor * node = gb_tmp->nodes[i]; + for (int k = 0; k < GGML_MAX_SRC; ++k) { + // insert new tensors recomputing src, reusing already made replacements, + // remember replacements: remember new tensors with mapping from corresponding gf nodes + // recurse for input tensors, + // unless (i.e. terminating when) input tensors are checkpoints + node->src[k] = ggml_recompute_graph_node(ctx, gf, replacements, node->src[k]); + } + // insert rewritten backward node with replacements made into resulting backward graph gb + ggml_build_forward_expand(gb, node); + } + + free_hash_map(replacements); +} + +struct ggml_tensor * llama_build_train_graphs( + struct my_llama_model * model, + struct ggml_allocr * alloc, + struct ggml_context * ctx, + struct ggml_cgraph * gf, + struct ggml_cgraph * gb, + struct ggml_cgraph * gb_tmp, struct ggml_tensor * * logits, struct ggml_tensor * tokens_input, struct ggml_tensor * targets, - void * compute_buf_0, - void * compute_buf_1, - size_t size_buf_0, - size_t size_buf_1, const int n_tokens, - const int n_batch) { - - ggml_set_scratch(ctx0, { 0, 0, nullptr, }); + const int n_batch, + const bool enable_flash_attn, + const bool enable_checkpointing) { + ggml_set_scratch(ctx, { 0, 0, nullptr, }); const int n_past = 0; const int N = n_tokens; - - gf->n_nodes = 0; - gf->n_leafs = 0; - gf->perf_runs = 0; - gf->perf_cycles = 0; - gf->perf_time_us = 0; - const auto & hparams = model->hparams; const int n_ctx = hparams.n_ctx; const int n_vocab = hparams.n_vocab; @@ -1445,476 +671,162 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train( const int n_layer = hparams.n_layer; const int n_head = hparams.n_head; const int n_rot = hparams.n_rot; - const int n_ff = get_n_ff(&hparams); - const int rope_mode = 0; + const int n_ff = hparams.n_ff; + const float f_norm_rms_eps = hparams.f_norm_rms_eps; + const float rope_freq_base = hparams.rope_freq_base; + const float rope_freq_scale = hparams.rope_freq_scale; - int last_buf = -1; - size_t buf_offs[2] = { 0, 0 }; - size_t buf_size[2] = { size_buf_0, - size_buf_1 }; - void * buf_data[2] = { compute_buf_0, - compute_buf_1 }; - auto use_buf = [ctx0, &last_buf, &buf_offs, &buf_size, &buf_data] (int buf) { - size_t last_offs = 0; - last_offs = ggml_set_scratch(ctx0, { 0, 0, nullptr, }); - if (last_buf >= 0) { - buf_offs[last_buf] = last_offs; - } - if (buf >= 0) { - size_t offs = buf_offs[buf]; - size_t size = buf_size[buf]; - void * data = buf_data[buf]; - ggml_set_scratch(ctx0, { offs, size, data, }); - } - last_buf = buf; - }; - - bool track_max_mem = false; - size_t buf_maxs[2] = { 0, 0 }; - - auto clr_buf = [ctx0, &last_buf, &buf_offs, &buf_size, &buf_data, &buf_maxs, track_max_mem] (int buf) { - if (buf < 0) return; - if (track_max_mem) { - size_t last_offs = 0; - last_offs = ggml_set_scratch(ctx0, { 0, 0, nullptr, }); - if (last_buf >= 0) { - buf_offs[last_buf] = last_offs; - buf_maxs[last_buf] = std::max(buf_maxs[last_buf], buf_offs[last_buf]); - } - } - buf_offs[buf] = 0; - if (track_max_mem && last_buf >= 0) { - size_t offs = buf_offs[last_buf]; - size_t size = buf_size[last_buf]; - void * data = buf_data[last_buf]; - ggml_set_scratch(ctx0, { offs, size, data, }); + auto set_name = [](struct ggml_tensor * t, const char * n) { + ggml_set_name(t, n); + if (t->grad) { + ggml_format_name(t->grad, "%s->grad", n); } }; + // rope has so much parameters that we make a custom function for it + auto rope = [ctx, n_rot, n_ctx, rope_freq_base, rope_freq_scale] + (struct ggml_tensor * t) -> struct ggml_tensor * { + // not capturing these, to silcence warnings + const int n_past = 0; + const int rope_mode = 0; - auto view__q = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * { - int64_t ne0 = n_embd/n_head; - int64_t ne1 = N; - int64_t ne2 = n_head; - int64_t ne3 = n_batch; - size_t nb0 = ggml_element_size(t); - size_t nb1 = nb0*ne0; - size_t nb2 = nb1*ne1; - size_t nb3 = nb2*ne2; - size_t offset = 0; - return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset); + return ggml_rope_custom(ctx, + t, n_past, n_rot, rope_mode, n_ctx, + rope_freq_base, rope_freq_scale); }; - auto view__k = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * { - int64_t ne0 = n_embd/n_head; - int64_t ne1 = N; - int64_t ne2 = n_head; - int64_t ne3 = n_batch; - size_t nb0 = ggml_element_size(t); - size_t nb1 = nb0*ne0; - size_t nb2 = nb1*ne1; - size_t nb3 = nb2*ne2; - size_t offset = nb3*ne3; - return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset); - }; + set_name(tokens_input, "tokens_input"); + set_name(targets, "targets"); - auto view__v = [ctx0, n_embd, n_head, N, n_batch] (struct ggml_tensor * t) -> struct ggml_tensor * { - int64_t ne0 = N; - int64_t ne1 = n_embd/n_head; - int64_t ne2 = n_head; - int64_t ne3 = n_batch; - size_t nb0 = ggml_element_size(t); - size_t nb1 = nb0*ne0; - size_t nb2 = nb1*ne1; - size_t nb3 = nb2*ne2; - size_t offset = 2*nb3*ne3; - return ggml_view_4d(ctx0, t, ne0, ne1, ne2, ne3, nb1, nb2, nb3, offset); - }; - - auto add_or_set = [ctx0] (struct ggml_tensor * a, struct ggml_tensor * b) -> struct ggml_tensor * { - if (a == NULL) { - return b; - } else { - return ggml_add_inplace(ctx0, a, b); - } - }; - - use_buf(-1); - - model->tok_embeddings->grad = NULL; - model->norm->grad = NULL; - model->output->grad = NULL; - - for (int il = 0; il < n_layer; ++il) { - struct my_llama_layer & layer = model->layers[il]; - layer.attention_norm->grad = NULL; - layer.wq->grad = NULL; - layer.wk->grad = NULL; - layer.wv->grad = NULL; - layer.wo->grad = NULL; - layer.ffn_norm->grad = NULL; - layer.w1->grad = NULL; - layer.w2->grad = NULL; - layer.w3->grad = NULL; - } - - clr_buf(0); - clr_buf(1); - - use_buf(-1); - - struct ggml_tensor * t00 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N*n_batch); assert_shape_1d(t00, N*n_batch); - memcpy(t00->data, tokens_input->data, ggml_element_size(t00)*N*n_batch); - - use_buf(-1); - - struct ggml_tensor * t01 = expand(gf, ggml_get_rows(ctx0, model->tok_embeddings, t00)); assert_shape_2d(t01, n_embd, N*n_batch); - - // need to remember these for the backward pass - std::vector t02L; t02L.resize(n_layer, NULL); - std::vector t03L; t03L.resize(n_layer, NULL); - std::vector t04L; t04L.resize(n_layer, NULL); - std::vector t05L; t05L.resize(n_layer, NULL); - std::vector t06L; t06L.resize(n_layer, NULL); - std::vector t07L; t07L.resize(n_layer, NULL); - std::vector t08L; t08L.resize(n_layer, NULL); - std::vector t09L; t09L.resize(n_layer, NULL); - std::vector t10L; t10L.resize(n_layer, NULL); - std::vector t11L; t11L.resize(n_layer, NULL); - std::vector t12L; t12L.resize(n_layer, NULL); - std::vector t13L; t13L.resize(n_layer, NULL); - std::vector t14L; t14L.resize(n_layer, NULL); - std::vector t15L; t15L.resize(n_layer, NULL); - std::vector t16L; t16L.resize(n_layer, NULL); - std::vector t17L; t17L.resize(n_layer, NULL); - std::vector t18L; t18L.resize(n_layer, NULL); - std::vector t19L; t19L.resize(n_layer, NULL); - std::vector t20L; t20L.resize(n_layer, NULL); - std::vector t21L; t21L.resize(n_layer, NULL); - std::vector t22L; t22L.resize(n_layer, NULL); - std::vector t23L; t23L.resize(n_layer, NULL); - std::vector t24L; t24L.resize(n_layer, NULL); - std::vector t25L; t25L.resize(n_layer, NULL); - std::vector t26L; t26L.resize(n_layer, NULL); - std::vector t27L; t27L.resize(n_layer, NULL); - std::vector t28L; t28L.resize(n_layer, NULL); - std::vector t29L; t29L.resize(n_layer, NULL); - std::vector t30L; t30L.resize(n_layer, NULL); + GGML_ASSERT(tokens_input->type == GGML_TYPE_I32); + struct ggml_tensor * t00 = ggml_reshape_1d(ctx, tokens_input, N*n_batch); set_name(t00, "t00"); assert_shape_1d(t00, N*n_batch); + struct ggml_tensor * t01 = ggml_get_rows(ctx, model->tok_embeddings, t00); set_name(t01, "t01"); assert_shape_2d(t01, n_embd, N*n_batch); struct ggml_tensor * cur = t01; + std::vector checkpoints; + checkpoints.push_back(tokens_input); + checkpoints.push_back(targets); + checkpoints.push_back(t00); + checkpoints.push_back(t01); + + struct ggml_tensor * kv_scale; + if (!enable_flash_attn) { + kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head)); + } + for (int il = 0; il < n_layer; ++il) { - clr_buf(0); struct my_llama_layer & layer = model->layers[il]; - // tensors with values necessary for backward pass are in persistent buf(-1) - // other tensors with buf(0) and buf(1) are only temporary needed, and their memory reused after layer is completed. - use_buf(-1); struct ggml_tensor * t02 = expand(gf, ggml_rms_norm (ctx0, cur, rms_norm_eps)); assert_shape_2d(t02, n_embd, N*n_batch); - use_buf( 0); struct ggml_tensor * t03 = expand(gf, ggml_repeat (ctx0, layer.attention_norm, t02)); assert_shape_2d(t03, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t04 = expand(gf, ggml_mul (ctx0, t02, t03)); assert_shape_2d(t04, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t05 = expand(gf, ggml_mul_mat (ctx0, layer.wq, t04)); assert_shape_2d(t05, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t06 = expand(gf, ggml_reshape_4d (ctx0, t05, n_embd/n_head, n_head, N, n_batch)); assert_shape_4d(t06, n_embd/n_head, n_head, N, n_batch); - use_buf(-1); struct ggml_tensor * t07 = expand(gf, ggml_rope_inplace (ctx0, t06, n_past, n_rot, rope_mode, 0)); assert_shape_4d(t07, n_embd/n_head, n_head, N, n_batch); - use_buf(-1); struct ggml_tensor * t08 = expand(gf, ggml_mul_mat (ctx0, layer.wk, t04)); assert_shape_2d(t08, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t09 = expand(gf, ggml_reshape_4d (ctx0, t08, n_embd/n_head, n_head, N, n_batch)); assert_shape_4d(t09, n_embd/n_head, n_head, N, n_batch); - use_buf(-1); struct ggml_tensor * t10 = expand(gf, ggml_rope_inplace (ctx0, t09, n_past, n_rot, rope_mode, 0)); assert_shape_4d(t10, n_embd/n_head, n_head, N, n_batch); - use_buf(-1); struct ggml_tensor * t11 = expand(gf, ggml_mul_mat (ctx0, t04, layer.wv)); assert_shape_2d(t11, N*n_batch, n_embd); - use_buf(-1); struct ggml_tensor * t12 = expand(gf, ggml_reshape_4d (ctx0, t11, N, n_batch, n_embd/n_head, n_head)); assert_shape_4d(t12, N, n_batch, n_embd/n_head, n_head); - use_buf(-1); struct ggml_tensor * t13 = expand(gf, ggml_permute (ctx0, t07, 0, 2, 1, 3)); assert_shape_4d(t13, n_embd/n_head, N, n_head, n_batch); - use_buf(-1); struct ggml_tensor * t14 = expand(gf, ggml_permute (ctx0, t10, 0, 2, 1, 3)); assert_shape_4d(t14, n_embd/n_head, N, n_head, n_batch); - use_buf(-1); struct ggml_tensor * t15 = expand(gf, ggml_permute (ctx0, t12, 0, 3, 1, 2)); assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch); - use_buf(-1); struct ggml_tensor * t16 = expand(gf, ggml_flash_attn (ctx0, t13, t14, t15, true)); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch); - use_buf( 0); struct ggml_tensor * t17 = expand(gf, ggml_permute (ctx0, t16, 0, 2, 1, 3)); assert_shape_4d(t17, n_embd/n_head, n_head, N, n_batch); - use_buf(-1); struct ggml_tensor * t18 = expand(gf, ggml_cont (ctx0, t17)); assert_shape_4d(t18, n_embd/n_head, n_head, N, n_batch); - use_buf(-1); struct ggml_tensor * t19 = expand(gf, ggml_reshape_2d (ctx0, t18, n_embd, N*n_batch)); assert_shape_2d(t19, n_embd, N*n_batch); - use_buf( 0); struct ggml_tensor * t20 = expand(gf, ggml_mul_mat (ctx0, layer.wo, t19)); assert_shape_2d(t20, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t21 = expand(gf, ggml_add (ctx0, t20, cur)); assert_shape_2d(t21, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t22 = expand(gf, ggml_rms_norm (ctx0, t21, rms_norm_eps)); assert_shape_2d(t22, n_embd, N*n_batch); - use_buf( 0); struct ggml_tensor * t23 = expand(gf, ggml_repeat (ctx0, layer.ffn_norm, t22)); assert_shape_2d(t23, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t24 = expand(gf, ggml_mul (ctx0, t23, t22)); assert_shape_2d(t24, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t25 = expand(gf, ggml_mul_mat (ctx0, layer.w3, t24)); assert_shape_2d(t25, n_ff, N*n_batch); - use_buf(-1); struct ggml_tensor * t26 = expand(gf, ggml_mul_mat (ctx0, layer.w1, t24)); assert_shape_2d(t26, n_ff, N*n_batch); - use_buf(-1); struct ggml_tensor * t27 = expand(gf, ggml_silu (ctx0, t26)); assert_shape_2d(t27, n_ff, N*n_batch); - use_buf(-1); struct ggml_tensor * t28 = expand(gf, ggml_mul (ctx0, t27, t25)); assert_shape_2d(t28, n_ff, N*n_batch); - use_buf( 0); struct ggml_tensor * t29 = expand(gf, ggml_mul_mat (ctx0, layer.w2, t28)); assert_shape_2d(t29, n_embd, N*n_batch); - use_buf(-1); struct ggml_tensor * t30 = expand(gf, ggml_add (ctx0, t21, t29)); assert_shape_2d(t30, n_embd, N*n_batch); - t02L[il] = t02; - t03L[il] = t03; - t04L[il] = t04; - t05L[il] = t05; - t06L[il] = t06; - t07L[il] = t07; - t08L[il] = t08; - t09L[il] = t09; - t10L[il] = t10; - t11L[il] = t11; - t12L[il] = t12; - t13L[il] = t13; - t14L[il] = t14; - t15L[il] = t15; - t16L[il] = t16; - t17L[il] = t17; - t18L[il] = t18; - t19L[il] = t19; - t20L[il] = t20; - t21L[il] = t21; - t22L[il] = t22; - t23L[il] = t23; - t24L[il] = t24; - t25L[il] = t25; - t26L[il] = t26; - t27L[il] = t27; - t28L[il] = t28; - t29L[il] = t29; - t30L[il] = t30; - - cur = t30; - } - clr_buf(0); - use_buf(0); - struct ggml_tensor * t31 = expand(gf, ggml_rms_norm (ctx0, cur, rms_norm_eps)); assert_shape_2d(t31, n_embd, N*n_batch); - struct ggml_tensor * t32 = expand(gf, ggml_repeat (ctx0, model->norm, t31)); assert_shape_2d(t32, n_embd, N*n_batch); - struct ggml_tensor * t33 = expand(gf, ggml_mul (ctx0, t32, t31)); assert_shape_2d(t33, n_embd, N*n_batch); - use_buf(-1); - struct ggml_tensor * t34 = expand(gf, ggml_mul_mat (ctx0, model->output, t33)); assert_shape_2d(t34, n_vocab, N*n_batch); - struct ggml_tensor * t35 = expand(gf, ggml_reshape_3d(ctx0, t34, n_vocab, N, n_batch)); assert_shape_3d(t35, n_vocab, N, n_batch); - struct ggml_tensor * t36 = expand(gf, ggml_cross_entropy_loss(ctx0, t35, targets)); assert_shape_1d(t36, 1); - - { - /* - tok_embeddings | grad_tok_embeddings = ggml_get_rows_back(grad_t01, t00) - L0_att_norm | grad_L0_att_norm = ggml_repeat_back(grad_t03L0, L0_att_norm.shape) - L0_wq | grad_L0_wq = ggml_out_prod(t04L0, grad_t05L0) - L0_wk | grad_L0_wk = ggml_out_prod(t04L0, grad_t08L0) - L0_wv | grad_L0_wv = ggml_out_prod(t04L0, ggml_transpose(grad_t11L0)) - L0_wo | grad_L0_wo = ggml_out_prod(t19L0, grad_t20L0) - L0_ffn_norm | grad_L0_ffn_norm = ggml_repeat_back(grad_t23L0, L0_ffn_norm.shape) - L0_w1 | grad_L0_w1 = ggml_out_prod(t24L0, grad_t26L0) - L0_w2 | grad_L0_w2 = ggml_out_prod(t28L0, grad_t29L0) - L0_w3 | grad_L0_w3 = ggml_out_prod(t24L0, grad_t25L0) - L1_att_norm | grad_L1_att_norm = ggml_repeat_back(grad_t03L1, L1_att_norm.shape) - L1_wq | grad_L1_wq = ggml_out_prod(t04L1, grad_t05L1) - L1_wk | grad_L1_wk = ggml_out_prod(t04L1, grad_t08L1) - L1_wv | grad_L1_wv = ggml_out_prod(t04L1, ggml_transpose(grad_t11L1)) - L1_wo | grad_L1_wo = ggml_out_prod(t19L1, grad_t20L1) - L1_ffn_norm | grad_L1_ffn_norm = ggml_repeat_back(grad_t23L1, L1_ffn_norm.shape) - L1_w1 | grad_L1_w1 = ggml_out_prod(t24L1, grad_t26L1) - L1_w2 | grad_L1_w2 = ggml_out_prod(t28L1, grad_t29L1) - L1_w3 | grad_L1_w3 = ggml_out_prod(t24L1, grad_t25L1) - norm | grad_norm = ggml_repeat_back(grad_t32, norm.shape) - output | grad_output = ggml_out_prod(t33, grad_t34) - | - t01 = ggml_get_rows(tok_embeddings, t00) | grad_t01 = grad_t21L0 + ggml_rms_norm_back(t01, grad_t02L0) - for layer: | - t02L0*= ggml_rms_norm (t01) | grad_t02L0 = ggml_mul(grad_t04L0, t03L0) - t03L0 = ggml_repeat (L0_att_norm, t02L0_shape) | grad_t03L0 = ggml_mul(grad_t04L0, t02L0) - t04L0*= ggml_mul (t02L0, t03L0) | grad_t04L0 = ggml_out_prod(L0_wv, grad_t11L0) + ggml_out_prod(L0_wk, ggml_transpose(grad_t08L0)) + ggml_out_prod(L0_wq, ggml_transpose(grad_t05L0)) - t05L0 = ggml_mul_mat (L0_wq, t04L0) | grad_t05L0 = ggml_reshape(grad_t06L0, t05L0_shape) - t06L0 = ggml_reshape_4d (t05L0, n_embd/n_head, n_head, N, n_batch) | grad_t06L0 = ggml_rope_back(grad_t07L0) - t07L0 = ggml_rope_inplace (t06L0) | grad_t07L0 = ggml_permute_back(grad_t13L0, 0, 2, 1, 3) = ggml_permute(grad_t13L0, 0, 2, 1, 3) - t08L0 = ggml_mul_mat (L0_wk, t04L0) | grad_t08L0 = ggml_reshape(grad_t09L0, t08L0_shape) - t09L0 = ggml_reshape_4d (t08L0, n_embd/n_head, n_head, N, n_batch) | grad_t09L0 = ggml_rope_back(grad_t10L0) - t10L0 = ggml_rope_inplace (t09L0) | grad_t10L0 = ggml_permute_back(grad_t14L0, 0, 2, 1, 3) = ggml_permute(grad_t14L0, 0, 2, 1, 3) - t11L0 = ggml_mul_mat (t04L0, L0_wv) | grad_t11L0 = ggml_reshape(grad_t12L0, t11L0_shape) - t12L0 = ggml_reshape_4d (t11L0, N, n_batch, n_embd/n_head, n_head) | grad_t12L0 = ggml_permute_back(grad_t15L0, 0, 3, 1, 2) = ggml_permute(grad_t15L0, 0, 2, 3, 1) - t13L0*= ggml_permute (t07L0, 0, 2, 1, 3) | grad_t13L0 = view__q(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0)) - t14L0*= ggml_permute (t10L0, 0, 2, 1, 3) | grad_t14L0 = view__k(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0)) - t15L0*= ggml_permute (t12L0, 0, 3, 1, 2) | grad_t15L0 = view__v(ggml_flash_attn_back(t13L0, t14L0, t15L0, grad_t16L0)) - t16L0 = ggml_flash_attn (t13L0, t14L0, t15L0) | grad_t16L0 = ggml_permute_back(grad_t17L0, 0, 2, 1, 3) = ggml_permute(grad_t17L0, 0, 2, 1, 3) - t17L0 = ggml_permute (t16L0, 0, 2, 1, 3) | grad_t17L0 = grad_t18L0 - t18L0 = ggml_cont (t17L0) | grad_t18L0 = ggml_reshape(grad_t19L0, t18L0_shape) - t19L0*= ggml_reshape_2d (t18L0, n_embd, N*n_batch) | grad_t19L0 = ggml_out_prod(L0_wo, ggml_transpose(grad_t20L0)) - t20L0 = ggml_mul_mat (L0_wo, t19L0) | grad_t20L0 = grad_t21L0 - t21L0*= ggml_add (t20L0, t01) | grad_t21L0 = grad_t30L0 + ggml_rms_norm_back(t21L0, grad_t22L0) - t22L0*= ggml_rms_norm (t21L0) | grad_t22L0 = ggml_mul(grad_t24L0, t23L0) - t23L0 = ggml_repeat (L0_ffn_norm, t22L0_shape) | grad_t23L0 = ggml_mul(grad_t24L0, t22L0) - t24L0*= ggml_mul (t23L0, t22L0) | grad_t24L0 = ggml_out_prod(L0_w1, ggml_transpose(grad_t26L0)) + ggml_out_prod(L0_w3, ggml_transpose(grad_t25L0)) - t25L0*= ggml_mul_mat (L0_w3, t24L0) | grad_t25L0 = ggml_mul(grad_t28L0, t27L0) - t26L0*= ggml_mul_mat (L0_w1, t24L0) | grad_t26L0 = ggml_silu_back(t26L0, grad_t27L0) - t27L0*= ggml_silu (t26L0) | grad_t27L0 = ggml_mul(grad_t28L0, t25L0) - t28L0*= ggml_mul (t27L0, t25L0) | grad_t28L0 = ggml_out_prod(L0_w2, ggml_transpose(grad_t29L0)) - t29L0 = ggml_mul_mat (L0_w2, t28L0) | grad_t29L0 = grad_t30L0 - t30L0*= ggml_add (t21L0, t29L0) | grad_t30L0 = ggml_rms_norm_back(t30L0, grad_t02L1) + grad_t21L1 - ^ - t02L1*= ggml_rms_norm (t30L0) | grad_t02L1 = ggml_mul(grad_t04L1, t03L1) - t03L1 = ggml_repeat (L1_att_norm, t02L1_shape) | grad_t03L1 = ggml_mul(grad_t04L1, t02L1) - t04L1*= ggml_mul (t02L1, t03L1) | grad_t04L1 = ggml_out_prod(L1_wv, grad_t11L1) + ggml_out_prod(L1_wk, ggml_transpose(grad_t08L1)) + ggml_out_prod(L1_wq, ggml_transpose(grad_t05L1)) - t05L1 = ggml_mul_mat (L1_wq, t04L1) | grad_t05L1 = ggml_reshape(grad_t06L1, t05L1_shape) - t06L1 = ggml_reshape_4d (t05L1, n_embd/n_head, n_head, N, n_batch) | grad_t06L1 = ggml_rope_back(grad_t07L1) - t07L1 = ggml_rope_inplace (t06L1) | grad_t07L1 = ggml_permute_back(grad_t13L1, 0, 2, 1, 3) = ggml_permute(grad_t13L1, 0, 2, 1, 3) - t08L1 = ggml_mul_mat (L1_wk, t04L1) | grad_t08L1 = ggml_reshape(grad_t09L1, t08L1_shape) - t09L1 = ggml_reshape_4d (t08L1, n_embd/n_head, n_head, N, n_batch) | grad_t09L1 = ggml_rope_back(grad_t10L1) - t10L1 = ggml_rope_inplace (t09L1) | grad_t10L1 = ggml_permute_back(grad_t14L1, 0, 2, 1, 3) = ggml_permute(grad_t14L1, 0, 2, 1, 3) - t11L1 = ggml_mul_mat (t04L1, L1_wv) | grad_t11L1 = ggml_reshape(grad_t12L1, t11L1_shape) - t12L1 = ggml_reshape_4d (t11L1, N, n_batch, n_embd/n_head, n_head) | grad_t12L1 = ggml_permute_back(grad_t15L1, 0, 3, 1, 2) = ggml_permute(grad_t15L1, 0, 2, 3, 1) - t13L1*= ggml_permute (t07L1, 0, 2, 1, 3) | grad_t13L1 = view__q(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1)) - t14L1*= ggml_permute (t10L1, 0, 2, 1, 3) | grad_t14L1 = view__k(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1)) - t15L1*= ggml_permute (t12L1, 0, 3, 1, 2) | grad_t15L1 = view__v(ggml_flash_attn_back(t13L1, t14L1, t15L1, grad_t16L1)) - t16L1 = ggml_flash_attn (t13L1, t14L1, t15L1) | grad_t16L1 = ggml_permute_back(grad_t17L1, 0, 2, 1, 3) = ggml_permute(grad_t17L1, 0, 2, 1, 3) - t17L1 = ggml_permute (t16L1, 0, 2, 1, 3) | grad_t17L1 = grad_t18L1 - t18L1 = ggml_cont (t17L1) | grad_t18L1 = ggml_reshape(grad_t19L1, t18L1_shape) - t19L1*= ggml_reshape_2d (t18L1, n_embd, N*n_batch) | grad_t19L1 = ggml_out_prod(L1_wo, ggml_transpose(grad_t20L1)) - t20L1 = ggml_mul_mat (L1_wo, t19L1) | grad_t20L1 = grad_t21L1 - t21L1*= ggml_add (t20L1, t30L0) | grad_t21L1 = grad_t30L1 + ggml_rms_norm_back(t21L1, grad_t22L1) - t22L1*= ggml_rms_norm (t21L1) | grad_t22L1 = ggml_mul(grad_t24L1, t23L1) - t23L1 = ggml_repeat (L1_ffn_norm, t22L1_shape) | grad_t23L1 = ggml_mul(grad_t24L1, t22L1) - t24L1*= ggml_mul (t23L1, t22L1) | grad_t24L1 = ggml_out_prod(L1_w1, ggml_transpose(grad_t26L1)) + ggml_out_prod(L1_w3, ggml_transpose(grad_t25L1)) - t25L1*= ggml_mul_mat (L1_w3, t24L1) | grad_t25L1 = ggml_mul(grad_t28L1, t27L1) - t26L1*= ggml_mul_mat (L1_w1, t24L1) | grad_t26L1 = ggml_silu_back(t26L1, grad_t27L1) - t27L1*= ggml_silu (t26L1) | grad_t27L1 = ggml_mul(grad_t28L1, t25L1) - t28L1*= ggml_mul (t27L1, t25L1) | grad_t28L1 = ggml_out_prod(L1_w2, ggml_transpose(grad_t29L1)) - t29L1 = ggml_mul_mat (L1_w2, t28L1) | grad_t29L1 = grad_t30L1 - t30L1*= ggml_add (t21L1, t29L1) | grad_t30L1 = ggml_rms_norm_back(t30L1, grad_t31) - ^ - t31 = ggml_rms_norm (t30L1) | grad_t31 = ggml_mul(grad_t33, t32) - t32 = ggml_repeat (norm, t31.shape) | grad_t32 = ggml_mul(grad_t33, t31) - t33 = ggml_mul (t32, t31) | grad_t33 = ggml_out_prod(output, ggml_transpose(grad_t34)) - t34 = ggml_mul_mat (output, t33) | grad_t34 = ggml_reshape(grad_t35, t34.shape) - t35 = ggml_reshape_3d (t34, n_vocab, N, n_batch) | grad_t35 = ggml_cross_entropy_loss_back(t35, targets, grad_t36) - t36 = ggml_cross_entropy_loss(t35, targets) | grad_t36 = 1 (optimizer) - tensors marked with * need to be stored until grad computation - tensors during grad computation are all temporary - */ - } - - *gb = *gf; - - // t36->grad gets set to one by optimizer, so we need the tensor. - // initialize it with 1.0f to make sure. - use_buf(-1); - t36->grad = expand(gb, ggml_new_f32(ctx0, 1.0f)); - - use_buf(0); - t35->grad = expand(gb, ggml_cross_entropy_loss_back(ctx0, t35, targets, t36->grad)); assert_shape_3d(t35->grad, n_vocab, N, n_batch); - t34->grad = expand(gb, ggml_reshape_2d (ctx0, t35->grad, n_vocab, N*n_batch)); assert_shape_2d(t34->grad, n_vocab, N*n_batch); - t33->grad = expand(gb, ggml_out_prod (ctx0, model->output, ggml_transpose(ctx0, t34->grad))); assert_shape_2d(t33->grad, n_embd, N*n_batch); - t32->grad = expand(gb, ggml_mul (ctx0, t33->grad, t31)); assert_shape_2d(t32->grad, n_embd, N*n_batch); - - use_buf(-1); - - model->norm->grad = expand(gb, add_or_set(model->norm->grad, ggml_repeat_back(ctx0, t32->grad, model->norm))); assert_shape_1d(model->norm->grad, n_embd); - model->output->grad = expand(gb, add_or_set(model->output->grad, ggml_out_prod(ctx0, t33, t34->grad))); assert_shape_2d(model->output->grad, n_embd, n_vocab); - - clr_buf(1); - use_buf(1); - t31->grad = expand(gb, ggml_mul(ctx0, t33->grad, t32)); assert_shape_2d(t31->grad, n_embd, N*n_batch); - - struct ggml_tensor * back_layer_inp = t31; - struct ggml_tensor * grad_layer_inp = NULL; - - for (int k = 0; k < n_layer; ++k) { - int il = n_layer-1-k; - struct my_llama_layer & layer = model->layers[il]; - - struct ggml_tensor * t02 = t02L[il]; - struct ggml_tensor * t03 = t03L[il]; - struct ggml_tensor * t04 = t04L[il]; - struct ggml_tensor * t05 = t05L[il]; - struct ggml_tensor * t06 = t06L[il]; - struct ggml_tensor * t07 = t07L[il]; - struct ggml_tensor * t08 = t08L[il]; - struct ggml_tensor * t09 = t09L[il]; - struct ggml_tensor * t10 = t10L[il]; - struct ggml_tensor * t11 = t11L[il]; - struct ggml_tensor * t12 = t12L[il]; - struct ggml_tensor * t13 = t13L[il]; - struct ggml_tensor * t14 = t14L[il]; - struct ggml_tensor * t15 = t15L[il]; - struct ggml_tensor * t16 = t16L[il]; - struct ggml_tensor * t17 = t17L[il]; - struct ggml_tensor * t18 = t18L[il]; - struct ggml_tensor * t19 = t19L[il]; - struct ggml_tensor * t20 = t20L[il]; - struct ggml_tensor * t21 = t21L[il]; - struct ggml_tensor * t22 = t22L[il]; - struct ggml_tensor * t23 = t23L[il]; - struct ggml_tensor * t24 = t24L[il]; - struct ggml_tensor * t25 = t25L[il]; - struct ggml_tensor * t26 = t26L[il]; - struct ggml_tensor * t27 = t27L[il]; - struct ggml_tensor * t28 = t28L[il]; - struct ggml_tensor * t29 = t29L[il]; - struct ggml_tensor * t30 = t30L[il]; - - clr_buf(0); - use_buf(0); - t30->grad = expand(gb, ggml_rms_norm_back(ctx0, t30, back_layer_inp->grad)); assert_shape_2d(t30->grad, n_embd, N*n_batch); - if (grad_layer_inp) { - t30->grad = expand(gb, ggml_add(ctx0, t30->grad, grad_layer_inp->grad)); assert_shape_2d(t30->grad, n_embd, N*n_batch); + struct ggml_tensor * t02 = ggml_rms_norm (ctx, cur, f_norm_rms_eps); set_name(t02, "t02"); assert_shape_2d(t02, n_embd, N*n_batch); + struct ggml_tensor * t03 = ggml_repeat (ctx, layer.attention_norm, t02); set_name(t03, "t03"); assert_shape_2d(t03, n_embd, N*n_batch); + struct ggml_tensor * t04 = ggml_mul (ctx, t03, t02); set_name(t04, "t04"); assert_shape_2d(t04, n_embd, N*n_batch); + struct ggml_tensor * t05 = ggml_mul_mat (ctx, layer.wq, t04); set_name(t05, "t05"); assert_shape_2d(t05, n_embd, N*n_batch); + struct ggml_tensor * t06 = ggml_reshape_4d (ctx, t05, n_embd/n_head, n_head, N, n_batch); set_name(t06, "t06"); assert_shape_4d(t06, n_embd/n_head, n_head, N, n_batch); + struct ggml_tensor * t07 = rope (t06); set_name(t07, "t07"); assert_shape_4d(t07, n_embd/n_head, n_head, N, n_batch); + struct ggml_tensor * t08 = ggml_mul_mat (ctx, layer.wk, t04); set_name(t08, "t08"); assert_shape_2d(t08, n_embd, N*n_batch); + struct ggml_tensor * t09 = ggml_reshape_4d (ctx, t08, n_embd/n_head, n_head, N, n_batch); set_name(t09, "t09"); assert_shape_4d(t09, n_embd/n_head, n_head, N, n_batch); + struct ggml_tensor * t10 = rope (t09); set_name(t10, "t10"); assert_shape_4d(t10, n_embd/n_head, n_head, N, n_batch); + struct ggml_tensor * t11 = ggml_mul_mat (ctx, t04, layer.wv); set_name(t11, "t11"); assert_shape_2d(t11, N*n_batch, n_embd); + struct ggml_tensor * t12 = ggml_reshape_4d (ctx, t11, N, n_batch, n_embd/n_head, n_head); set_name(t12, "t12"); assert_shape_4d(t12, N, n_batch, n_embd/n_head, n_head); + struct ggml_tensor * t13 = ggml_permute (ctx, t07, 0, 2, 1, 3); set_name(t13, "t13"); assert_shape_4d(t13, n_embd/n_head, N, n_head, n_batch); + struct ggml_tensor * t14 = ggml_permute (ctx, t10, 0, 2, 1, 3); set_name(t14, "t14"); assert_shape_4d(t14, n_embd/n_head, N, n_head, n_batch); + struct ggml_tensor * t15 = ggml_permute (ctx, t12, 0, 3, 1, 2); set_name(t15, "t15"); assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch); + struct ggml_tensor * t16; + if (enable_flash_attn) { + t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch); + } else { + struct ggml_tensor * t16_0 = ggml_mul_mat (ctx, t14, t13); set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch); + struct ggml_tensor * t16_1 = ggml_scale_inplace (ctx, t16_0, kv_scale); set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch); + struct ggml_tensor * t16_2 = ggml_diag_mask_inf_inplace(ctx, t16_1, n_past); set_name(t16_2, "t16_2"); assert_shape_4d(t16_2, N, N, n_head, n_batch); + struct ggml_tensor * t16_3 = ggml_soft_max_inplace (ctx, t16_2); set_name(t16_3, "t16_3"); assert_shape_4d(t16_3, N, N, n_head, n_batch); + t16 = ggml_mul_mat(ctx, t15, t16_3); set_name(t16, "t16"); assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch); } - clr_buf(1); - t29->grad = t30->grad; assert_shape_2d(t29->grad, n_embd, N*n_batch); - t28->grad = expand(gb, ggml_out_prod(ctx0, layer.w2, ggml_transpose(ctx0, t29->grad))); assert_shape_2d(t28->grad, n_ff, N*n_batch); - t27->grad = expand(gb, ggml_mul(ctx0, t28->grad, t25)); assert_shape_2d(t27->grad, n_ff, N*n_batch); - t26->grad = expand(gb, ggml_silu_back(ctx0, t26, t27->grad)); assert_shape_2d(t26->grad, n_ff, N*n_batch); - t25->grad = expand(gb, ggml_mul(ctx0, t28->grad, t27)); assert_shape_2d(t25->grad, n_ff, N*n_batch); - t24->grad = expand(gb, ggml_add_inplace(ctx0, - ggml_out_prod(ctx0, layer.w1, ggml_transpose(ctx0, t26->grad)), - ggml_out_prod(ctx0, layer.w3, ggml_transpose(ctx0, t25->grad)))); assert_shape_2d(t24->grad, n_embd, N*n_batch); - t23->grad = expand(gb, ggml_mul(ctx0, t24->grad, t22)); assert_shape_2d(t23->grad, n_embd, N*n_batch); - t22->grad = expand(gb, ggml_mul(ctx0, t24->grad, ggml_repeat(ctx0, layer.ffn_norm, t24->grad))); assert_shape_2d(t22->grad, n_embd, N*n_batch); - use_buf(1); - t21->grad = expand(gb, ggml_add(ctx0, t30->grad, ggml_rms_norm_back(ctx0, t21, t22->grad))); assert_shape_2d(t21->grad, n_embd, N*n_batch); - grad_layer_inp = t21; - use_buf(0); - t20->grad = t21->grad; assert_shape_2d(t20->grad, n_embd, N*n_batch); - t19->grad = expand(gb, ggml_out_prod(ctx0, layer.wo, ggml_transpose(ctx0, t20->grad))); assert_shape_2d(t19->grad, n_embd, N*n_batch); - t18->grad = expand(gb, ggml_reshape_4d(ctx0, t19->grad, n_embd/n_head, n_head, N, n_batch)); assert_shape_4d(t18->grad, n_embd/n_head, n_head, N, n_batch); - t17->grad = t18->grad; assert_shape_4d(t17->grad, n_embd/n_head, n_head, N, n_batch); - t16->grad = expand(gb, ggml_permute(ctx0, t17->grad, 0, 2, 1, 3)); assert_shape_4d(t16->grad, n_embd/n_head, N, n_head, n_batch); - struct ggml_tensor * flash_attn = expand(gb, ggml_flash_attn_back(ctx0, t13, t14, t15, t16->grad, true)); assert_shape_4d(flash_attn, n_embd/n_head, N*3, n_head, n_batch); - t15->grad = expand(gb, view__v(flash_attn)); assert_shape_4d(t15->grad, N, n_embd/n_head, n_head, n_batch); - t14->grad = expand(gb, view__k(flash_attn)); assert_shape_4d(t14->grad, n_embd/n_head, N, n_head, n_batch); - t13->grad = expand(gb, view__q(flash_attn)); assert_shape_4d(t13->grad, n_embd/n_head, N, n_head, n_batch); - t12->grad = expand(gb, ggml_permute(ctx0, t15->grad, 0, 2, 3, 1)); assert_shape_4d(t12->grad, N, n_batch, n_embd/n_head, n_head); - t11->grad = expand(gb, ggml_reshape_2d(ctx0, ggml_cont(ctx0, t12->grad), N*n_batch, n_embd)); assert_shape_2d(t11->grad, N*n_batch, n_embd); - t10->grad = expand(gb, ggml_permute(ctx0, t14->grad, 0, 2, 1, 3)); assert_shape_4d(t10->grad, n_embd/n_head, n_head, N, n_batch); - t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode, n_ctx, 10000.0f, 1.0f, 0.0f, false)); assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch); - t08->grad = expand(gb, ggml_reshape_2d(ctx0, t09->grad, n_embd, N*n_batch)); assert_shape_2d(t08->grad, n_embd, N*n_batch); - t07->grad = expand(gb, ggml_permute(ctx0, t13->grad, 0, 2, 1, 3)); assert_shape_4d(t07->grad, n_embd/n_head, n_head, N, n_batch); - t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode, n_ctx, 10000.0f, 1.0f, 0.0f, false)); assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch); - t05->grad = expand(gb, ggml_reshape_2d(ctx0, t06->grad, n_embd, N*n_batch)); assert_shape_2d(t05->grad, n_embd, N*n_batch); - t04->grad = expand(gb, ggml_add_inplace(ctx0, - ggml_add_inplace(ctx0, - ggml_out_prod(ctx0, layer.wv, t11->grad), - ggml_out_prod(ctx0, layer.wk, ggml_transpose(ctx0, t08->grad))), - ggml_out_prod(ctx0, layer.wq, ggml_transpose(ctx0, t05->grad)))); assert_shape_2d(t04->grad, n_embd, N*n_batch); - t03->grad = expand(gb, ggml_mul(ctx0, t04->grad, t02)); assert_shape_2d(t04->grad, n_embd, N*n_batch); - use_buf(1); - t02->grad = expand(gb, ggml_mul(ctx0, t04->grad, ggml_repeat(ctx0, layer.attention_norm, t02))); assert_shape_2d(t02->grad, n_embd, N*n_batch); - back_layer_inp = t02; - // use_buf(0); - - use_buf(-1); - layer.attention_norm->grad = expand(gb, add_or_set(layer.attention_norm->grad, ggml_repeat_back(ctx0, t03->grad, layer.attention_norm))); assert_shape_1d(layer.attention_norm->grad, n_embd); - layer.wq->grad = expand(gb, add_or_set(layer.wq->grad, ggml_out_prod(ctx0, t04, t05->grad))); assert_shape_2d(layer.wq->grad, n_embd, n_embd); - layer.wk->grad = expand(gb, add_or_set(layer.wk->grad, ggml_out_prod(ctx0, t04, t08->grad))); assert_shape_2d(layer.wk->grad, n_embd, n_embd); - layer.wv->grad = expand(gb, add_or_set(layer.wv->grad, ggml_out_prod(ctx0, t04, ggml_transpose(ctx0, t11->grad)))); assert_shape_2d(layer.wv->grad, n_embd, n_embd); - layer.wo->grad = expand(gb, add_or_set(layer.wo->grad, ggml_out_prod(ctx0, t19, t20->grad))); assert_shape_2d(layer.wo->grad, n_embd, n_embd); - layer.ffn_norm->grad = expand(gb, add_or_set(layer.ffn_norm->grad, ggml_repeat_back(ctx0, t23->grad, layer.ffn_norm))); assert_shape_1d(layer.ffn_norm->grad, n_embd); - layer.w1->grad = expand(gb, add_or_set(layer.w1->grad, ggml_out_prod(ctx0, t24, t26->grad))); assert_shape_2d(layer.w1->grad, n_embd, n_ff); - layer.w2->grad = expand(gb, add_or_set(layer.w2->grad, ggml_out_prod(ctx0, t28, t29->grad))); assert_shape_2d(layer.w2->grad, n_ff, n_embd); - layer.w3->grad = expand(gb, add_or_set(layer.w3->grad, ggml_out_prod(ctx0, t24, t25->grad))); assert_shape_2d(layer.w3->grad, n_embd, n_ff); - // use_buf(0); + struct ggml_tensor * t17 = ggml_permute (ctx, t16, 0, 2, 1, 3); set_name(t17, "t17"); assert_shape_4d(t17, n_embd/n_head, n_head, N, n_batch); + struct ggml_tensor * t18 = ggml_cont (ctx, t17); set_name(t18, "t18"); assert_shape_4d(t18, n_embd/n_head, n_head, N, n_batch); + struct ggml_tensor * t19 = ggml_reshape_2d (ctx, t18, n_embd, N*n_batch); set_name(t19, "t19"); assert_shape_2d(t19, n_embd, N*n_batch); + struct ggml_tensor * t20 = ggml_mul_mat (ctx, layer.wo, t19); set_name(t20, "t20"); assert_shape_2d(t20, n_embd, N*n_batch); + struct ggml_tensor * t21 = ggml_add (ctx, t20, cur); set_name(t21, "t21"); assert_shape_2d(t21, n_embd, N*n_batch); + struct ggml_tensor * t22 = ggml_rms_norm (ctx, t21, f_norm_rms_eps); set_name(t22, "t22"); assert_shape_2d(t22, n_embd, N*n_batch); + struct ggml_tensor * t23 = ggml_repeat (ctx, layer.ffn_norm, t22); set_name(t23, "t23"); assert_shape_2d(t23, n_embd, N*n_batch); + struct ggml_tensor * t24 = ggml_mul (ctx, t23, t22); set_name(t24, "t24"); assert_shape_2d(t24, n_embd, N*n_batch); + struct ggml_tensor * t25 = ggml_mul_mat (ctx, layer.w3, t24); set_name(t25, "t25"); assert_shape_2d(t25, n_ff, N*n_batch); + struct ggml_tensor * t26 = ggml_mul_mat (ctx, layer.w1, t24); set_name(t26, "t26"); assert_shape_2d(t26, n_ff, N*n_batch); + struct ggml_tensor * t27 = ggml_silu (ctx, t26); set_name(t27, "t27"); assert_shape_2d(t27, n_ff, N*n_batch); + struct ggml_tensor * t28 = ggml_mul (ctx, t27, t25); set_name(t28, "t28"); assert_shape_2d(t28, n_ff, N*n_batch); + struct ggml_tensor * t29 = ggml_mul_mat (ctx, layer.w2, t28); set_name(t29, "t29"); assert_shape_2d(t29, n_embd, N*n_batch); + struct ggml_tensor * t30 = ggml_add (ctx, t29, t21); set_name(t30, "t30"); assert_shape_2d(t30, n_embd, N*n_batch); + cur = t30; + checkpoints.push_back(cur); + } + struct ggml_tensor * t31 = ggml_rms_norm (ctx, cur, f_norm_rms_eps); set_name(t31, "t31"); assert_shape_2d(t31, n_embd, N*n_batch); + struct ggml_tensor * t32 = ggml_repeat (ctx, model->norm, t31); set_name(t32, "t32"); assert_shape_2d(t32, n_embd, N*n_batch); + struct ggml_tensor * t33 = ggml_mul (ctx, t32, t31); set_name(t33, "t33"); assert_shape_2d(t33, n_embd, N*n_batch); + struct ggml_tensor * t34 = ggml_mul_mat (ctx, model->output, t33); set_name(t34, "t34"); assert_shape_2d(t34, n_vocab, N*n_batch); + struct ggml_tensor * t35 = ggml_reshape_3d (ctx, t34, n_vocab, N, n_batch); set_name(t35, "t35"); assert_shape_3d(t35, n_vocab, N, n_batch); + struct ggml_tensor * t36 = ggml_cross_entropy_loss(ctx, t35, targets); set_name(t36, "t36"); assert_shape_1d(t36, 1); + + checkpoints.push_back(t31); + checkpoints.push_back(t32); + checkpoints.push_back(t33); + checkpoints.push_back(t34); + checkpoints.push_back(t35); + checkpoints.push_back(t36); + + ggml_build_forward_expand(gf, t36); + + if (enable_checkpointing) { + ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size()); + } else { + *gb = *gf; + ggml_build_backward_expand(ctx, gf, gb, true); + } + + if (alloc) { + // make sure some tensors are not reallocated by inserting new temporary nodes depending on them + int n_leafs_before = gb->n_leafs; + int n_nodes_before = gb->n_nodes; + struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f); + // output tensors + ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one)); + ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one)); + // input gradient + ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one)); + GGML_ASSERT(t36->grad->data == NULL && !ggml_is_view(t36->grad)); + ggml_allocr_alloc(alloc, t36->grad); + // gradient tensors (will be set to zero by ggml_graph_reset) + // pinning these produces large unnecessary memory overhead, which will be resolved by PR 2632 + for (int i = 0; i < gf->n_nodes; ++i) { + if (!gf->grads[i]) continue; + if (gf->grads[i]->data == NULL && !ggml_is_view(gf->grads[i])) { + ggml_allocr_alloc(alloc, gf->grads[i]); + } + ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, gf->grads[i], one)); + } + // allocating checkpoints in one block to reduce memory fragmentation + // note: they will be freed in reverse order + for (int i = 0; i < (int) checkpoints.size(); ++i) { + if (checkpoints[i]->data == NULL && !ggml_is_view(checkpoints[i])) { + ggml_allocr_alloc(alloc, checkpoints[i]); + } + } + + //int n_leafs_after = gb->n_leafs; + //int n_nodes_after = gb->n_nodes; + + ggml_allocr_alloc_graph(alloc, gb); + + // remove the additional nodes and leafs + for (int i = n_leafs_before; i < gb->n_leafs; ++i) { + gb->leafs[i] = NULL; + } + for (int i = n_nodes_before; i < gb->n_nodes; ++i) { + gb->nodes[i] = NULL; + } + gb->n_leafs = n_leafs_before; + gb->n_nodes = n_nodes_before; } - clr_buf(0); - use_buf(0); - t01->grad = expand(gb, ggml_add_inplace(ctx0, grad_layer_inp->grad, ggml_rms_norm_back(ctx0, t01, back_layer_inp->grad))); assert_shape_2d(t01->grad, n_embd, N*n_batch); - use_buf(-1); - model->tok_embeddings->grad = expand(gb, ggml_get_rows_back(ctx0, t01->grad, t00, model->tok_embeddings)); assert_shape_2d(model->tok_embeddings->grad, n_embd, n_vocab); - // clr_buf(1); - // clr_buf(0); *logits = t35; - - if (track_max_mem) { - printf("%s: max size compute buf0: %zu\n", __func__, buf_maxs[0]); - printf("%s: max size compute buf1: %zu\n", __func__, buf_maxs[1]); - } - - // now that all grads are created, set the graph leafs and grads - graph_set_leafs_grads(gf); - graph_set_leafs_grads(gb); - return t36; } @@ -1962,42 +874,6 @@ void print_matrix(struct ggml_tensor * probs) { } } - -void print_token(struct llama_context * ctx, llama_token token) { - printf("%s", llama_token_to_piece(ctx, token).c_str()); -} - -void print_tokens(struct llama_context* ctx, struct ggml_tensor * tokens) { - for (int i=0; ine[0]; ++i) { - int token = ggml_get_i32_1d(tokens, i); - print_token(ctx, token); - } -} - -void print_tokens_batch(struct llama_context* ctx, struct ggml_tensor * tokens) { - for (int i1=0; i1ne[1]; ++i1) { - //int num_newline = 0; - for (int i0=0; i0ne[0]; ++i0) { - int token = get_i32_2d(tokens, i0, i1); - print_token(ctx, token); - // bool isnl = (token == llama_token_nl()); - // if (isnl) { - // ++num_newline; - // } - // if (isnl) { - // if (num_newline < 2) { - // print_token(ctx, token); - // } else { - // printf("\\n"); - // } - // } else { - // print_token(ctx, token); - // } - } - printf("\n--\n"); - } -} - void get_example_targets(struct llama_context * lctx, const int * train_samples, size_t n_train_samples, const llama_token * train_data, size_t n_train_data, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * target_logits, struct ggml_tensor * target_probs) { int n_tokens = tokens_input->ne[0]; int n_vocab = target_logits->ne[0]; @@ -2033,51 +909,27 @@ void get_example_targets_batch(struct llama_context * lctx, const int * train_sa ggml_set_f32(target_logits, -1.0f/n_vocab); ggml_set_f32(target_probs, 0.0f); + // printf("%s: example_id=%d n_batch=%d n_train_samples=%zu\n", __func__, example_id, n_batch, n_train_samples); for (int k=0; kne[0]; - int n_vocab = target_logits->ne[0]; - for (int i=0; i chars(len); - read_raw(chars.data(), len); - return std::string(chars.data(), len); - } - - void write_raw(const void * ptr, size_t size) { - if (size == 0) { - return; - } - errno = 0; - size_t ret = std::fwrite(ptr, size, 1, fp); - if (ret != 1) { - throw std::runtime_error(format("write error: %s", strerror(errno))); - } - } - - void write_u32(std::uint32_t val) { - write_raw(&val, sizeof(val)); - } - - ~llama_file() { - if (fp) { - std::fclose(fp); - } - } -}; - int tokenize_file(struct llama_context * lctx, const char * filename, std::vector& out) { - struct llama_file f(filename, "rb"); + FILE * fp = std::fopen(filename, "rb"); + if (fp == NULL) { + return 0; + } + +#ifdef _WIN32 + GGML_ASSERT(_fseeki64(fp, (__int64) 0, SEEK_END) == 0); +#else + GGML_ASSERT(std::fseek(fp, (long) 0, SEEK_END) == 0); +#endif + + size_t size = 0; +#ifdef _WIN32 + __int64 ret = _ftelli64(fp); + size = ret; +#else + long ret = std::ftell(fp); + size = ret; +#endif + +#ifdef _WIN32 + GGML_ASSERT(_fseeki64(fp, (__int64) 0, SEEK_SET) == 0); +#else + GGML_ASSERT(std::fseek(fp, (long) 0, SEEK_SET) == 0); +#endif std::vector buf; - buf.resize(f.size+1); + buf.resize(size+1); + out.resize(size+1); - f.read_raw(buf.data(), f.size); - buf[f.size] = '\0'; + if (std::fread(buf.data(), size, 1, fp) != 1) { + throw std::runtime_error(std::string("unexpectedly reached end of file")); + } + if (ferror(fp)) { + throw std::runtime_error(format("read error: %s", strerror(errno))); + } + + buf[size] = '\0'; int n_tokens = llama_tokenize(lctx, buf.data(), out.data(), out.size(), false); if (n_tokens < 0) { out.resize(-n_tokens); - llama_tokenize(lctx, buf.data(), out.data(), out.size(), false); + n_tokens = llama_tokenize(lctx, buf.data(), out.data(), out.size(), false); } + GGML_ASSERT(n_tokens >= 0); + out.resize(n_tokens); bool verify = false; if (verify) { @@ -2238,438 +1040,466 @@ void shuffle_ints(int * begin, int * end) { }); } -struct my_llama_sampler_params { - float temp = 0.0f; // <= 0.0 disabled - int top_k = 20; // <= 0 to use vocab size - float top_p = 0.95f; // 1.0 = disabled - float tfs_z = 1.00f; // 1.0 = disabled - float typical_p = 1.00f; // 1.0 = disabled - int repeat_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size) - float repeat_penalty = 1.0f; // 1.0 = disabled - float alpha_presence = 0.0f; // 0.0 = disabled - float alpha_frequency = 0.0f; // 0.0 = disabled - int mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0 - float mirostat_tau = 5.00f; // target entropy - float mirostat_eta = 0.10f; // learning rate - bool penalize_nl = true; // consider newlines as a repeatable token -}; - -struct my_llama_sampler { - struct llama_context * ctx = NULL; - my_llama_sampler_params params; - - int n_vocab = 0; - int n_ctx = 0; - - float mirostat_mu; - - std::vector candidates; - llama_token_data_array candidates_p; - -}; - -void init_sampler(struct my_llama_sampler * sampler, struct llama_context * ctx) { - sampler->ctx = ctx; - sampler->n_vocab = llama_n_vocab(sampler->ctx); - sampler->n_ctx = llama_n_ctx(sampler->ctx); - sampler->mirostat_mu = 2.0f * sampler->params.mirostat_tau; +#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \ +{ \ + const std::string skey(key); \ + const int kid = gguf_find_key(ctx, skey.c_str()); \ + if (kid >= 0) { \ + enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \ + if (ktype != (type)) { \ + throw std::runtime_error(format("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype))); \ + } \ + (dst) = func(ctx, kid); \ + } else if (req) { \ + throw std::runtime_error(format("key not found in model: %s", skey.c_str())); \ + } \ } -llama_token sample(struct my_llama_sampler * sampler, float * logits, const llama_token * last_tokens, int n_last_tokens) { - GGML_ASSERT(sampler->ctx != NULL); - struct llama_context * ctx = sampler->ctx; +bool are_same_layout(struct ggml_tensor * a, struct ggml_tensor * b) { + GGML_ASSERT(a != NULL); + GGML_ASSERT(b != NULL); + GGML_ASSERT(a->type == b->type); + GGML_ASSERT(ggml_are_same_shape(a, b)); + GGML_ASSERT(ggml_is_contiguous(a) && ggml_is_contiguous(b)); - sampler->candidates.resize(sampler->n_vocab); - for (llama_token token_id = 0; token_id < sampler->n_vocab; ++token_id) { - sampler->candidates[token_id].id = token_id; - sampler->candidates[token_id].logit = logits[token_id]; - sampler->candidates[token_id].p = 0.0; + return true; +} + +void read_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name) { + if (dst == NULL) { + return; } + struct ggml_tensor * t = ggml_get_tensor(ctx, name); + GGML_ASSERT(are_same_layout(dst, t)); + memcpy(dst->data, t->data, ggml_nbytes(t)); - llama_token_data_array * candidates_p = & sampler->candidates_p; - - candidates_p->data = sampler->candidates.data(); - candidates_p->size = sampler->candidates.size(); - candidates_p->sorted = false; - - const auto params = sampler->params; - - // Apply penalties - const float nl_logit = logits[llama_token_nl(ctx)]; - - const int n_last = std::min(std::min(n_last_tokens, params.repeat_last_n), sampler->n_ctx); - - llama_sample_repetition_penalty( - ctx, - candidates_p, - last_tokens + n_last_tokens - n_last, - n_last, - params.repeat_penalty); - llama_sample_frequency_and_presence_penalties( - ctx, - candidates_p, - last_tokens + n_last_tokens - n_last, - n_last, - params.alpha_frequency, - params.alpha_presence); - - if (!params.penalize_nl) { - logits[llama_token_nl(ctx)] = nl_logit; + if (strlen(ggml_get_name(dst)) == 0) { + ggml_set_name(dst, name); } +} - llama_token token = 0; - if (params.temp <= 0) { - // Greedy sampling - token = llama_sample_token_greedy(ctx, candidates_p); +void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt) { + // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read + + uint32_t file_version; + GGUF_GET_KEY(fctx, file_version, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_FILE_VERSION); + GGML_ASSERT(file_version == 0); + + GGUF_GET_KEY(fctx, opt->params.past, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT); + GGUF_GET_KEY(fctx, opt->iter, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_ITERATION_COUNT); + GGUF_GET_KEY(fctx, opt->just_initialized, gguf_get_val_bool, GGUF_TYPE_BOOL, true, LLM_KV_OPTIMIZER_JUST_INITIALIZED); + + uint64_t nx; + GGUF_GET_KEY(fctx, nx, gguf_get_val_u64, GGUF_TYPE_UINT64, true, LLM_KV_OPTIMIZER_PARAMETER_COUNT); + opt->nx = (size_t) nx; + + // don't call ggml_opt_init until optimizer type and optimizer specific parameters are know + + std::string opt_type; + GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE); + if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) { + opt->params.type = GGML_OPT_ADAM; + + GGUF_GET_KEY(fctx, opt->adam.fx_best, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS); + GGUF_GET_KEY(fctx, opt->adam.fx_prev, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS); + GGUF_GET_KEY(fctx, opt->adam.n_no_improvement, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT); + + GGML_ASSERT(opt->ctx != NULL); + ggml_opt_init(opt->ctx, opt, opt->params, opt->nx); + + read_tensor_by_name(opt->adam.m, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS); + read_tensor_by_name(opt->adam.v, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS); + read_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES); + } else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) { + opt->params.type = GGML_OPT_LBFGS; + + GGUF_GET_KEY(fctx, opt->params.lbfgs.m, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT); + GGUF_GET_KEY(fctx, opt->lbfgs.fx_best, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS); + GGUF_GET_KEY(fctx, opt->lbfgs.step, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP); + GGUF_GET_KEY(fctx, opt->lbfgs.j, gguf_get_val_i32, GGUF_TYPE_INT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J); + GGUF_GET_KEY(fctx, opt->lbfgs.k, gguf_get_val_i32, GGUF_TYPE_INT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K); + GGUF_GET_KEY(fctx, opt->lbfgs.end, gguf_get_val_i32, GGUF_TYPE_INT32, true, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END); + GGUF_GET_KEY(fctx, opt->lbfgs.n_no_improvement, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT); + + GGML_ASSERT(opt->ctx != NULL); + ggml_opt_init(opt->ctx, opt, opt->params, opt->nx); + + read_tensor_by_name(opt->lbfgs.x, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS); + read_tensor_by_name(opt->lbfgs.xp, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS); + read_tensor_by_name(opt->lbfgs.g, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS); + read_tensor_by_name(opt->lbfgs.gp, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS); + read_tensor_by_name(opt->lbfgs.d, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION); + read_tensor_by_name(opt->lbfgs.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES); + read_tensor_by_name(opt->lbfgs.lmal, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA); + read_tensor_by_name(opt->lbfgs.lmys, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS); + read_tensor_by_name(opt->lbfgs.lms, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S); + read_tensor_by_name(opt->lbfgs.lmy, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y); } else { - if (params.mirostat == 1) { - int mirostat_m = 100; - llama_sample_temperature(ctx, candidates_p, params.temp); - token = llama_sample_token_mirostat(ctx, candidates_p, params.mirostat_tau, params.mirostat_eta, mirostat_m, &sampler->mirostat_mu); - } else if (params.mirostat == 2) { - llama_sample_temperature(ctx, candidates_p, params.temp); - token = llama_sample_token_mirostat_v2(ctx, candidates_p, params.mirostat_tau, params.mirostat_eta, &sampler->mirostat_mu); - } else { - // Temperature sampling - llama_sample_top_k (ctx, candidates_p, params.top_k, 1); - llama_sample_tail_free (ctx, candidates_p, params.tfs_z, 1); - llama_sample_typical (ctx, candidates_p, params.typical_p, 1); - - llama_sample_top_p (ctx, candidates_p, params.top_p, 1); - llama_sample_temperature (ctx, candidates_p, params.temp); - token = llama_sample_token(ctx, candidates_p); - } - } - return token; -} - -void set_logits_masked(struct ggml_tensor * logits, std::vector& mask, float value) { - GGML_ASSERT(logits->ne[0] == (int64_t) mask.size()); - for (int i2 = 0; i2 < logits->ne[2]; ++i2) { - for (int i1 = 0; i1 < logits->ne[1]; ++i1) { - for (int i0 = 0; i0 < logits->ne[0]; ++i0) { - if (!mask[i0]) continue; - float * ptr = (float *) ((char *) logits->data + i2*logits->nb[2] + i1*logits->nb[1] + i0*logits->nb[0]); - *ptr = value; - } - } + throw std::runtime_error("unknown optimizer type\n"); } } -void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) { - if (tensor == NULL) { - file->write_u32(0); - file->write_u32(0); - file->write_u32(GGML_TYPE_F32); - file->seek((0-file->tell()) & 31, SEEK_CUR); - return; - } - const char * name = ggml_get_name(tensor); - uint32_t name_len = strlen(name); - uint32_t nd = tensor->n_dims; - uint32_t ne[4] = { (uint32_t)tensor->ne[0], - (uint32_t)tensor->ne[1], - (uint32_t)tensor->ne[2], - (uint32_t)tensor->ne[3] }; - file->write_u32(nd); - file->write_u32(name_len); - file->write_u32(tensor->type); - file->write_raw(ne, sizeof(ne[0]) * nd); - file->write_raw(name, name_len); - file->seek((0-file->tell()) & 31, SEEK_CUR); - file->write_raw(tensor->data, ggml_nbytes(tensor)); -} - -void read_tensor(struct llama_file * file, struct ggml_tensor * tensor) { - int32_t nd = file->read_u32(); - GGML_ASSERT(nd == tensor->n_dims); - - uint32_t name_len = file->read_u32(); - enum ggml_type type = (enum ggml_type) file->read_u32(); - GGML_ASSERT(type == tensor->type); - - uint32_t ne[4]; - file->read_raw(ne, sizeof(ne[0]) * nd); - for (int i=0; ine[i]); - } - - std::string name = file->read_string(name_len); - GGML_ASSERT(strncmp(ggml_get_name(tensor), name.c_str(), sizeof(tensor->name)-1) == 0); - - file->seek((0-file->tell()) & 31, SEEK_CUR); - file->read_raw(tensor->data, ggml_nbytes(tensor)); -} - -void write_opt_context(struct llama_file * file, struct ggml_opt_context * opt) { - const uint32_t version = 0; - GGML_ASSERT(opt->nx >= 0); - GGML_ASSERT(opt->iter >= 0); - file->write_u32(version); - file->write_raw(&opt->params, sizeof(opt->params)); - file->write_raw(&opt->nx, sizeof(opt->nx)); - file->write_raw(&opt->iter, sizeof(opt->iter)); - file->write_u32((uint32_t) opt->just_initialized); - switch (opt->params.type) { - case GGML_OPT_ADAM: - { - GGML_ASSERT(opt->adam.x != NULL); - write_tensor(file, opt->adam.x); - write_tensor(file, opt->adam.g1); - write_tensor(file, opt->adam.g2); - write_tensor(file, opt->adam.m); - write_tensor(file, opt->adam.v); - write_tensor(file, opt->adam.mh); - write_tensor(file, opt->adam.vh); - write_tensor(file, opt->adam.pf); - file->write_raw(&opt->adam.fx_best, sizeof(opt->adam.fx_best)); - file->write_raw(&opt->adam.fx_prev, sizeof(opt->adam.fx_prev)); - file->write_raw(&opt->adam.n_no_improvement, sizeof(opt->adam.n_no_improvement)); - } break; - case GGML_OPT_LBFGS: - { - GGML_ASSERT(opt->adam.x != NULL); - write_tensor(file, opt->lbfgs.x); - write_tensor(file, opt->lbfgs.xp); - write_tensor(file, opt->lbfgs.g); - write_tensor(file, opt->lbfgs.gp); - write_tensor(file, opt->lbfgs.d); - write_tensor(file, opt->lbfgs.pf); - write_tensor(file, opt->lbfgs.lmal); - write_tensor(file, opt->lbfgs.lmys); - write_tensor(file, opt->lbfgs.lms); - write_tensor(file, opt->lbfgs.lmy); - file->write_raw(&opt->lbfgs.fx_best, sizeof(opt->lbfgs.fx_best)); - file->write_raw(&opt->lbfgs.step, sizeof(opt->lbfgs.step)); - file->write_raw(&opt->lbfgs.j, sizeof(opt->lbfgs.j)); - file->write_raw(&opt->lbfgs.k, sizeof(opt->lbfgs.k)); - file->write_raw(&opt->lbfgs.end, sizeof(opt->lbfgs.end)); - file->write_raw(&opt->lbfgs.n_no_improvement, sizeof(opt->lbfgs.n_no_improvement)); - } break; - } -} - -void read_opt_context(struct llama_file * file, struct ggml_context * ctx, struct ggml_opt_context * opt) { - uint32_t version = file->read_u32(); - GGML_ASSERT(version == 0); - - file->read_raw(&opt->params, sizeof(opt->params)); - file->read_raw(&opt->nx, sizeof(opt->nx)); - ggml_opt_init(ctx, opt, opt->params, opt->nx); - - file->read_raw(&opt->iter, sizeof(opt->iter)); - opt->just_initialized = (bool) file->read_u32(); +void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt) { + gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_FILE_VERSION, 0); + gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, opt->params.past); + gguf_set_val_u64(fctx, LLM_KV_OPTIMIZER_PARAMETER_COUNT, (uint64_t) opt->nx); + gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_ITERATION_COUNT, opt->iter); + gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized); switch (opt->params.type) { case GGML_OPT_ADAM: { - read_tensor(file, opt->adam.x); - read_tensor(file, opt->adam.g1); - read_tensor(file, opt->adam.g2); - read_tensor(file, opt->adam.m); - read_tensor(file, opt->adam.v); - read_tensor(file, opt->adam.mh); - read_tensor(file, opt->adam.vh); - if (opt->adam.pf) { read_tensor(file, opt->adam.pf); } - file->read_raw(&opt->adam.fx_best, sizeof(opt->adam.fx_best)); - file->read_raw(&opt->adam.fx_prev, sizeof(opt->adam.fx_prev)); - file->read_raw(&opt->adam.n_no_improvement, sizeof(opt->adam.n_no_improvement)); + gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM); + gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, opt->adam.fx_best); + gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, opt->adam.fx_prev); + gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, opt->adam.n_no_improvement); + + ggml_set_name(opt->adam.m, LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS); + ggml_set_name(opt->adam.v, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS); + if (opt->adam.pf) { + ggml_set_name(opt->adam.pf, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES); + } + + gguf_add_tensor(fctx, opt->adam.m); + gguf_add_tensor(fctx, opt->adam.v); + if (opt->adam.pf) { + gguf_add_tensor(fctx, opt->adam.pf); + } } break; case GGML_OPT_LBFGS: { - GGML_ASSERT(opt->adam.x != NULL); - read_tensor(file, opt->lbfgs.x); - read_tensor(file, opt->lbfgs.xp); - read_tensor(file, opt->lbfgs.g); - read_tensor(file, opt->lbfgs.gp); - read_tensor(file, opt->lbfgs.d); - if (opt->lbfgs.pf) { read_tensor(file, opt->lbfgs.pf); } - read_tensor(file, opt->lbfgs.lmal); - read_tensor(file, opt->lbfgs.lmys); - read_tensor(file, opt->lbfgs.lms); - read_tensor(file, opt->lbfgs.lmy); - file->read_raw(&opt->lbfgs.fx_best, sizeof(opt->lbfgs.fx_best)); - file->read_raw(&opt->lbfgs.step, sizeof(opt->lbfgs.step)); - file->read_raw(&opt->lbfgs.j, sizeof(opt->lbfgs.j)); - file->read_raw(&opt->lbfgs.k, sizeof(opt->lbfgs.k)); - file->read_raw(&opt->lbfgs.end, sizeof(opt->lbfgs.end)); - file->read_raw(&opt->lbfgs.n_no_improvement, sizeof(opt->lbfgs.n_no_improvement)); + gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS); + gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m); + gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, opt->lbfgs.fx_best); + gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, opt->lbfgs.step); + gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, opt->lbfgs.j); + gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, opt->lbfgs.k); + gguf_set_val_i32(fctx, LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, opt->lbfgs.end); + gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, opt->lbfgs.n_no_improvement); + + ggml_set_name(opt->lbfgs.x, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS); + ggml_set_name(opt->lbfgs.xp, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS); + ggml_set_name(opt->lbfgs.g, LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS); + ggml_set_name(opt->lbfgs.gp, LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS); + ggml_set_name(opt->lbfgs.d, LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION); + if (opt->lbfgs.pf) { + ggml_set_name(opt->lbfgs.pf, LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES); + } + ggml_set_name(opt->lbfgs.lmal, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA); + ggml_set_name(opt->lbfgs.lmys, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS); + ggml_set_name(opt->lbfgs.lms, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S); + ggml_set_name(opt->lbfgs.lmy, LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y); + + gguf_add_tensor(fctx, opt->lbfgs.x); + gguf_add_tensor(fctx, opt->lbfgs.xp); + gguf_add_tensor(fctx, opt->lbfgs.g); + gguf_add_tensor(fctx, opt->lbfgs.gp); + gguf_add_tensor(fctx, opt->lbfgs.d); + if (opt->lbfgs.pf) { + gguf_add_tensor(fctx, opt->lbfgs.pf); + } + gguf_add_tensor(fctx, opt->lbfgs.lmal); + gguf_add_tensor(fctx, opt->lbfgs.lmys); + gguf_add_tensor(fctx, opt->lbfgs.lms); + gguf_add_tensor(fctx, opt->lbfgs.lmy); } break; } } -void save_checkpoint(struct my_llama_model * model, struct ggml_opt_context * opt, const char * filename) { - struct llama_file file(filename, "wb"); - if (file.fp == NULL) { - return; +void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model) { + // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read + std::string arch; + + std::vector keybuf; + keybuf.resize(512); + auto kv = [&arch, &keybuf](const char * key) -> const char * { + snprintf(keybuf.data(), keybuf.size(), key, arch.c_str()); + return keybuf.data(); + }; + + std::vector tn_buf; + tn_buf.resize(GGML_MAX_NAME); + auto tn = [&tn_buf](const char * key) -> const char * { + snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key); + return tn_buf.data(); + }; + auto tni = [&tn_buf](const char * key, int bid) -> const char * { + snprintf(tn_buf.data(), tn_buf.size(), key, bid); + std::string s = tn_buf.data(); + snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str()); + return tn_buf.data(); + }; + + GGUF_GET_KEY(fctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE); + GGML_ASSERT(arch == "llama"); + + uint32_t ftype_u; + GGUF_GET_KEY(fctx, ftype_u, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_GENERAL_FILE_TYPE); + GGML_ASSERT((enum llama_ftype) ftype_u == LLAMA_FTYPE_ALL_F32); + + // n_ctx was not saved in earlier checkpoint file versions, so we make it optional here + GGUF_GET_KEY(fctx, model->hparams.n_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH)); + + GGUF_GET_KEY(fctx, model->hparams.n_embd, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH)); + GGUF_GET_KEY(fctx, model->hparams.n_ff, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH)); + GGUF_GET_KEY(fctx, model->hparams.n_head, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT)); + GGUF_GET_KEY(fctx, model->hparams.n_layer, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT)); + + model->hparams.n_rot = model->hparams.n_embd / model->hparams.n_head; + GGUF_GET_KEY(fctx, model->hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT)); + + float rope_freq_scale = 1.0f; + GGUF_GET_KEY(fctx, model->hparams.f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); + GGUF_GET_KEY(fctx, model->hparams.rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE)); + GGUF_GET_KEY(fctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); + if (rope_freq_scale != 1.0f) { + model->hparams.rope_freq_scale = 1.0f / rope_freq_scale; } - const uint32_t magic = 'ggcp'; - const uint32_t version = 0; + init_model(model); - file.write_u32(magic); - file.write_u32(version); - file.write_u32(model->train_its); - file.write_u32(model->train_samples); - file.write_u32(model->train_tokens); - file.write_u32(model->hparams.n_vocab); - file.write_u32(model->hparams.n_embd); - file.write_u32(model->hparams.n_mult); - file.write_u32(model->hparams.n_head); - file.write_u32(model->hparams.n_layer); - file.write_u32(model->hparams.n_rot); - - write_tensor(&file, model->tok_embeddings); - write_tensor(&file, model->norm); - write_tensor(&file, model->output); + read_tensor_by_name(model->tok_embeddings, f_ggml_ctx, tn(LLM_TENSOR_TOKEN_EMBD)); + read_tensor_by_name(model->norm, f_ggml_ctx, tn(LLM_TENSOR_OUTPUT_NORM)); + read_tensor_by_name(model->output, f_ggml_ctx, tn(LLM_TENSOR_OUTPUT)); for (uint32_t i = 0; i < model->hparams.n_layer; ++i) { auto & layer = model->layers[i]; - write_tensor(&file, layer.attention_norm); - write_tensor(&file, layer.wq); - write_tensor(&file, layer.wk); - write_tensor(&file, layer.wv); - write_tensor(&file, layer.wo); - write_tensor(&file, layer.ffn_norm); - write_tensor(&file, layer.w1); - write_tensor(&file, layer.w2); - write_tensor(&file, layer.w3); + read_tensor_by_name(layer.attention_norm, f_ggml_ctx, tni(LLM_TENSOR_ATTN_NORM, i)); + read_tensor_by_name(layer.wq, f_ggml_ctx, tni(LLM_TENSOR_ATTN_Q, i)); + read_tensor_by_name(layer.wk, f_ggml_ctx, tni(LLM_TENSOR_ATTN_K, i)); + read_tensor_by_name(layer.wv, f_ggml_ctx, tni(LLM_TENSOR_ATTN_V, i)); + read_tensor_by_name(layer.wo, f_ggml_ctx, tni(LLM_TENSOR_ATTN_OUT, i)); + read_tensor_by_name(layer.ffn_norm, f_ggml_ctx, tni(LLM_TENSOR_FFN_NORM, i)); + read_tensor_by_name(layer.w1, f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i)); + read_tensor_by_name(layer.w2, f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i)); + read_tensor_by_name(layer.w3, f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i)); } - - write_opt_context(&file, opt); } -bool load_checkpoint(struct my_llama_model * model, struct ggml_opt_context * opt, const char * filename, bool init) { - struct llama_file file(filename, "rb"); +void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model) { + const char * arch = "llama"; + enum llama_ftype ftype = LLAMA_FTYPE_ALL_F32; - uint32_t magic; - uint32_t version; + std::vector keybuf; + keybuf.resize(512); + auto kv = [arch, &keybuf](const char * key) -> const char * { + snprintf(keybuf.data(), keybuf.size(), key, arch); + return keybuf.data(); + }; - uint32_t train_its = 0; - uint32_t train_samples = 0; - uint32_t train_tokens = 0; + // set arch + gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch); + gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype); - if (file.fp) { - printf("%s: Loading model from '%s'.\n", __func__, filename); - magic = file.read_u32(); - GGML_ASSERT(magic == 'ggcp'); - version = file.read_u32(); - GGML_ASSERT(version == 0); - train_its = file.read_u32(); - train_samples = file.read_u32(); - train_tokens = file.read_u32(); - model->hparams.n_vocab = file.read_u32(); - model->hparams.n_embd = file.read_u32(); - model->hparams.n_mult = file.read_u32(); - model->hparams.n_head = file.read_u32(); - model->hparams.n_layer = file.read_u32(); - model->hparams.n_rot = file.read_u32(); - print_params(&model->hparams); - } + // set hparams + gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH), model->hparams.n_ctx ); + gguf_set_val_u32(fctx, kv(LLM_KV_EMBEDDING_LENGTH), model->hparams.n_embd ); + gguf_set_val_u32(fctx, kv(LLM_KV_FEED_FORWARD_LENGTH), model->hparams.n_ff ); + gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT), model->hparams.n_head ); + gguf_set_val_u32(fctx, kv(LLM_KV_BLOCK_COUNT), model->hparams.n_layer ); + gguf_set_val_u32(fctx, kv(LLM_KV_ROPE_DIMENSION_COUNT), model->hparams.n_rot ); - if (init) { - init_model(model); - } + gguf_set_val_f32(fctx, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS), model->hparams.f_norm_rms_eps ); + gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_FREQ_BASE), model->hparams.rope_freq_base ); // TODO load in llama.cpp + gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_SCALE_LINEAR), 1.0f / model->hparams.rope_freq_scale ); - if (file.fp) { - model->train_its = train_its; - model->train_samples = train_samples; - model->train_tokens = train_tokens; - } + // set vocab by copying from vocab_model gguf file + { + struct gguf_init_params params = { + /*.no_alloc = */ false, + /*.ctx = */ NULL, + }; + struct gguf_context * vctx = gguf_init_from_file(fn_vocab_model, params); - printf("%s: Training iterations: %u.\n", __func__, model->train_its); - printf("%s: Training samples: %u.\n", __func__, model->train_samples); - printf("%s: Training tokens: %u.\n", __func__, model->train_tokens); + const int token_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_LIST)); + if (token_idx == -1) { + throw std::runtime_error("cannot find tokenizer vocab in model file\n"); + } + const uint32_t n_vocab = gguf_get_arr_n(vctx, token_idx); - if (file.fp) { - read_tensor(&file, model->tok_embeddings); - read_tensor(&file, model->norm); - read_tensor(&file, model->output); - - for (uint32_t i = 0; i < model->hparams.n_layer; ++i) { - auto & layer = model->layers[i]; - - read_tensor(&file, layer.attention_norm); - read_tensor(&file, layer.wq); - read_tensor(&file, layer.wk); - read_tensor(&file, layer.wv); - read_tensor(&file, layer.wo); - read_tensor(&file, layer.ffn_norm); - read_tensor(&file, layer.w1); - read_tensor(&file, layer.w2); - read_tensor(&file, layer.w3); + const int score_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_SCORES)); + if (score_idx == -1) { + throw std::runtime_error("cannot find tokenizer scores in model file\n"); } - read_opt_context(&file, model->ctx, opt); + const float * scores = (const float * ) gguf_get_arr_data(vctx, score_idx); + + const int toktype_idx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE)); + if (toktype_idx == -1) { + throw std::runtime_error("cannot find token type list in GGUF file\n"); + } + + const int * toktypes = (const int * ) gguf_get_arr_data(vctx, toktype_idx); + + std::string tokenizer_name; + GGUF_GET_KEY(vctx, tokenizer_name, gguf_get_val_str, GGUF_TYPE_STRING, true, kv(LLM_KV_TOKENIZER_MODEL)); + + gguf_set_val_str(fctx, kv(LLM_KV_TOKENIZER_MODEL), tokenizer_name.c_str()); + gguf_set_arr_data(fctx, kv(LLM_KV_TOKENIZER_SCORES), GGUF_TYPE_FLOAT32, scores, n_vocab); + gguf_set_arr_data(fctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE), GGUF_TYPE_INT32, toktypes, n_vocab); + + int32_t special_bos_id = 1; + int32_t special_eos_id = 2; + int32_t special_unk_id = 0; + int32_t special_sep_id = -1; + int32_t special_pad_id = -1; + if (tokenizer_name == "llama") { + // default special tokens + special_bos_id = 1; + special_eos_id = 2; + special_unk_id = 0; + special_sep_id = -1; + special_pad_id = -1; + } else if (tokenizer_name == "gpt2") { + // read and copy bpe merges + const int merges_keyidx = gguf_find_key(vctx, kv(LLM_KV_TOKENIZER_MERGES)); + if (merges_keyidx == -1) { + throw std::runtime_error("cannot find tokenizer merges in model file\n"); + } + + const int n_merges = gguf_get_arr_n(vctx, merges_keyidx); + + std::vector merges; + merges.resize(n_merges); + for (int i = 0; i < n_merges; i++) { + merges[i] = gguf_get_arr_str(vctx, merges_keyidx, i); + } + gguf_set_arr_str(fctx, kv(LLM_KV_TOKENIZER_MERGES), merges.data(), n_merges); + + // default special tokens + special_bos_id = 11; + special_eos_id = 11; + special_unk_id = -1; + special_sep_id = -1; + special_pad_id = -1; + } else { + fprintf(stderr, "%s: unknown tokenizer: '%s'", __func__, tokenizer_name.c_str()); + fprintf(stderr, "%s: using default tokenizer: 'llama'", __func__); + } + + std::vector tokens; + tokens.resize(n_vocab); + for (uint32_t i = 0; i < n_vocab; i++) { + tokens[i] = gguf_get_arr_str(vctx, token_idx, i); + } + gguf_set_arr_str(fctx, kv(LLM_KV_TOKENIZER_LIST), tokens.data(), n_vocab); + + GGUF_GET_KEY(vctx, special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID)); + GGUF_GET_KEY(vctx, special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID)); + GGUF_GET_KEY(vctx, special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID)); + GGUF_GET_KEY(vctx, special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID)); + GGUF_GET_KEY(vctx, special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID)); + + gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_BOS_ID), special_bos_id); + gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_EOS_ID), special_eos_id); + gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_UNK_ID), special_unk_id); + gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_SEP_ID), special_sep_id); + gguf_set_val_u32(fctx, kv(LLM_KV_TOKENIZER_PAD_ID), special_pad_id); + + gguf_free(vctx); } - return (file.fp != NULL); + // add tensors + gguf_add_tensor(fctx, model->tok_embeddings); + gguf_add_tensor(fctx, model->norm); + gguf_add_tensor(fctx, model->output); + for (uint32_t i = 0; i < model->hparams.n_layer; ++i) { + auto & layer = model->layers[i]; + + + gguf_add_tensor(fctx, layer.attention_norm); + gguf_add_tensor(fctx, layer.wq); + gguf_add_tensor(fctx, layer.wk); + gguf_add_tensor(fctx, layer.wv); + gguf_add_tensor(fctx, layer.wo); + gguf_add_tensor(fctx, layer.ffn_norm); + gguf_add_tensor(fctx, layer.w1); + gguf_add_tensor(fctx, layer.w2); + gguf_add_tensor(fctx, layer.w3); + } } -void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, const char * filename) { - struct llama_file file(filename, "wb"); - if (file.fp == NULL) { - return; +void save_llama_model_file(const char * filename, const char * fn_vocab_model, struct my_llama_model * model) { + struct gguf_context * fctx = gguf_init_empty(); + + save_llama_model_gguf(fctx, fn_vocab_model, model); + + // write file + const bool only_meta = false; + gguf_write_to_file(fctx, filename, only_meta); + gguf_free(fctx); +} + +void load_checkpoint_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct ggml_opt_context * opt) { + load_llama_model_gguf(fctx, f_ggml_ctx, model); + + uint32_t file_version; + GGUF_GET_KEY(fctx, file_version, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_FILE_VERSION); + GGML_ASSERT(file_version == 0); + + GGUF_GET_KEY(fctx, model->train_its, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_ITERATION_COUNT); + GGUF_GET_KEY(fctx, model->train_samples, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_SAMPLE_COUNT); + GGUF_GET_KEY(fctx, model->train_tokens, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_TOKEN_COUNT); + + load_opt_context_gguf(fctx, f_ggml_ctx, opt); +} + +void save_checkpoint_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model, struct ggml_opt_context * opt) { + save_llama_model_gguf(fctx, fn_vocab_model, model); + + gguf_set_val_u32(fctx, LLM_KV_TRAINING_FILE_VERSION, 0); + gguf_set_val_u32(fctx, LLM_KV_TRAINING_ITERATION_COUNT, model->train_its); + gguf_set_val_u32(fctx, LLM_KV_TRAINING_SAMPLE_COUNT, model->train_samples); + gguf_set_val_u32(fctx, LLM_KV_TRAINING_TOKEN_COUNT, model->train_tokens); + + save_opt_context_gguf(fctx, opt); +} + +bool load_checkpoint_file(const char * filename, struct my_llama_model * model, struct ggml_opt_context * opt) { + struct ggml_context * f_ggml_ctx; + struct gguf_init_params params; + params.no_alloc = false; + params.ctx = &f_ggml_ctx; + struct gguf_context * fctx = gguf_init_from_file(filename, params); + if (fctx == NULL) { + return false; } -#pragma message("TODO: implement file saving using gguf") - (void) vocab; - (void) model; -// // write_magic -// file.write_u32(LLAMA_FILE_MAGIC); // magic -// file.write_u32(LLAMA_FILE_VERSION); // version -// // write_hparams -// file.write_u32(model->hparams.n_vocab); -// file.write_u32(model->hparams.n_embd); -// file.write_u32(model->hparams.n_mult); -// file.write_u32(model->hparams.n_head); -// file.write_u32(model->hparams.n_layer); -// file.write_u32(model->hparams.n_rot); -// file.write_u32(LLAMA_FTYPE_ALL_F32); -// // write_vocab -// uint32_t n_vocab = model->hparams.n_vocab; -// for (uint32_t i = 0; i < n_vocab; i++) { -// const auto & token_data = vocab->id_to_token.at(i); -// file.write_u32((uint32_t) token_data.tok.size()); -// file.write_raw(token_data.tok.data(), token_data.tok.size()); -// file.write_raw(&token_data.score, sizeof(token_data.score)); -// } -// // write tensors -// write_tensor(&file, model->tok_embeddings); -// write_tensor(&file, model->norm); -// write_tensor(&file, model->output); -// for (uint32_t i = 0; i < model->hparams.n_layer; ++i) { -// auto & layer = model->layers[i]; -// -// write_tensor(&file, layer.attention_norm); -// write_tensor(&file, layer.wq); -// write_tensor(&file, layer.wk); -// write_tensor(&file, layer.wv); -// write_tensor(&file, layer.wo); -// write_tensor(&file, layer.ffn_norm); -// write_tensor(&file, layer.w1); -// write_tensor(&file, layer.w2); -// write_tensor(&file, layer.w3); -// } + load_checkpoint_gguf(fctx, f_ggml_ctx, model, opt); + + return true; } -float cosine_decay(const int decay_steps, const float alpha, int step) { +void save_checkpoint_file(const char * filename, const char * fn_vocab_model, struct my_llama_model * model, struct ggml_opt_context * opt) { + struct gguf_context * fctx = gguf_init_empty(); + + save_checkpoint_gguf(fctx, fn_vocab_model, model, opt); + + // write file + const bool only_meta = false; + gguf_write_to_file(fctx, filename, only_meta); + gguf_free(fctx); +} + +float cosine_decay(const int decay_steps, const float minimum, int step) { if (step > decay_steps) { step = decay_steps; } const float cosine_decay = 0.50f*(1.0f + cosf(3.14159265359f*step/decay_steps)); - const float decay = (1 - alpha)*cosine_decay + alpha; + const float decay = (1 - minimum)*cosine_decay + minimum; return decay; } -float cosine_decay_restart(int decay_steps, const float alpha, int step, float restart_step_mult) { - while (step > decay_steps) { - step -= decay_steps; - decay_steps = (int) restart_step_mult * decay_steps; +float cosine_decay_restart(int decay_steps, const float minimum, int step, float restart_step_mult, bool enable_restart) { + if (enable_restart) { + while (step > decay_steps) { + step -= decay_steps; + decay_steps = (int) restart_step_mult * decay_steps; + } } - return cosine_decay(decay_steps, alpha, step); + return cosine_decay(decay_steps, minimum, step); } struct train_params { @@ -2683,39 +1513,51 @@ struct train_params { int n_ctx; int n_embd; - int n_mult; int n_head; int n_layer; - int n_rotmax; + int n_ff; int n_threads; int n_batch; int n_examples; - int n_predict; + + float f_norm_rms_eps; + float rope_freq_base; + float rope_freq_scale; int print_info_interval; - int print_details_interval; bool samples_start_after_nl; bool use_adam; bool use_flash; - bool use_scratch; + bool use_checkpointing; + bool use_alloc; // only adam int warmup; int cos_decay_steps; float cos_decay_restart; - float cos_decay_alpha; + float cos_decay_min; + bool enable_restart; + + int opt_past; + float opt_delta; + int opt_max_no_improvement; int lbfgs_n_iter; int adam_n_iter; float adam_alpha; + float adam_min_alpha; float adam_decay; + int adam_decay_min_ndim; + float adam_beta1; + float adam_beta2; + float adam_gclip; + float adam_eps_f; int mem_model_gb; int mem_compute_gb; int mem_compute0_gb; - int mem_compute1_gb; }; struct train_params get_default_train_params() { @@ -2730,40 +1572,51 @@ struct train_params get_default_train_params() { params.n_ctx = 128; params.n_embd = 256; - params.n_mult = 256; params.n_head = 8; params.n_layer = 16; - params.n_rotmax = 64; + params.n_ff = 768; params.n_threads = 6; params.n_batch = 8; - params.n_examples = 8; - params.n_predict = 1024; + params.n_examples = 1; + + params.f_norm_rms_eps = 1e-5; + params.rope_freq_base = 10000.0f; + params.rope_freq_scale = 1.0f; params.print_info_interval = 1; - params.print_details_interval = 2; params.samples_start_after_nl = false; params.use_adam = true; params.use_flash = true; - params.use_scratch = true; + params.use_checkpointing = true; + params.use_alloc = true; + + params.opt_past = 0; + params.opt_delta = 1e-5f; + params.opt_max_no_improvement = 0; // only adam params.warmup = 100; params.cos_decay_steps = 1000; params.cos_decay_restart = 1.1f; - params.cos_decay_alpha = 0.0f; + params.cos_decay_min = 0.1f; + params.enable_restart = false; - params.lbfgs_n_iter = 16; - params.adam_n_iter = 16; - params.adam_alpha = 1e-3f; - params.adam_decay = 1e-3f; + params.lbfgs_n_iter = 256; + params.adam_n_iter = 256; + params.adam_alpha = 1e-3f; + params.adam_min_alpha = 0; + params.adam_decay = 1e-1f; + params.adam_decay_min_ndim = 2; + params.adam_beta1 = 0.9f; + params.adam_beta2 = 0.999f; + params.adam_gclip = 1.0f; + params.adam_eps_f = 0.0f; - params.mem_model_gb = 2; + params.mem_model_gb = 2; params.mem_compute_gb = 24; params.mem_compute0_gb = 8; - params.mem_compute1_gb = 2; - return params; } @@ -2780,35 +1633,47 @@ void train_print_usage(int /*argc*/, char ** argv, const struct train_params * p fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for -1)\n"); fprintf(stderr, " -c N, --ctx N Context size used during training (default %d)\n", params->n_ctx); fprintf(stderr, " --embd N Embedding size used for new models (default %d)\n", params->n_embd); - fprintf(stderr, " --mult N Mult size used for new models, influences feedforward size. (default %d)\n", params->n_mult); + fprintf(stderr, " --ff N Feedforward size used for new models. (default %d)\n", params->n_ff); fprintf(stderr, " --head N Number of heads for new models (default %d)\n", params->n_head); fprintf(stderr, " --layer N Number of layers for new models (default %d)\n", params->n_layer); - fprintf(stderr, " --rotmax N Maximal number Rope dimensions for new models (default %d)\n", params->n_rotmax); + fprintf(stderr, " --norm-rms-eps F RMS-Norm epsilon value (default %f)\n", params->f_norm_rms_eps); + fprintf(stderr, " --rope-freq-base F Frequency base for ROPE (default %f)\n", params->rope_freq_base); + fprintf(stderr, " --rope-freq-scale F Frequency scale for ROPE (default %f)\n", params->rope_freq_scale); fprintf(stderr, " -t N, --threads N Number of threads (default %d)\n", params->n_threads); fprintf(stderr, " -b N, --batch N Parallel batch size (default %d)\n", params->n_batch); fprintf(stderr, " -n N, --examples N Number of examples to train (default %d)\n", params->n_examples); - fprintf(stderr, " --predict N Number of tokens to generate after training (default %d)\n", params->n_predict); fprintf(stderr, " --print-info-interval N Print infos during training each N examples (default %d)\n", params->print_info_interval); - fprintf(stderr, " --print-details-interval N Print details during training each N examples (default %d)\n", params->print_details_interval); fprintf(stderr, " --samples-after-nl Training samples start after newlines. (default %s)\n", params->samples_start_after_nl ? "on" : "off"); fprintf(stderr, " --use-lbfgs Use LBFGS optimizer instead of default Adam\n"); fprintf(stderr, " --use-adam Use Adam optimizer (default)\n"); - fprintf(stderr, " --no-flash Don't use flash attention.\n"); + fprintf(stderr, " --no-flash Don't use flash attention \n"); fprintf(stderr, " --use-flash Use flash attention (default)\n"); - fprintf(stderr, " --no-scratch Don't use scratch buffers\n"); - fprintf(stderr, " --use-scratch Use scratch buffers (default)\n"); - fprintf(stderr, " --warmup N Number of warmup steps (default %d)\n", params->warmup); - fprintf(stderr, " --cos-decay-steps N Number of cosine decay steps (default %d)\n", params->cos_decay_steps); - fprintf(stderr, " --cos-decay-restart N Increase of cosine decay steps after restart (default %f)\n", params->cos_decay_restart); - fprintf(stderr, " --cos-decay-alpha N Cosine decay alpha (default %f)\n", params->cos_decay_alpha); - fprintf(stderr, " --lbfgs-iter N Maximum number of LBFGS optimization iterations for each batch (default %d)\n", params->lbfgs_n_iter); + fprintf(stderr, " --no-checkpointing Don't use gradient checkpointing\n"); + fprintf(stderr, " --use-checkpointing Use gradient checkpointing (default)\n"); + fprintf(stderr, " --no-alloc Don't use allocator\n"); + fprintf(stderr, " --use-alloc Use allocator (default)\n"); + fprintf(stderr, " --warmup N Only for Adam optimizer. Number of warmup steps (default %d)\n", params->warmup); + fprintf(stderr, " --cos-decay-steps N Only for Adam optimizer. Number of cosine decay steps (default %d)\n", params->cos_decay_steps); + fprintf(stderr, " --cos-decay-restart N Only for Adam optimizer. Increase of cosine decay steps after restart (default %f)\n", params->cos_decay_restart); + fprintf(stderr, " --cos-decay-min N Only for Adam optimizer. Cosine decay minimum (default %f)\n", params->cos_decay_min); + fprintf(stderr, " --enable-restart N Only for Adam optimizer. Enable restarts of cos-decay %s\n", params->enable_restart ? "(default)" : ""); + fprintf(stderr, " --disable-restart N Only for Adam optimizer. Disable restarts of cos-decay %s\n", !params->enable_restart ? "(default)" : ""); + fprintf(stderr, " --opt-past N Number of optimization iterations to track for delta convergence test. Disabled when zero. (default %d)\n", params->opt_past); + fprintf(stderr, " --opt-delta N Maximum delta for delta convergence test. Disabled when <= zero. (default %f)\n", params->opt_delta); + fprintf(stderr, " --opt-max-no-improvement N Maximum number of optimization iterations with no improvement. Disabled when <= zero. (default %d)\n", params->opt_max_no_improvement); + fprintf(stderr, " --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. (default %f)\n", params->adam_eps_f); fprintf(stderr, " --adam-iter N Maximum number of Adam optimization iterations for each batch (default %d)\n", params->adam_n_iter); fprintf(stderr, " --adam-alpha N Adam learning rate alpha (default %f)\n", params->adam_alpha); + fprintf(stderr, " --adam-min-alpha N Adam minimum learning rate alpha - including warmup phase (default %f)\n", params->adam_min_alpha); fprintf(stderr, " --adam-decay N AdamW weight decay. Values greater zero enable AdamW instead of regular Adam. (default %f)\n", params->adam_decay); + fprintf(stderr, " --adam-decay-min-ndim N Minimum number of tensor dimensions to apply AdamW weight decay. Weight decay is not applied to tensors with less n_dims. (default %d)\n", params->adam_decay_min_ndim); + fprintf(stderr, " --adam-beta1 N AdamW beta1 in interval [0,1). How much to smooth the first moment of gradients. (default %f)\n", params->adam_beta1); + fprintf(stderr, " --adam-beta2 N AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2); + fprintf(stderr, " --adam-gclip N AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip); + fprintf(stderr, " --lbfgs-iter N Maximum number of LBFGS optimization iterations for each batch (default %d)\n", params->lbfgs_n_iter); fprintf(stderr, " --mem-model N Memory to allocate for model and cache in gigabytes. (default %d)\n", params->mem_model_gb); fprintf(stderr, " --mem-compute N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute_gb); - fprintf(stderr, " --mem-compute0 N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute0_gb); - fprintf(stderr, " --mem-compute1 N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute1_gb); + fprintf(stderr, " --mem-compute0 N Memory to allocate for automatic memory allocator in gigabytes. (default %d)\n", params->mem_compute0_gb); fprintf(stderr, "\n"); } @@ -2872,12 +1737,12 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) { break; } params->n_embd = std::stoi(argv[i]); - } else if (arg == "--mult") { + } else if (arg == "--ff") { if (++i >= argc) { invalid_param = true; break; } - params->n_mult = std::stoi(argv[i]); + params->n_ff = std::stoi(argv[i]); } else if (arg == "--head") { if (++i >= argc) { invalid_param = true; @@ -2890,12 +1755,24 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) { break; } params->n_layer = std::stoi(argv[i]); - } else if (arg == "--rotmax") { + } else if (arg == "--norm-rms-eps") { if (++i >= argc) { invalid_param = true; break; } - params->n_rotmax = std::stoi(argv[i]); + params->f_norm_rms_eps = std::stof(argv[i]); + } else if (arg == "--rope-freq-base") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->rope_freq_base = std::stof(argv[i]); + } else if (arg == "--rope-freq-scale") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->rope_freq_scale = std::stof(argv[i]); } else if (arg == "-t" || arg == "--threads") { if (++i >= argc) { invalid_param = true; @@ -2914,24 +1791,12 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) { break; } params->n_examples = std::stoi(argv[i]); - } else if (arg == "--predict") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_predict = std::stoi(argv[i]); } else if (arg == "--print-info-interval") { if (++i >= argc) { invalid_param = true; break; } params->print_info_interval = std::stoi(argv[i]); - } else if (arg == "--print-details-interval") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->print_details_interval = std::stoi(argv[i]); } else if (arg == "--samples-after-nl") { params->samples_start_after_nl = true; } else if (arg == "--use-lbfgs") { @@ -2942,10 +1807,14 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) { params->use_flash = false; } else if (arg == "--use-flash") { params->use_flash = true; - } else if (arg == "--no-scratch") { - params->use_scratch = false; - } else if (arg == "--use-scratch") { - params->use_scratch = true; + } else if (arg == "--no-checkpointing") { + params->use_checkpointing = false; + } else if (arg == "--use-checkpointing") { + params->use_checkpointing = true; + } else if (arg == "--no-alloc") { + params->use_alloc = false; + } else if (arg == "--use-alloc") { + params->use_alloc = true; } else if (arg == "--warmup") { if (++i >= argc) { invalid_param = true; @@ -2964,18 +1833,40 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) { break; } params->cos_decay_restart = std::stof(argv[i]); - } else if (arg == "--cos-decay-alpha") { + } else if (arg == "--cos-decay-min") { if (++i >= argc) { invalid_param = true; break; } - params->cos_decay_alpha = std::stof(argv[i]); - } else if (arg == "--lbfgs-iter") { + params->cos_decay_min = std::stof(argv[i]); + } else if (arg == "--enable-restart") { + params->enable_restart = true; + } else if (arg == "--disable-restart") { + params->enable_restart = false; + } else if (arg == "--opt-past") { if (++i >= argc) { invalid_param = true; break; } - params->lbfgs_n_iter = std::stoi(argv[i]); + params->opt_past = std::stoi(argv[i]); + } else if (arg == "--opt-delta") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->opt_delta = std::stof(argv[i]); + } else if (arg == "--opt-max-no-improvement") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->opt_max_no_improvement = std::stoi(argv[i]); + } else if (arg == "--adam-epsf") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->adam_eps_f = std::stof(argv[i]); } else if (arg == "--adam-iter") { if (++i >= argc) { invalid_param = true; @@ -2988,12 +1879,48 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) { break; } params->adam_alpha = std::stof(argv[i]); + } else if (arg == "--adam-min-alpha") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->adam_min_alpha = std::stof(argv[i]); } else if (arg == "--adam-decay") { if (++i >= argc) { invalid_param = true; break; } params->adam_decay = std::stof(argv[i]); + } else if (arg == "--adam-decay-min-ndim") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->adam_decay_min_ndim = std::stoi(argv[i]); + } else if (arg == "--adam-beta1") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->adam_beta1 = std::stof(argv[i]); + } else if (arg == "--adam-beta2") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->adam_beta2 = std::stof(argv[i]); + } else if (arg == "--adam-gclip") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->adam_gclip = std::stof(argv[i]); + } else if (arg == "--lbfgs-iter") { + if (++i >= argc) { + invalid_param = true; + break; + } + params->lbfgs_n_iter = std::stoi(argv[i]); } else if (arg == "--mem-model") { if (++i >= argc) { invalid_param = true; @@ -3012,12 +1939,6 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) { break; } params->mem_compute0_gb = std::stoi(argv[i]); - } else if (arg == "--mem-compute1") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->mem_compute1_gb = std::stoi(argv[i]); } else if (arg == "-h" || arg == "--help") { train_print_usage(argc, argv, &default_params); exit(0); @@ -3036,6 +1957,63 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) { return true; } +struct opt_callback_data { + struct train_params * params; + struct ggml_opt_context * opt; + struct llama_context * lctx; + llama_token * tokens_data; + size_t tokens_size; + int * samples_data; + size_t samples_size; + int shuffle_countdown; + struct ggml_tensor * tokens_input; + struct ggml_tensor * target_logits; + struct ggml_tensor * target_probs; +}; + +void opt_callback(void * vdata, float * sched) { + struct opt_callback_data * data = (struct opt_callback_data *) vdata; + struct train_params * params = data->params; + struct ggml_opt_context * opt = data->opt; + int n_batch = params->n_batch; + + *sched = (opt->iter < params->warmup) + ? (float) opt->iter / (float) params->warmup + : cosine_decay_restart( + params->cos_decay_steps, + params->cos_decay_min, + opt->iter - params->warmup, + params->cos_decay_restart, + params->enable_restart); + float min_sched = params->adam_min_alpha / params->adam_alpha; + *sched = min_sched + *sched * (1.0f - min_sched); + + int impr_plot = std::isnan(opt->loss_after) ? 0 : -(int)(1 + (opt->loss_before - opt->loss_after) * 10.0f + 0.5f); + printf("%s: iter=%*d, sched=%f loss0=%f loss=%f | improvement: %*d>\n", __func__, 6, opt->iter, *sched, opt->loss_before, opt->loss_after, impr_plot, (int)0); + + if (data->shuffle_countdown < n_batch) { + printf("%s: reshuffle samples\n", __func__); + shuffle_ints(data->samples_data, data->samples_data + data->samples_size); + for (int i = 0; i < (int) data->samples_size; ++i) { + GGML_ASSERT(data->samples_data[i]+params->n_ctx-1 < (int) data->tokens_size); + } + data->shuffle_countdown = data->samples_size; + } + + get_example_targets_batch( + data->lctx, + data->samples_data, + data->samples_size, + data->tokens_data, + data->tokens_size, + opt->iter, + data->tokens_input, + data->target_logits, + data->target_probs); + + data->shuffle_countdown -= n_batch; +} + int main(int argc, char ** argv) { struct train_params params = get_default_train_params(); @@ -3055,18 +2033,6 @@ int main(int argc, char ** argv) { struct llama_model * lmodel = llama_load_model_from_file(params.fn_vocab_model, llama_params); struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_params); - struct llama_vocab vocab; - { - const int n_vocab = llama_n_vocab(lctx); - vocab.id_to_token.resize(n_vocab); - for (int i=0; i train_tokens; if (tokenize_file(lctx, params.fn_train_data, train_tokens) < 0) { @@ -3078,10 +2044,14 @@ int main(int argc, char ** argv) { model.hparams.n_vocab = llama_n_vocab(lctx); model.hparams.n_ctx = params.n_ctx; model.hparams.n_embd = params.n_embd; - model.hparams.n_mult = params.n_mult; model.hparams.n_head = params.n_head; model.hparams.n_layer = params.n_layer; - model.hparams.n_rot = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head); + model.hparams.n_ff = params.n_ff; + // llama.cpp requires n_rot to be exactly n_embd / n_head + model.hparams.n_rot = model.hparams.n_embd / model.hparams.n_head; + model.hparams.f_norm_rms_eps = params.f_norm_rms_eps; + model.hparams.rope_freq_base = params.rope_freq_base; + model.hparams.rope_freq_scale = params.rope_freq_scale; print_params(&model.hparams); @@ -3103,19 +2073,12 @@ int main(int argc, char ** argv) { } printf("%s: number of unique tokens: %d\n", __func__, n_unique_tokens); - struct my_llama_kv_cache kv_self; - - struct ggml_init_params lcparams; lcparams.mem_size = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb); lcparams.mem_buffer = NULL; lcparams.no_alloc = false; model.ctx = ggml_init(lcparams); - kv_self.ctx = model.ctx; - - my_llama_sampler sampler; - int n_tokens = model.hparams.n_ctx; int n_vocab = model.hparams.n_vocab; @@ -3126,24 +2089,38 @@ int main(int argc, char ** argv) { struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM); struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS); - opt_params_adam.print_forward_graph = false; + opt_params_adam.print_forward_graph = false; opt_params_adam.print_backward_graph = false; - opt_params_adam.n_threads = params.n_threads; - opt_params_adam.adam.n_iter = params.adam_n_iter; - opt_params_adam.adam.sched = 1.0f; - opt_params_adam.adam.alpha = params.adam_alpha; - opt_params_adam.adam.decay = params.adam_decay; + opt_params_adam.n_threads = params.n_threads; + opt_params_adam.past = params.opt_past; + opt_params_adam.delta = params.opt_delta; + opt_params_adam.max_no_improvement = params.opt_max_no_improvement; + opt_params_adam.adam.n_iter = params.adam_n_iter; + opt_params_adam.adam.sched = 1.0f; + opt_params_adam.adam.alpha = params.adam_alpha; + opt_params_adam.adam.decay = params.adam_decay; + opt_params_adam.adam.decay_min_ndim = params.adam_decay_min_ndim; + opt_params_adam.adam.beta1 = params.adam_beta1; + opt_params_adam.adam.beta2 = params.adam_beta2; + opt_params_adam.adam.gclip = params.adam_gclip; + opt_params_adam.adam.eps_f = params.adam_eps_f; - opt_params_lbfgs.print_forward_graph = false; + opt_params_lbfgs.print_forward_graph = false; opt_params_lbfgs.print_backward_graph = false; - opt_params_lbfgs.n_threads = params.n_threads; - opt_params_lbfgs.lbfgs.n_iter = params.lbfgs_n_iter; + opt_params_lbfgs.n_threads = params.n_threads; + opt_params_adam.past = params.opt_past; + opt_params_adam.delta = params.opt_delta; + opt_params_adam.max_no_improvement = params.opt_max_no_improvement; + opt_params_lbfgs.lbfgs.n_iter = params.lbfgs_n_iter; opt->ctx = model.ctx; opt->params = params.use_adam ? opt_params_adam : opt_params_lbfgs; printf("%s: init model\n", __func__); - bool existed = load_checkpoint(&model, opt, params.fn_checkpoint_in, true); + bool existed = load_checkpoint_file(params.fn_checkpoint_in, &model, opt); + if (!existed) { + init_model(&model); + } set_param_model(&model); opt->params = params.use_adam ? opt_params_adam : opt_params_lbfgs; @@ -3156,11 +2133,7 @@ int main(int argc, char ** argv) { randomize_model(&model, params.seed, 0.0f, 1.0f, -1.0f, +1.0f); } - init_kv_cache(&kv_self, &model, 1); - // init_kv_cache(&kv_self, &model, n_batch); - init_sampler(&sampler, lctx); - - printf("used_mem model+cache: %zu bytes\n", ggml_used_mem(model.ctx)); + printf("used_mem model: %zu bytes\n", ggml_used_mem(model.ctx)); // ggml_print_tensor_objects(model.ctx); // TODO: use std::vector intead of "new" @@ -3168,9 +2141,13 @@ int main(int argc, char ** argv) { uint8_t * compute_addr = new uint8_t[compute_size]; size_t size_buf_0 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute0_gb); - size_t size_buf_1 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute1_gb); uint8_t * compute_buf_0 = new uint8_t[size_buf_0]; - uint8_t * compute_buf_1 = new uint8_t[size_buf_1]; + + ggml_allocr * alloc = NULL; + if (params.use_alloc) { + static const size_t tensor_alignment = 32; + alloc = ggml_allocr_new(compute_buf_0, size_buf_0, tensor_alignment); + } GGML_ASSERT(n_tokens < (int) train_tokens.size()); std::vector train_samples; @@ -3185,10 +2162,23 @@ int main(int argc, char ** argv) { GGML_ASSERT(train_samples[i]+n_tokens-1 < (int) train_tokens.size()); } - std::vector work_buffer; - printf("%s: begin training\n", __func__); + struct opt_callback_data opt_cb_data; + opt_cb_data.params = ¶ms; + opt_cb_data.opt = opt; + opt_cb_data.lctx = lctx; + opt_cb_data.tokens_data = train_tokens.data(); + opt_cb_data.tokens_size = train_tokens.size(); + opt_cb_data.samples_data = train_samples.data(); + opt_cb_data.samples_size = train_samples.size(); + opt_cb_data.shuffle_countdown = train_samples.size(); + opt_cb_data.tokens_input = NULL; + opt_cb_data.target_logits = NULL; + opt_cb_data.target_probs = NULL; + + int64_t t0 = ggml_time_ms(); + for (int ex = 0; ex < params.n_examples; ++ex) { if (ex*n_batch >= (int) train_samples.size()) { shuffle_ints(train_samples.data(), train_samples.data() + train_samples.size()); @@ -3198,198 +2188,110 @@ int main(int argc, char ** argv) { } struct ggml_init_params cparams = { - /*.mem_size =*/ compute_size, - /*.mem_buffer =*/ compute_addr, - /*.no_alloc =*/ false, + compute_size, // mem_size + compute_addr, // mem_buffer + false, // no_alloc }; struct ggml_context * ctx0 = ggml_init(cparams); - struct ggml_tensor * after_opt_best_samples = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_batch); + ggml_set_no_alloc(ctx0, false); + + // don't use alloc for input tensors, so we can safely fill them with data + //struct ggml_tensor * after_opt_best_samples = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_batch); //struct ggml_tensor * after_opt_probs = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab, n_tokens, n_batch); struct ggml_tensor * tokens_input = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_batch); struct ggml_tensor * target_logits = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab, n_tokens, n_batch); struct ggml_tensor * target_probs = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_vocab, n_tokens, n_batch); + ggml_set_no_alloc(ctx0, (alloc != NULL)); + + if (alloc) { + ggml_allocr_reset(alloc); + } + + opt_cb_data.tokens_input = tokens_input; + opt_cb_data.target_logits = target_logits; + opt_cb_data.target_probs = target_probs; + int n_past = 0; - struct ggml_tensor * gfbuf = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32) + (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0)); - struct ggml_tensor * gbbuf = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32) + (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0)); - - memset(gfbuf->data, 0, ggml_nbytes(gfbuf)); - memset(gbbuf->data, 0, ggml_nbytes(gbbuf)); - - struct ggml_cgraph * gf = (struct ggml_cgraph *) gfbuf->data; - struct ggml_cgraph * gb = (struct ggml_cgraph *) gbbuf->data; - - - get_example_targets_batch(lctx, train_samples.data(), train_samples.size(), train_tokens.data(), train_tokens.size(), ex, tokens_input, target_logits, target_probs); + struct ggml_cgraph * gf = ggml_new_graph(ctx0); + struct ggml_cgraph * gb = ggml_new_graph(ctx0); + struct ggml_cgraph * gb_tmp = params.use_checkpointing + ? ggml_new_graph(ctx0) + : NULL; GGML_ASSERT(n_past == 0); struct ggml_tensor * loss = NULL; struct ggml_tensor * logits = NULL; - if (params.use_scratch) { - loss = forward_batch_wo_cache_flash_attn_train( - &model, ctx0, - gf, gb, - &logits, tokens_input, target_probs, - compute_buf_0, compute_buf_1, - size_buf_0, size_buf_1, - n_tokens, n_batch); - } else if (params.use_flash) { - logits = forward_batch_wo_cache_flash_attn(&model, ctx0, gf, tokens_input, n_tokens, n_batch); - loss = cross_entropy_loss(ctx0, logits, target_probs); - ggml_build_forward_expand(gf, loss); - *gb = ggml_build_backward(ctx0, gf, true); - } else { - logits = forward_batch_wo_cache(&model, ctx0, gf, tokens_input, n_tokens, n_batch); - loss = cross_entropy_loss(ctx0, logits, target_probs); - ggml_build_forward_expand(gf, loss); - *gb = ggml_build_backward(ctx0, gf, true); - } - - ggml_graph_compute_helper(work_buffer, gf, params.n_threads); + loss = llama_build_train_graphs( + &model, alloc, ctx0, + gf, gb, gb_tmp, + &logits, tokens_input, target_probs, + n_tokens, n_batch, + params.use_flash, + params.use_checkpointing + ); size_t used_mem_before_opt = ggml_used_mem(ctx0); - float error_before_opt = ggml_get_f32_1d(loss, 0); - opt->params.adam.sched = (opt->iter < params.warmup) ? (float) opt->iter / (float) params.warmup : cosine_decay_restart( params.cos_decay_steps, - params.cos_decay_alpha, + params.cos_decay_min, opt->iter - params.warmup, - params.cos_decay_restart); + params.cos_decay_restart, + params.enable_restart); + + float min_sched = params.adam_min_alpha / params.adam_alpha; + opt->params.adam.sched = min_sched + opt->params.adam.sched * (1.0f - min_sched); printf("%s: opt->params.adam.sched %.5f\n", __func__, opt->params.adam.sched); - ggml_opt_resume_g(ctx0, opt, loss, gf, gb); + ggml_opt_resume_g(ctx0, opt, loss, gf, gb, &opt_callback, (void *) &opt_cb_data); size_t used_mem_after_opt = ggml_used_mem(ctx0); + int n_iter = params.use_adam ? params.adam_n_iter : params.lbfgs_n_iter; model.train_its = opt->iter; - model.train_samples += n_batch; - model.train_tokens += n_batch * n_tokens; - - ggml_graph_compute_helper(work_buffer, gf, params.n_threads); - - float error_after_opt = ggml_get_f32_1d(loss, 0); + model.train_samples += n_batch * n_iter; + model.train_tokens += n_batch * n_tokens * n_iter; if (params.print_info_interval > 0 && ex % params.print_info_interval == 0) { printf("Example %d, opt iter %d\n", ex, opt->iter); - printf("error_before_opt: %.6f\n", error_before_opt); - printf("error_after_opt: %.6f\n", error_after_opt); + printf("error_before_opt: %.6f\n", opt->loss_before); + printf("error_after_opt: %.6f\n", opt->loss_after); printf("used_mem_before_opt: %zu bytes\n", used_mem_before_opt); printf("used_mem_after_opt: %zu bytes\n", used_mem_after_opt); } - if (params.print_details_interval > 0 && ex % params.print_details_interval == 0) { - // set_logits_masked(logits, token_notavail, -1e9); - for (int i=0; idata + i*logits->nb[2] + k*logits->nb[1]), - (llama_token *) ((char *) tokens_input->data + i*tokens_input->nb[1]), - k); - * ((int32_t *) ((char *) after_opt_best_samples->data + i*after_opt_best_samples->nb[1] + k*after_opt_best_samples->nb[0])) = token; - } - } - - // printf("probabilities after optimization:\n"); - // print_matrix(after_opt_probs); - printf("Example:\n---\n"); - print_tokens_batch(lctx, tokens_input); - printf("\n---\n"); - - // printf("best samples after optimization:\n---\n"); - printf("samples after optimization:\n---\n"); - print_tokens_batch(lctx, after_opt_best_samples); - printf("\n---\n"); - } - ggml_free(ctx0); } + int64_t t1 = ggml_time_ms(); + int64_t d = t1-t0; + double dd = (double) d * 1e-3; + printf("%s: total training time=%f seconds\n", __func__, dd); + if (params.n_examples > 0) { - save_checkpoint(&model, opt, params.fn_checkpoint_out); + save_checkpoint_file(params.fn_checkpoint_out, params.fn_vocab_model, &model, opt); } if (strlen(params.fn_model_out) > 0) { - save_as_llama_model(&vocab, &model, params.fn_model_out); + save_llama_model_file(params.fn_model_out, params.fn_vocab_model, &model); } - { - int n_gen = params.n_predict; - int sample_ctx = n_tokens - n_tokens/8; - - sampler.params.temp = 0.2f; - sampler.params.repeat_penalty = 1.1f; - sampler.params.mirostat = 2; - init_sampler(&sampler, lctx); - - printf("Generating %d tokens.\n", n_gen); - - struct ggml_tensor * tokens_input = ggml_new_tensor_1d(model.ctx, GGML_TYPE_I32, n_tokens); - struct ggml_tensor * target_logits = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, n_vocab, n_tokens); - struct ggml_tensor * target_probs = ggml_new_tensor_2d(model.ctx, GGML_TYPE_F32, n_vocab, n_tokens); - - get_example_targets(lctx, train_samples.data(), train_samples.size(), train_tokens.data(), train_tokens.size(), rand()%train_samples.size(), tokens_input, target_logits, target_probs); - for (int i=sample_ctx; idata + (sample_ctx-1)*logits->nb[1]), - (llama_token *) tokens_input->data, - sample_ctx-1); - //int token = ggml_get_i32_1d(best_samples, sample_ctx-1); - - // print_row(probs, sample_at); - print_token(lctx, token); - - lshift_examples(tokens_input, target_logits, target_probs, 1); - ggml_set_i32_1d(tokens_input, 0, 0); - ggml_set_i32_1d(tokens_input, sample_ctx-1, token); - - ggml_free(ctx0); - } + if (alloc) { + ggml_allocr_free(alloc); } delete[] compute_addr; delete[] compute_buf_0; - delete[] compute_buf_1; - + ggml_free(model.ctx); llama_free(lctx); llama_free_model(lmodel); - ggml_free(model.ctx); - return 0; } diff --git a/ggml-alloc.c b/ggml-alloc.c index 140e9a2a7..63beb1d4e 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -107,6 +107,10 @@ static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct g } void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +#ifdef GGML_ALLOCATOR_DEBUG + GGML_ASSERT(ggml_is_view(tensor) == false); // views generally get data pointer from one of their sources + GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated +#endif size_t size = ggml_allocator_get_alloc_size(alloc, tensor); size = aligned_offset(NULL, size, alloc->alignment); diff --git a/ggml.c b/ggml.c index dadb30757..9a787863d 100644 --- a/ggml.c +++ b/ggml.c @@ -123,6 +123,8 @@ typedef void * thread_ret_t; #define GGML_GELU_FP16 #define GGML_GELU_QUICK_FP16 #define GGML_SILU_FP16 +// #define GGML_CROSS_ENTROPY_EXP_FP16 +// #define GGML_FLASH_ATTN_EXP_FP16 #define GGML_SOFT_MAX_UNROLL 4 #define GGML_VEC_DOT_UNROLL 2 @@ -186,8 +188,8 @@ typedef void * thread_ret_t; // #if defined(_MSC_VER) || defined(__MINGW32__) -#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN) -#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr) +#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN) +#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr) #else inline static void * ggml_aligned_malloc(size_t size) { void * aligned_memory = NULL; @@ -212,8 +214,8 @@ inline static void * ggml_aligned_malloc(size_t size) { } return aligned_memory; } -#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size) -#define GGML_ALIGNED_FREE(ptr) free(ptr) +#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size) +#define GGML_ALIGNED_FREE(ptr) free(ptr) #endif #define UNUSED GGML_UNUSED @@ -5857,7 +5859,8 @@ struct ggml_tensor * ggml_rms_norm_inplace( struct ggml_tensor * ggml_rms_norm_back( struct ggml_context * ctx, struct ggml_tensor * a, - struct ggml_tensor * b) { + struct ggml_tensor * b, + float eps) { bool is_node = false; if (a->grad) { @@ -5867,6 +5870,8 @@ struct ggml_tensor * ggml_rms_norm_back( struct ggml_tensor * result = ggml_dup_tensor(ctx, a); + ggml_set_op_params(result, &eps, sizeof(eps)); + result->op = GGML_OP_RMS_NORM_BACK; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; @@ -9443,6 +9448,8 @@ static void ggml_compute_forward_div_f32( #ifdef GGML_USE_ACCELERATE + UNUSED(ggml_vec_div_f32); + vDSP_vdiv( (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, @@ -10749,7 +10756,8 @@ static void ggml_compute_forward_rms_norm_back_f32( GGML_TENSOR_BINARY_OP_LOCALS; - const float eps = 1e-6f; // TODO: make this a parameter + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); // TODO: optimize for (int64_t i03 = 0; i03 < ne03; i03++) { @@ -12139,6 +12147,7 @@ static void ggml_compute_forward_soft_max_back_f32( // dx = J * dy // dxk = sum_i(Jki * dyi) // dxk = sum_i(-yk*yi * dyi) - (-yk*yk)*dyk + (yk - yk*yk)*dyk + // dxk = sum_i(-yk*yi * dyi) + yk*yk*dyk + yk*dyk - yk*yk*dyk // dxk = sum_i(-yk*yi * dyi) + yk*dyk // dxk = -yk * sum_i(yi * dyi) + yk*dyk // dxk = -yk * dot(y, dy) + yk*dyk @@ -13929,7 +13938,7 @@ static void ggml_compute_forward_flash_attn_f32( vvexpf(S, S, &Mup); ggml_vec_sum_f32(Mup, &sum, S); #else - uint16_t scvt[GGML_SOFT_MAX_UNROLL]; + uint16_t scvt[GGML_SOFT_MAX_UNROLL]; UNUSED(scvt); ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 }; for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) { @@ -13939,9 +13948,13 @@ static void ggml_compute_forward_flash_attn_f32( if (SS[j] == -INFINITY) { SS[j] = 0.0f; } else { +#ifndef GGML_FLASH_ATTN_EXP_FP16 + const float val = expf(SS[j] - max); +#else ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max); memcpy(&scvt[j], &s, sizeof(uint16_t)); const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]); +#endif sump[j] += (ggml_float)val; SS[j] = val; } @@ -14519,7 +14532,7 @@ static void ggml_compute_forward_flash_attn_back_f32( vvexpf(SM, SM, &Mup); ggml_vec_sum_f32(Mup, &sum, SM); #else - uint16_t scvt[GGML_SOFT_MAX_UNROLL]; + uint16_t scvt[GGML_SOFT_MAX_UNROLL]; UNUSED(scvt); ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 }; for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) { @@ -14530,9 +14543,13 @@ static void ggml_compute_forward_flash_attn_back_f32( if (SR[j] == -INFINITY) { SW[j] = 0.0f; } else { +#ifndef GGML_FLASH_ATTN_EXP_FP16 + const float val = expf(SR[j] - max); +#else ggml_fp16_t s = GGML_FP32_TO_FP16(SR[j] - max); memcpy(&scvt[j], &s, sizeof(uint16_t)); const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]); +#endif sump[j] += (ggml_float)val; SW[j] = val; } @@ -15270,6 +15287,8 @@ static void ggml_compute_forward_cross_entropy_loss_f32( const int nc = src0->ne[0]; const int nr = ggml_nrows(src0); + GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc)); + if (params->type == GGML_TASK_INIT) { if (ith == 0) { memset(sums, 0, sizeof(float) * (nth + nth * nc)); @@ -15281,7 +15300,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32( if (ith == 0) { float * dp = (float *) dst->data; ggml_vec_sum_f32(nth, dp, sums); - dp[0] *= -1.0f; + dp[0] *= -1.0f / (float) nr; } return; } @@ -15298,7 +15317,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32( for (int i1 = ir0; i1 < ir1; i1++) { float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]); float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]); - float * st = (float *) params->wdata + nth + ith*nc; + float * st = ((float *) params->wdata) + nth + ith*nc; #ifndef NDEBUG for (int i = 0; i < nc; ++i) { @@ -15313,15 +15332,19 @@ static void ggml_compute_forward_cross_entropy_loss_f32( float max = -INFINITY; ggml_vec_max_f32(nc, &max, s0); - uint16_t scvt; + uint16_t scvt; UNUSED(scvt); for (int i = 0; i < nc; i++) { if (s0[i] == -INFINITY) { st[i] = 0.0f; } else { - // const float val = (s0[i] == -INFINITY) ? 0.0 : exp(s0[i] - max); +#ifndef GGML_CROSS_ENTROPY_EXP_FP16 + const float s = s0[i] - max; + const float val = expf(s); +#else ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max); memcpy(&scvt, &s, sizeof(scvt)); const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]); +#endif sum += (ggml_float)val; st[i] = val; } @@ -15337,7 +15360,9 @@ static void ggml_compute_forward_cross_entropy_loss_f32( ggml_vec_log_f32(nc, st, st); ggml_vec_mul_f32(nc, st, st, s1); - ggml_vec_sum_f32(nc, sums + ith, st); + float st_sum = 0; + ggml_vec_sum_f32(nc, &st_sum, st); + sums[ith] += st_sum; #ifndef NDEBUG for (int i = 0; i < nc; ++i) { @@ -15387,7 +15412,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( return; } - const float eps = 1e-9f; + const double eps = 1e-9; // TODO: handle transposed/permuted matrices const int64_t nc = src0->ne[0]; @@ -15406,7 +15431,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( float * ds0 = (float *)((char *) dst->data + i1*dst->nb[1]); float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]); float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]); - float * sm = (float *) params->wdata + ith*nc; #ifndef NDEBUG for (int i = 0; i < nc; ++i) { @@ -15415,54 +15439,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( assert(!isnan(s1[i])); } #endif - // step by step explanation: - { - //float * sums = (float *) params->wdata; - - // forward pass with annotated gradients from backward pass - // (built by going in reverse operation order, adding to gradients of current operation args) - // st0 = exp(s0-max(s0)) grad[st0] = grad[st1]*(1.0 - eps)/sum - // from softmax_back: grad[s0] = st1_k * (grad[st1]_k - dot(st1, grad[st1])) - // ggml_vec_scale_f32(nc, st, sum); // st1 = st0*/sum = softmax(s0) grad[st1] = grad[st2]*(1.0 - eps) - // ggml_vec_scale_f32(nc, st, (1.0f - eps)); // st2 = st1*(1.0 - eps) grad[st2] = grad[st3] - // ggml_vec_add1_f32(nc, st, st, eps); // st3 = st2 + eps grad[st3] = grad[st4]/st3 - // ggml_vec_log_f32(nc, st, st); // st4 = log(st3) grad[st4] = grad[st5] * s1 - // ggml_vec_mul_f32(nc, st, st, s1); // st5 = st4 * s1 grad[st5] = grad[sums[ith]] - // ggml_vec_sum_f32(nc, sums + ith, st); // sums[ith] = st5 grad[sums[ith]] = grad[cross_entropy_loss] = -grad[cel] - - // substitute into grad[st1], because we can reuse softmax_back from this point on - // grad[st1] = -grad[cel]*s1*(1.0 - eps)/(eps + softmax(s0)*(1.0 - eps)) - // postorder: - // grad[st1] := softmax(s0) - // grad[st1] := grad[st1]*(1.0 - eps) - // grad[st1] := grad[st1] + eps - // grad[st1] := s1 / grad[st1] - // grad[st1] := grad[st1]*(1.0-eps)*-grad[cel] - - // src0 gradients by going through softmax_back - // grad[s0] = st1_k * (grad[st1]_k - dot(st1, grad[st1])) - // from softmax_back: - // dxk = yk * (dyk - dot(y, dy)) - // dot_y_dy := dot(y, dy) - // dx := dy - // dx := dx - dot_y_dy - // dx := dx * y - // postorder: - // dot_st1_dst1 := dot(st1, grad[st1]) - // grad[s0] := grad[st1] - // grad[s0] := grad[s0] - dot_st1_dst1 - // grad[s0] := grad[s0] * st1 - - // prepend postorder from grad[st1] directly using grad[s0] as memory location, as we will grad[s0] := grad[st1] - // sm := softmax(s0) - // grad[s0] := sm*(1.0 - eps) - // grad[s0] := grad[s0] + eps - // grad[s0] := s1 / grad[s0] - // grad[s0] := grad[s0]*(1.0-eps)*-grad[cel] - // dot_st1_dst1 := dot(sm, grad[s0]) - // grad[s0] := grad[s0] - dot_st1_dst1 - // grad[s0] := grad[s0] * sm - } // soft_max ggml_float sum = 0.0; @@ -15470,39 +15446,37 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( float max = -INFINITY; ggml_vec_max_f32(nc, &max, s0); - uint16_t scvt; + uint16_t scvt; UNUSED(scvt); for (int i = 0; i < nc; i++) { if (s0[i] == -INFINITY) { - sm[i] = 0.0f; + ds0[i] = 0.0f; } else { - // const float val = (s0[i] == -INFINITY) ? 0.0 : exp(s0[i] - max); +#ifndef GGML_CROSS_ENTROPY_EXP_FP16 + const float s = s0[i] - max; + const float val = expf(s); +#else ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max); memcpy(&scvt, &s, sizeof(scvt)); const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]); +#endif sum += (ggml_float)val; - sm[i] = val; + ds0[i] = val; } } assert(sum > 0.0); - sum = 1.0/sum; + sum = (1.0 - eps)/sum; } - float dot_st1_dst1 = 0; - ggml_vec_scale_f32(nc, sm, sum); - ggml_vec_cpy_f32 (nc, ds0, sm); - ggml_vec_scale_f32(nc, ds0, (1.0f - eps)); - ggml_vec_add1_f32 (nc, ds0, ds0, eps); - ggml_vec_div_f32 (nc, ds0, s1, ds0); - ggml_vec_scale_f32(nc, ds0, -(1.0f - eps)*d[0]); - ggml_vec_dot_f32 (nc, &dot_st1_dst1, sm, ds0); - ggml_vec_acc1_f32 (nc, ds0, -dot_st1_dst1); - ggml_vec_mul_f32 (nc, ds0, ds0, sm); + // grad(src0) = (softmax(src0) - src1) * grad(cross_entropy_loss(src0, src1)) / nr + ggml_vec_scale_f32(nc, ds0, sum); + ggml_vec_add1_f32(nc, ds0, ds0, eps); + ggml_vec_sub_f32(nc, ds0, ds0, s1); + ggml_vec_scale_f32(nc, ds0, d[0] / (float) nr); + #ifndef NDEBUG for (int i = 0; i < nc; ++i) { - assert(!isnan(sm[i])); - assert(!isinf(sm[i])); assert(!isnan(ds0[i])); assert(!isinf(ds0[i])); } @@ -16057,9 +16031,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { // necessary for llama if (src0->grad) { + float eps; + memcpy(&eps, tensor->op_params, sizeof(float)); + src0->grad = ggml_add_impl(ctx, src0->grad, - ggml_rms_norm_back(ctx, src0, tensor->grad), + ggml_rms_norm_back(ctx, src0, tensor->grad, eps), inplace); } } break; @@ -16827,9 +16804,7 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) { return result; } -struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep) { - struct ggml_cgraph result = *gf; - +void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep) { GGML_ASSERT(gf->n_nodes > 0); // if we are keeping the gradient graph, we have to detach the gradient nodes from the original graph @@ -16853,15 +16828,19 @@ struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cg } } - for (int i = gf->n_nodes - 1; i >= 0; i--) { + for (int i = 0; i < gf->n_nodes; i++) { struct ggml_tensor * node = gf->nodes[i]; if (node->is_param) { GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node); - ggml_build_forward_expand(&result, node->grad); + ggml_build_forward_expand(gb, node->grad); } } +} +struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep) { + struct ggml_cgraph result = *gf; + ggml_build_backward_expand(ctx, gf, &result, keep); return result; } @@ -17537,10 +17516,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { case GGML_OP_CROSS_ENTROPY_LOSS_BACK: { n_tasks = n_threads; - - size_t cur = ggml_type_size(node->type)*node->src[0]->ne[0]*n_tasks; - - work_size = MAX(work_size, cur); } break; case GGML_OP_NONE: { @@ -18418,14 +18393,16 @@ static enum ggml_opt_result ggml_opt_adam( struct ggml_opt_params params, struct ggml_tensor * f, struct ggml_cgraph * gf, - struct ggml_cgraph * gb) { + struct ggml_cgraph * gb, + ggml_opt_callback callback, + void * callback_data) { GGML_ASSERT(ggml_is_scalar(f)); // these will store the parameters we want to optimize struct ggml_tensor * ps[GGML_MAX_PARAMS]; int np = 0; - int nx = 0; + int64_t nx = 0; for (int i = 0; i < gf->n_nodes; ++i) { if (gf->nodes[i]->is_param) { GGML_PRINT_DEBUG("found param %d: grad->op = %d\n", np, gf->nodes[i]->grad->op); @@ -18444,31 +18421,32 @@ static enum ggml_opt_result ggml_opt_adam( } // constants - const float sched = params.adam.sched; - const float decay = params.adam.decay * sched; - const float alpha = params.adam.alpha * sched; + float sched = params.adam.sched; + const float alpha = params.adam.alpha; + const float decay = params.adam.decay * alpha; const float beta1 = params.adam.beta1; const float beta2 = params.adam.beta2; const float eps = params.adam.eps; + const float gclip = params.adam.gclip; + const int decay_min_ndim = params.adam.decay_min_ndim; - float * x = opt->adam.x->data; // view of the parameters - float * g1 = opt->adam.g1->data; // gradient - float * g2 = opt->adam.g2->data; // gradient squared float * m = opt->adam.m->data; // first moment float * v = opt->adam.v->data; // second moment - float * mh = opt->adam.mh->data; // first moment hat - float * vh = opt->adam.vh->data; // second moment hat float * pf = params.past > 0 ? opt->adam.pf->data : NULL; // past function values - // update view - ggml_opt_get_params(np, ps, x); + if (callback) { + callback(callback_data, &sched); + } // compute the function value ggml_graph_reset (gf); ggml_set_f32 (f->grad, 1.0f); - ggml_graph_compute_with_ctx(ctx, gb, params.n_threads); + struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads); + struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size); + cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs; + ggml_graph_compute(gb, &cplan); opt->adam.fx_prev = ggml_get_f32_1d(f, 0); opt->adam.fx_best = opt->adam.fx_prev; @@ -18476,6 +18454,9 @@ static enum ggml_opt_result ggml_opt_adam( pf[opt->iter % params.past] = opt->adam.fx_prev; } + opt->loss_before = opt->adam.fx_prev; + opt->loss_after = opt->adam.fx_prev; + // initialize if (opt->just_initialized) { opt->adam.n_no_improvement = 0; @@ -18508,50 +18489,55 @@ static enum ggml_opt_result ggml_opt_adam( UNUSED(t_start_cpu); { - // update the gradient - ggml_opt_get_grad(np, ps, g1); + float gnorm = 1.0f; + if (gclip > 0.0f) { + // gradient clipping + ggml_float sum = 0.0; + for (int p = 0; p < np; ++p) { + const int64_t ne = ggml_nelements(ps[p]); + for (int64_t j = 0; j < ne; ++j) { + float g = ggml_get_f32_1d(ps[p]->grad, j); + sum += (ggml_float)(g*g); + } + } + ggml_float norm = sqrt(sum); + if (norm > (ggml_float) gclip) { + gnorm = (float) ((ggml_float) gclip / norm); + } + } + const float beta1h = alpha*sched/(1.0f - powf(beta1, opt->iter)); + const float beta2h = 1.0f/(1.0f - powf(beta2, opt->iter)); + int64_t i = 0; + for (int p = 0; p < np; ++p) { + const int64_t ne = ggml_nelements(ps[p]); + const float p_decay = ((ps[p]->n_dims >= decay_min_ndim) ? decay : 0.0f) * sched; + for (int64_t j = 0; j < ne; ++j) { + float x = ggml_get_f32_1d(ps[p], j); + float g = ggml_get_f32_1d(ps[p]->grad, j)*gnorm; + m[i] = m[i]*beta1 + g*(1.0f - beta1); + v[i] = v[i]*beta2 + g*g*(1.0f - beta2); + float mh = m[i]*beta1h; + float vh = v[i]*beta2h; + vh = sqrtf(vh) + eps; + x = x*(1.0f - p_decay) - mh/vh; + ggml_set_f32_1d(ps[p], j, x); + ++i; + } + } + } - // m_t = beta1*m_t-1 + (1 - beta1)*g_t - ggml_vec_scale_f32(nx, m, beta1); - ggml_vec_mad_f32 (nx, m, g1, 1.0f - beta1); - - // g2 = g1^2 - ggml_vec_sqr_f32 (nx, g2, g1); - - // v_t = beta2*v_t-1 + (1 - beta2)*g_t^2 - ggml_vec_scale_f32(nx, v, beta2); - ggml_vec_mad_f32 (nx, v, g2, 1.0f - beta2); - - // m^hat = m_t / (1 - beta1^t) - // v^hat = v_t / (1 - beta2^t) - // x_t = x_t-1 - sched*(alpha*m^hat/(sqrt(v^hat) + eps) + decay*x_t-1) - // x_t = x_t-1 - sched*alpha*m^hat/(sqrt(v^hat) + eps) - sched*decay*x_t-1 - // x_t = x_t-1*(1-sched*decay) - sched*alpha*m^hat/(sqrt(v^hat) + eps) - // x_t = x_t-1*(1-sched*decay) + sched*decay*(-alpha/decay)*m^hat/(sqrt(v^hat) + eps) - // x_t = mix(x_t-1, (-alpha/decay)*m^hat/(sqrt(v^hat) + eps), sched*decay) - ggml_vec_cpy_f32 (nx, mh, m); - ggml_vec_cpy_f32 (nx, vh, v); - - ggml_vec_scale_f32(nx, mh, alpha/(1.0f - powf(beta1, opt->iter))); - ggml_vec_scale_f32(nx, vh, 1.0f/(1.0f - powf(beta2, opt->iter))); - - ggml_vec_sqrt_f32 (nx, vh, vh); - ggml_vec_acc1_f32 (nx, vh, eps); - - ggml_vec_div_f32 (nx, mh, mh, vh); - ggml_vec_scale_f32(nx, x, 1.0f - decay); - ggml_vec_sub_f32 (nx, x, x, mh); - - // update the parameters - ggml_opt_set_params(np, ps, x); + if (callback) { + callback(callback_data, &sched); } ggml_graph_reset (gf); ggml_set_f32 (f->grad, 1.0f); - ggml_graph_compute_with_ctx(ctx, gb, params.n_threads); + ggml_graph_compute(gb, &cplan); const float fx = ggml_get_f32_1d(f, 0); + opt->loss_after = fx; + // check convergence if (fabsf(fx - fx_prev[0])/fx < params.adam.eps_f) { @@ -18620,7 +18606,6 @@ struct ggml_lbfgs_iteration_data { }; static enum ggml_opt_result linesearch_backtracking( - struct ggml_context * ctx, const struct ggml_opt_params * params, int nx, float * x, @@ -18632,8 +18617,11 @@ static enum ggml_opt_result linesearch_backtracking( struct ggml_tensor * f, struct ggml_cgraph * gf, struct ggml_cgraph * gb, + struct ggml_cplan * cplan, const int np, - struct ggml_tensor * ps[]) { + struct ggml_tensor * ps[], + ggml_opt_callback callback, + void * callback_data) { int count = 0; float width = 0.0f; @@ -18662,6 +18650,12 @@ static enum ggml_opt_result linesearch_backtracking( dgtest = params->lbfgs.ftol*dginit; while (true) { + if (callback) { + // LBFG-S does not support learning rate -> ignore learning schedule + float sched = 0; + callback(callback_data, &sched); + } + ggml_vec_cpy_f32(nx, x, xp); ggml_vec_mad_f32(nx, x, d, *step); @@ -18672,7 +18666,7 @@ static enum ggml_opt_result linesearch_backtracking( ggml_graph_reset (gf); ggml_set_f32 (f->grad, 1.0f); - ggml_graph_compute_with_ctx(ctx, gb, params->n_threads); + ggml_graph_compute(gb, cplan); ggml_opt_get_grad(np, ps, g); @@ -18732,7 +18726,9 @@ static enum ggml_opt_result ggml_opt_lbfgs( struct ggml_opt_params params, struct ggml_tensor * f, struct ggml_cgraph * gf, - struct ggml_cgraph * gb) { + struct ggml_cgraph * gb, + ggml_opt_callback callback, + void * callback_data) { if (params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_WOLFE || params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE) { if (params.lbfgs.wolfe <= params.lbfgs.ftol || 1.f <= params.lbfgs.wolfe) { @@ -18764,6 +18760,10 @@ static enum ggml_opt_result ggml_opt_lbfgs( opt->iter = iter; } + struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads); + struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size); + cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs; + float * x = opt->lbfgs.x->data; // current parameters float * xp = opt->lbfgs.xp->data; // previous parameters float * g = opt->lbfgs.g->data; // current gradient @@ -18785,6 +18785,12 @@ static enum ggml_opt_result ggml_opt_lbfgs( float * lm_s = opt->lbfgs.lms->data; float * lm_y = opt->lbfgs.lmy->data; + if (callback) { + // LBFG-S does not support learning rate -> ignore learning schedule + float sched = 0; + callback(callback_data, &sched); + } + // evaluate the function value and its gradient { ggml_opt_set_params(np, ps, x); @@ -18792,11 +18798,14 @@ static enum ggml_opt_result ggml_opt_lbfgs( ggml_graph_reset (gf); ggml_set_f32 (f->grad, 1.0f); - ggml_graph_compute_with_ctx(ctx, gb, params.n_threads); + ggml_graph_compute(gb, &cplan); ggml_opt_get_grad(np, ps, g); fx = ggml_get_f32_1d(f, 0); + + opt->loss_before = fx; + opt->loss_after = fx; } // search direction = -gradient @@ -18851,7 +18860,7 @@ static enum ggml_opt_result ggml_opt_lbfgs( ggml_vec_cpy_f32(nx, xp, x); ggml_vec_cpy_f32(nx, gp, g); - ls = linesearch_backtracking(ctx, ¶ms, nx, x, &fx, g, d, step, xp, f, gf, gb, np, ps); + ls = linesearch_backtracking(¶ms, nx, x, &fx, g, d, step, xp, f, gf, gb, &cplan, np, ps, callback, callback_data); if (ls < 0) { // linesearch failed - go back to the previous point and return @@ -18861,6 +18870,8 @@ static enum ggml_opt_result ggml_opt_lbfgs( return ls; } + opt->loss_after = fx; + ggml_vec_norm_f32(nx, &xnorm, x); ggml_vec_norm_f32(nx, &gnorm, g); @@ -18918,7 +18929,7 @@ static enum ggml_opt_result ggml_opt_lbfgs( // ys = y^t \cdot s -> 1 / \rho. // yy = y^t \cdot y. // - ggml_vec_dot_f32(nx, &ys, &lm_y[end[0]*nx], &lm_s[end[0] *nx]); + ggml_vec_dot_f32(nx, &ys, &lm_y[end[0]*nx], &lm_s[end[0]*nx]); ggml_vec_dot_f32(nx, &yy, &lm_y[end[0]*nx], &lm_y[end[0]*nx]); lm_ys[end[0]] = ys; @@ -18981,13 +18992,15 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) { .adam = { .n_iter = 10000, .sched = 1.000f, - .decay = 0.001f, + .decay = 0.0f, + .decay_min_ndim = 2, .alpha = 0.001f, .beta1 = 0.9f, .beta2 = 0.999f, .eps = 1e-8f, .eps_f = 1e-5f, .eps_g = 1e-3f, + .gclip = 0.0f, }, }; } break; @@ -19037,23 +19050,13 @@ GGML_API void ggml_opt_init( switch (opt->params.type) { case GGML_OPT_ADAM: { - opt->adam.x = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx); - opt->adam.g1 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx); - opt->adam.g2 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx); opt->adam.m = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx); opt->adam.v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx); - opt->adam.mh = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx); - opt->adam.vh = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, nx); opt->adam.pf = params.past > 0 ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, params.past) : NULL; - ggml_set_zero(opt->adam.x); - ggml_set_zero(opt->adam.g1); - ggml_set_zero(opt->adam.g2); ggml_set_zero(opt->adam.m); ggml_set_zero(opt->adam.v); - ggml_set_zero(opt->adam.mh); - ggml_set_zero(opt->adam.vh); if (opt->adam.pf) { ggml_set_zero(opt->adam.pf); } @@ -19137,7 +19140,7 @@ enum ggml_opt_result ggml_opt_resume( *gf = ggml_build_forward (f); *gb = ggml_build_backward(ctx, gf, true); - return ggml_opt_resume_g(ctx, opt, f, gf, gb); + return ggml_opt_resume_g(ctx, opt, f, gf, gb, NULL, NULL); } enum ggml_opt_result ggml_opt_resume_g( @@ -19145,7 +19148,9 @@ enum ggml_opt_result ggml_opt_resume_g( struct ggml_opt_context * opt, struct ggml_tensor * f, struct ggml_cgraph * gf, - struct ggml_cgraph * gb) { + struct ggml_cgraph * gb, + ggml_opt_callback callback, + void * callback_data) { // build forward + backward compute graphs enum ggml_opt_result result = GGML_OPT_OK; @@ -19153,11 +19158,11 @@ enum ggml_opt_result ggml_opt_resume_g( switch (opt->params.type) { case GGML_OPT_ADAM: { - result = ggml_opt_adam(ctx, opt, opt->params, f, gf, gb); + result = ggml_opt_adam(ctx, opt, opt->params, f, gf, gb, callback, callback_data); } break; case GGML_OPT_LBFGS: { - result = ggml_opt_lbfgs(ctx, opt, opt->params, f, gf, gb); + result = ggml_opt_lbfgs(ctx, opt, opt->params, f, gf, gb, callback, callback_data); } break; } @@ -19612,7 +19617,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p // read the kv pairs { - ctx->kv = GGML_ALIGNED_MALLOC(ctx->header.n_kv * sizeof(struct gguf_kv)); + ctx->kv = malloc(ctx->header.n_kv * sizeof(struct gguf_kv)); for (uint32_t i = 0; i < ctx->header.n_kv; ++i) { struct gguf_kv * kv = &ctx->kv[i]; @@ -19695,7 +19700,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p // read the tensor infos { - ctx->infos = GGML_ALIGNED_MALLOC(ctx->header.n_tensors * sizeof(struct gguf_tensor_info)); + ctx->infos = malloc(ctx->header.n_tensors * sizeof(struct gguf_tensor_info)); for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) { struct gguf_tensor_info * info = &ctx->infos[i]; @@ -19896,7 +19901,7 @@ void gguf_free(struct gguf_context * ctx) { } } - GGML_ALIGNED_FREE(ctx->kv); + free(ctx->kv); } if (ctx->infos) { @@ -19908,7 +19913,7 @@ void gguf_free(struct gguf_context * ctx) { } } - GGML_ALIGNED_FREE(ctx->infos); + free(ctx->infos); } GGML_ALIGNED_FREE(ctx); diff --git a/ggml.h b/ggml.h index 4ef3d5253..8b410cc85 100644 --- a/ggml.h +++ b/ggml.h @@ -952,11 +952,11 @@ extern "C" { // a - x // b - dy - // TODO: update with configurable eps GGML_API struct ggml_tensor * ggml_rms_norm_back( struct ggml_context * ctx, struct ggml_tensor * a, - struct ggml_tensor * b); + struct ggml_tensor * b, + float eps); // A: n columns, m rows // B: n columns, p rows (i.e. we transpose it internally) @@ -1612,7 +1612,8 @@ extern "C" { struct ggml_tensor * tensor); - GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor); + GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor); + GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor); GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep); @@ -1677,6 +1678,8 @@ extern "C" { GGML_LINESEARCH_INVALID_PARAMETERS, }; + typedef void (*ggml_opt_callback)(void * data, float * sched); + // optimization parameters // // see ggml.c (ggml_opt_default_params) for default values @@ -1712,12 +1715,14 @@ extern "C" { float sched; // schedule multiplier (fixed, decay or warmup) float decay; // weight decay for AdamW, use 0.0f to disable + int decay_min_ndim; // minimum number of tensor dimension to apply weight decay float alpha; // learning rate float beta1; float beta2; float eps; // epsilon for numerical stability float eps_f; // epsilon for convergence test float eps_g; // epsilon for convergence test + float gclip; // gradient clipping } adam; // LBFGS parameters @@ -1745,14 +1750,12 @@ extern "C" { bool just_initialized; + float loss_before; + float loss_after; + struct { - struct ggml_tensor * x; // view of the parameters - struct ggml_tensor * g1; // gradient - struct ggml_tensor * g2; // gradient squared struct ggml_tensor * m; // first moment struct ggml_tensor * v; // second moment - struct ggml_tensor * mh; // first moment hat - struct ggml_tensor * vh; // second moment hat struct ggml_tensor * pf; // past function values float fx_best; float fx_prev; @@ -1789,10 +1792,10 @@ extern "C" { // initialize optimizer context GGML_API void ggml_opt_init( - struct ggml_context * ctx, + struct ggml_context * ctx, struct ggml_opt_context * opt, - struct ggml_opt_params params, - int64_t nx); + struct ggml_opt_params params, + int64_t nx); // continue optimizing the function defined by the tensor f GGML_API enum ggml_opt_result ggml_opt_resume( @@ -1806,7 +1809,9 @@ extern "C" { struct ggml_opt_context * opt, struct ggml_tensor * f, struct ggml_cgraph * gf, - struct ggml_cgraph * gb); + struct ggml_cgraph * gb, + ggml_opt_callback callback, + void * callback_data); // // quantization diff --git a/llama.cpp b/llama.cpp index 11697ee65..7cb468538 100644 --- a/llama.cpp +++ b/llama.cpp @@ -6248,7 +6248,6 @@ const char * llama_print_system_info(void) { } void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) { - fprintf(stream, "\n"); fprintf(stream, "###########\n"); fprintf(stream, "# Timings #\n"); @@ -6264,10 +6263,10 @@ void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx) { fprintf(stream, "n_eval: %d # number of tokens generated (excluding the first one)\n", ctx->n_eval); fprintf(stream, "n_p_eval: %d # number of tokens processed in batches at the beginning\n", ctx->n_p_eval); fprintf(stream, "n_sample: %d # number of sampled tokens\n", ctx->n_sample); - fprintf(stream, "t_eval_us: %ld # total microseconds spent generating tokens\n", ctx->t_eval_us); - fprintf(stream, "t_load_us: %ld # total microseconds spent loading the model\n", ctx->t_load_us); - fprintf(stream, "t_p_eval_us: %ld # total microseconds spent prompt processing\n", ctx->t_p_eval_us); - fprintf(stream, "t_sample_us: %ld # total microseconds spent sampling\n", ctx->t_sample_us); + fprintf(stream, "t_eval_us: %" PRId64 " # total microseconds spent generating tokens\n", ctx->t_eval_us); + fprintf(stream, "t_load_us: %" PRId64 " # total microseconds spent loading the model\n", ctx->t_load_us); + fprintf(stream, "t_p_eval_us: %" PRId64 " # total microseconds spent prompt processing\n", ctx->t_p_eval_us); + fprintf(stream, "t_sample_us: %" PRId64 " # total microseconds spent sampling\n", ctx->t_sample_us); fprintf(stream, "ts_eval: %.2f # tokens / second during generation\n", 1.0e6 * ctx->n_eval / ctx->t_eval_us); fprintf(stream, "ts_p_eval: %.2f # tokens / second during prompt processing\n", diff --git a/tests/test-grad0.cpp b/tests/test-grad0.cpp index 75a698d73..468cde66a 100644 --- a/tests/test-grad0.cpp +++ b/tests/test-grad0.cpp @@ -275,14 +275,14 @@ static bool check_gradient( ggml_graph_compute_with_ctx(ctx0, &gf, n_threads); - const float f0 = ggml_get_f32_1d(f, 0); + const double f0 = ggml_get_f32_1d(f, 0); ggml_set_f32_1d(x[i], k, xm); ggml_graph_compute_with_ctx(ctx0, &gf, n_threads); - const float f1 = ggml_get_f32_1d(f, 0); - const float g0 = (f0 - f1)/(2.0f*eps); + const double f1 = ggml_get_f32_1d(f, 0); + const double g0 = (f0 - f1)/(2.0*(double) eps); ggml_set_f32_1d(x[i], k, x0); @@ -292,10 +292,10 @@ static bool check_gradient( ggml_graph_compute_with_ctx(ctx0, &gb, n_threads); - const float g1 = ggml_get_f32_1d(x[i]->grad, k); + const double g1 = ggml_get_f32_1d(x[i]->grad, k); - const float error_abs = fabsf(g0 - g1); - const float error_rel = g0 != 0 ? fabsf(g0 - g1)/fabsf(g0) : 0; + const double error_abs = fabs(g0 - g1); + const double error_rel = g0 != 0 ? fabs(g0 - g1)/fabs(g0) : 0; if (error_abs > max_error_abs || error_rel > max_error_rel) { printf("%s: ndims=%d, i=%d, k=%d, x0=%f, xm=%f, xp=%f, f0=%f, f1=%f, g0=%f, g1=%f, eps=%f, error_abs=%f, error_rel=%f\n", @@ -531,7 +531,7 @@ int main(int argc, const char ** argv) { struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqrt(ctx0, x[0])); - check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-1f); + check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, 2e-2f, 1e-1f); } } @@ -1345,9 +1345,18 @@ int main(int argc, const char ** argv) { x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f); ggml_set_param(ctx0, x[0]); - struct ggml_tensor * f = ggml_sum(ctx0, ggml_soft_max(ctx0, x[0])); + float eps = 1e-6f; + // dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work + // instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0) + struct ggml_tensor * f = ggml_sum(ctx0, + ggml_log(ctx0, + ggml_add1(ctx0, + ggml_scale(ctx0, + ggml_soft_max(ctx0, x[0]), + ggml_new_f32(ctx0, 1.0f - eps)), + ggml_new_f32(ctx0, eps)))); - check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY); + check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY); } } @@ -1358,15 +1367,26 @@ int main(int argc, const char ** argv) { int64_t ne2[4]; get_random_dims(ne2, 4); - for (int ndims = 1; ndims <= 3; ++ndims) { - x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f); + for (int ndims = 1; ndims <= 4; ++ndims) { + x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -0.1f, 0.1f); x[1] = get_random_tensor_f32(ctx0, ndims, ne2, 0.0f, 1.0f); + // the second argument to cross_entropy_loss must sum up to 1 for each row + int nr = ggml_nrows(x[1]); + int nc = ggml_nelements(x[1]) / nr; + for (int ir = 0; ir < nr; ++ir) { + float sum = 0; + for (int ic = 0; ic < nc; ++ic) { + sum += ((float *) x[1]->data)[ic + ir*nc]; + } + for (int ic = 0; ic < nc; ++ic) { + ((float *) x[1]->data)[ic + ir*nc] /= sum; + } + } ggml_set_param(ctx0, x[0]); - struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1])); + struct ggml_tensor * f = ggml_cross_entropy_loss(ctx0, x[0], x[1]); - check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-1f, 1e-2f, INFINITY); - // finite differences regularly fails! + check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-4f, 1e-3f, INFINITY); } } @@ -1473,7 +1493,7 @@ int main(int argc, const char ** argv) { struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0))); - check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f); + check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY); } } } @@ -1514,7 +1534,7 @@ int main(int argc, const char ** argv) { struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0))); - check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f); + check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY); } } } From 95b6e5212f5e4e1419de1d833d7f8d788f9f2227 Mon Sep 17 00:00:00 2001 From: Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com> Date: Mon, 28 Aug 2023 23:33:27 -0700 Subject: [PATCH 085/147] added `struct` to llama_dump_timing_info_yaml's `llama_context` (#2857) fixes C compat. --- llama.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.h b/llama.h index b38d3be20..6e5e1df63 100644 --- a/llama.h +++ b/llama.h @@ -521,7 +521,7 @@ extern "C" { // If this is not called, or NULL is supplied, everything is output on stderr. LLAMA_API void llama_log_set(llama_log_callback log_callback, void * user_data); - LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const llama_context * ctx); + LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx); #ifdef __cplusplus } From 611363ac791435497e66278dfe31ac8a4e11fa4f Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 29 Aug 2023 10:50:30 +0300 Subject: [PATCH 086/147] scripts : add pipefail --- scripts/qnt-all.sh | 1 + scripts/run-all-perf.sh | 1 + scripts/run-all-ppl.sh | 1 + 3 files changed, 3 insertions(+) diff --git a/scripts/qnt-all.sh b/scripts/qnt-all.sh index 1376e4194..b4c2a159e 100755 --- a/scripts/qnt-all.sh +++ b/scripts/qnt-all.sh @@ -20,6 +20,7 @@ fi model="$1" out="../tmp/results-${model}" +set -o pipefail set -e mkdir -p ${out} diff --git a/scripts/run-all-perf.sh b/scripts/run-all-perf.sh index 7391e3dd5..6384e364d 100755 --- a/scripts/run-all-perf.sh +++ b/scripts/run-all-perf.sh @@ -20,6 +20,7 @@ fi model="$1" out="../tmp/results-${model}" +set -o pipefail set -e mkdir -p ${out} diff --git a/scripts/run-all-ppl.sh b/scripts/run-all-ppl.sh index f643ca3ae..e04d61d7f 100755 --- a/scripts/run-all-ppl.sh +++ b/scripts/run-all-ppl.sh @@ -17,6 +17,7 @@ if [ ! -z "$3" ]; then args="$3" fi +set -o pipefail set -e model="$1" From 3a007648f230ea37d6cca5e63850f04ebb12d2cf Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 29 Aug 2023 11:33:46 +0300 Subject: [PATCH 087/147] metal : add option to disable debug logs (close #2764) --- CMakeLists.txt | 2 +- Makefile | 2 +- ggml-metal.m | 71 +++++++++++++++++++++++--------------------------- 3 files changed, 35 insertions(+), 40 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index ba008bcc6..1eae2d670 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -301,7 +301,7 @@ if (LLAMA_METAL) set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h) add_compile_definitions(GGML_USE_METAL) - add_compile_definitions(GGML_METAL_NDEBUG) + #add_compile_definitions(GGML_METAL_NDEBUG) # get full path to the file #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/") diff --git a/Makefile b/Makefile index e60821dd5..a64374e7d 100644 --- a/Makefile +++ b/Makefile @@ -305,7 +305,7 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h endif # LLAMA_HIPBLAS ifdef LLAMA_METAL - CFLAGS += -DGGML_USE_METAL -DGGML_METAL_NDEBUG + CFLAGS += -DGGML_USE_METAL #-DGGML_METAL_NDEBUG CXXFLAGS += -DGGML_USE_METAL LDFLAGS += -framework Foundation -framework Metal -framework MetalKit OBJS += ggml-metal.o diff --git a/ggml-metal.m b/ggml-metal.m index ad2ee8cf5..e929c4b07 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -11,6 +11,7 @@ #define MIN(a, b) ((a) < (b) ? (a) : (b)) #define MAX(a, b) ((a) > (b) ? (a) : (b)) +// TODO: temporary - reuse llama.cpp logging #ifdef GGML_METAL_NDEBUG #define metal_printf(...) #else @@ -113,7 +114,7 @@ static NSString * const msl_library_source = @"see metal.metal"; @end struct ggml_metal_context * ggml_metal_init(int n_cb) { - fprintf(stderr, "%s: allocating\n", __func__); + metal_printf("%s: allocating\n", __func__); struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context)); @@ -132,7 +133,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error]; if (error) { - fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]); + metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; } } @@ -146,11 +147,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"]; NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]]; NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; - fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]); + metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]); NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error]; if (error) { - fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]); + metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; } @@ -162,7 +163,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error]; #endif if (error) { - fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]); + metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]); return NULL; } } @@ -174,11 +175,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { #define GGML_METAL_ADD_KERNEL(name) \ ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \ ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \ - fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \ + metal_printf("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \ (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \ (int) ctx->pipeline_##name.threadExecutionWidth); \ if (error) { \ - fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ + metal_printf("%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ return NULL; \ } @@ -230,19 +231,19 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { #undef GGML_METAL_ADD_KERNEL } - fprintf(stderr, "%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); - fprintf(stderr, "%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false"); + metal_printf("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); + metal_printf("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false"); if (ctx->device.maxTransferRate != 0) { - fprintf(stderr, "%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0); + metal_printf("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0); } else { - fprintf(stderr, "%s: maxTransferRate = built-in GPU\n", __func__); + metal_printf("%s: maxTransferRate = built-in GPU\n", __func__); } return ctx; } void ggml_metal_free(struct ggml_metal_context * ctx) { - fprintf(stderr, "%s: deallocating\n", __func__); + metal_printf("%s: deallocating\n", __func__); #define GGML_METAL_DEL_KERNEL(name) \ [ctx->function_##name release]; \ [ctx->pipeline_##name release]; @@ -311,7 +312,7 @@ void * ggml_metal_host_malloc(size_t n) { void * data = NULL; const int result = posix_memalign((void **) &data, getpagesize(), n); if (result != 0) { - fprintf(stderr, "%s: error: posix_memalign failed\n", __func__); + metal_printf("%s: error: posix_memalign failed\n", __func__); return NULL; } @@ -339,7 +340,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) { // Metal buffer based on the host memory pointer // static id ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) { - //fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach); + //metal_printf("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach); const int64_t tsize = ggml_nbytes(t); @@ -350,13 +351,13 @@ static id ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) { *offs = (size_t) ioffs; - //fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs); + //metal_printf("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs); return ctx->buffers[i].metal; } } - fprintf(stderr, "%s: error: buffer is nil\n", __func__); + metal_printf("%s: error: buffer is nil\n", __func__); return nil; } @@ -368,7 +369,7 @@ bool ggml_metal_add_buffer( size_t size, size_t max_size) { if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) { - fprintf(stderr, "%s: too many buffers\n", __func__); + metal_printf("%s: too many buffers\n", __func__); return false; } @@ -378,7 +379,7 @@ bool ggml_metal_add_buffer( const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data; if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) { - fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name); + metal_printf("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name); return false; } } @@ -399,11 +400,11 @@ bool ggml_metal_add_buffer( ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil]; if (ctx->buffers[ctx->n_buffers].metal == nil) { - fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0); + metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0); return false; } - fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0); + metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0); ++ctx->n_buffers; } else { @@ -423,27 +424,27 @@ bool ggml_metal_add_buffer( ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil]; if (ctx->buffers[ctx->n_buffers].metal == nil) { - fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0); + metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0); return false; } - fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i); + metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i); if (i + size_step < size) { - fprintf(stderr, "\n"); + metal_printf("\n"); } ++ctx->n_buffers; } } - fprintf(stderr, ", (%8.2f / %8.2f)", + metal_printf(", (%8.2f / %8.2f)", ctx->device.currentAllocatedSize / 1024.0 / 1024.0, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) { - fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n"); + metal_printf(", warning: current allocated size is greater than the recommended max working set size\n"); } else { - fprintf(stderr, "\n"); + metal_printf("\n"); } } @@ -453,8 +454,6 @@ bool ggml_metal_add_buffer( void ggml_metal_set_tensor( struct ggml_metal_context * ctx, struct ggml_tensor * t) { - metal_printf("%s: set input for tensor '%s'\n", __func__, t->name); - size_t offs; id id_dst = ggml_metal_get_buffer(ctx, t, &offs); @@ -464,8 +463,6 @@ void ggml_metal_set_tensor( void ggml_metal_get_tensor( struct ggml_metal_context * ctx, struct ggml_tensor * t) { - metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name); - size_t offs; id id_src = ggml_metal_get_buffer(ctx, t, &offs); @@ -560,15 +557,13 @@ void ggml_metal_graph_find_concurrency( } if (ctx->concur_list_len > GGML_MAX_CONCUR) { - fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__); + metal_printf("%s: too many elements for metal ctx->concur_list!\n", __func__); } } void ggml_metal_graph_compute( struct ggml_metal_context * ctx, struct ggml_cgraph * gf) { - metal_printf("%s: evaluating graph\n", __func__); - @autoreleasepool { // if there is ctx->concur_list, dispatch concurrently @@ -616,7 +611,7 @@ void ggml_metal_graph_compute( continue; } - metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op)); + //metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op)); struct ggml_tensor * src0 = gf->nodes[i]->src[0]; struct ggml_tensor * src1 = gf->nodes[i]->src[1]; @@ -764,7 +759,7 @@ void ggml_metal_graph_compute( } break; default: { - fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op)); + metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op)); GGML_ASSERT(false); } } break; @@ -923,7 +918,7 @@ void ggml_metal_graph_compute( } break; default: { - fprintf(stderr, "Asserting on type %d\n",(int)src0t); + metal_printf("Asserting on type %d\n",(int)src0t); GGML_ASSERT(false && "not implemented"); } }; @@ -1161,7 +1156,7 @@ void ggml_metal_graph_compute( } break; default: { - fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op)); + metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op)); GGML_ASSERT(false); } } @@ -1186,7 +1181,7 @@ void ggml_metal_graph_compute( MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status]; if (status != MTLCommandBufferStatusCompleted) { - fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status); + metal_printf("%s: command buffer %d failed with status %lu\n", __func__, i, status); GGML_ASSERT(false); } } From d4b5e16c32ba9c5fa6bbd035e80a99c113050cde Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Tue, 29 Aug 2023 04:42:41 -0400 Subject: [PATCH 088/147] make : fix clang tests build, add missing examples (#2859) * make : do not pass headers to the compiler This fixes building tests with clang. * make : add missing examples * make : fix build-info.h dependencies --- Makefile | 49 ++++++++++++++++++++++++++++++++----------------- 1 file changed, 32 insertions(+), 17 deletions(-) diff --git a/Makefile b/Makefile index a64374e7d..02ba3e36d 100644 --- a/Makefile +++ b/Makefile @@ -1,5 +1,5 @@ # Define the default target now so that it is always the first target -BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple server embd-input-test gguf llama-bench +BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam_search # Binaries only useful for tests TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1 @@ -356,7 +356,7 @@ OBJS += ggml-alloc.o llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h $(CXX) $(CXXFLAGS) -c $< -o $@ -common.o: common/common.cpp common/common.h +common.o: common/common.cpp common/common.h build-info.h $(CXX) $(CXXFLAGS) -c $< -o $@ console.o: common/console.cpp common/console.h @@ -369,7 +369,7 @@ libllama.so: llama.o ggml.o $(OBJS) $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) clean: - rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch convert-llama2c-to-ggml embd-input-test gguf llama-bench build-info.h $(TEST_TARGETS) + rm -vf *.o *.so *.dll benchmark-matmult build-info.h $(BUILD_TARGETS) $(TEST_TARGETS) # # Examples @@ -409,18 +409,33 @@ $(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-in embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput -gguf: examples/gguf/gguf.cpp build-info.h ggml.o llama.o $(OBJS) +gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp build-info.h ggml.o llama.o common.o $(OBJS) +train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp build-info.h ggml.o llama.o $(OBJS) +convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) +baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + +beam_search: examples/beam_search/beam_search.cpp build-info.h ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + +ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))' +BUILD_TARGETS += metal +endif + +ifdef LLAMA_METAL +metal: examples/metal/metal.cpp ggml.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) +endif + build-info.h: $(wildcard .git/index) scripts/build-info.sh @sh scripts/build-info.sh > $@.tmp @if ! cmp -s $@.tmp $@; then \ @@ -443,34 +458,34 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o common.o grammar-parser.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS) - $(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) From 74e0caeb82fc9db77fa2cc93070bb919a9a935dd Mon Sep 17 00:00:00 2001 From: Jhen-Jie Hong Date: Tue, 29 Aug 2023 17:30:10 +0800 Subject: [PATCH 089/147] readme : add react-native binding (#2869) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index bf3eb0b76..8d54a558d 100644 --- a/README.md +++ b/README.md @@ -113,6 +113,7 @@ as the main playground for developing new features for the [ggml](https://github - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp) - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s) - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj) +- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn) **UI:** From bcce96ba4dd95482824700c4ce2455fe8c49055a Mon Sep 17 00:00:00 2001 From: jameswu2014 <545426914@qq.com> Date: Tue, 29 Aug 2023 17:48:41 +0800 Subject: [PATCH 090/147] convert.py : fix baichuan7B support (#2870) * [Fix]: convert.py support baichuan7B * convert.py : fix trailing whitespaces --------- Co-authored-by: Georgi Gerganov --- convert.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/convert.py b/convert.py index a15e6ccd2..3f0a1c932 100755 --- a/convert.py +++ b/convert.py @@ -469,7 +469,7 @@ class UnquantizedTensor(Tensor): def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': r = self.ndarray.shape[0] // 3 - return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head)) + return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head)) def part(self, n_part: int) -> 'UnquantizedTensor': r = self.ndarray.shape[0] // 3 @@ -952,9 +952,10 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel: #tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"] elif f"model.layers.{i}.self_attn.W_pack.weight" in model: print(f"Unpacking and permuting layer {i}") - tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head) - tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv) + tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head) + tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head) tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2) + del tmp[f"model.layers.{i}.self_attn.W_pack.weight"] else: break From 53885d7256909ec3e2176cdc2477f3986c15ec69 Mon Sep 17 00:00:00 2001 From: maddes8cht <55592906+maddes8cht@users.noreply.github.com> Date: Tue, 29 Aug 2023 15:51:02 +0200 Subject: [PATCH 091/147] py : fix "usage" messages (#2873) convert-to-gguf python scripts --- convert-falcon-hf-to-gguf.py | 2 +- convert-gptneox-hf-to-gguf.py | 2 +- convert-llama-7b-pth-to-gguf.py | 2 +- convert-llama-hf-to-gguf.py | 2 +- 4 files changed, 4 insertions(+), 4 deletions(-) diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py index 411cbf682..168bcf17f 100755 --- a/convert-falcon-hf-to-gguf.py +++ b/convert-falcon-hf-to-gguf.py @@ -48,7 +48,7 @@ def count_model_parts(dir_model: str) -> int: if len(sys.argv) < 3: - print("Usage: convert-h5-to-ggml.py dir-model ftype\n") + print(f"Usage: python {sys.argv[0]} dir-model ftype\n") print(" ftype == 0 -> float32") print(" ftype == 1 -> float16") sys.exit(1) diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py index 6eeff5bb1..d9c42d76b 100755 --- a/convert-gptneox-hf-to-gguf.py +++ b/convert-gptneox-hf-to-gguf.py @@ -50,7 +50,7 @@ def count_model_parts(dir_model: str) -> int: if len(sys.argv) < 3: - print("Usage: convert-h5-to-ggml.py dir-model ftype\n") + print(f"Usage: python {sys.argv[0]} dir-model ftype\n") print(" ftype == 0 -> float32") print(" ftype == 1 -> float16") sys.exit(1) diff --git a/convert-llama-7b-pth-to-gguf.py b/convert-llama-7b-pth-to-gguf.py index f103f5f61..2ab082383 100755 --- a/convert-llama-7b-pth-to-gguf.py +++ b/convert-llama-7b-pth-to-gguf.py @@ -32,7 +32,7 @@ def count_model_parts(dir_model: str) -> int: if len(sys.argv) < 3: - print("Usage: convert-h5-to-ggml.py dir-model ftype\n") + print(f"Usage: python {sys.argv[0]} dir-model ftype\n") print(" ftype == 0 -> float32") print(" ftype == 1 -> float16") diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py index 08fde238b..b00810dbb 100755 --- a/convert-llama-hf-to-gguf.py +++ b/convert-llama-hf-to-gguf.py @@ -44,7 +44,7 @@ def count_model_parts(dir_model: str) -> int: if len(sys.argv) < 3: - print("Usage: convert-h5-to-ggml.py dir-model ftype\n") + print(f"Usage: python {sys.argv[0]} dir-model ftype\n") print(" ftype == 0 -> float32") print(" ftype == 1 -> float16") From e37e69dcc3d52f21222a63cafed2a71b3f6b53c6 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Tue, 29 Aug 2023 23:55:03 +0300 Subject: [PATCH 092/147] 10X faster BPE tokenizer (#2876) * 10X faster BPE tokenizer * Remove comment that no longer applies --------- Co-authored-by: Iwan Kawrakow --- llama.cpp | 19 +++++++++---------- 1 file changed, 9 insertions(+), 10 deletions(-) diff --git a/llama.cpp b/llama.cpp index 7cb468538..fcd6f276a 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3211,7 +3211,7 @@ private: struct llm_bigram_bpe { struct comparator { - bool operator()(llm_bigram_bpe & l, llm_bigram_bpe & r) { + bool operator()(const llm_bigram_bpe & l, const llm_bigram_bpe & r) const { return l.rank > r.rank || (l.rank == r.rank && l.left > r.left); } }; @@ -3359,23 +3359,22 @@ private: } // probably not 100% correct - // TODO: this is quite slow - how to make it more efficient? - static std::vector bpe_gpt2_preprocess(std::string text) { + static std::vector bpe_gpt2_preprocess(const std::string & text) { std::vector words; // ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53 const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"; const std::regex re(pattern); - std::smatch m; - while (std::regex_search(text, m, re)) { - for (auto x : m) { - words.push_back(x); - } - text = m.suffix(); + auto words_begin = std::sregex_iterator(text.begin(), text.end(), re); + auto words_end = std::sregex_iterator(); + auto n_words = std::distance(words_begin, words_end); + words.reserve(n_words); + for (auto it = words_begin; it != words_end; ++it) { + words.push_back(it->str()); } - return words; + } const llama_vocab & vocab; From fa3582f509a2715e80a473e79f88dcd1ebff44c2 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Tue, 29 Aug 2023 23:55:45 +0300 Subject: [PATCH 093/147] Tell users attmepting to run perplexity with too few tokens to use more (#2882) Closes #2858 Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 15 +++++++++++++++ 1 file changed, 15 insertions(+) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index aeb774c5f..7c02b6d40 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -142,6 +142,14 @@ results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) fprintf(stderr, "%s: tokenizing the input ..\n", __func__); std::vector tokens = ::llama_tokenize(ctx, params.prompt, add_bos); + + if (int(tokens.size()) < 2*params.n_ctx) { + fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*params.n_ctx, + params.n_ctx); + fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size()); + return {std::move(tokens), 0., {}, {}}; + } + std::vector logit_history; std::vector prob_history; @@ -274,6 +282,13 @@ results_perplexity perplexity(llama_context * ctx, const gpt_params & params) { auto tim2 = std::chrono::high_resolution_clock::now(); fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast(tim2-tim1).count()); + if (int(tokens.size()) < 2*params.n_ctx) { + fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*params.n_ctx, + params.n_ctx); + fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size()); + return {std::move(tokens), 0., {}, {}}; + } + std::vector logit_history; logit_history.resize(tokens.size()); From c03a243abf9f30889f31fefdfa94fe9d7034820c Mon Sep 17 00:00:00 2001 From: slaren Date: Tue, 29 Aug 2023 23:17:34 +0200 Subject: [PATCH 094/147] remove outdated references to -eps and -gqa from README (#2881) --- README.md | 2 -- 1 file changed, 2 deletions(-) diff --git a/README.md b/README.md index 8d54a558d..a880fd29f 100644 --- a/README.md +++ b/README.md @@ -729,8 +729,6 @@ python3 convert.py pygmalion-7b/ --outtype q4_1 - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGML) - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGML) - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGML) -- Specify `-eps 1e-5` for best generation quality -- Specify `-gqa 8` for 70B models to work ### Verifying the model files From 06abf8eebabe086ca4003dee2754ab45032cd3fd Mon Sep 17 00:00:00 2001 From: slaren Date: Tue, 29 Aug 2023 23:24:42 +0200 Subject: [PATCH 095/147] ggml : add view_src and view_offs to ggml_tensor for views (#2874) * ggml : add view_src and view_offs * update ggml-alloc to use view_src * update ggml_diag_mask to work correctly with automatic inplace * exclude other ops that set an inplace flag from automatic inplace --- ggml-alloc.c | 53 ++---------- ggml.c | 233 +++++++++++++++++++++++---------------------------- ggml.h | 5 +- 3 files changed, 113 insertions(+), 178 deletions(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index 63beb1d4e..f07a4a217 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -321,8 +321,7 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) { //////////// compute graph allocator static bool ggml_is_view(struct ggml_tensor * t) { - return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE || - t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY; + return t->view_src != NULL; } static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) { @@ -340,28 +339,6 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml return true; } -static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) { - switch (t->op) { - case GGML_OP_PERMUTE: - case GGML_OP_RESHAPE: - case GGML_OP_TRANSPOSE: - case GGML_OP_VIEW: - return t->src[0]; - case GGML_OP_CPY: - return t->src[1]; - default: - return NULL; - } -} - -static struct ggml_tensor * get_view_source(struct ggml_tensor * t) { - struct ggml_tensor * parent = t; - do { - parent = get_view_parent(parent); - } while (ggml_is_view(parent)); - return parent; -} - static bool ggml_op_can_inplace(enum ggml_op op) { switch (op) { case GGML_OP_SCALE: @@ -369,7 +346,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) { case GGML_OP_DIAG_MASK_INF: case GGML_OP_ADD: case GGML_OP_ADD1: - case GGML_OP_ACC: case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: @@ -379,7 +355,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) { case GGML_OP_UNARY: case GGML_OP_ROPE: case GGML_OP_RMS_NORM: - case GGML_OP_SET: case GGML_OP_SOFT_MAX: case GGML_OP_CONT: return true; @@ -393,24 +368,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) struct hash_node * ht = alloc->hash_table; if (node->data == NULL) { if (ggml_is_view(node)) { - size_t offset; - switch(node->op) { - case GGML_OP_VIEW: - memcpy(&offset, node->op_params, sizeof(size_t)); - node->data = (char *) node->src[0]->data + offset; - break; - case GGML_OP_PERMUTE: - case GGML_OP_RESHAPE: - case GGML_OP_TRANSPOSE: - node->data = node->src[0]->data; - break; - case GGML_OP_CPY: - node->data = node->src[1]->data; - break; - default: - GGML_ASSERT(!"unknown view op"); - break; - } + assert(node->view_src->data != NULL); + node->data = (char *)node->view_src->data + node->view_offs; } else { // see if we can reuse a parent's buffer (inplace) if (ggml_op_can_inplace(node->op)) { @@ -430,7 +389,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) struct hash_node * p_hn = hash_get(ht, parent); if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) { if (ggml_is_view(parent)) { - struct ggml_tensor * view_src = get_view_source(parent); + struct ggml_tensor * view_src = parent->view_src; struct hash_node * view_src_hn = hash_get(ht, view_src); if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) { // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite @@ -472,7 +431,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n( struct ggml_tensor * node = gf->nodes[i]; if (ggml_is_view(node)) { - struct ggml_tensor * view_src = get_view_source(node); + struct ggml_tensor * view_src = node->view_src; hash_get(ht, view_src)->n_views += 1; } @@ -557,7 +516,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n( if (p_hn->n_children == 0 && p_hn->n_views == 0) { if (ggml_is_view(parent)) { - struct ggml_tensor * view_src = get_view_source(parent); + struct ggml_tensor * view_src = parent->view_src; struct hash_node * view_src_hn = hash_get(ht, view_src); view_src_hn->n_views -= 1; AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views); diff --git a/ggml.c b/ggml.c index 9a787863d..46ce4a581 100644 --- a/ggml.c +++ b/ggml.c @@ -4104,16 +4104,11 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) { } size_t ggml_nbytes(const struct ggml_tensor * tensor) { - static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); - - // this should handle cases where the tensor is not contiguous in memory - // probaby just: - // - // return tensor->ne[3]*tensor->nb[3] - // - // is enough, but just in case, adding the second part - - return MAX(tensor->ne[3]*tensor->nb[3], (ggml_nelements(tensor)*ggml_type_size(tensor->type))/ggml_blck_size(tensor->type)); + size_t nbytes = tensor->ne[0]*tensor->nb[0]/ggml_blck_size(tensor->type); + for (int i = 1; i < GGML_MAX_DIMS; ++i) { + nbytes += (tensor->ne[i] - 1)*tensor->nb[i]; + } + return nbytes; } size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) { @@ -4567,36 +4562,51 @@ static struct ggml_tensor * ggml_new_tensor_impl( enum ggml_type type, int n_dims, const int64_t * ne, - void * data) { + struct ggml_tensor * view_src, + size_t view_offs) { assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS); - size_t data_size = 0; + // find the base tensor and absolute offset + if (view_src != NULL && view_src->view_src != NULL) { + view_offs += view_src->view_offs; + view_src = view_src->view_src; + } - if (data == NULL && !ctx->no_alloc) { - data_size += ggml_type_size(type)*(ne[0]/ggml_blck_size(type)); - for (int i = 1; i < n_dims; i++) { - data_size *= ne[i]; + size_t data_size = ggml_type_size(type)*(ne[0]/ggml_blck_size(type)); + for (int i = 1; i < n_dims; i++) { + data_size *= ne[i]; + } + + GGML_ASSERT(view_src == NULL || data_size + view_offs <= ggml_nbytes(view_src)); + + void * data = view_src != NULL ? view_src->data : NULL; + if (data != NULL) { + data = (char *) data + view_offs; + } + + size_t obj_alloc_size = 0; + + if (view_src == NULL && ctx->no_alloc == false) { + if (ctx->scratch.data != NULL) { + // allocate tensor data in the scratch buffer + if (ctx->scratch.offs + data_size > ctx->scratch.size) { + GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n", + __func__, ctx->scratch.offs + data_size, ctx->scratch.size); + assert(false); + return NULL; + } + + data = (char * const) ctx->scratch.data + ctx->scratch.offs; + + ctx->scratch.offs += data_size; + } else { + // allocate tensor data in the context's memory pool + obj_alloc_size = data_size; } } - if (ctx->scratch.data != NULL && data == NULL) { - // allocate tensor data in the scratch buffer - if (ctx->scratch.offs + data_size > ctx->scratch.size) { - GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n", - __func__, ctx->scratch.offs + data_size, ctx->scratch.size); - assert(false); - return NULL; - } - - data = (char * const) ctx->scratch.data + ctx->scratch.offs; - - ctx->scratch.offs += data_size; - - data_size = 0; - } - - struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + data_size); + struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size); // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here @@ -4616,7 +4626,9 @@ static struct ggml_tensor * ggml_new_tensor_impl( /*.perf_runs =*/ 0, /*.perf_cycles =*/ 0, /*.perf_time_us =*/ 0, - /*.data =*/ (data == NULL && !ctx->no_alloc) ? (void *)(result + 1) : data, + /*.view_src =*/ view_src, + /*.view_offs =*/ view_offs, + /*.data =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data, /*.name =*/ { 0 }, /*.extra =*/ NULL, /*.padding =*/ { 0 }, @@ -4640,28 +4652,12 @@ static struct ggml_tensor * ggml_new_tensor_impl( return result; } -static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) { - GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings - assert(params_size <= GGML_MAX_OP_PARAMS); - memcpy(tensor->op_params, params, params_size); -} - -static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) { - assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t)); - return ((const int32_t *)(tensor->op_params))[i]; -} - -static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) { - assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t)); - ((int32_t *)(tensor->op_params))[i] = value; -} - struct ggml_tensor * ggml_new_tensor( struct ggml_context * ctx, enum ggml_type type, int n_dims, const int64_t * ne) { - return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL); + return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL, 0); } struct ggml_tensor * ggml_new_tensor_1d( @@ -4726,7 +4722,23 @@ struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) { } struct ggml_tensor * ggml_dup_tensor(struct ggml_context * ctx, const struct ggml_tensor * src) { - return ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, NULL); + return ggml_new_tensor(ctx, src->type, src->n_dims, src->ne); +} + +static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) { + GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings + assert(params_size <= GGML_MAX_OP_PARAMS); + memcpy(tensor->op_params, params, params_size); +} + +static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_t i) { + assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t)); + return ((const int32_t *)(tensor->op_params))[i]; +} + +static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) { + assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t)); + ((int32_t *)(tensor->op_params))[i] = value; } struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) { @@ -5012,14 +5024,13 @@ struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char * struct ggml_tensor * ggml_view_tensor( struct ggml_context * ctx, - const struct ggml_tensor * src) { - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src->data); + struct ggml_tensor * src) { + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src, 0); ggml_format_name(result, "%s (view)", src->name); - result->nb[0] = src->nb[0]; - result->nb[1] = src->nb[1]; - result->nb[2] = src->nb[2]; - result->nb[3] = src->nb[3]; + for (int i = 0; i < GGML_MAX_DIMS; i++) { + result->nb[i] = src->nb[i]; + } return result; } @@ -5592,7 +5603,7 @@ struct ggml_tensor * ggml_repeat_back( // ggml_concat -struct ggml_tensor* ggml_concat( +struct ggml_tensor * ggml_concat( struct ggml_context* ctx, struct ggml_tensor* a, struct ggml_tensor* b) { @@ -6201,7 +6212,7 @@ struct ggml_tensor * ggml_reshape( //GGML_ASSERT(false); } - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a->data); + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); result->op = GGML_OP_RESHAPE; @@ -6225,7 +6236,7 @@ struct ggml_tensor * ggml_reshape_1d( } const int64_t ne[1] = { ne0 }; - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a->data); + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); result->op = GGML_OP_RESHAPE; @@ -6250,7 +6261,7 @@ struct ggml_tensor * ggml_reshape_2d( } const int64_t ne[2] = { ne0, ne1 }; - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a->data); + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); result->op = GGML_OP_RESHAPE; @@ -6276,7 +6287,7 @@ struct ggml_tensor * ggml_reshape_3d( } const int64_t ne[3] = { ne0, ne1, ne2 }; - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a->data); + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); result->op = GGML_OP_RESHAPE; @@ -6286,7 +6297,6 @@ struct ggml_tensor * ggml_reshape_3d( return result; } - struct ggml_tensor * ggml_reshape_4d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -6304,7 +6314,7 @@ struct ggml_tensor * ggml_reshape_4d( } const int64_t ne[4] = { ne0, ne1, ne2, ne3 }; - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a->data); + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a, 0); ggml_format_name(result, "%s (reshaped)", a->name); result->op = GGML_OP_RESHAPE; @@ -6314,34 +6324,12 @@ struct ggml_tensor * ggml_reshape_4d( return result; } -// ggml_view_1d - -static struct ggml_tensor * ggml_view_tensor_offset( +static struct ggml_tensor * ggml_view_impl( struct ggml_context * ctx, struct ggml_tensor * a, int n_dims, const int64_t * ne, size_t offset) { - // don't calculate an offset from an unallocated tensor - void * data = NULL; - if (a->data != NULL) { - data = (char *) a->data + offset; - } - - struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, data); - - ggml_format_name(result, "%s (view)", a->name); - - ggml_set_op_params(result, &offset, sizeof(offset)); - - return result; -} - -struct ggml_tensor * ggml_view_1d( - struct ggml_context * ctx, - struct ggml_tensor * a, - int64_t ne0, - size_t offset) { bool is_node = false; @@ -6349,7 +6337,10 @@ struct ggml_tensor * ggml_view_1d( is_node = true; } - struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 1, &ne0, offset); + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, a, offset); + ggml_format_name(result, "%s (view)", a->name); + + ggml_set_op_params(result, &offset, sizeof(offset)); result->op = GGML_OP_VIEW; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; @@ -6358,6 +6349,19 @@ struct ggml_tensor * ggml_view_1d( return result; } +// ggml_view_1d + +struct ggml_tensor * ggml_view_1d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + size_t offset) { + + struct ggml_tensor * result = ggml_view_impl(ctx, a, 1, &ne0, offset); + + return result; +} + // ggml_view_2d struct ggml_tensor * ggml_view_2d( @@ -6368,24 +6372,14 @@ struct ggml_tensor * ggml_view_2d( size_t nb1, size_t offset) { - bool is_node = false; + const int64_t ne[2] = { ne0, ne1 }; - if (a->grad) { - is_node = true; - } - - const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, 1, 1 }; - - struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 2, ne, offset); + struct ggml_tensor * result = ggml_view_impl(ctx, a, 2, ne, offset); result->nb[1] = nb1; result->nb[2] = result->nb[1]*ne1; result->nb[3] = result->nb[2]; - result->op = GGML_OP_VIEW; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - return result; } @@ -6401,24 +6395,14 @@ struct ggml_tensor * ggml_view_3d( size_t nb2, size_t offset) { - bool is_node = false; + const int64_t ne[3] = { ne0, ne1, ne2 }; - if (a->grad) { - is_node = true; - } - - const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 }; - - struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 3, ne, offset); + struct ggml_tensor * result = ggml_view_impl(ctx, a, 3, ne, offset); result->nb[1] = nb1; result->nb[2] = nb2; result->nb[3] = result->nb[2]*ne2; - result->op = GGML_OP_VIEW; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - return result; } @@ -6436,24 +6420,14 @@ struct ggml_tensor * ggml_view_4d( size_t nb3, size_t offset) { - bool is_node = false; + const int64_t ne[4] = { ne0, ne1, ne2, ne3 }; - if (a->grad) { - is_node = true; - } - - const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, ne3 }; - - struct ggml_tensor * result = ggml_view_tensor_offset(ctx, a, 4, ne, offset); + struct ggml_tensor * result = ggml_view_impl(ctx, a, 4, ne, offset); result->nb[1] = nb1; result->nb[2] = nb2; result->nb[3] = nb3; - result->op = GGML_OP_VIEW; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - return result; } @@ -6640,7 +6614,7 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - int32_t params[] = { n_past, inplace ? 1 : 0 }; + int32_t params[] = { n_past }; ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_DIAG_MASK_INF; @@ -6657,7 +6631,6 @@ struct ggml_tensor * ggml_diag_mask_inf( return ggml_diag_mask_inf_impl(ctx, a, n_past, false); } - struct ggml_tensor * ggml_diag_mask_inf_inplace( struct ggml_context * ctx, struct ggml_tensor * a, @@ -6680,7 +6653,7 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - int32_t params[] = { n_past, inplace ? 1 : 0 }; + int32_t params[] = { n_past }; ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_DIAG_MASK_ZERO; @@ -11935,8 +11908,8 @@ static void ggml_compute_forward_diag_mask_f32( const int ith = params->ith; const int nth = params->nth; - const int n_past = ((int32_t *) dst->op_params)[0]; - const bool inplace = (bool)((int32_t *) dst->op_params)[1]; + const int n_past = ((int32_t *) dst->op_params)[0]; + const bool inplace = src0->data == dst->data; GGML_ASSERT(n_past >= 0); diff --git a/ggml.h b/ggml.h index 8b410cc85..c936823d6 100644 --- a/ggml.h +++ b/ggml.h @@ -479,6 +479,9 @@ extern "C" { int64_t perf_cycles; int64_t perf_time_us; + struct ggml_tensor * view_src; + size_t view_offs; + void * data; char name[GGML_MAX_NAME]; @@ -661,7 +664,7 @@ extern "C" { GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value); GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src); - GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src); + GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src); GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name); From 849408957c687cde4ab32c147107f643fc55130b Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Wed, 30 Aug 2023 02:20:26 -0400 Subject: [PATCH 096/147] tests : add a C compliance test (#2848) * tests : add a C compliance test * make : build C compliance test by default * make : fix clean and make sure C test fails on clang * make : move -Werror=implicit-int to CFLAGS --- CMakeLists.txt | 1 + Makefile | 9 ++++++--- tests/CMakeLists.txt | 5 +++++ tests/test-c.c | 3 +++ 4 files changed, 15 insertions(+), 3 deletions(-) create mode 100644 tests/test-c.c diff --git a/CMakeLists.txt b/CMakeLists.txt index 1eae2d670..d6c1b3b33 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -402,6 +402,7 @@ if (LLAMA_ALL_WARNINGS) -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes + -Werror=implicit-int ) set(cxx_flags -Wall diff --git a/Makefile b/Makefile index 02ba3e36d..44e68b7fc 100644 --- a/Makefile +++ b/Makefile @@ -1,5 +1,5 @@ # Define the default target now so that it is always the first target -BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam_search +BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam_search tests/test-c.o # Binaries only useful for tests TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1 @@ -64,7 +64,7 @@ endif # warnings CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \ - -Wmissing-prototypes + -Wmissing-prototypes -Werror=implicit-int CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar # OS specific @@ -369,7 +369,7 @@ libllama.so: llama.o ggml.o $(OBJS) $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) clean: - rm -vf *.o *.so *.dll benchmark-matmult build-info.h $(BUILD_TARGETS) $(TEST_TARGETS) + rm -vf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h $(BUILD_TARGETS) $(TEST_TARGETS) # # Examples @@ -489,3 +489,6 @@ tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml tests/test-tokenizer-1: tests/test-tokenizer-1.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + +tests/test-c.o: tests/test-c.c llama.h + $(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@ diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index ca1f39d31..483210d7b 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -37,3 +37,8 @@ llama_build_and_test_executable(test-grammar-parser.cpp) llama_build_and_test_executable(test-llama-grammar.cpp) llama_build_and_test_executable(test-grad0.cpp) # SLOW # llama_build_and_test_executable(test-opt.cpp) # SLOW + +# dummy executable - not installed +get_filename_component(TEST_TARGET test-c.c NAME_WE) +add_executable(${TEST_TARGET} test-c.c) +target_link_libraries(${TEST_TARGET} PRIVATE llama) diff --git a/tests/test-c.c b/tests/test-c.c new file mode 100644 index 000000000..a05071080 --- /dev/null +++ b/tests/test-c.c @@ -0,0 +1,3 @@ +#include "llama.h" + +int main(void) {} From 8341a25957b319a03d4a811176cd5ad7f2b0fbd4 Mon Sep 17 00:00:00 2001 From: staviq Date: Wed, 30 Aug 2023 08:29:32 +0200 Subject: [PATCH 097/147] main : log file (#2748) * initial, base LOG macro * add *.log to .gitignore * added basic log file handler * reverted log auto endline to better mimic printf * remove atomics and add dynamic log target * log_enable/disable, LOG_TEE, basic usage doc * update .gitignore * mv include to common, params, help msg * log tostring helpers, token vectors pretty prints * main: replaced fprintf/LOG_TEE, some trace logging * LOG_DISABLE_LOGS compile flag, wrapped f in macros * fix LOG_TEELN and configchecker * stub LOG_DUMP_CMDLINE for WIN32 for now * fix msvc * cleanup main.cpp:273 * fix stray whitespace after master sync * log : fix compile warnings - do not use C++20 stuff - use PRIu64 to print uint64_t - avoid string copies by using const ref - fix ", ##__VA_ARGS__" warnings - compare strings with == and != * log : do not append to existing log + disable file line func by default * log : try to fix Windows build * main : wip logs * main : add trace log * review: macro f lowercase, str append to sstream * review: simplify ifs and str comparisons * fix MSVC, formatting, FMT/VAL placeholders * review: if/else cleanup * review: if/else cleanup (2) * replace _ prefix with _impl suffix --------- Co-authored-by: Georgi Gerganov --- .gitignore | 1 + Makefile | 7 +- common/common.cpp | 22 ++ common/common.h | 3 + common/log.h | 643 +++++++++++++++++++++++++++++++++++++++++ examples/chat.sh | 2 +- examples/main/main.cpp | 288 +++++++++++------- 7 files changed, 859 insertions(+), 107 deletions(-) create mode 100644 common/log.h diff --git a/.gitignore b/.gitignore index 7a3f3fff4..54ea2b522 100644 --- a/.gitignore +++ b/.gitignore @@ -5,6 +5,7 @@ *.bin *.exe *.dll +*.log .DS_Store .build/ .cache/ diff --git a/Makefile b/Makefile index 44e68b7fc..c8b8a92d7 100644 --- a/Makefile +++ b/Makefile @@ -326,6 +326,11 @@ k_quants.o: k_quants.c k_quants.h $(CC) $(CFLAGS) -c $< -o $@ endif # LLAMA_NO_K_QUANTS +ifdef LLAMA_DISABLE_LOGS + CFLAGS += -DLOG_DISABLE_LOGS + CXXFLAGS += -DLOG_DISABLE_LOGS +endif # LLAMA_DISABLE_LOGS + # # Print build information # @@ -356,7 +361,7 @@ OBJS += ggml-alloc.o llama.o: llama.cpp ggml.h ggml-alloc.h ggml-cuda.h ggml-metal.h llama.h $(CXX) $(CXXFLAGS) -c $< -o $@ -common.o: common/common.cpp common/common.h build-info.h +common.o: common/common.cpp common/common.h build-info.h common/log.h $(CXX) $(CXXFLAGS) -c $< -o $@ console.o: common/console.cpp common/console.h diff --git a/common/common.cpp b/common/common.cpp index 90fe2e84e..ed09fc27d 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -480,6 +480,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { } } else if (arg == "-h" || arg == "--help") { gpt_print_usage(argc, argv, default_params); +#ifndef LOG_DISABLE_LOGS + log_print_usage(); +#endif // LOG_DISABLE_LOGS exit(0); } else if (arg == "--random-prompt") { params.random_prompt = true; @@ -519,6 +522,25 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { std::istreambuf_iterator(), std::back_inserter(params.grammar) ); +#ifndef LOG_DISABLE_LOGS + // Parse args for logging parameters + } else if ( log_param_single_parse( argv[i] ) ) { + // Do nothing, log_param_single_parse automatically does it's thing + // and returns if a match was found and parsed. + } else if ( log_param_pair_parse( /*check_but_dont_parse*/ true, argv[i] ) ) { + // We have a matching known parameter requiring an argument, + // now we need to check if there is anything after this argv + // and flag invalid_param or parse it. + if (++i >= argc) { + invalid_param = true; + break; + } + if( !log_param_pair_parse( /*check_but_dont_parse*/ false, argv[i-1], argv[i]) ) { + invalid_param = true; + break; + } + // End of Parse args for logging parameters +#endif // LOG_DISABLE_LOGS } else { fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); gpt_print_usage(argc, argv, default_params); diff --git a/common/common.h b/common/common.h index c15373144..5a379688e 100644 --- a/common/common.h +++ b/common/common.h @@ -4,6 +4,9 @@ #include "llama.h" +#define LOG_NO_FILE_LINE_FUNCTION +#include "log.h" + #include #include #include diff --git a/common/log.h b/common/log.h new file mode 100644 index 000000000..c1364187d --- /dev/null +++ b/common/log.h @@ -0,0 +1,643 @@ +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include + +// -------------------------------- +// +// Basic usage: +// +// -------- +// +// The LOG() and LOG_TEE() macros are ready to go by default +// they do not require any initialization. +// +// LOGLN() and LOG_TEELN() are variants which automatically +// include \n character at the end of the log string. +// +// LOG() behaves exactly like printf, by default writing to a logfile. +// LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ). +// +// Default logfile is named +// "llama..log" +// Default LOG_TEE() secondary output target is +// stderr +// +// Logs can be dynamically disabled or enabled using functions: +// log_disable() +// and +// log_enable() +// +// A log target can be changed with: +// log_set_target( string ) +// creating and opening, or re-opening a file by string filename +// or +// log_set_target( FILE* ) +// allowing to point at stderr, stdout, or any valid FILE* file handler. +// +// -------- +// +// End of Basic usage. +// +// -------------------------------- + +// Specifies a log target. +// default uses log_handler() with "llama.log" log file +// this can be changed, by defining LOG_TARGET +// like so: +// +// #define LOG_TARGET (a valid FILE*) +// #include "log.h" +// +// or it can be simply redirected to stdout or stderr +// like so: +// +// #define LOG_TARGET stderr +// #include "log.h" +// +// The log target can also be redirected to a diffrent function +// like so: +// +// #define LOG_TARGET log_handler_diffrent() +// #include "log.h" +// +// FILE* log_handler_diffrent() +// { +// return stderr; +// } +// +// or: +// +// #define LOG_TARGET log_handler_another_one("somelog.log") +// #include "log.h" +// +// FILE* log_handler_another_one(char*filename) +// { +// static FILE* logfile = nullptr; +// (...) +// if( !logfile ) +// { +// fopen(...) +// } +// (...) +// return logfile +// } +// +#ifndef LOG_TARGET + #define LOG_TARGET log_handler() +#endif + +#ifndef LOG_TEE_TARGET + #define LOG_TEE_TARGET stderr +#endif + +// Utility to obtain "pid" like unique process id and use it when creating log files. +inline std::string log_get_pid() +{ + static std::string pid; + if (pid.empty()) + { + // std::this_thread::get_id() is the most portable way of obtaining a "process id" + // it's not the same as "pid" but is unique enough to solve multiple instances + // trying to write to the same log. + std::stringstream ss; + ss << std::this_thread::get_id(); + pid = ss.str(); + } + + return pid; +} + +// Utility function for generating log file names with unique id based on thread id. +// invocation with log_filename_generator( "llama", "log" ) creates a string "llama..log" +// where the number is a runtime id of the current thread. + +#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension) + +// INTERNAL, DO NOT USE +inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension) +{ + std::stringstream buf; + + buf << log_file_basename; + buf << "."; + buf << log_get_pid(); + buf << "."; + buf << log_file_extension; + + return buf.str(); +} + +#ifndef LOG_DEFAULT_FILE_NAME + #define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log") +#endif + +// Utility for turning #define values into string literals +// so we can have a define for stderr and +// we can print "stderr" instead of literal stderr, etc. +#define LOG_STRINGIZE1(s) #s +#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s) + +#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET) + +// Allows disabling timestamps. +// in order to disable, define LOG_NO_TIMESTAMPS +// like so: +// +// #define LOG_NO_TIMESTAMPS +// #include "log.h" +// +#ifndef LOG_NO_TIMESTAMPS + #ifndef _WIN32 + #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] " + #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast>(std::chrono::system_clock::now().time_since_epoch())).count() + #else + #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] " + #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast>(std::chrono::system_clock::now().time_since_epoch())).count() + #endif +#else + #define LOG_TIMESTAMP_FMT "%s" + #define LOG_TIMESTAMP_VAL ,"" +#endif + +#ifdef LOG_TEE_TIMESTAMPS + #ifndef _WIN32 + #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] " + #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast>(std::chrono::system_clock::now().time_since_epoch())).count() + #else + #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] " + #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast>(std::chrono::system_clock::now().time_since_epoch())).count() + #endif +#else + #define LOG_TEE_TIMESTAMP_FMT "%s" + #define LOG_TEE_TIMESTAMP_VAL ,"" +#endif + +// Allows disabling file/line/function prefix +// in order to disable, define LOG_NO_FILE_LINE_FUNCTION +// like so: +// +// #define LOG_NO_FILE_LINE_FUNCTION +// #include "log.h" +// +#ifndef LOG_NO_FILE_LINE_FUNCTION + #ifndef _WIN32 + #define LOG_FLF_FMT "[%24s:%5d][%24s] " + #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__ + #else + #define LOG_FLF_FMT "[%24s:%5ld][%24s] " + #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__ + #endif +#else + #define LOG_FLF_FMT "%s" + #define LOG_FLF_VAL ,"" +#endif + +#ifdef LOG_TEE_FILE_LINE_FUNCTION + #ifndef _WIN32 + #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] " + #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__ + #else + #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] " + #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__ + #endif +#else + #define LOG_TEE_FLF_FMT "%s" + #define LOG_TEE_FLF_VAL ,"" +#endif + +// Utility for synchronizing log configuration state +// since std::optional was introduced only in c++17 +enum LogTriState +{ + LogTriStateSame, + LogTriStateFalse, + LogTriStateTrue +}; + +// INTERNAL, DO NOT USE +// USE LOG() INSTEAD +// +#ifndef _WIN32 + #define LOG_IMPL(str, ...) \ + { \ + if (LOG_TARGET != nullptr) \ + { \ + fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \ + fflush(LOG_TARGET); \ + } \ + } +#else + #define LOG_IMPL(str, ...) \ + { \ + if (LOG_TARGET != nullptr) \ + { \ + fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \ + fflush(LOG_TARGET); \ + } \ + } +#endif + +// INTERNAL, DO NOT USE +// USE LOG_TEE() INSTEAD +// +#ifndef _WIN32 + #define LOG_TEE_IMPL(str, ...) \ + { \ + if (LOG_TARGET != nullptr) \ + { \ + fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \ + fflush(LOG_TARGET); \ + } \ + if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr) \ + { \ + fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \ + fflush(LOG_TEE_TARGET); \ + } \ + } +#else + #define LOG_TEE_IMPL(str, ...) \ + { \ + if (LOG_TARGET != nullptr) \ + { \ + fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \ + fflush(LOG_TARGET); \ + } \ + if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr) \ + { \ + fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \ + fflush(LOG_TEE_TARGET); \ + } \ + } +#endif + +// The '\0' as a last argument, is a trick to bypass the silly +// "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro" +// so we can have a single macro which can be called just like printf. + +// Main LOG macro. +// behaves like printf, and supports arguments the exact same way. +// +#ifndef _WIN32 + #define LOG(...) LOG_IMPL(__VA_ARGS__, "") +#else + #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "") +#endif + +// Main TEE macro. +// does the same as LOG +// and +// simultaneously writes stderr. +// +// Secondary target can be changed just like LOG_TARGET +// by defining LOG_TEE_TARGET +// +#ifndef _WIN32 + #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "") +#else + #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "") +#endif + +// LOG macro variants with auto endline. +#ifndef _WIN32 + #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n") + #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n") +#else + #define LOGLN(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "\n") + #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "\n") +#endif + +// INTERNAL, DO NOT USE +inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr) +{ + static bool _initialized{false}; + static bool _disabled{(filename.empty() && target == nullptr)}; + static std::string log_current_filename{filename}; + static FILE *log_current_target{target}; + static FILE *logfile = nullptr; + + if (change) + { + if (disable == LogTriStateTrue) + { + // Disable primary target + _disabled = true; + } + // If previously disabled, only enable, and keep previous target + else if (disable == LogTriStateFalse) + { + _disabled = false; + } + // Otherwise, process the arguments + else if (log_current_filename != filename || log_current_target != target) + { + _initialized = false; + } + } + + if (_initialized) + { + if (_disabled) + { + // Log is disabled + return nullptr; + } + + // with fallback in case something went wrong + return logfile ? logfile : stderr; + } + + // do the (re)initialization + if (target != nullptr) + { + if (logfile != nullptr && logfile != stdout && logfile != stderr) + { + fclose(logfile); + } + + log_current_filename = LOG_DEFAULT_FILE_NAME; + log_current_target = target; + + logfile = target; + } + else + { + if (log_current_filename != filename) + { + if (logfile != nullptr && logfile != stdout && logfile != stderr) + { + fclose(logfile); + } + } + + logfile = fopen(filename.c_str(), "w"); + } + + if (!logfile) + { + // Verify whether the file was opened, otherwise fallback to stderr + logfile = stderr; + + fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno)); + fflush(stderr); + + // At this point we let the init flag be to true below, and let the target fallback to stderr + // otherwise we would repeatedly fopen() which was already unsuccessful + } + + _initialized = true; + + return logfile ? logfile : stderr; +} + +// INTERNAL, DO NOT USE +inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME) +{ + return log_handler1_impl(change, disable, filename, target); +} + +// Disables logs entirely at runtime. +// Makes LOG() and LOG_TEE() produce no output, +// untill enabled back. +#define log_disable() log_disable_impl() + +// INTERNAL, DO NOT USE +inline FILE *log_disable_impl() +{ + return log_handler1_impl(true, LogTriStateTrue); +} + +// Enables logs at runtime. +#define log_enable() log_enable_impl() + +// INTERNAL, DO NOT USE +inline FILE *log_enable_impl() +{ + return log_handler1_impl(true, LogTriStateFalse); +} + +// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*) +#define log_set_target(target) log_set_target_impl(target) + +// INTERNAL, DO NOT USE +inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); } +inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); } + +// INTERNAL, DO NOT USE +inline FILE *log_handler() { return log_handler1_impl(); } + +inline void log_test() +{ + log_disable(); + LOG("01 Hello World to nobody, because logs are disabled!\n") + log_enable(); + LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET)) + LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n") + log_set_target(stderr); + LOG("04 Hello World to stderr!\n") + LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n") + log_set_target(LOG_DEFAULT_FILE_NAME); + LOG("06 Hello World to default log file!\n") + log_set_target(stdout); + LOG("07 Hello World to stdout!\n") + log_set_target(LOG_DEFAULT_FILE_NAME); + LOG("08 Hello World to default log file again!\n") + log_disable(); + LOG("09 Hello World _1_ into the void!\n") + log_enable(); + LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n") + log_disable(); + log_set_target("llama.anotherlog.log"); + LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n") + log_enable(); + LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n") + log_set_target("llama.yetanotherlog.log"); + LOG("13 Hello World this time in yet new file?\n") + log_set_target(log_filename_generator("llama_autonamed", "log")); + LOG("14 Hello World in log with generated filename!\n") +#ifdef _WIN32 + LOG_TEE("15 Hello msvc TEE without arguments\n") + LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test") + LOG_TEELN("17 Hello msvc TEELN without arguments\n") + LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test") + LOG("19 Hello msvc LOG without arguments\n") + LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test") + LOGLN("21 Hello msvc LOGLN without arguments\n") + LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test") +#endif +} + +inline bool log_param_single_parse(const std::string & param) +{ + if ( param == "--log-test") + { + log_test(); + return true; + } + + if ( param == "--log-disable") + { + log_disable(); + return true; + } + + if ( param == "--log-enable") + { + log_enable(); + return true; + } + + return false; +} + +inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string()) +{ + if ( param == "--log-file") + { + if (!check_but_dont_parse) + { + log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log")); + } + + return true; + } + + return false; +} + +inline void log_print_usage() +{ + fprintf(stdout, "log options:\n"); + /* format + fprintf(stdout, " -h, --help show this help message and exit\n");*/ + /* spacing + fprintf(stdout, "__-param----------------Description\n");*/ + fprintf(stdout, " --log-test Run simple logging test\n"); + fprintf(stdout, " --log-disable Disable trace logs\n"); + fprintf(stdout, " --log-enable Enable trace logs\n"); + fprintf(stdout, " --log-file Specify a log filename (without extension)\n"); + fprintf(stdout, " Log file will be tagged with unique ID and written as \"..log\"\n"); /* */ +} + +#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv) + +// INTERNAL, DO NOT USE +inline void log_dump_cmdline_impl(int argc, char **argv) +{ + std::stringstream buf; + for (int i = 0; i < argc; ++i) + { + if (std::string(argv[i]).find(' ') != std::string::npos) + { + buf << " \"" << argv[i] <<"\""; + } + else + { + buf << " " << argv[i]; + } + } + LOGLN("Cmd:%s", buf.str().c_str()) +} + +#define log_tostr(var) log_var_to_string_impl(var).c_str() + +inline std::string log_var_to_string_impl(bool var) +{ + return var ? "true" : "false"; +} + +inline std::string log_var_to_string_impl(std::string var) +{ + return var; +} + +inline std::string log_var_to_string_impl(const std::vector & var) +{ + std::stringstream buf; + buf << "[ "; + bool first = true; + for (auto e : var) + { + if (first) + { + first = false; + } + else + { + buf << ", "; + } + buf << std::to_string(e); + } + buf << " ]"; + + return buf.str(); +} + +#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens) \ + [&tokens, &ctx]() \ + { \ + std::stringstream buf; \ + buf << "[ "; \ + \ + bool first = true; \ + for (const auto &token : tokens) \ + { \ + if (!first) \ + buf << ", "; \ + else \ + first = false; \ + \ + auto detokenized = llama_token_to_piece(ctx, token); \ + \ + detokenized.erase( \ + std::remove_if( \ + detokenized.begin(), \ + detokenized.end(), \ + [](const unsigned char c) { return !std::isprint(c); }), \ + detokenized.end()); \ + \ + buf \ + << "'" << detokenized << "'" \ + << ":" << std::to_string(token); \ + } \ + buf << " ]"; \ + \ + return buf.str(); \ + }() \ + .c_str() + +#ifdef LOG_DISABLE_LOGS + +#undef LOG +#define LOG(...) // dummy stub +#undef LOGLN +#define LOGLN(...) // dummy stub + +#undef LOG_TEE +#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__); // convert to normal fprintf + +#undef LOG_TEELN +#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__); // convert to normal fprintf + +#undef LOG_DISABLE +#define LOG_DISABLE() // dummy stub + +#undef LOG_ENABLE +#define LOG_ENABLE() // dummy stub + +#undef LOG_ENABLE +#define LOG_ENABLE() // dummy stub + +#undef LOG_SET_TARGET +#define LOG_SET_TARGET(...) // dummy stub + +#undef LOG_DUMP_CMDLINE +#define LOG_DUMP_CMDLINE(...) // dummy stub + +#endif // LOG_DISABLE_LOGS diff --git a/examples/chat.sh b/examples/chat.sh index 9a928ef05..d567acecd 100755 --- a/examples/chat.sh +++ b/examples/chat.sh @@ -11,6 +11,6 @@ cd .. # # "--keep 48" is based on the contents of prompts/chat-with-bob.txt # -./main -m ./models/7B/ggml-model-q4_0.bin -c 512 -b 1024 -n 256 --keep 48 \ +./main -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \ --repeat_penalty 1.0 --color -i \ -r "User:" -f prompts/chat-with-bob.txt diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 89cc4f602..7117db4b0 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -4,6 +4,7 @@ #endif #include "common.h" + #include "console.h" #include "llama.h" #include "build-info.h" @@ -112,6 +113,15 @@ int main(int argc, char ** argv) { return 1; } +#ifndef LOG_DISABLE_LOGS + log_set_target(log_filename_generator("main", "log")); + LOG_TEE("Log start\n"); + log_dump_cmdline(argc,argv); +#endif // LOG_DISABLE_LOGS + + // TODO: Dump params ? + //LOG("Params perplexity: %s\n", LOG_TOSTR(params.perplexity)); + // save choice to use color for later // (note for later: this is a slightly awkward choice) console::init(params.simple_io, params.use_color); @@ -134,34 +144,35 @@ int main(int argc, char ** argv) { } if (params.rope_freq_base != 10000.0) { - fprintf(stderr, "%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base); + LOG_TEE("%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base); } if (params.rope_freq_scale != 1.0) { - fprintf(stderr, "%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale); + LOG_TEE("%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale); } if (params.n_ctx > 2048) { // TODO: determine the actual max context of the model (e.g. 4096 for LLaMA v2) and use that instead of 2048 - fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx); + LOG_TEE("%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx); } else if (params.n_ctx < 8) { - fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__); + LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__); params.n_ctx = 8; } - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); + LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } - fprintf(stderr, "%s: seed = %u\n", __func__, params.seed); + LOG_TEE("%s: seed = %u\n", __func__, params.seed); std::mt19937 rng(params.seed); if (params.random_prompt) { params.prompt = gpt_random_prompt(rng); } + LOG("%s: llama backend init\n", __func__); llama_backend_init(params.numa); llama_model * model; @@ -171,6 +182,7 @@ int main(int argc, char ** argv) { g_ctx = &ctx; // load the model and apply lora adapter, if any + LOG("%s: load the model and apply lora adapter, if any\n", __func__); std::tie(model, ctx) = llama_init_from_gpt_params(params); if (params.cfg_scale > 1.f) { struct llama_context_params lparams = llama_context_params_from_gpt_params(params); @@ -178,14 +190,14 @@ int main(int argc, char ** argv) { } if (model == NULL) { - fprintf(stderr, "%s: error: unable to load model\n", __func__); + LOG_TEE("%s: error: unable to load model\n", __func__); return 1; } // print system information { - fprintf(stderr, "\n"); - fprintf(stderr, "system_info: n_threads = %d / %d | %s\n", + LOG_TEE("\n"); + LOG_TEE("system_info: n_threads = %d / %d | %s\n", params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info()); } @@ -193,7 +205,7 @@ int main(int argc, char ** argv) { // uncomment the "used_mem" line in llama.cpp to see the results if (params.mem_test) { { - fprintf(stderr, "%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx); + LOG_TEE("%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx); const std::vector tmp(params.n_batch, llama_token_bos(ctx)); llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads); @@ -219,7 +231,7 @@ int main(int argc, char ** argv) { std::vector session_tokens; if (!path_session.empty()) { - fprintf(stderr, "%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str()); + LOG_TEE("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str()); // fopen to check for existing session FILE * fp = std::fopen(path_session.c_str(), "rb"); @@ -229,33 +241,38 @@ int main(int argc, char ** argv) { session_tokens.resize(params.n_ctx); size_t n_token_count_out = 0; if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) { - fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str()); + LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str()); return 1; } session_tokens.resize(n_token_count_out); llama_set_rng_seed(ctx, params.seed); - fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size()); + LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size()); } else { - fprintf(stderr, "%s: session file does not exist, will create\n", __func__); + LOG_TEE("%s: session file does not exist, will create\n", __func__); } } - // Add BOS if SPM tokenizer const bool add_bos = llama_vocab_type(ctx) == LLAMA_VOCAB_TYPE_SPM; + LOG("add_bos: %d\n", add_bos); - // tokenize the prompt std::vector embd_inp; if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) { + LOG("tokenize the prompt\n"); embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos); } else { + LOG("use session tokens\n"); embd_inp = session_tokens; } + LOG("prompt: \"%s\"\n", log_tostr(params.prompt)); + LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp)); + // Should not run without any tokens if (embd_inp.empty()) { embd_inp.push_back(llama_token_bos(ctx)); + LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp)); } // Tokenize negative prompt @@ -263,23 +280,31 @@ int main(int argc, char ** argv) { int guidance_offset = 0; int original_prompt_len = 0; if (ctx_guidance) { + LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(params.cfg_negative_prompt)); + guidance_inp = ::llama_tokenize(ctx_guidance, params.cfg_negative_prompt, add_bos); + LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp)); std::vector original_inp = ::llama_tokenize(ctx, params.prompt, add_bos); + LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp)); + original_prompt_len = original_inp.size(); guidance_offset = (int)guidance_inp.size() - original_prompt_len; + LOG("original_prompt_len: %s", log_tostr(original_prompt_len)); + LOG("guidance_offset: %s", log_tostr(guidance_offset)); } const int n_ctx = llama_n_ctx(ctx); + LOG("n_ctx: %d\n", n_ctx); if ((int) embd_inp.size() > n_ctx - 4) { - fprintf(stderr, "%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4); + LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4); return 1; } // debug message about similarity of saved session, if applicable size_t n_matching_session_tokens = 0; - if (session_tokens.size()) { + if (session_tokens.size() > 0) { for (llama_token id : session_tokens) { if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) { break; @@ -287,22 +312,27 @@ int main(int argc, char ** argv) { n_matching_session_tokens++; } if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) { - fprintf(stderr, "%s: using full prompt from session file\n", __func__); + LOG_TEE("%s: using full prompt from session file\n", __func__); } else if (n_matching_session_tokens >= embd_inp.size()) { - fprintf(stderr, "%s: session file has exact match for prompt!\n", __func__); + LOG_TEE("%s: session file has exact match for prompt!\n", __func__); } else if (n_matching_session_tokens < (embd_inp.size() / 2)) { - fprintf(stderr, "%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n", + LOG_TEE("%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n", __func__, n_matching_session_tokens, embd_inp.size()); } else { - fprintf(stderr, "%s: session file matches %zu / %zu tokens of prompt\n", + LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n", __func__, n_matching_session_tokens, embd_inp.size()); } } + LOGLN( + "recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu, embd_inp.size() %zu", + log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size(), embd_inp.size()); + // if we will use the cache for the full prompt without reaching the end of the cache, force // reevaluation of the last token token to recalculate the cached logits - if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && - session_tokens.size() > embd_inp.size()) { + if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) { + LOGLN("recalculate the cached logits (do): session_tokens.resize( %zu )", embd_inp.size() - 1); + session_tokens.resize(embd_inp.size() - 1); } @@ -315,6 +345,9 @@ int main(int argc, char ** argv) { const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos); const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false); + LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx)); + LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx)); + // in instruct mode, we inject a prefix and a suffix to each input by the user if (params.instruct) { params.interactive_first = true; @@ -327,30 +360,30 @@ int main(int argc, char ** argv) { } if (params.verbose_prompt) { - fprintf(stderr, "\n"); - fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str()); - fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); + LOG_TEE("\n"); + LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str()); + LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); for (int i = 0; i < (int) embd_inp.size(); i++) { - fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str()); + LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str()); } if (ctx_guidance) { - fprintf(stderr, "\n"); - fprintf(stderr, "%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str()); - fprintf(stderr, "%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size()); + LOG_TEE("\n"); + LOG_TEE("%s: negative prompt: '%s'\n", __func__, params.cfg_negative_prompt.c_str()); + LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size()); for (int i = 0; i < (int) guidance_inp.size(); i++) { - fprintf(stderr, "%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str()); + LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str()); } } if (params.n_keep > 0) { - fprintf(stderr, "%s: static prompt based on n_keep: '", __func__); + LOG_TEE("%s: static prompt based on n_keep: '", __func__); for (int i = 0; i < params.n_keep; i++) { - fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_inp[i]).c_str()); + LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str()); } - fprintf(stderr, "'\n"); + LOG_TEE("'\n"); } - fprintf(stderr, "\n"); + LOG_TEE("\n"); } if (params.interactive) { @@ -367,30 +400,30 @@ int main(int argc, char ** argv) { SetConsoleCtrlHandler(reinterpret_cast(console_ctrl_handler), true); #endif - fprintf(stderr, "%s: interactive mode on.\n", __func__); + LOG_TEE("%s: interactive mode on.\n", __func__); if (params.antiprompt.size()) { - for (auto antiprompt : params.antiprompt) { - fprintf(stderr, "Reverse prompt: '%s'\n", antiprompt.c_str()); + for (const auto & antiprompt : params.antiprompt) { + LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str()); } } if (params.input_prefix_bos) { - fprintf(stderr, "Input prefix with BOS\n"); + LOG_TEE("Input prefix with BOS\n"); } if (!params.input_prefix.empty()) { - fprintf(stderr, "Input prefix: '%s'\n", params.input_prefix.c_str()); + LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str()); } if (!params.input_suffix.empty()) { - fprintf(stderr, "Input suffix: '%s'\n", params.input_suffix.c_str()); + LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str()); } } - fprintf(stderr, "sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n", + LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n", params.repeat_last_n, params.repeat_penalty, params.presence_penalty, params.frequency_penalty, params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau); - fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep); - fprintf(stderr, "\n\n"); + LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep); + LOG_TEE("\n\n"); grammar_parser::parse_state parsed_grammar; llama_grammar * grammar = NULL; @@ -400,14 +433,14 @@ int main(int argc, char ** argv) { if (parsed_grammar.rules.empty()) { return 1; } - fprintf(stderr, "%s: grammar:\n", __func__); + LOG_TEE("%s: grammar:\n", __func__); grammar_parser::print_grammar(stderr, parsed_grammar); - fprintf(stderr, "\n"); + LOG_TEE("\n"); { auto it = params.logit_bias.find(llama_token_eos(ctx)); if (it != params.logit_bias.end() && it->second == -INFINITY) { - fprintf(stderr, "%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__); + LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__); } } @@ -430,11 +463,11 @@ int main(int argc, char ** argv) { " - To return control without starting a new line, end your input with '/'.\n" " - If you want to submit another line, end your input with '\\'.\n"; } - fprintf(stderr, "== Running in interactive mode. ==\n" + LOG_TEE("== Running in interactive mode. ==\n"); #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) - " - Press Ctrl+C to interject at any time.\n" + LOG_TEE( " - Press Ctrl+C to interject at any time.\n"); #endif - "%s\n", control_message); + LOG_TEE( "%s\n", control_message); is_interacting = params.interactive_first; } @@ -459,8 +492,9 @@ int main(int argc, char ** argv) { std::vector embd; std::vector embd_guidance; - // do one empty run to warm up the model { + LOG("warming up the model with an empty run\n"); + const std::vector tmp = { llama_token_bos(ctx), }; llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads); llama_reset_timings(ctx); @@ -471,15 +505,17 @@ int main(int argc, char ** argv) { if (embd.size() > 0) { // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via // --prompt or --file which uses the same value. - auto max_embd_size = n_ctx - 4; + int max_embd_size = n_ctx - 4; + // Ensure the input doesn't exceed the context size by truncating embd if necessary. - if ((int)embd.size() > max_embd_size) { - auto skipped_tokens = embd.size() - max_embd_size; + if ((int) embd.size() > max_embd_size) { + const int skipped_tokens = (int) embd.size() - max_embd_size; + embd.resize(max_embd_size); + console::set_display(console::error); - printf("<>", skipped_tokens, skipped_tokens != 1 ? "s" : ""); + printf("<>", skipped_tokens, skipped_tokens != 1 ? "s" : ""); console::set_display(console::reset); fflush(stdout); - embd.resize(max_embd_size); } // infinite text generation via context swapping @@ -488,28 +524,26 @@ int main(int argc, char ** argv) { // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches if (n_past + (int) embd.size() + std::max(0, guidance_offset) > n_ctx) { if (params.n_predict == -2) { - fprintf(stderr, "\n\n%s: context full, stopping generation\n", __func__); + LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict); break; } const int n_left = n_past - params.n_keep; + LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d\n", n_past, n_left, n_ctx, params.n_keep); + // always keep the first token - BOS - n_past = std::max(1, params.n_keep); + n_past = std::max(1, params.n_keep); n_past_guidance = std::max(1, params.n_keep + guidance_offset); + LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance); + // insert n_left/2 tokens at the start of embd from last_n_tokens embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size()); - // stop saving session if we run out of context - path_session.clear(); + LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd)); - //printf("\n---\n"); - //printf("resetting: '"); - //for (int i = 0; i < (int) embd.size(); i++) { - // printf("%s", llama_token_to_piece(ctx, embd[i])); - //} - //printf("'\n"); - //printf("\n---\n"); + LOG("clear session path\n"); + path_session.clear(); } // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past) @@ -539,7 +573,7 @@ int main(int argc, char ** argv) { if (ctx_guidance) { int input_size = 0; - llama_token* input_buf = NULL; + llama_token * input_buf = NULL; if (n_past_guidance < (int) guidance_inp.size()) { // Guidance context should have the same data with these modifications: @@ -555,22 +589,19 @@ int main(int argc, char ** argv) { ); } - input_buf = embd_guidance.data(); + input_buf = embd_guidance.data(); input_size = embd_guidance.size(); - //fprintf(stderr, "\n---------------------\n"); - //for (int i = 0; i < (int) embd_guidance.size(); i++) { - //fprintf(stderr, "%s", llama_token_to_piece(ctx, embd_guidance[i])); - //} - //fprintf(stderr, "\n---------------------\n"); + + LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance)); } else { - input_buf = embd.data(); + input_buf = embd.data(); input_size = embd.size(); } for (int i = 0; i < input_size; i += params.n_batch) { int n_eval = std::min(input_size - i, params.n_batch); if (llama_eval(ctx_guidance, input_buf + i, n_eval, n_past_guidance, params.n_threads)) { - fprintf(stderr, "%s : failed to eval\n", __func__); + LOG_TEE("%s : failed to eval\n", __func__); return 1; } @@ -583,11 +614,17 @@ int main(int argc, char ** argv) { if (n_eval > params.n_batch) { n_eval = params.n_batch; } + + LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd)); + if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) { - fprintf(stderr, "%s : failed to eval\n", __func__); + LOG_TEE("%s : failed to eval\n", __func__); return 1; } + n_past += n_eval; + + LOG("n_past = %d\n", n_past); } if (embd.size() > 0 && !path_session.empty()) { @@ -600,7 +637,6 @@ int main(int argc, char ** argv) { embd_guidance.clear(); if ((int) embd_inp.size() <= n_consumed && !is_interacting) { - // out of user input, sample next token const float temp = params.temp; const int32_t top_k = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k; const float top_p = params.top_p; @@ -619,6 +655,8 @@ int main(int argc, char ** argv) { if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) { need_to_save_session = false; llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size()); + + LOG("saved session to %s\n", path_session.c_str()); } llama_token id = 0; @@ -638,55 +676,68 @@ int main(int argc, char ** argv) { candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); } - llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + llama_token_data_array cur_p = { candidates.data(), candidates.size(), false }; if (ctx_guidance) { - llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale); + llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale); } // Apply penalties float nl_logit = logits[llama_token_nl(ctx)]; auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx); - llama_sample_repetition_penalty(ctx, &candidates_p, + llama_sample_repetition_penalty(ctx, &cur_p, last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, last_n_repeat, repeat_penalty); - llama_sample_frequency_and_presence_penalties(ctx, &candidates_p, + llama_sample_frequency_and_presence_penalties(ctx, &cur_p, last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, last_n_repeat, alpha_frequency, alpha_presence); if (!penalize_nl) { - for (size_t idx = 0; idx < candidates_p.size; idx++) { - if (candidates_p.data[idx].id == llama_token_nl(ctx)) { - candidates_p.data[idx].logit = nl_logit; + for (size_t idx = 0; idx < cur_p.size; idx++) { + if (cur_p.data[idx].id == llama_token_nl(ctx)) { + cur_p.data[idx].logit = nl_logit; break; } } } if (grammar != NULL) { - llama_sample_grammar(ctx, &candidates_p, grammar); + llama_sample_grammar(ctx, &cur_p, grammar); } if (temp <= 0) { // Greedy sampling - id = llama_sample_token_greedy(ctx, &candidates_p); + id = llama_sample_token_greedy(ctx, &cur_p); } else { if (mirostat == 1) { static float mirostat_mu = 2.0f * mirostat_tau; const int mirostat_m = 100; - llama_sample_temperature(ctx, &candidates_p, temp); - id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu); + llama_sample_temperature(ctx, &cur_p, temp); + id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu); } else if (mirostat == 2) { static float mirostat_mu = 2.0f * mirostat_tau; - llama_sample_temperature(ctx, &candidates_p, temp); - id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu); + llama_sample_temperature(ctx, &cur_p, temp); + id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu); } else { // Temperature sampling - llama_sample_top_k(ctx, &candidates_p, top_k, 1); - llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1); - llama_sample_typical(ctx, &candidates_p, typical_p, 1); - llama_sample_top_p(ctx, &candidates_p, top_p, 1); - llama_sample_temperature(ctx, &candidates_p, temp); - id = llama_sample_token(ctx, &candidates_p); + llama_sample_top_k (ctx, &cur_p, top_k, 1); + llama_sample_tail_free (ctx, &cur_p, tfs_z, 1); + llama_sample_typical (ctx, &cur_p, typical_p, 1); + llama_sample_top_p (ctx, &cur_p, top_p, 1); + llama_sample_temperature(ctx, &cur_p, temp); + + { + const int n_top = 10; + LOG("top %d candidates:\n", n_top); + + for (int i = 0; i < n_top; i++) { + const llama_token id = cur_p.data[i].id; + LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p); + } + } + + id = llama_sample_token(ctx, &cur_p); + + LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str()); } } // printf("`%d`", candidates_p.size); @@ -697,9 +748,10 @@ int main(int argc, char ** argv) { last_n_tokens.erase(last_n_tokens.begin()); last_n_tokens.push_back(id); + + LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_n_tokens)); } - // add it to the context embd.push_back(id); // echo this to console @@ -707,8 +759,11 @@ int main(int argc, char ** argv) { // decrement remaining sampling budget --n_remain; + + LOG("n_remain: %d\n", n_remain); } else { // some user input remains from prompt or interaction, forward it to processing + LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed); while ((int) embd_inp.size() > n_consumed) { embd.push_back(embd_inp[n_consumed]); last_n_tokens.erase(last_n_tokens.begin()); @@ -736,13 +791,12 @@ int main(int argc, char ** argv) { fflush(stdout); } // reset color to default if we there is no pending user input - if (input_echo && (int)embd_inp.size() == n_consumed) { + if (input_echo && (int) embd_inp.size() == n_consumed) { console::set_display(console::reset); } // if not currently processing queued inputs; if ((int) embd_inp.size() <= n_consumed) { - // check for reverse prompt if (params.antiprompt.size()) { std::string last_output; @@ -760,7 +814,7 @@ int main(int argc, char ** argv) { ? last_output.length() - static_cast(antiprompt.length() + extra_padding) : 0; - if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) { + if (last_output.find(antiprompt, search_start_pos) != std::string::npos) { if (params.interactive) { is_interacting = true; console::set_display(console::user_input); @@ -770,10 +824,16 @@ int main(int argc, char ** argv) { break; } } + + if (is_antiprompt) { + LOG("found antiprompt: %s\n", last_output.c_str()); + } } // deal with end of text token in interactive mode if (last_n_tokens.back() == llama_token_eos(ctx)) { + LOG("found EOS token\n"); + if (params.interactive) { if (params.antiprompt.size() != 0) { // tokenize and inject first reverse prompt @@ -792,16 +852,20 @@ int main(int argc, char ** argv) { } if (n_past > 0 && is_interacting) { + LOG("waiting for user input\n"); + if (params.instruct) { printf("\n> "); } if (params.input_prefix_bos) { + LOG("adding input prefix BOS token\n"); embd_inp.push_back(llama_token_bos(ctx)); } std::string buffer; if (!params.input_prefix.empty()) { + LOG("appending input prefix: '%s'\n", params.input_prefix.c_str()); buffer += params.input_prefix; printf("%s", buffer.c_str()); } @@ -821,23 +885,30 @@ int main(int argc, char ** argv) { if (buffer.length() > 1) { // append input suffix if any if (!params.input_suffix.empty()) { + LOG("appending input suffix: '%s'\n", params.input_suffix.c_str()); buffer += params.input_suffix; printf("%s", params.input_suffix.c_str()); } + LOG("buffer: '%s'\n", buffer.c_str()); + const size_t original_size = embd_inp.size(); // instruct mode: insert instruction prefix if (params.instruct && !is_antiprompt) { + LOG("inserting instruction prefix\n"); n_consumed = embd_inp.size(); embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end()); } - auto line_inp = ::llama_tokenize(ctx, buffer, false); + const auto line_inp = ::llama_tokenize(ctx, buffer, false); + LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp)); + embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end()); // instruct mode: insert response suffix if (params.instruct) { + LOG("inserting instruction suffix\n"); embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end()); } @@ -848,6 +919,9 @@ int main(int argc, char ** argv) { } n_remain -= line_inp.size(); + LOG("n_remain: %d\n", n_remain); + } else { + LOG("empty line, passing control back\n"); } input_echo = false; // do not echo this again @@ -871,7 +945,7 @@ int main(int argc, char ** argv) { // end of text token if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !(params.instruct || params.interactive)) { - fprintf(stderr, " [end of text]\n"); + LOG_TEE(" [end of text]\n"); break; } @@ -884,7 +958,7 @@ int main(int argc, char ** argv) { } if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) { - fprintf(stderr, "\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str()); + LOG_TEE("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str()); llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size()); } @@ -900,5 +974,9 @@ int main(int argc, char ** argv) { } llama_backend_free(); +#ifndef LOG_DISABLE_LOGS + LOG_TEE("Log end\n") +#endif // LOG_DISABLE_LOGS + return 0; } From ad9ddcff6ef322db5cf13785bd7c856b610d242e Mon Sep 17 00:00:00 2001 From: chaihahaha Date: Wed, 30 Aug 2023 14:50:55 +0800 Subject: [PATCH 098/147] llm.vim : stop generation at multiple linebreaks, bind to (#2879) --- examples/llm.vim | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/examples/llm.vim b/examples/llm.vim index 594a28549..d580a3d00 100644 --- a/examples/llm.vim +++ b/examples/llm.vim @@ -8,7 +8,7 @@ function! Llm() let buffer_content = join(getline(1, '$'), "\n") " Create the JSON payload - let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":10,"stream": v:false} + let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":256,"stop": ["\n\n\n"],"stream": v:false} let json_payload.prompt = buffer_content " Define the curl command @@ -25,3 +25,4 @@ function! Llm() endfunction command! Llm call Llm() +noremap :Llm From dc07dc492ef9640bbb82904d7c7679f7bdcf6d76 Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Wed, 30 Aug 2023 02:25:50 -0600 Subject: [PATCH 099/147] convert : various script cleanups/fixes + merges and special token handling (#2842) * convert: Fix permute calls and method/func definitions * Cleanups for gguf-py * Minor types cleanups. * Initial implementation of handling merges and special tokens * convert: Handle special tokens and merges in vocab only mode convert: Vocab only mode no longer requires loading model tensors * gguf: Refactor tensor name mapping * convert: Fix type hint for special_token_types in SpecialVocab * Use common special vocab handling in various conversion scripts * First pass at implementing suggested changes * Second pass * gguf: SpecialVocab: Fix issue with special token content not in a dict gguf: SpecialVocab: Allow skipping handling of merges * convert-falcon-hf-to-gguf: Support --vocab-only option, bail out if no tokenizer.json * convert-gptneox-hf-to-gguf and convert: Only handle merges for BPE tokenizer * gguf: SpecialVocab: Actually set load_merges in object * Uniform args parsing and vocab only mode for convert examples * convert.py: Set gpt2 as tokenizer model when using BPE * Squish last type warning in gguf.py - yay! --- convert-falcon-hf-to-gguf.py | 170 +++++----- convert-gptneox-hf-to-gguf.py | 175 ++++------ convert-llama-7b-pth-to-gguf.py | 200 +++++------- convert-llama-ggmlv3-to-gguf.py | 28 +- convert-llama-hf-to-gguf.py | 203 +++++------- convert-lora-to-ggml.py | 6 +- convert.py | 142 ++++---- gguf-py/gguf/gguf.py | 551 +++++++++++++++++++------------- gguf-py/gguf/py.typed | 0 gguf-py/pyproject.toml | 1 + 10 files changed, 728 insertions(+), 748 deletions(-) create mode 100644 gguf-py/gguf/py.typed diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py index 168bcf17f..0fdea70e1 100755 --- a/convert-falcon-hf-to-gguf.py +++ b/convert-falcon-hf-to-gguf.py @@ -8,6 +8,7 @@ import struct import json import numpy as np import torch +import argparse from typing import Any, List from pathlib import Path @@ -32,11 +33,10 @@ def bytes_to_unicode(): bs.append(b) cs.append(2**8+n) n += 1 - cs = [chr(n) for n in cs] - return dict(zip(bs, cs)) + return dict(zip(bs, (chr(n) for n in cs))) -def count_model_parts(dir_model: str) -> int: +def count_model_parts(dir_model: Path) -> int: num_parts = 0 for filename in os.listdir(dir_model): if filename.startswith("pytorch_model-"): @@ -47,17 +47,22 @@ def count_model_parts(dir_model: str) -> int: return num_parts -if len(sys.argv) < 3: - print(f"Usage: python {sys.argv[0]} dir-model ftype\n") - print(" ftype == 0 -> float32") - print(" ftype == 1 -> float16") +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") + parser.add_argument("ftype", type=int, choices=[0, 1], help="output format - use 0 for float32, 1 for float16", default = 1) + return parser.parse_args() + +args = parse_args() + +dir_model = args.model +ftype = args.ftype +if not dir_model.is_dir(): + print(f'Error: {args.model} is not a directory', file = sys.stderr) sys.exit(1) - -# output in the same directory as the model -dir_model = sys.argv[1] -last_dir = os.path.basename(os.path.normpath(dir_model)) - # possible tensor data types # ftype == 0 -> float32 # ftype == 1 -> float16 @@ -65,25 +70,21 @@ last_dir = os.path.basename(os.path.normpath(dir_model)) # map from ftype to string ftype_str = ["f32", "f16"] -ftype = 1 -if len(sys.argv) > 2: - ftype = int(sys.argv[2]) - if ftype < 0 or ftype > 1: - print("Invalid ftype: " + str(ftype)) +if args.outfile is not None: + fname_out = args.outfile +else: + # output in the same directory as the model by default + fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - sys.exit(1) +print("gguf: loading model "+dir_model.name) -fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf" - -print("gguf: loading model "+last_dir) - -with open(dir_model + "/config.json", "r", encoding="utf-8") as f: +with open(dir_model / "config.json", "r", encoding="utf-8") as f: hparams = json.load(f) if hparams["architectures"][0] != "RWForCausalLM": print("Model architecture not supported: " + hparams["architectures"][0]) - sys.exit() + sys.exit(1) # get number of model parts num_parts = count_model_parts(dir_model) @@ -113,77 +114,58 @@ gguf_writer.add_file_type(ftype) print("gguf: get tokenizer metadata") -tokens: List[str] = [] +tokens: List[bytearray] = [] scores: List[float] = [] toktypes: List[int] = [] -merges: List[str] = [] +tokenizer_json_file = dir_model / 'tokenizer.json' +if not tokenizer_json_file.is_file(): + print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr) + sys.exit(1) -if Path(dir_model + "/tokenizer.json").is_file(): - # gpt2 tokenizer - gguf_writer.add_tokenizer_model("gpt2") +# gpt2 tokenizer +gguf_writer.add_tokenizer_model("gpt2") - print("gguf: get gpt2 tokenizer merges") +with open(tokenizer_json_file, "r", encoding="utf-8") as f: + tokenizer_json = json.load(f) - with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f: - tokenizer_json = json.load(f) - merges = tokenizer_json["model"]["merges"] +print("gguf: get gpt2 tokenizer vocab") - gguf_writer.add_token_merges(merges) +vocab_size = len(tokenizer_json["model"]["vocab"]) - print("gguf: get gpt2 tokenizer vocab") +# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py +tokenizer = AutoTokenizer.from_pretrained(dir_model) - vocab_size = len(tokenizer_json["model"]["vocab"]) +reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} +byte_encoder = bytes_to_unicode() +byte_decoder = {v: k for k, v in byte_encoder.items()} - # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py - tokenizer = AutoTokenizer.from_pretrained(dir_model) +for i in range(vocab_size): + if i in reverse_vocab: + try: + text = bytearray([byte_decoder[c] for c in reverse_vocab[i]]) + except KeyError: + text = bytearray() + for c in reverse_vocab[i]: + if ord(c) < 256: # single byte character + text.append(byte_decoder[ord(c)]) + else: # multibyte special token character + text.extend(c.encode('utf-8')) + else: + print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.") + pad_token = f"[PAD{i}]".encode("utf8") + text = bytearray(pad_token) - reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} - byte_encoder = bytes_to_unicode() - byte_decoder = {v: k for k, v in byte_encoder.items()} + tokens.append(text) + scores.append(0.0) # dymmy + toktypes.append(gguf.TokenType.NORMAL) # dummy - for i in range(vocab_size): - if i in reverse_vocab: - try: - text = bytearray([byte_decoder[c] for c in reverse_vocab[i]]) - except KeyError: - text = bytearray() - for c in reverse_vocab[i]: - if ord(c) < 256: # single byte character - text.append(byte_decoder[ord(c)]) - else: # multibyte special token character - text.extend(c.encode('utf-8')) - else: - print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.") - pad_token = f"[PAD{i}]".encode("utf8") - text = bytearray(pad_token) - - tokens.append(text) - scores.append(0.0) # dymmy - toktypes.append(gguf.TokenType.NORMAL) # dummy - - gguf_writer.add_token_list(tokens) - gguf_writer.add_token_scores(scores) - gguf_writer.add_token_types(toktypes) - -print("gguf: get special token ids") -# Look for special tokens in config.json - -if "bos_token_id" in hparams and hparams["bos_token_id"] != None: - gguf_writer.add_bos_token_id(hparams["bos_token_id"]) - -if "eos_token_id" in hparams and hparams["eos_token_id"] != None: - gguf_writer.add_eos_token_id(hparams["eos_token_id"]) - -if "unk_token_id" in hparams and hparams["unk_token_id"] != None: - gguf_writer.add_unk_token_id(hparams["unk_token_id"]) - -if "sep_token_id" in hparams and hparams["sep_token_id"] != None: - gguf_writer.add_sep_token_id(hparams["sep_token_id"]) - -if "pad_token_id" in hparams and hparams["pad_token_id"] != None: - gguf_writer.add_pad_token_id(hparams["pad_token_id"]) +gguf_writer.add_token_list(tokens) +gguf_writer.add_token_scores(scores) +gguf_writer.add_token_types(toktypes) +special_vocab = gguf.SpecialVocab(dir_model, load_merges = True) +special_vocab.add_to_gguf(gguf_writer) # TENSORS @@ -199,15 +181,17 @@ head_dim = hparams["hidden_size"] // n_head print("gguf: get tensor metadata") if num_parts == 0: - part_names = ("pytorch_model.bin",) + part_names = iter(("pytorch_model.bin",)) else: part_names = ( f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) ) for part_name in part_names: + if args.vocab_only: + break print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") + model_part = torch.load(dir_model / part_name, map_location="cpu") for name in model_part.keys(): data = model_part[name] @@ -238,11 +222,8 @@ for part_name in part_names: data = data.squeeze().numpy() # map tensor names - if name.endswith(".weight") and name[:-7] in tensor_map: - name = tensor_map[name[:-7]] + ".weight" - elif name.endswith(".bias") and name[:-5] in tensor_map: - name = tensor_map[name[:-5]] + ".bias" - else: + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) + if new_name is None: print("Can not map tensor '" + name + "'") sys.exit() @@ -261,19 +242,20 @@ for part_name in part_names: if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - gguf_writer.add_tensor(name, data) + gguf_writer.add_tensor(new_name, data) print("gguf: write header") gguf_writer.write_header_to_file() print("gguf: write metadata") gguf_writer.write_kv_data_to_file() -print("gguf: write tensors") -gguf_writer.write_tensors_to_file() +if not args.vocab_only: + print("gguf: write tensors") + gguf_writer.write_tensors_to_file() gguf_writer.close() -print("gguf: model successfully exported to '" + fname_out + "'") +print(f"gguf: model successfully exported to '{fname_out}'") print("") diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py index d9c42d76b..38e71e03b 100755 --- a/convert-gptneox-hf-to-gguf.py +++ b/convert-gptneox-hf-to-gguf.py @@ -8,6 +8,7 @@ import struct import json import numpy as np import torch +import argparse from typing import Any, List from pathlib import Path @@ -34,11 +35,10 @@ def bytes_to_unicode(): bs.append(b) cs.append(2**8+n) n += 1 - cs = [chr(n) for n in cs] - return dict(zip(bs, cs)) + return dict(zip(bs, (chr(n) for n in cs))) -def count_model_parts(dir_model: str) -> int: +def count_model_parts(dir_model: Path) -> int: num_parts = 0 for filename in os.listdir(dir_model): if filename.startswith("pytorch_model-"): @@ -49,17 +49,22 @@ def count_model_parts(dir_model: str) -> int: return num_parts -if len(sys.argv) < 3: - print(f"Usage: python {sys.argv[0]} dir-model ftype\n") - print(" ftype == 0 -> float32") - print(" ftype == 1 -> float16") +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") + parser.add_argument("ftype", type=int, choices=[0, 1], help="output format - use 0 for float32, 1 for float16", default = 1) + return parser.parse_args() + +args = parse_args() + +dir_model = args.model +ftype = args.ftype +if not dir_model.is_dir(): + print(f'Error: {args.model} is not a directory', file = sys.stderr) sys.exit(1) - -# output in the same directory as the model -dir_model = sys.argv[1] -last_dir = os.path.basename(os.path.normpath(dir_model)) - # possible tensor data types # ftype == 0 -> float32 # ftype == 1 -> float16 @@ -67,19 +72,15 @@ last_dir = os.path.basename(os.path.normpath(dir_model)) # map from ftype to string ftype_str = ["f32", "f16"] -ftype = 1 -if len(sys.argv) > 2: - ftype = int(sys.argv[2]) - if ftype < 0 or ftype > 1: - print("Invalid ftype: " + str(ftype)) +if args.outfile is not None: + fname_out = args.outfile +else: + # output in the same directory as the model by default + fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - sys.exit(1) +print("gguf: loading model "+dir_model.name) -fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf" - -print("gguf: loading model "+last_dir) - -with open(dir_model + "/config.json", "r", encoding="utf-8") as f: +with open(dir_model / "config.json", "r", encoding="utf-8") as f: hparams = json.load(f) if hparams["architectures"][0] != "GPTNeoXForCausalLM": @@ -97,7 +98,7 @@ print("gguf: get model metadata") block_count = hparams["num_hidden_layers"] -gguf_writer.add_name(last_dir) +gguf_writer.add_name(dir_model.name) gguf_writer.add_context_length(hparams["max_position_embeddings"]) gguf_writer.add_embedding_length(hparams["hidden_size"]) gguf_writer.add_block_count(block_count) @@ -111,86 +112,52 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"]) print("gguf: get tokenizer metadata") -tokens: List[str] = [] -merges: List[str] = [] +tokens: List[bytearray] = [] +tokenizer_json_file = dir_model / 'tokenizer.json' +if not tokenizer_json_file.is_file(): + print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr) + sys.exit(1) -if Path(dir_model + "/tokenizer.json").is_file(): - # gpt2 tokenizer - gguf_writer.add_tokenizer_model("gpt2") +# gpt2 tokenizer +gguf_writer.add_tokenizer_model("gpt2") - print("gguf: get gpt2 tokenizer merges") +with open(tokenizer_json_file, "r", encoding="utf-8") as f: + tokenizer_json = json.load(f) - with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f: - tokenizer_json = json.load(f) - merges = tokenizer_json["model"]["merges"] +print("gguf: get gpt2 tokenizer vocab") - gguf_writer.add_token_merges(merges) +vocab_size = len(tokenizer_json["model"]["vocab"]) - print("gguf: get gpt2 tokenizer vocab") +# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py +tokenizer = AutoTokenizer.from_pretrained(dir_model) - vocab_size = len(tokenizer_json["model"]["vocab"]) +reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} +byte_encoder = bytes_to_unicode() +byte_decoder = {v: k for k, v in byte_encoder.items()} - # ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py - tokenizer = AutoTokenizer.from_pretrained(dir_model) +for i in range(vocab_size): + if i in reverse_vocab: + try: + text = bytearray([byte_decoder[c] for c in reverse_vocab[i]]) + except KeyError: + text = bytearray() + for c in reverse_vocab[i]: + if ord(c) < 256: # single byte character + text.append(byte_decoder[ord(c)]) + else: # multibyte special token character + text.extend(c.encode('utf-8')) + else: + print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.") + pad_token = f"[PAD{i}]".encode("utf8") + text = bytearray(pad_token) - reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} - byte_encoder = bytes_to_unicode() - byte_decoder = {v: k for k, v in byte_encoder.items()} + tokens.append(text) - for i in range(vocab_size): - if i in reverse_vocab: - try: - text = bytearray([byte_decoder[c] for c in reverse_vocab[i]]) - except KeyError: - text = bytearray() - for c in reverse_vocab[i]: - if ord(c) < 256: # single byte character - text.append(byte_decoder[ord(c)]) - else: # multibyte special token character - text.extend(c.encode('utf-8')) - else: - print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.") - pad_token = f"[PAD{i}]".encode("utf8") - text = bytearray(pad_token) - - tokens.append(text) - - gguf_writer.add_token_list(tokens) - - if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file(): - print("gguf: get special token ids") - - with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f: - tokenizer_config = json.load(f) - - # find special token ids - - if "bos_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["bos_token"]: - gguf_writer.add_bos_token_id(key["id"]) - - if "eos_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["eos_token"]: - gguf_writer.add_eos_token_id(key["id"]) - - if "unk_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["unk_token"]: - gguf_writer.add_unk_token_id(key["id"]) - - if "sep_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["sep_token"]: - gguf_writer.add_sep_token_id(key["id"]) - - if "pad_token" in tokenizer_config: - for key in tokenizer_json["added_tokens"]: - if key["content"] == tokenizer_config["pad_token"]: - gguf_writer.add_pad_token_id(key["id"]) +gguf_writer.add_token_list(tokens) +special_vocab = gguf.SpecialVocab(dir_model, load_merges = True) +special_vocab.add_to_gguf(gguf_writer) # TENSORS @@ -200,13 +167,15 @@ tensor_map = gguf.get_tensor_name_map(ARCH,block_count) print("gguf: get tensor metadata") if num_parts == 0: - part_names = ("pytorch_model.bin",) + part_names = iter(("pytorch_model.bin",)) else: part_names = ( f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) ) for part_name in part_names: + if args.vocab_only: + break print("gguf: loading model part '" + part_name + "'") model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") @@ -226,11 +195,8 @@ for part_name in part_names: data = data.squeeze().numpy() # map tensor names - if name.endswith(".weight") and name[:-7] in tensor_map: - name = tensor_map[name[:-7]] + ".weight" - elif name.endswith(".bias") and name[:-5] in tensor_map: - name = tensor_map[name[:-5]] + ".bias" - else: + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) + if new_name is None: print("Can not map tensor '" + name + "'") sys.exit() @@ -249,19 +215,20 @@ for part_name in part_names: if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - gguf_writer.add_tensor(name, data) + gguf_writer.add_tensor(new_name, data) print("gguf: write header") gguf_writer.write_header_to_file() print("gguf: write metadata") gguf_writer.write_kv_data_to_file() -print("gguf: write tensors") -gguf_writer.write_tensors_to_file() +if not args.vocab_only: + print("gguf: write tensors") + gguf_writer.write_tensors_to_file() gguf_writer.close() -print("gguf: model successfully exported to '" + fname_out + "'") +print(f"gguf: model successfully exported to '{fname_out}'") print("") diff --git a/convert-llama-7b-pth-to-gguf.py b/convert-llama-7b-pth-to-gguf.py index 2ab082383..6e973a116 100755 --- a/convert-llama-7b-pth-to-gguf.py +++ b/convert-llama-7b-pth-to-gguf.py @@ -10,8 +10,9 @@ import struct import json import numpy as np import torch +import argparse -from typing import Any, List +from typing import Any, List, TypeAlias from pathlib import Path from sentencepiece import SentencePieceProcessor @@ -20,7 +21,7 @@ from sentencepiece import SentencePieceProcessor NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]' -def count_model_parts(dir_model: str) -> int: +def count_model_parts(dir_model: Path) -> int: num_parts = 0 for filename in os.listdir(dir_model): if filename.startswith("consolidated."): @@ -31,19 +32,22 @@ def count_model_parts(dir_model: str) -> int: return num_parts -if len(sys.argv) < 3: - print(f"Usage: python {sys.argv[0]} dir-model ftype\n") - print(" ftype == 0 -> float32") - print(" ftype == 1 -> float16") +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="Convert a PyTorch 7B LLaMA model to a GGML compatible file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") + parser.add_argument("ftype", type=int, choices=[0, 1], help="output format - use 0 for float32, 1 for float16", default = 1) + return parser.parse_args() +args = parse_args() + +dir_model = args.model +ftype = args.ftype +if not dir_model.is_dir(): + print(f'Error: {args.model} is not a directory', file = sys.stderr) sys.exit(1) - -# output in the same directory as the model -dir_model = sys.argv[1] -last_dir = os.path.basename(os.path.normpath(dir_model)) - - # possible tensor data types # ftype == 0 -> float32 # ftype == 1 -> float16 @@ -51,19 +55,15 @@ last_dir = os.path.basename(os.path.normpath(dir_model)) # map from ftype to string ftype_str = ["f32", "f16"] -ftype = 1 -if len(sys.argv) > 2: - ftype = int(sys.argv[2]) - if ftype < 0 or ftype > 1: - print("Invalid ftype: " + str(ftype)) +if args.outfile is not None: + fname_out = args.outfile +else: + # output in the same directory as the model by default + fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - sys.exit(1) +print("gguf: loading model "+dir_model.name) -fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf" - -print("gguf: loading model "+last_dir) - -with open(dir_model + "/config.json", "r", encoding="utf-8") as f: +with open(dir_model / "config.json", "r", encoding="utf-8") as f: hparams = json.load(f) if hparams["architectures"][0] != "LlamaForCausalLM": @@ -107,7 +107,7 @@ else: sys.exit() -gguf_writer.add_name(last_dir) +gguf_writer.add_name(dir_model.name) gguf_writer.add_source_hf_repo(hf_repo) gguf_writer.add_tensor_data_layout("Meta AI original pth") gguf_writer.add_context_length(ctx_length) @@ -133,109 +133,60 @@ tokens: List[bytes] = [] scores: List[float] = [] toktypes: List[int] = [] -if Path(dir_model + "/tokenizer.model").is_file(): - # vocab type sentencepiece - print("gguf: get sentencepiece tokenizer vocab and scores") +tokenizer_model_file = dir_model / 'tokenizer.model' +if not tokenizer_model_file.is_file(): + print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr) + sys.exit(1) - tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model") +# vocab type sentencepiece +print("gguf: get sentencepiece tokenizer vocab and scores") - for i in range(tokenizer.vocab_size()): - text: bytes - score: float +tokenizer = SentencePieceProcessor(str(tokenizer_model_file)) - piece = tokenizer.id_to_piece(i) - text = piece.encode("utf-8") - score = tokenizer.get_score(i) +for i in range(tokenizer.vocab_size()): + text: bytes + score: float - toktype = 1 # defualt to normal token type - if tokenizer.is_unknown(i): - toktype = 2 - if tokenizer.is_control(i): - toktype = 3 + piece = tokenizer.id_to_piece(i) + text = piece.encode("utf-8") + score = tokenizer.get_score(i) - # toktype = 4 is user-defined = tokens from added_tokens.json + toktype = 1 # defualt to normal token type + if tokenizer.is_unknown(i): + toktype = 2 + if tokenizer.is_control(i): + toktype = 3 - if tokenizer.is_unused(i): - toktype = 5 - if tokenizer.is_byte(i): - toktype = 6 + # toktype = 4 is user-defined = tokens from added_tokens.json - tokens.append(text) - scores.append(score) - toktypes.append(toktype) + if tokenizer.is_unused(i): + toktype = 5 + if tokenizer.is_byte(i): + toktype = 6 - if Path(dir_model + "/added_tokens.json").is_file(): - with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f: - addtokens_json = json.load(f) + tokens.append(text) + scores.append(score) + toktypes.append(toktype) - print("gguf: get added tokens") +added_tokens_file = dir_model / 'added_tokens.json' +if added_tokens_file.is_file(): + with open(added_tokens_file, "r", encoding="utf-8") as f: + addtokens_json = json.load(f) - for key in addtokens_json: - tokens.append( key.encode("utf-8") ) - scores.append(-1000.0) - toktypes.append(4) # user-defined token type + print("gguf: get added tokens") - gguf_writer.add_tokenizer_model("llama") - gguf_writer.add_token_list(tokens) - gguf_writer.add_token_scores(scores) - gguf_writer.add_token_types(toktypes) + for key in addtokens_json: + tokens.append( key.encode("utf-8") ) + scores.append(-1000.0) + toktypes.append(4) # user-defined token type +gguf_writer.add_tokenizer_model("llama") +gguf_writer.add_token_list(tokens) +gguf_writer.add_token_scores(scores) +gguf_writer.add_token_types(toktypes) -print("gguf: get special token ids") - -if Path(dir_model + "/tokenizer.json").is_file(): - # Look for special tokens in tokenizer.json if it exists - - with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f: - tokenizer = json.load(f) - - if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file(): - - with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f: - tokenizer_config = json.load(f) - - if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["bos_token"]["content"]: - gguf_writer.add_bos_token_id(key["id"]) - - if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["eos_token"]["content"]: - gguf_writer.add_eos_token_id(key["id"]) - - if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["unk_token"]["content"]: - gguf_writer.add_unk_token_id(key["id"]) - - if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["sep_token"]["content"]: - gguf_writer.add_sep_token_id(key["id"]) - - if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["pad_token"]["content"]: - gguf_writer.add_pad_token_id(key["id"]) -else: - # If no tokenizer.json: Look for special tokens in config.json - - if "bos_token_id" in hparams and hparams["bos_token_id"] != None: - gguf_writer.add_bos_token_id(hparams["bos_token_id"]) - - if "eos_token_id" in hparams and hparams["eos_token_id"] != None: - gguf_writer.add_eos_token_id(hparams["eos_token_id"]) - - if "unk_token_id" in hparams and hparams["unk_token_id"] != None: - gguf_writer.add_unk_token_id(hparams["unk_token_id"]) - - if "sep_token_id" in hparams and hparams["sep_token_id"] != None: - gguf_writer.add_sep_token_id(hparams["sep_token_id"]) - - if "pad_token_id" in hparams and hparams["pad_token_id"] != None: - gguf_writer.add_pad_token_id(hparams["pad_token_id"]) - +special_vocab = gguf.SpecialVocab(dir_model) +special_vocab.add_to_gguf(gguf_writer) # TENSORS @@ -247,6 +198,8 @@ print("gguf: get tensor metadata") part_names = (f"consolidated.{n:02}.pth" for n in range(0, num_parts)) for part_name in part_names: + if args.vocab_only: + break print("gguf: loading model part '" + part_name + "'") model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") @@ -266,11 +219,8 @@ for part_name in part_names: data = data.squeeze().numpy() # map tensor names - if name.endswith(".weight") and name[:-7] in tensor_map: - name = tensor_map[name[:-7]] + ".weight" - elif name.endswith(".bias") and name[:-5] in tensor_map: - name = tensor_map[name[:-5]] + ".bias" - else: + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) + if new_name is None: print("Can not map tensor '" + name + "'") sys.exit() @@ -289,20 +239,20 @@ for part_name in part_names: if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - gguf_writer.add_tensor(name, data) + gguf_writer.add_tensor(new_name, data) print("gguf: write header") gguf_writer.write_header_to_file() print("gguf: write metadata") gguf_writer.write_kv_data_to_file() -print("gguf: write tensors") -gguf_writer.write_tensors_to_file() +if not args.vocab_only: + print("gguf: write tensors") + gguf_writer.write_tensors_to_file() gguf_writer.close() - -print("gguf: model successfully exported to '" + fname_out + "'") +print(f"gguf: model successfully exported to '{fname_out}'") print("") diff --git a/convert-llama-ggmlv3-to-gguf.py b/convert-llama-ggmlv3-to-gguf.py index 3bf93627d..c8e7f1761 100755 --- a/convert-llama-ggmlv3-to-gguf.py +++ b/convert-llama-ggmlv3-to-gguf.py @@ -75,7 +75,7 @@ class Tensor: self.dims = () self.dtype = None self.start_offset = 0 - self.len_bytes = 0 + self.len_bytes = np.int64(0) def load(self, data, offset): orig_offset = offset @@ -134,13 +134,14 @@ class GGMLV3Model: return offset class GGMLToGGUF: - def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None): + def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None): hp = ggml_model.hyperparameters self.model = ggml_model self.data = data self.cfg = cfg self.params_override = params_override self.vocab_override = vocab_override + self.special_vocab = special_vocab if params_override is not None: n_kv_head = params_override.n_head_kv else: @@ -162,6 +163,8 @@ class GGMLToGGUF: gguf_writer = gguf.GGUFWriter(self.cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False) self.add_params(gguf_writer) self.add_vocab(gguf_writer) + if self.special_vocab is not None: + self.special_vocab.add_to_gguf(gguf_writer) self.add_tensors(gguf_writer) print(" gguf: write header") gguf_writer.write_header_to_file() @@ -259,20 +262,13 @@ class GGMLToGGUF: gguf_writer.add_eos_token_id(2) def add_tensors(self, gguf_writer): - nm = self.name_map + tensor_map = self.name_map data = self.data print(f'* Adding {len(self.model.tensors)} tensor(s)') for tensor in self.model.tensors: name = str(tensor.name, 'UTF-8') - if name.endswith('.weight'): - name = name[:-7] - suffix = '.weight' - elif name.endswith('.bias'): - name = name[:-5] - suffix = '.bias' - mapped_name = nm.get(name) + mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) assert mapped_name is not None, f'Bad name {name}' - mapped_name += suffix tempdims = list(tensor.dims[:]) if len(tempdims) > 1: temp = tempdims[1] @@ -302,8 +298,10 @@ def handle_metadata(cfg, hp): else: raise ValueError('Unable to load metadata') vocab = convert.load_vocab(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, cfg.vocabtype) + # FIXME: Respect cfg.vocab_dir? + svocab = gguf.SpecialVocab(cfg.model_metadata_dir) convert.check_vocab_size(params, vocab) - return (params, vocab) + return (params, vocab, svocab) def handle_args(): parser = argparse.ArgumentParser(description = 'Convert GGMLv3 models to GGUF') @@ -330,14 +328,16 @@ def main(): print(f'* GGML model hyperparameters: {model.hyperparameters}') vocab_override = None params_override = None + special_vocab = None if cfg.model_metadata_dir is not None: - (params_override, vocab_override) = handle_metadata(cfg, model.hyperparameters) + (params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters) print('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.') print(f'* Overriding params: {params_override}') print(f'* Overriding vocab: {vocab_override}') + print(f'* Special vocab: {special_vocab}') else: print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n') - converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override) + converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override, special_vocab = special_vocab) converter.save() print(f'* Successful completion. Output saved to: {cfg.output}') diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py index b00810dbb..ab94b5eab 100755 --- a/convert-llama-hf-to-gguf.py +++ b/convert-llama-hf-to-gguf.py @@ -8,8 +8,9 @@ import struct import json import numpy as np import torch +import argparse -from typing import Any, List, Optional +from typing import Any, List, Optional, TypeAlias from pathlib import Path from sentencepiece import SentencePieceProcessor @@ -43,40 +44,38 @@ def count_model_parts(dir_model: str) -> int: return num_parts -if len(sys.argv) < 3: - print(f"Usage: python {sys.argv[0]} dir-model ftype\n") - print(" ftype == 0 -> float32") - print(" ftype == 1 -> float16") +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") + parser.add_argument("ftype", type=int, choices=[0, 1], help="output format - use 0 for float32, 1 for float16", default = 1) + return parser.parse_args() +args = parse_args() + +dir_model = args.model +ftype = args.ftype +if not dir_model.is_dir(): + print(f'Error: {args.model} is not a directory', file = sys.stderr) sys.exit(1) - -# output in the same directory as the model -dir_model = sys.argv[1] -last_dir = os.path.basename(os.path.normpath(dir_model)) - - # possible tensor data types # ftype == 0 -> float32 # ftype == 1 -> float16 - # map from ftype to string ftype_str = ["f32", "f16"] -ftype = 1 -if len(sys.argv) > 2: - ftype = int(sys.argv[2]) - if ftype < 0 or ftype > 1: - print("Invalid ftype: " + str(ftype)) +if args.outfile is not None: + fname_out = args.outfile +else: + # output in the same directory as the model by default + fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - sys.exit(1) +print("gguf: loading model "+dir_model.name) -fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf" - -print("gguf: loading model "+last_dir) - -with open(dir_model + "/config.json", "r", encoding="utf-8") as f: +with open(dir_model / "config.json", "r", encoding="utf-8") as f: hparams = json.load(f) if hparams["architectures"][0] != "LlamaForCausalLM": @@ -115,7 +114,7 @@ else: sys.exit() -gguf_writer.add_name(last_dir) +gguf_writer.add_name(dir_model.name) gguf_writer.add_source_hf_repo(hf_repo) gguf_writer.add_tensor_data_layout("Meta AI original pth") gguf_writer.add_context_length(ctx_length) @@ -141,110 +140,61 @@ tokens: List[bytes] = [] scores: List[float] = [] toktypes: List[int] = [] -if Path(dir_model + "/tokenizer.model").is_file(): - # vocab type sentencepiece - print("gguf: get sentencepiece tokenizer vocab, scores and token types") +tokenizer_model_file = dir_model / 'tokenizer.model' +if not tokenizer_model_file.is_file(): + print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr) + sys.exit(1) - tokenizer = SentencePieceProcessor(dir_model + "/tokenizer.model") +# vocab type sentencepiece +print("gguf: get sentencepiece tokenizer vocab, scores and token types") - for i in range(tokenizer.vocab_size()): - text: bytes - score: float +tokenizer = SentencePieceProcessor(str(tokenizer_model_file)) - piece = tokenizer.id_to_piece(i) - text = piece.encode("utf-8") - score = tokenizer.get_score(i) +for i in range(tokenizer.vocab_size()): + text: bytes + score: float - toktype = 1 # defualt to normal token type - if tokenizer.is_unknown(i): - toktype = 2 - if tokenizer.is_control(i): - toktype = 3 + piece = tokenizer.id_to_piece(i) + text = piece.encode("utf-8") + score = tokenizer.get_score(i) - # toktype = 4 is user-defined = tokens from added_tokens.json + toktype = 1 # defualt to normal token type + if tokenizer.is_unknown(i): + toktype = 2 + if tokenizer.is_control(i): + toktype = 3 - if tokenizer.is_unused(i): - toktype = 5 - if tokenizer.is_byte(i): - toktype = 6 + # toktype = 4 is user-defined = tokens from added_tokens.json - tokens.append(text) - scores.append(score) - toktypes.append(toktype) + if tokenizer.is_unused(i): + toktype = 5 + if tokenizer.is_byte(i): + toktype = 6 - if Path(dir_model + "/added_tokens.json").is_file(): - with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f: - addtokens_json = json.load(f) + tokens.append(text) + scores.append(score) + toktypes.append(toktype) - print("gguf: get added tokens") +added_tokens_file = dir_model / 'added_tokens.json' +if added_tokens_file.is_file(): + with open(added_tokens_file, "r", encoding="utf-8") as f: + addtokens_json = json.load(f) - for key in addtokens_json: - tokens.append( key.encode("utf-8") ) - scores.append(-1000.0) - toktypes.append(4) # user-defined token type + print("gguf: get added tokens") + + for key in addtokens_json: + tokens.append( key.encode("utf-8") ) + scores.append(-1000.0) + toktypes.append(4) # user-defined token type - gguf_writer.add_tokenizer_model("llama") - gguf_writer.add_token_list(tokens) - gguf_writer.add_token_scores(scores) - gguf_writer.add_token_types(toktypes) - - -print("gguf: get special token ids") - -if Path(dir_model + "/tokenizer.json").is_file(): - # Look for special tokens in tokenizer.json if it exists - - with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f: - tokenizer = json.load(f) - - if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file(): - - with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f: - tokenizer_config = json.load(f) - - if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["bos_token"]["content"]: - gguf_writer.add_bos_token_id(key["id"]) - - if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["eos_token"]["content"]: - gguf_writer.add_eos_token_id(key["id"]) - - if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["unk_token"]["content"]: - gguf_writer.add_unk_token_id(key["id"]) - - if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["sep_token"]["content"]: - gguf_writer.add_sep_token_id(key["id"]) - - if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None: - for key in tokenizer["added_tokens"]: - if key["content"] == tokenizer_config["pad_token"]["content"]: - gguf_writer.add_pad_token_id(key["id"]) -else: - # If no tokenizer.json: Look for special tokens in config.json - - if "bos_token_id" in hparams and hparams["bos_token_id"] != None: - gguf_writer.add_bos_token_id(hparams["bos_token_id"]) - - if "eos_token_id" in hparams and hparams["eos_token_id"] != None: - gguf_writer.add_eos_token_id(hparams["eos_token_id"]) - - if "unk_token_id" in hparams and hparams["unk_token_id"] != None: - gguf_writer.add_unk_token_id(hparams["unk_token_id"]) - - if "sep_token_id" in hparams and hparams["sep_token_id"] != None: - gguf_writer.add_sep_token_id(hparams["sep_token_id"]) - - if "pad_token_id" in hparams and hparams["pad_token_id"] != None: - gguf_writer.add_pad_token_id(hparams["pad_token_id"]) +gguf_writer.add_tokenizer_model("llama") +gguf_writer.add_token_list(tokens) +gguf_writer.add_token_scores(scores) +gguf_writer.add_token_types(toktypes) +special_vocab = gguf.SpecialVocab(dir_model) +special_vocab.add_to_gguf(gguf_writer) # TENSORS @@ -254,13 +204,15 @@ tensor_map = gguf.get_tensor_name_map(ARCH,block_count) print("gguf: get tensor metadata") if num_parts == 0: - part_names = ("pytorch_model.bin",) + part_names = iter(("pytorch_model.bin",)) else: part_names = ( f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) ) for part_name in part_names: + if args.vocab_only: + break print("gguf: loading model part '" + part_name + "'") model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") @@ -286,11 +238,8 @@ for part_name in part_names: data = reverse_hf_permute(data, head_count, head_count_kv) # map tensor names - if name.endswith(".weight") and name[:-7] in tensor_map: - name = tensor_map[name[:-7]] + ".weight" - elif name.endswith(".bias") and name[:-5] in tensor_map: - name = tensor_map[name[:-5]] + ".bias" - else: + new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) + if new_name is None: print("Can not map tensor '" + name + "'") sys.exit() @@ -309,20 +258,20 @@ for part_name in part_names: if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - gguf_writer.add_tensor(name, data) + gguf_writer.add_tensor(new_name, data) print("gguf: write header") gguf_writer.write_header_to_file() print("gguf: write metadata") gguf_writer.write_kv_data_to_file() -print("gguf: write tensors") -gguf_writer.write_tensors_to_file() +if not args.vocab_only: + print("gguf: write tensors") + gguf_writer.write_tensors_to_file() gguf_writer.close() - -print("gguf: model successfully exported to '" + fname_out + "'") +print(f"gguf: model successfully exported to '{fname_out}'") print("") diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index a94a7d0af..a00339b47 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -4,7 +4,7 @@ import os import re import struct import sys -from typing import Any, Dict, Sequence, TextIO +from typing import Any, Dict, Sequence, BinaryIO import numpy as np import torch @@ -46,7 +46,7 @@ def translate_tensor_name(t: str) -> str: sys.exit(1) -def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None: +def write_file_header(fout: BinaryIO, params: Dict[str, Any]) -> None: fout.write(b"ggla"[::-1]) # magic (ggml lora) fout.write(struct.pack("i", 1)) # file version fout.write(struct.pack("i", params["r"])) @@ -60,7 +60,7 @@ def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None: def write_tensor_header( - self, name: str, shape: Sequence[int], data_type: np.dtype + self, name: str, shape: Sequence[int], data_type: np.dtype[Any] ) -> None: sname = name.encode("utf-8") fout.write( diff --git a/convert.py b/convert.py index 3f0a1c932..448b6f0f3 100755 --- a/convert.py +++ b/convert.py @@ -25,7 +25,7 @@ import numpy as np from abc import ABCMeta, abstractmethod from dataclasses import dataclass from pathlib import Path -from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, TypeVar, Union) +from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, Type, TypeVar, Union) from sentencepiece import SentencePieceProcessor # type: ignore if TYPE_CHECKING: @@ -299,8 +299,10 @@ class Params: params = Params.loadHFTransformerJson(model_plus.model, hf_config_path) elif orig_config_path.exists(): params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path) - else: + elif model_plus.format != 'none': params = Params.guessed(model_plus.model) + else: + raise ValueError('Cannot guess params when model format is none') params.path_model = model_plus.paths[0].parent @@ -353,7 +355,7 @@ class BpeVocab: yield from self.added_tokens() def __repr__(self) -> str: - return f"BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>" + return f"" class SentencePieceVocab: @@ -416,7 +418,6 @@ class SentencePieceVocab: Vocab = Union[BpeVocab, SentencePieceVocab] - # # data loading # TODO: reuse (probably move to gguf.py?) @@ -439,14 +440,14 @@ class Tensor(metaclass=ABCMeta): @abstractmethod def permute(self, n_head: int, n_head_kv: int) -> 'Tensor': ... @abstractmethod - def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': ... + def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> 'UnquantizedTensor': ... @abstractmethod def part(self, n_part: int) -> 'UnquantizedTensor': ... @abstractmethod def to_ggml(self) -> 'GGMLCompatibleTensor': ... -def bf16_to_fp32(bf16_arr: np.ndarray) -> np.ndarray: +def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray: assert bf16_arr.dtype == np.uint16, f"Input array should be of dtype uint16, but got {bf16_arr.dtype}" fp32_arr = bf16_arr.astype(np.uint32) << 16 return fp32_arr.view(np.float32) @@ -467,9 +468,9 @@ class UnquantizedTensor(Tensor): def to_ggml(self) -> 'UnquantizedTensor': return self - def permute_part(self, n_part: int, n_head: int) -> 'UnquantizedTensor': + def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> 'UnquantizedTensor': r = self.ndarray.shape[0] // 3 - return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head)) + return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head_kv)) def part(self, n_part: int) -> 'UnquantizedTensor': r = self.ndarray.shape[0] // 3 @@ -531,7 +532,7 @@ LazyModel = Dict[str, LazyTensor] class ModelPlus: model: LazyModel paths: List[Path] # Where this was read from. - format: Literal['ggml', 'torch', 'safetensors'] + format: Literal['ggml', 'torch', 'safetensors', 'none'] vocab: Optional[Vocab] # For GGML models (which have vocab built in), the vocab. @@ -597,12 +598,12 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTe return lazy_tensor.load().permute(n_head, n_head_kv) return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description) -def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int) -> LazyTensor: +def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor: def load() -> Tensor: - return lazy_tensor.load().permute_part(n_part, n_head) + return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv) s = lazy_tensor.shape.copy() s[0] = s[0] // 3 - return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description) + return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description) def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor: def load() -> Tensor: @@ -657,7 +658,7 @@ class LazyUnpickler(pickle.Unpickler): description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}' return LazyStorage(load=load, kind=pid[1], description=description) - # @staticmethod + @staticmethod def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any, # pyright: ignore[reportSelfClsParameterName] requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor: @@ -669,13 +670,15 @@ class LazyUnpickler(pickle.Unpickler): description = f'pickled storage_offset={storage_offset} in {storage.description}' return LazyTensor(load, list(size), storage.kind.data_type, description) - # @staticmethod + @staticmethod def rebuild_from_type_v2(func, new_type, args, state): return func(*args) - CLASSES: Dict[Any, Any] = { - ('torch._tensor', '_rebuild_from_type_v2'): rebuild_from_type_v2, - ('torch._utils', '_rebuild_tensor_v2'): lazy_rebuild_tensor_v2, + CLASSES: Dict[Tuple[str, str], Any] = { + # getattr used here as a workaround for mypy not being smart enough to detrmine + # the staticmethods have a __func__ attribute. + ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'), + ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'), ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16), ('torch', 'HalfStorage'): LazyStorageKind(DT_F16), ('torch', 'FloatStorage'): LazyStorageKind(DT_F32), @@ -751,7 +754,7 @@ def lazy_load_file(path: Path) -> ModelPlus: In = TypeVar('In') Out = TypeVar('Out') -def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, factory: Callable = ThreadPoolExecutor) -> Iterable[Out]: +def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, use_processpool_executor: bool = False) -> Iterable[Out]: '''Parallel map, but with backpressure. If the caller doesn't call `next` fast enough, this will stop calling `func` at some point rather than letting results pile up in memory. Specifically, there is a max of one @@ -760,7 +763,12 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc yield from map(func, iterable) # Not reached. iterable = iter(iterable) - with factory(max_workers = max_workers) as executor: + executor_class: Union[Type[ThreadPoolExecutor], Type[ProcessPoolExecutor]] + if use_processpool_executor: + executor_class = ProcessPoolExecutor + else: + executor_class = ThreadPoolExecutor + with executor_class(max_workers = max_workers) as executor: futures: List[concurrent.futures.Future[Out]] = [] done = False for _ in range(concurrency): @@ -838,11 +846,19 @@ class OutputFile: scores.append(score) toktypes.append(toktype) - self.gguf.add_tokenizer_model("llama") + if isinstance(vocab, SentencePieceVocab): + self.gguf.add_tokenizer_model("llama") + elif isinstance(vocab, BpeVocab): + self.gguf.add_tokenizer_model("gpt2") + else: + raise ValueError(f'Unknown vocab type: Not BpeVocab or SentencePieceVocab') self.gguf.add_token_list(tokens) self.gguf.add_token_scores(scores) self.gguf.add_token_types(toktypes) + def add_meta_special_vocab(self, svocab: gguf.SpecialVocab) -> None: + svocab.add_to_gguf(self.gguf) + def add_tensor_info(self, name: str, tensor: LazyTensor) -> None: n_elements = int(np.prod(tensor.shape)) raw_dtype = getattr(tensor.data_type, 'ggml_type', None) @@ -861,7 +877,7 @@ class OutputFile: self.gguf.close() @staticmethod - def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab) -> None: + def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None: check_vocab_size(params, vocab) of = OutputFile(fname_out) @@ -869,6 +885,8 @@ class OutputFile: # meta data of.add_meta_arch(params) of.add_meta_vocab(vocab) + of.add_meta_special_vocab(svocab) + of.write_meta() of.close() @@ -887,7 +905,7 @@ class OutputFile: return dt.quantize(arr) @staticmethod - def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, concurrency: int = DEFAULT_CONCURRENCY) -> None: + def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None: check_vocab_size(params, vocab) of = OutputFile(fname_out) @@ -895,6 +913,7 @@ class OutputFile: # meta data of.add_meta_arch(params) of.add_meta_vocab(vocab) + of.add_meta_special_vocab(svocab) # tensor info for name, lazy_tensor in model.items(): @@ -906,7 +925,7 @@ class OutputFile: # tensor data ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency) if ftype == GGMLFileType.MostlyQ8_0: - ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, factory = ProcessPoolExecutor) + ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, use_processpool_executor = True) else: ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner) @@ -939,7 +958,8 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM for (name, tensor) in model.items()} def convert_model_names(model: LazyModel, params: Params) -> LazyModel: - tmap = gguf.get_tensor_name_map(ARCH, params.n_layer) + tmap = gguf.TensorNameMap(ARCH, params.n_layer) + should_skip: Set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, [])) tmp = model @@ -952,8 +972,8 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel: #tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"] elif f"model.layers.{i}.self_attn.W_pack.weight" in model: print(f"Unpacking and permuting layer {i}") - tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head) - tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head) + tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head) + tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv) tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2) del tmp[f"model.layers.{i}.self_attn.W_pack.weight"] else: @@ -961,23 +981,16 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel: out: LazyModel = {} for name, lazy_tensor in model.items(): - name_new = name - - if name in tmap: - name_new = tmap[name] - elif name.endswith(".weight") and name[:-7] in tmap: - name_new = tmap[name[:-7]] + ".weight" - elif name.endswith(".bias") and name[:-5] in tmap: - name_new = tmap[name[:-5]] + ".bias" - else: + tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None) + if name_new is None: raise Exception(f"Unexpected tensor name: {name}") - if gguf.should_skip_tensor_TMP(ARCH, params.n_layer, name_new): + if tensor_type in should_skip: print(f"skipping tensor {name_new}") continue - else: - print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}") - out[name_new] = lazy_tensor + + print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}") + out[name_new] = lazy_tensor return out @@ -1117,8 +1130,16 @@ def main(args_in: Optional[List[str]] = None) -> None: if args.dump_single: model_plus = lazy_load_file(args.model) do_dump_model(model_plus) + return - model_plus = load_some_model(args.model) + if not args.vocab_only: + model_plus = load_some_model(args.model) + else: + model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None) + + if args.dump: + do_dump_model(model_plus) + return params = Params.load(model_plus) if params.n_ctx == -1: @@ -1140,33 +1161,34 @@ def main(args_in: Optional[List[str]] = None) -> None: vocab: Vocab if args.vocab_only: - vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype) assert args.outfile, "need --outfile if using --vocab-only" + # FIXME: Try to respect vocab_dir somehow? + vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype) + special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe') outfile = args.outfile - OutputFile.write_vocab_only(outfile, params, vocab) + OutputFile.write_vocab_only(outfile, params, vocab, special_vocab) print(f"Wrote {outfile}") + return + + if model_plus.vocab is not None and args.vocab_dir is None: + vocab = model_plus.vocab else: - if args.dump: - do_dump_model(model_plus) - return + vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent + vocab = load_vocab(vocab_dir, args.vocabtype) + # FIXME: Try to respect vocab_dir somehow? + special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe') - if model_plus.vocab is not None and args.vocab_dir is None: - vocab = model_plus.vocab - else: - vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent - vocab = load_vocab(vocab_dir, args.vocabtype) + model = model_plus.model + model = convert_model_names(model, params) + ftype = pick_output_type(model, args.outtype) + model = convert_to_output_type(model, ftype) + outfile = args.outfile or default_outfile(model_plus.paths, ftype) - model = model_plus.model - model = convert_model_names(model, params) - ftype = pick_output_type(model, args.outtype) - model = convert_to_output_type(model, ftype) - outfile = args.outfile or default_outfile(model_plus.paths, ftype) + params.ftype = ftype + print(f"Writing {outfile}, format {ftype}") - params.ftype = ftype - print(f"Writing {outfile}, format {ftype}") - - OutputFile.write_all(outfile, ftype, params, model, vocab, concurrency = args.concurrency) - print(f"Wrote {outfile}") + OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency) + print(f"Wrote {outfile}") if __name__ == '__main__': diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py index 838a2c0f8..de3edbc99 100644 --- a/gguf-py/gguf/gguf.py +++ b/gguf-py/gguf/gguf.py @@ -4,9 +4,13 @@ import sys import struct import tempfile import numpy as np +import json +import os +from pathlib import Path from enum import IntEnum, auto -from typing import Any, IO, List, Optional +from io import BufferedWriter +from typing import Any, BinaryIO, Callable, IO, Dict, List, Optional, Sequence, Tuple, Union # # constants @@ -71,35 +75,35 @@ KEY_TOKENIZER_RWKV = "tokenizer.rwkv.world" class MODEL_ARCH(IntEnum): - LLAMA = auto() - FALCON = auto() - GPT2 = auto() - GPTJ = auto() - GPTNEOX = auto() - MPT = auto() + LLAMA : int = auto() + FALCON : int = auto() + GPT2 : int = auto() + GPTJ : int = auto() + GPTNEOX: int = auto() + MPT : int = auto() class MODEL_TENSOR(IntEnum): - TOKEN_EMBD = auto() - POS_EMBD = auto() - OUTPUT = auto() - OUTPUT_NORM = auto() - ROPE_FREQS = auto() - ATTN_Q = auto() - ATTN_K = auto() - ATTN_V = auto() - ATTN_QKV = auto() - ATTN_OUT = auto() - ATTN_NORM = auto() - ATTN_NORM_2 = auto() - ATTN_ROT_EMBD = auto() - FFN_GATE = auto() - FFN_DOWN = auto() - FFN_UP = auto() - FFN_NORM = auto() + TOKEN_EMBD : int = auto() + POS_EMBD : int = auto() + OUTPUT : int = auto() + OUTPUT_NORM : int = auto() + ROPE_FREQS : int = auto() + ATTN_Q : int = auto() + ATTN_K : int = auto() + ATTN_V : int = auto() + ATTN_QKV : int = auto() + ATTN_OUT : int = auto() + ATTN_NORM : int = auto() + ATTN_NORM_2 : int = auto() + ATTN_ROT_EMBD: int = auto() + FFN_GATE : int = auto() + FFN_DOWN : int = auto() + FFN_UP : int = auto() + FFN_NORM : int = auto() -MODEL_ARCH_NAMES = { +MODEL_ARCH_NAMES: Dict[MODEL_ARCH, str] = { MODEL_ARCH.LLAMA: "llama", MODEL_ARCH.FALCON: "falcon", MODEL_ARCH.GPT2: "gpt2", @@ -108,7 +112,7 @@ MODEL_ARCH_NAMES = { MODEL_ARCH.MPT: "mpt", } -MODEL_TENSOR_NAMES = { +MODEL_TENSOR_NAMES: Dict[MODEL_ARCH, Dict[MODEL_TENSOR, str]] = { MODEL_ARCH.LLAMA: { MODEL_TENSOR.TOKEN_EMBD: "token_embd", MODEL_TENSOR.OUTPUT_NORM: "output_norm", @@ -154,7 +158,7 @@ MODEL_TENSOR_NAMES = { } # tensors that will not be serialized -MODEL_TENSOR_SKIP = { +MODEL_TENSOR_SKIP: Dict[MODEL_ARCH, List[MODEL_TENSOR]] = { MODEL_ARCH.LLAMA: [ MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, @@ -162,167 +166,198 @@ MODEL_TENSOR_SKIP = { } -# TODO: the following helper functions should be removed -# instead, get_tensor_name_map should return tuples of (name, MODEL_TENSOR) -# however, my Python is very bad, and I couldn't figure out how to do this, hence these functions -# REMOVE -def should_skip_tensor_TMP(arch: MODEL_ARCH, n_blocks: int, name: str) -> bool: - for skip in MODEL_TENSOR_SKIP.get(arch, []): - for i in range(n_blocks): - if name == MODEL_TENSOR_NAMES[arch][skip].format(bid=i): - return True +class TensorNameMap: + mappings_cfg: Dict[MODEL_TENSOR, Tuple[str, ...]] = { + # Token embeddings + MODEL_TENSOR.TOKEN_EMBD: ( + "gpt_neox.embed_in", # gptneox + "transformer.wte", # gpt2 mpt + "transformer.word_embeddings", # falcon + "model.embed_tokens", # llama-hf + "tok_embeddings", # llama-pth + ), - return False + # Position embeddings + MODEL_TENSOR.POS_EMBD: ( + "transformer.wpe", # gpt2 + ), + # Output + MODEL_TENSOR.OUTPUT: ( + "embed_out", # gptneox + "lm_head", # gpt2 mpt falcon llama-hf + "output", # llama-pth + ), -def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> dict: - tensor_map = {} + # Output norm + MODEL_TENSOR.OUTPUT_NORM: ( + "gpt_neox.final_layer_norm", # gptneox + "transformer.ln_f", # gpt2 falcon + "model.norm", # llama-hf + "norm", # llama-pth + ), - # Token embeddings - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.TOKEN_EMBD, None) + # Rope frequencies + MODEL_TENSOR.ROPE_FREQS: ( + "rope.freqs", # llama-pth + ), + } - tensor_map["gpt_neox.embed_in"] = mapped_to # gptneox - tensor_map["transformer.wte"] = mapped_to # gpt2 mpt - tensor_map["transformer.word_embeddings"] = mapped_to # falcon - tensor_map["model.embed_tokens"] = mapped_to # llama-hf - tensor_map["tok_embeddings"] = mapped_to # llama-pth - - # Position embeddings - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.POS_EMBD, None) - - tensor_map["transformer.wpe"] = mapped_to # gpt2 - - # Output - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.OUTPUT, None) - - tensor_map["embed_out"] = mapped_to # gptneox - tensor_map["lm_head"] = mapped_to # gpt2 mpt falcon llama-hf - tensor_map["output"] = mapped_to # llama-pth - - # Output norm - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.OUTPUT_NORM, None) - - tensor_map["gpt_neox.final_layer_norm"] = mapped_to # gptneox - tensor_map["transformer.ln_f"] = mapped_to # gpt2 falcon - tensor_map["transformer.norm_f"] = mapped_to # mpt - tensor_map["model.norm"] = mapped_to # llama-hf - tensor_map["norm"] = mapped_to # llama-pth - - # Rope frequencies - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ROPE_FREQS, None) - - tensor_map["rope.freqs"] = mapped_to # llama-pth - - # Attention and feed-forward blocks - for i in range(0, n_blocks): + block_mappings_cfg: Dict[MODEL_TENSOR, Tuple[str, ...]] = { # Attention norm - # TODO: is there are simpler way to write these 2 lines in Python? - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_NORM, None) - mapped_to = mapped_to.format(bid=i) if mapped_to else None - - tensor_map["gpt_neox.layers."+str(i)+".input_layernorm"] = mapped_to # gptneox - tensor_map["transformer.h."+str(i)+".ln_1"] = mapped_to # gpt2 - tensor_map["transformer.blocks."+str(i)+".norm_1"] = mapped_to # mpt - tensor_map["transformer.h."+str(i)+".input_layernorm"] = mapped_to # falcon7b - tensor_map["transformer.h."+str(i)+".ln_mlp"] = mapped_to # falcon40b - tensor_map["model.layers."+str(i)+".input_layernorm"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".attention_norm"] = mapped_to # llama-pth + MODEL_TENSOR.ATTN_NORM: ( + "gpt_neox.layers.{bid}.input_layernorm", # gptneox + "transformer.h.{bid}.ln_1", # gpt2 + "transformer.blocks.{bid}.norm_1", # mpt + "transformer.h.{bid}.input_layernorm", # falcon7b + "transformer.h.{bid}.ln_mlp", # falcon40b + "model.layers.{bid}.input_layernorm", # llama-hf + "layers.{bid}.attention_norm", # llama-pth + ), # Attention norm 2 - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_NORM_2, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["transformer.h."+str(i)+".ln_attn"] = mapped_to # falcon40b + MODEL_TENSOR.ATTN_NORM_2: ( + "transformer.h.{bid}.ln_attn", # falcon40b + ), # Attention query-key-value - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_QKV, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["gpt_neox.layers."+str(i)+".attention.query_key_value"] = mapped_to # gptneox - tensor_map["transformer.h."+str(i)+".attn.c_attn"] = mapped_to # gpt2 - tensor_map["transformer.blocks."+str(i)+".attn.Wqkv"] = mapped_to # mpt - tensor_map["transformer.h."+str(i)+".self_attention.query_key_value"] = mapped_to # falcon + MODEL_TENSOR.ATTN_QKV: ( + "gpt_neox.layers.{bid}.attention.query_key_value", # gptneox + "transformer.h.{bid}.attn.c_attn", # gpt2 + "transformer.blocks.{bid}.attn.Wqkv", # mpt + "transformer.h.{bid}.self_attention.query_key_value", # falcon + ), # Attention query - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_Q, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["model.layers."+str(i)+".self_attn.q_proj"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".attention.wq"] = mapped_to # llama-pth + MODEL_TENSOR.ATTN_Q: ( + "model.layers.{bid}.self_attn.q_proj", # llama-hf + "layers.{bid}.attention.wq", # llama-pth + ), # Attention key - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_K, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["model.layers."+str(i)+".self_attn.k_proj"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".attention.wk"] = mapped_to # llama-pth + MODEL_TENSOR.ATTN_K: ( + "model.layers.{bid}.self_attn.k_proj", # llama-hf + "layers.{bid}.attention.wk", # llama-pth + ), # Attention value - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_V, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["model.layers."+str(i)+".self_attn.v_proj"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".attention.wv"] = mapped_to # llama-pth + MODEL_TENSOR.ATTN_V: ( + "model.layers.{bid}.self_attn.v_proj", # llama-hf + "layers.{bid}.attention.wv", # llama-pth + ), # Attention output - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_OUT, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["gpt_neox.layers."+str(i)+".attention.dense"] = mapped_to # gptneox - tensor_map["transformer.h."+str(i)+".attn.c_proj"] = mapped_to # gpt2 - tensor_map["transformer.blocks."+str(i)+".attn.out_proj"] = mapped_to # mpt - tensor_map["transformer.h."+str(i)+".self_attention.dense"] = mapped_to # falcon - tensor_map["model.layers."+str(i)+".self_attn.o_proj"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".attention.wo"] = mapped_to # llama-pth + MODEL_TENSOR.ATTN_OUT: ( + "gpt_neox.layers.{bid}.attention.dense", # gptneox + "transformer.h.{bid}.attn.c_proj", # gpt2 + "transformer.blocks.{bid}.attn.out_proj", # mpt + "transformer.h.{bid}.self_attention.dense", # falcon + "model.layers.{bid}.self_attn.o_proj", # llama-hf + "layers.{bid}.attention.wo", # llama-pth + ), # Rotary embeddings - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.ATTN_ROT_EMBD, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["model.layers."+str(i)+".self_attn.rotary_emb.inv_freq"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".attention.inner_attention.rope.freqs"] = mapped_to # llama-pth + MODEL_TENSOR.ATTN_ROT_EMBD: ( + "model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf + "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth + ), # Feed-forward norm - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_NORM, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["gpt_neox.layers."+str(i)+".post_attention_layernorm"] = mapped_to # gptneox - tensor_map["transformer.h."+str(i)+".ln_2"] = mapped_to # gpt2 - tensor_map["transformer.blocks."+str(i)+".norm_2"] = mapped_to # mpt - tensor_map["model.layers."+str(i)+".post_attention_layernorm"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".ffn_norm"] = mapped_to # llama-pth + MODEL_TENSOR.FFN_NORM: ( + "gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox + "transformer.h.{bid}.ln_2", # gpt2 + "transformer.blocks.{bid}.norm_2", # mpt + "model.layers.{bid}.post_attention_layernorm", # llama-hf + "layers.{bid}.ffn_norm", # llama-pth + ), # Feed-forward up - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_UP, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["gpt_neox.layers."+str(i)+".mlp.dense_h_to_4h"] = mapped_to # gptneox - tensor_map["transformer.h."+str(i)+".mlp.c_fc"] = mapped_to # gpt2 - tensor_map["transformer.blocks."+str(i)+".ffn.up_proj"] = mapped_to # mpt - tensor_map["transformer.h."+str(i)+".mlp.dense_h_to_4h"] = mapped_to # falcon - tensor_map["model.layers."+str(i)+".mlp.up_proj"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".feed_forward.w3"] = mapped_to # llama-pth + MODEL_TENSOR.FFN_UP: ( + "gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox + "transformer.h.{bid}.mlp.c_fc", # gpt2 + "transformer.blocks.{bid}.ffn.up_proj", # mpt + "transformer.h.{bid}.mlp.dense_h_to_4h", # falcon + "model.layers.{bid}.mlp.up_proj", # llama-hf + "layers.{bid}.feed_forward.w3", # llama-pth + ), # Feed-forward gate - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_GATE, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None - - tensor_map["model.layers."+str(i)+".mlp.gate_proj"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".feed_forward.w1"] = mapped_to # llama-pth + MODEL_TENSOR.FFN_GATE: ( + "model.layers.{bid}.mlp.gate_proj", # llama-hf + "layers.{bid}.feed_forward.w1", # llama-pth + ), # Feed-forward down - mapped_to = MODEL_TENSOR_NAMES[arch].get(MODEL_TENSOR.FFN_DOWN, None) - mapped_to = mapped_to.format(bid=i) if mapped_to is not None else None + MODEL_TENSOR.FFN_DOWN: ( + "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox + "transformer.h.{bid}.mlp.c_proj", # gpt2 + "transformer.blocks.{bid}.ffn.down_proj", # mpt + "transformer.h.{bid}.mlp.dense_4h_to_h", # falcon + "model.layers.{bid}.mlp.down_proj", # llama-hf + "layers.{bid}.feed_forward.w2", # llama-pth + ), + } - tensor_map["gpt_neox.layers."+str(i)+".mlp.dense_4h_to_h"] = mapped_to # gptneox - tensor_map["transformer.h."+str(i)+".mlp.c_proj"] = mapped_to # gpt2 - tensor_map["transformer.blocks."+str(i)+".ffn.down_proj"] = mapped_to # mpt - tensor_map["transformer.h."+str(i)+".mlp.dense_4h_to_h"] = mapped_to # falcon - tensor_map["model.layers."+str(i)+".mlp.down_proj"] = mapped_to # llama-hf - tensor_map["layers."+str(i)+".feed_forward.w2"] = mapped_to # llama-pth + mapping: Dict[str, Tuple[MODEL_TENSOR, str]] - return tensor_map + tensor_names: Dict[MODEL_TENSOR, str] + def __init__(self, arch: MODEL_ARCH, n_blocks: int): + mapping = self.mapping = {} + tensor_names = self.tensor_names = MODEL_TENSOR_NAMES[arch] + for tensor, keys in self.mappings_cfg.items(): + tensor_name = tensor_names.get(tensor) + if tensor_name is None: + continue + for key in keys: + mapping[key] = (tensor, tensor_name) + for bid in range(n_blocks): + for tensor, keys in self.block_mappings_cfg.items(): + tensor_name = tensor_names.get(tensor) + if tensor_name is None: + continue + tensor_name = tensor_name.format(bid = bid) + for key in keys: + key = key.format(bid = bid) + mapping[key] = (tensor, tensor_name) + + def get_type_and_name(self, key: str, try_suffixes: Sequence[str]) -> Optional[Tuple[MODEL_TENSOR, str]]: + result = self.mapping.get(key) + if result is not None: + return result + for suffix in try_suffixes: + if key.endswith(suffix): + result = self.mapping.get(key[:-len(suffix)]) + if result is not None: + return (result[0], result[1] + suffix) + return None + + def get_name(self, key: str, try_suffixes: Sequence[str]) -> Optional[str]: + result = self.get_type_and_name(key, try_suffixes = try_suffixes) + if result is None: + return None + return result[1] + + def get_type(self, key: str, try_suffixes: Sequence[str]) -> Optional[MODEL_TENSOR]: + result = self.get_type_and_name(key, try_suffixes = try_suffixes) + if result is None: + return None + return result[0] + + def __getitem__(self, key: str) -> str: + try: + return self.mapping[key][1] + except KeyError: + raise KeyError(key) + + def __contains__(self, key: str) -> bool: + return key in self.mapping + + def __repr__(self) -> str: + return repr(self.mapping) + +def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap: + return TensorNameMap(arch, n_blocks) class TokenType(IntEnum): NORMAL = 1 @@ -388,15 +423,21 @@ class GGUFValueType(IntEnum): class GGUFWriter: - def __init__(self, path: str, arch: str, use_temp_file = True): + fout: BufferedWriter + arch: str + offset_tensor = 0 + data_alignment = GGUF_DEFAULT_ALIGNMENT + kv_data = b"" + kv_data_count = 0 + ti_data = b"" + ti_data_count = 0 + use_temp_file: bool + temp_file: Optional[tempfile.SpooledTemporaryFile[bytes]] = None + tensors: List[Tuple[np.ndarray[Any, Any], int]] + + def __init__(self, path: Union[os.PathLike[str], str], arch: str, use_temp_file = True): self.fout = open(path, "wb") self.arch = arch - self.offset_tensor = 0 - self.data_alignment = GGUF_DEFAULT_ALIGNMENT - self.kv_data = b"" - self.kv_data_count = 0 - self.ti_data = b"" - self.ti_data_count = 0 self.add_architecture() self.use_temp_file = use_temp_file self.tensors = [] @@ -470,14 +511,27 @@ class GGUFWriter: self.add_key(key) self.add_val(val, GGUFValueType.STRING) - def add_array(self, key: str, val: list): - if not isinstance(val, list): - raise ValueError("Value must be a list for array type") + def add_array(self, key: str, val: Sequence[Any]): + if not isinstance(val, Sequence): + raise ValueError("Value must be a sequence for array type") self.add_key(key) self.add_val(val, GGUFValueType.ARRAY) - def add_val(self: str, val: Any, vtype: GGUFValueType = None, add_vtype: bool = True): + _simple_value_packing = { + GGUFValueType.UINT8: " 0: + ltype = GGUFValueType.get_type(val[0]) + if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]): + raise ValueError("All items in a GGUF array should be of the same type") + self.kv_data += struct.pack(" int: return ((x + n - 1) // n) * n - def add_tensor_info(self, name: str, tensor_shape: np.ndarray, tensor_dtype: np.dtype, tensor_nbytes: int, raw_dtype: Optional[GGMLQuantizationType] = None): + def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: Union[np.dtype[np.float16], np.dtype[np.float32]], tensor_nbytes: int, raw_dtype: Optional[GGMLQuantizationType] = None): assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now" encoded_name = name.encode("utf8") @@ -544,16 +580,18 @@ class GGUFWriter: self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment) self.ti_data_count += 1 - def add_tensor(self, name: str, tensor: np.ndarray, raw_shape: Optional[np.ndarray] = None, raw_dtype: Optional[GGMLQuantizationType] = None): - if self.use_temp_file and not hasattr(self, "temp_file"): - self.temp_file = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024) - self.temp_file.seek(0) + def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Optional[Sequence[int]] = None, raw_dtype: Optional[GGMLQuantizationType] = None): + if self.use_temp_file and self.temp_file is None: + fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024) + fp.seek(0) + self.temp_file = fp - self.add_tensor_info(name, raw_shape if raw_shape is not None else tensor.shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype) + shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape + self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype) pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes - if not self.use_temp_file: + if self.temp_file is None: self.tensors.append((tensor, pad)) return @@ -562,25 +600,22 @@ class GGUFWriter: if pad != 0: self.temp_file.write(bytes([0] * pad)) - def write_tensor_data(self, tensor: np.ndarray): - pad = GGUFWriter.ggml_pad(self.fout.tell(), self.data_alignment) - self.fout.tell() + def write_padding(self, fp: BinaryIO, n: int, align: Optional[int] = None): + pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n if pad != 0: - self.fout.write(bytes([0] * pad)) + fp.write(bytes([0] * pad)) + def write_tensor_data(self, tensor: np.ndarray[Any, Any]): + self.write_padding(self.fout, self.fout.tell()) tensor.tofile(self.fout) - - pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes - if pad != 0: - self.fout.write(bytes([0] * pad)) + self.write_padding(self.fout, tensor.nbytes) def write_tensors_to_file(self): self.write_ti_data_to_file() - pad = GGUFWriter.ggml_pad(self.fout.tell(), self.data_alignment) - self.fout.tell() - if pad != 0: - self.fout.write(bytes([0] * pad)) + self.write_padding(self.fout, self.fout.tell()) - if not self.use_temp_file: + if self.temp_file is None: for (currtensor, currpad) in self.tensors: currtensor.tofile(self.fout) if currpad != 0: @@ -654,10 +689,6 @@ class GGUFWriter: self.add_bool( KEY_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use) - def add_tensor_data_layout(self, layout: str): - self.add_string( - KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout) - def add_head_count(self, count: int): self.add_uint32( KEY_ATTENTION_HEAD_COUNT.format(arch=self.arch), count) @@ -695,16 +726,16 @@ class GGUFWriter: def add_tokenizer_model(self, model: str): self.add_string(KEY_TOKENIZER_MODEL, model) - def add_token_list(self, tokens: List): + def add_token_list(self, tokens: Union[Sequence[str], Sequence[bytes], Sequence[bytearray]]): self.add_array(KEY_TOKENIZER_LIST, tokens) - def add_token_merges(self, merges: List): + def add_token_merges(self, merges: Union[Sequence[str], Sequence[bytes], Sequence[bytearray]]): self.add_array(KEY_TOKENIZER_MERGES, merges) - def add_token_types(self, types: List[int]): + def add_token_types(self, types: Union[Sequence[TokenType], Sequence[int]]): self.add_array(KEY_TOKENIZER_TOKEN_TYPE, types) - def add_token_scores(self, scores: List[float]): + def add_token_scores(self, scores: Sequence[float]): self.add_array(KEY_TOKENIZER_SCORES, scores) def add_bos_token_id(self, id: int): @@ -723,6 +754,84 @@ class GGUFWriter: self.add_uint32(KEY_TOKENIZER_PAD_ID, id) +class SpecialVocab: + load_merges: bool = False + merges: List[str] = [] + special_token_types: Tuple[str, ...] = tuple(('bos', 'eos', 'unk', 'sep', 'pad')) + special_token_ids: Dict[str, int] = {} + + def __init__(self, path: Path, load_merges: bool = False, special_token_types: Optional[Tuple[str, ...]] = None): + self.special_token_ids = {} + self.load_merges = load_merges + if special_token_types is not None: + self.special_token_types = special_token_types + self.load(path) + + def load(self, path: Path): + if not self.try_load_from_tokenizer_json(path): + self.try_load_from_config_json(path) + + def try_load_from_tokenizer_json(self, path: Path) -> bool: + tokenizer_file = path / 'tokenizer.json' + if not tokenizer_file.is_file(): + return False + with open(tokenizer_file, 'r', encoding = 'utf-8') as f: + tokenizer = json.load(f) + if self.load_merges: + merges = tokenizer.get('model', {}).get('merges') + if isinstance(merges, list) and len(merges) > 0 and isinstance(merges[0], str): + self.merges = merges + tokenizer_config_file = path / 'tokenizer_config.json' + added_tokens = tokenizer.get('added_tokens') + if added_tokens is None or not tokenizer_config_file.is_file(): + return True + with open(tokenizer_config_file, 'r', encoding = 'utf-8') as f: + tokenizer_config = json.load(f) + for typ in self.special_token_types: + entry = tokenizer_config.get(f'{typ}_token') + if isinstance(entry, str): + tc_content = entry + elif isinstance(entry, dict): + entry_content = entry.get('content') + if not isinstance(entry_content, str): + continue + tc_content = entry_content + else: + continue + for maybe_token_id in (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content): + if isinstance(maybe_token_id, int): + self.special_token_ids[typ] = maybe_token_id + break + return True + + def try_load_from_config_json(self, path: Path) -> bool: + config_file = path / 'config.json' + if not config_file.is_file(): + return False + with open(config_file, 'r', encoding = 'utf-8') as f: + config = json.load(f) + for typ in self.special_token_types: + maybe_token_id = config.get(f'{typ}_token_id') + if isinstance(maybe_token_id, int): + self.special_token_ids[typ] = maybe_token_id + return True + + def add_to_gguf(self, gw: GGUFWriter): + if len(self.merges) > 0: + print(f'gguf: Adding {len(self.merges)} merge(s).') + gw.add_token_merges(self.merges) + for typ, tokid in self.special_token_ids.items(): + handler: Optional[Callable[[int], None]] = getattr(gw, f'add_{typ}_token_id', None) + if handler is None: + print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping') + continue + print(f'gguf: Setting special token type {typ} to {tokid}') + handler(tokid) + + def __repr__(self): + return f'' + + # Example usage: if __name__ == "__main__": # Example usage with a file diff --git a/gguf-py/gguf/py.typed b/gguf-py/gguf/py.typed new file mode 100644 index 000000000..e69de29bb diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index cc70e28b7..c66b069f9 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -5,6 +5,7 @@ description = "Write ML models in GGUF for GGML" authors = ["GGML "] packages = [ {include = "gguf"}, + {include = "gguf/py.typed"}, ] readme = "README.md" homepage = "https://ggml.ai" From 35092fb54712d032860f3976a6fc1ae1f84a4a28 Mon Sep 17 00:00:00 2001 From: Gilad S Date: Wed, 30 Aug 2023 11:40:12 +0300 Subject: [PATCH 100/147] docs : add `node-llama-cpp` to `README.md` (#2885) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index a880fd29f..d727b0554 100644 --- a/README.md +++ b/README.md @@ -107,7 +107,7 @@ as the main playground for developing new features for the [ggml](https://github - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python) - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp) -- Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node) +- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp), [hlhr202/llama-node](https://github.com/hlhr202/llama-node) - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb) - Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp) - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp) From 950929442070874d45561d2a4b68b010457767de Mon Sep 17 00:00:00 2001 From: alonfaraj Date: Wed, 30 Aug 2023 12:42:51 +0300 Subject: [PATCH 101/147] make : add test and update CI (#2897) * build ci: run make test * makefile: - add all - add test * enable tests/test-tokenizer-0-llama * fix path to model * remove gcc-8 from macos build test * Update Makefile * Update Makefile --- .github/workflows/build.yml | 12 ++++++++++++ Makefile | 17 +++++++++++++++++ 2 files changed, 29 insertions(+) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 10320ad1f..20fd8c2b5 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -41,6 +41,12 @@ jobs: run: | CC=gcc-8 make + - name: Test + id: make_test + run: | + CC=gcc-8 make tests + make test + ubuntu-latest-cmake: runs-on: ubuntu-latest @@ -157,6 +163,12 @@ jobs: run: | make + - name: Test + id: make_test + run: | + make tests + make test + macOS-latest-cmake: runs-on: macos-latest diff --git a/Makefile b/Makefile index c8b8a92d7..bd2d92869 100644 --- a/Makefile +++ b/Makefile @@ -6,6 +6,23 @@ TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-dou default: $(BUILD_TARGETS) +test: + @echo "Running tests..." + @for test_target in $(TEST_TARGETS); do \ + if [ "$$test_target" = "tests/test-tokenizer-0-llama" ]; then \ + ./$$test_target $(CURDIR)/models/ggml-vocab-llama.gguf; \ + elif [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \ + continue; \ + elif [ "$$test_target" = "tests/test-tokenizer-1" ]; then \ + continue; \ + else \ + ./$$test_target; \ + fi; \ + done + @echo "All tests have been run." + +all: $(BUILD_TARGETS) $(TEST_TARGETS) + ifndef UNAME_S UNAME_S := $(shell uname -s) endif From 0d1c706181cd31e7f368dd14eeb16c1a2569e4df Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=2E=20Yusuf=20Sar=C4=B1g=C3=B6z?= Date: Wed, 30 Aug 2023 12:47:40 +0300 Subject: [PATCH 102/147] gguf : add workflow for Pypi publishing (#2896) * gguf : add workflow for Pypi publishing * gguf : add workflow for Pypi publishing * fix trailing whitespace --- .github/workflows/gguf-publish.yml | 43 ++++++++++++++++++++++++++++++ gguf-py/README.md | 23 +++++++++++++--- 2 files changed, 63 insertions(+), 3 deletions(-) create mode 100644 .github/workflows/gguf-publish.yml diff --git a/.github/workflows/gguf-publish.yml b/.github/workflows/gguf-publish.yml new file mode 100644 index 000000000..a6289e335 --- /dev/null +++ b/.github/workflows/gguf-publish.yml @@ -0,0 +1,43 @@ +# This workflow will upload a Python Package using Twine when a GGUF release is created +# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries + +# See `gguf-py/README.md` for how to make a release. + +# This workflow uses actions that are not certified by GitHub. +# They are provided by a third-party and are governed by +# separate terms of service, privacy policy, and support +# documentation. + +name: Upload Python Package + +on: + workflow_dispatch: + push: + # Pattern matched against refs/tags + tags: + - 'gguf-v*' # Push events to every version tag + + +jobs: + deploy: + + runs-on: ubuntu-latest + + steps: + - uses: actions/checkout@v2 + - name: Set up Python + uses: actions/setup-python@v2 + with: + python-version: '3.9.x' + - name: Install dependencies + run: | + cd gguf-py + python -m pip install poetry + poetry install + + - name: Build package + run: poetry build + - name: Publish package + uses: pypa/gh-action-pypi-publish@release/v1 + with: + password: ${{ secrets.PYPI_API_TOKEN }} diff --git a/gguf-py/README.md b/gguf-py/README.md index 03ad306ec..ffe25c495 100644 --- a/gguf-py/README.md +++ b/gguf-py/README.md @@ -27,8 +27,25 @@ In this case, upgrade Pip to the latest: pip install --upgrade pip ``` -## Publishing -To publish the package, you need to have `twine` and `build` installed: +## Automatic publishing with CI + +There's a GitHub workflow to make a release automatically upon creation of tags in a specified format. + +1. Bump the version in `pyproject.toml`. +2. Create a tag named `gguf-vx.x.x` where `x.x.x` is the semantic version number. + +```sh +git tag -a gguf-v1.0.0 -m "Version 1.0 release" +``` + +3. Push the tags. + +```sh +git push origin --tags +``` + +## Manual publishing +If you want to publish the package manually for any reason, you need to have `twine` and `build` installed: ```sh pip install build twine @@ -36,7 +53,7 @@ pip install build twine Then, folow these steps to release a new version: -1. Update the version in `pyproject.toml`. +1. Bump the version in `pyproject.toml`. 2. Build the package: ```sh From c90d135eb433cf0d40fb95e46a48d1391d2352b5 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 30 Aug 2023 12:52:46 +0300 Subject: [PATCH 103/147] examples : fix underscore in beam-search + .gitignore (close #2900) --- .gitignore | 3 +++ Makefile | 4 ++-- examples/CMakeLists.txt | 2 +- examples/{beam_search => beam-search}/CMakeLists.txt | 4 ++-- .../beam_search.cpp => beam-search/beam-search.cpp} | 0 5 files changed, 8 insertions(+), 5 deletions(-) rename examples/{beam_search => beam-search}/CMakeLists.txt (78%) rename examples/{beam_search/beam_search.cpp => beam-search/beam-search.cpp} (100%) diff --git a/.gitignore b/.gitignore index 54ea2b522..8b5f45a2d 100644 --- a/.gitignore +++ b/.gitignore @@ -42,6 +42,9 @@ models-mnt /gguf-llama-simple /libllama.so /llama-bench +/baby-llama +/beam-search +/save-load-state build-info.h arm_neon.h compile_commands.json diff --git a/Makefile b/Makefile index bd2d92869..b750540fe 100644 --- a/Makefile +++ b/Makefile @@ -1,5 +1,5 @@ # Define the default target now so that it is always the first target -BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam_search tests/test-c.o +BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch convert-llama2c-to-ggml simple save-load-state server embd-input-test gguf llama-bench baby-llama beam-search tests/test-c.o # Binaries only useful for tests TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1 @@ -446,7 +446,7 @@ llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o co baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -beam_search: examples/beam_search/beam_search.cpp build-info.h ggml.o llama.o common.o $(OBJS) +beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))' diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 94b785224..6e65eb087 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -25,7 +25,7 @@ else() add_subdirectory(simple) add_subdirectory(embd-input) add_subdirectory(llama-bench) - add_subdirectory(beam_search) + add_subdirectory(beam-search) if (LLAMA_METAL) add_subdirectory(metal) endif() diff --git a/examples/beam_search/CMakeLists.txt b/examples/beam-search/CMakeLists.txt similarity index 78% rename from examples/beam_search/CMakeLists.txt rename to examples/beam-search/CMakeLists.txt index b29e01092..e44a74975 100644 --- a/examples/beam_search/CMakeLists.txt +++ b/examples/beam-search/CMakeLists.txt @@ -1,5 +1,5 @@ -set(TARGET beam_search) -add_executable(${TARGET} beam_search.cpp) +set(TARGET beam-search) +add_executable(${TARGET} beam-search.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/beam_search/beam_search.cpp b/examples/beam-search/beam-search.cpp similarity index 100% rename from examples/beam_search/beam_search.cpp rename to examples/beam-search/beam-search.cpp From b532a69b2fd08067f34f32f37a2fd9b37678a34a Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 30 Aug 2023 13:29:40 +0300 Subject: [PATCH 104/147] convert.py : use dir name to name the llama --- convert.py | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/convert.py b/convert.py index 448b6f0f3..a7f4c2d75 100755 --- a/convert.py +++ b/convert.py @@ -811,10 +811,12 @@ class OutputFile: def add_meta_arch(self, params: Params) -> None: name = "LLaMA" + + # TODO: better logic to determine model name if (params.n_ctx == 4096): name = "LLaMA v2" - if params.path_model: - name = str(params.path_model.parent).split('/')[-1] + elif params.path_model: + name = str(params.path_model.parent).split('/')[-1] self.gguf.add_name (name) self.gguf.add_context_length (params.n_ctx) @@ -839,8 +841,7 @@ class OutputFile: tokens = [] scores = [] toktypes = [] - # NOTE: `all_tokens` returns the the base vocabulary and added tokens - # TODO: add special tokens? + # NOTE: `all_tokens` returns the base vocabulary and added tokens for text, score, toktype in vocab.all_tokens(): tokens.append(text) scores.append(score) From 71d6975559acfd6c8407a4ef8275a9979c737765 Mon Sep 17 00:00:00 2001 From: Henri Vasserman Date: Wed, 30 Aug 2023 19:14:53 +0300 Subject: [PATCH 105/147] [Docker] fix tools.sh argument passing. (#2884) * [Docker] fix tools.sh argument passing. This should allow passing multiple arguments to containers with the full image that are using the tools.sh frontend. Fix from https://github.com/ggerganov/llama.cpp/issues/2535#issuecomment-1697091734 --- .devops/tools.sh | 11 ++++------- 1 file changed, 4 insertions(+), 7 deletions(-) diff --git a/.devops/tools.sh b/.devops/tools.sh index 2787c21fe..9d999315f 100755 --- a/.devops/tools.sh +++ b/.devops/tools.sh @@ -7,15 +7,12 @@ arg1="$1" # Shift the arguments to remove the first one shift -# Join the remaining arguments into a single string -arg2="$@" - if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then - python3 ./convert.py "$arg2" + python3 ./convert.py "$@" elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then - ./quantize "$arg2" + ./quantize "$@" elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then - ./main "$arg2" + ./main "$@" elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then echo "Converting PTH to GGML..." for i in `ls $1/$2/ggml-model-f16.bin*`; do @@ -27,7 +24,7 @@ elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then fi done elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then - ./server "$arg2" + ./server "$@" else echo "Unknown command: $arg1" echo "Available commands: " From 8afe2280009ecbfc9de2c93b8f41283dc810609a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Wed, 30 Aug 2023 21:46:19 +0200 Subject: [PATCH 106/147] CUDA: mul_mat_q=true llama_context_params default (#2912) --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index fcd6f276a..95ee6ffe4 100644 --- a/llama.cpp +++ b/llama.cpp @@ -5287,7 +5287,7 @@ struct llama_context_params llama_context_default_params() { /*.progress_callback =*/ nullptr, /*.progress_callback_user_data =*/ nullptr, /*.low_vram =*/ false, - /*.mul_mat_q =*/ false, + /*.mul_mat_q =*/ true, /*.f16_kv =*/ true, /*.logits_all =*/ false, /*.vocab_only =*/ false, From 92d0b751a77a089e650983e9f1564ef4d31b32b9 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Thu, 31 Aug 2023 01:02:23 -0400 Subject: [PATCH 107/147] convert : fix python 3.8 support, modernize type annotations (#2916) * convert : fix python 3.8 support * convert : sort imports * convert : fix required parameters in convert-llama-ggmlv3-to-gguf * convert : fix mypy errors in convert-llama-ggmlv3-to-gguf * convert : use PEP 585 generics and PEP 604 unions Now that we have `from __future__ import annotations`, we can use this modern syntax in Python 3.7 instead of restricting support to Python 3.9 or 3.10 respectively. * gguf.py : a tuple is already a tuple * add mypy.ini * convert : add necessary `type: ignore` comments * gguf-py: bump version --- convert-falcon-hf-to-gguf.py | 25 +++--- convert-gptneox-hf-to-gguf.py | 22 ++--- convert-llama-7b-pth-to-gguf.py | 31 ++++--- convert-llama-ggmlv3-to-gguf.py | 18 ++-- convert-llama-hf-to-gguf.py | 33 +++---- convert-lora-to-ggml.py | 8 +- convert.py | 149 ++++++++++++++++---------------- gguf-py/gguf/gguf.py | 68 ++++++++------- gguf-py/pyproject.toml | 2 +- mypy.ini | 5 ++ 10 files changed, 193 insertions(+), 168 deletions(-) create mode 100644 mypy.ini diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py index 0fdea70e1..ec786ff67 100755 --- a/convert-falcon-hf-to-gguf.py +++ b/convert-falcon-hf-to-gguf.py @@ -1,18 +1,21 @@ #!/usr/bin/env python3 # HF falcon--> gguf conversion -import gguf -import os -import sys -import struct +from __future__ import annotations + +import argparse import json +import os +import struct +import sys +from pathlib import Path +from typing import Any + +import gguf import numpy as np import torch -import argparse +from transformers import AutoTokenizer # type: ignore[import] -from typing import Any, List -from pathlib import Path -from transformers import AutoTokenizer def bytes_to_unicode(): # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py @@ -114,9 +117,9 @@ gguf_writer.add_file_type(ftype) print("gguf: get tokenizer metadata") -tokens: List[bytearray] = [] -scores: List[float] = [] -toktypes: List[int] = [] +tokens: list[bytearray] = [] +scores: list[float] = [] +toktypes: list[int] = [] tokenizer_json_file = dir_model / 'tokenizer.json' if not tokenizer_json_file.is_file(): diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py index 38e71e03b..852123d99 100755 --- a/convert-gptneox-hf-to-gguf.py +++ b/convert-gptneox-hf-to-gguf.py @@ -1,18 +1,20 @@ #!/usr/bin/env python3 # HF gptneox--> gguf conversion -import gguf -import os -import sys -import struct +from __future__ import annotations + +import argparse import json +import os +import struct +import sys +from pathlib import Path +from typing import Any + +import gguf import numpy as np import torch -import argparse - -from typing import Any, List -from pathlib import Path -from transformers import AutoTokenizer +from transformers import AutoTokenizer # type: ignore[import] # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py @@ -112,7 +114,7 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"]) print("gguf: get tokenizer metadata") -tokens: List[bytearray] = [] +tokens: list[bytearray] = [] tokenizer_json_file = dir_model / 'tokenizer.json' if not tokenizer_json_file.is_file(): diff --git a/convert-llama-7b-pth-to-gguf.py b/convert-llama-7b-pth-to-gguf.py index 6e973a116..6574c11dd 100755 --- a/convert-llama-7b-pth-to-gguf.py +++ b/convert-llama-7b-pth-to-gguf.py @@ -3,22 +3,25 @@ # Only models with a single datafile are supported, like 7B # HF files required in the model dir: config.json tokenizer_config.json tokenizer.json tokenizer.model -import gguf -import os -import sys -import struct +from __future__ import annotations + +import argparse import json +import os +import struct +import sys +from pathlib import Path +from typing import TYPE_CHECKING, Any + +import gguf import numpy as np import torch -import argparse +from sentencepiece import SentencePieceProcessor # type: ignore[import] -from typing import Any, List, TypeAlias -from pathlib import Path -from sentencepiece import SentencePieceProcessor +if TYPE_CHECKING: + from typing import TypeAlias -#NDArray = np.ndarray[Any, Any] -# compatible with python < 3.9 -NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]' +NDArray: TypeAlias = 'np.ndarray[Any, Any]' def count_model_parts(dir_model: Path) -> int: @@ -129,9 +132,9 @@ if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in print("gguf: get tokenizer metadata") -tokens: List[bytes] = [] -scores: List[float] = [] -toktypes: List[int] = [] +tokens: list[bytes] = [] +scores: list[float] = [] +toktypes: list[int] = [] tokenizer_model_file = dir_model / 'tokenizer.model' if not tokenizer_model_file.is_file(): diff --git a/convert-llama-ggmlv3-to-gguf.py b/convert-llama-ggmlv3-to-gguf.py index c8e7f1761..3f39bc39e 100755 --- a/convert-llama-ggmlv3-to-gguf.py +++ b/convert-llama-ggmlv3-to-gguf.py @@ -1,10 +1,14 @@ #!/usr/bin/env python3 -import sys, struct, math, argparse +from __future__ import annotations + +import argparse +import math +import struct +import sys from pathlib import Path -import numpy as np - import gguf +import numpy as np # Note: Does not support GGML_QKK_64 QK_K = 256 @@ -72,7 +76,7 @@ class Vocab: class Tensor: def __init__(self): self.name = None - self.dims = () + self.dims: tuple[int, ...] = () self.dtype = None self.start_offset = 0 self.len_bytes = np.int64(0) @@ -119,7 +123,7 @@ class GGMLV3Model: offset += hp.load(data, offset) vocab = Vocab() offset += vocab.load(data, offset, hp.n_vocab) - tensors = [] + tensors: list[Tensor] = [] tensor_map = {} while offset < len(data): tensor = Tensor() @@ -305,8 +309,8 @@ def handle_metadata(cfg, hp): def handle_args(): parser = argparse.ArgumentParser(description = 'Convert GGMLv3 models to GGUF') - parser.add_argument('--input', '-i', type = Path, help = 'Input GGMLv3 filename') - parser.add_argument('--output', '-o', type = Path, help ='Output GGUF filename') + parser.add_argument('--input', '-i', type = Path, required = True, help = 'Input GGMLv3 filename') + parser.add_argument('--output', '-o', type = Path, required = True, help ='Output GGUF filename') parser.add_argument('--name', help = 'Set model name') parser.add_argument('--desc', help = 'Set model description') parser.add_argument('--gqa', type = int, default = 1, help = 'grouped-query attention factor (use 8 for LLaMA2 70B)') diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py index ab94b5eab..c453c83c3 100755 --- a/convert-llama-hf-to-gguf.py +++ b/convert-llama-hf-to-gguf.py @@ -1,28 +1,31 @@ #!/usr/bin/env python3 # HF llama --> gguf conversion -import gguf -import os -import sys -import struct +from __future__ import annotations + +import argparse import json +import os +import struct +import sys +from pathlib import Path +from typing import TYPE_CHECKING, Any + +import gguf import numpy as np import torch -import argparse +from sentencepiece import SentencePieceProcessor # type: ignore[import] -from typing import Any, List, Optional, TypeAlias -from pathlib import Path -from sentencepiece import SentencePieceProcessor +if TYPE_CHECKING: + from typing import TypeAlias -#NDArray = np.ndarray[Any, Any] -# compatible with python < 3.9 -NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]' +NDArray: TypeAlias = 'np.ndarray[Any, Any]' # reverse HF permute back to original pth layout # https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py -def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: Optional[int] = None) -> NDArray: +def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray: if n_kv_head is not None and n_head != n_kv_head: n_head //= n_kv_head @@ -136,9 +139,9 @@ if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in print("gguf: get tokenizer metadata") -tokens: List[bytes] = [] -scores: List[float] = [] -toktypes: List[int] = [] +tokens: list[bytes] = [] +scores: list[float] = [] +toktypes: list[int] = [] tokenizer_model_file = dir_model / 'tokenizer.model' if not tokenizer_model_file.is_file(): diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index a00339b47..a937410dd 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -1,15 +1,17 @@ #!/usr/bin/env python3 +from __future__ import annotations + import json import os import re import struct import sys -from typing import Any, Dict, Sequence, BinaryIO +from typing import Any, BinaryIO, Sequence import numpy as np import torch -NUMPY_TYPE_TO_FTYPE: Dict[str, int] = {"float32": 0, "float16": 1} +NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1} HF_SUBLAYER_TO_GGML = { @@ -46,7 +48,7 @@ def translate_tensor_name(t: str) -> str: sys.exit(1) -def write_file_header(fout: BinaryIO, params: Dict[str, Any]) -> None: +def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None: fout.write(b"ggla"[::-1]) # magic (ggml lora) fout.write(struct.pack("i", 1)) # file version fout.write(struct.pack("i", params["r"])) diff --git a/convert.py b/convert.py index a7f4c2d75..9a39edb99 100755 --- a/convert.py +++ b/convert.py @@ -1,9 +1,8 @@ #!/usr/bin/env python3 +from __future__ import annotations -import gguf import argparse import concurrent.futures -from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor import copy import enum import faulthandler @@ -20,21 +19,23 @@ import struct import sys import time import zipfile -import numpy as np - from abc import ABCMeta, abstractmethod +from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor from dataclasses import dataclass from pathlib import Path -from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Generator, Iterable, List, Literal, Optional, Sequence, Set, Tuple, Type, TypeVar, Union) -from sentencepiece import SentencePieceProcessor # type: ignore +from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar + +import gguf +import numpy as np +from sentencepiece import SentencePieceProcessor # type: ignore[import] if TYPE_CHECKING: - from typing_extensions import TypeAlias + from typing import TypeAlias if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'): faulthandler.register(signal.SIGUSR1) -NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]' +NDArray: TypeAlias = 'np.ndarray[Any, Any]' ARCH=gguf.MODEL_ARCH.LLAMA NAMES=gguf.MODEL_TENSOR_NAMES[ARCH] @@ -47,8 +48,8 @@ DEFAULT_CONCURRENCY = 8 @dataclass(frozen=True) class DataType: name: str - dtype: 'np.dtype[Any]' - valid_conversions: List[str] + dtype: np.dtype[Any] + valid_conversions: list[str] def elements_to_bytes(self, n_elements: int) -> int: return n_elements * self.dtype.itemsize @@ -65,7 +66,7 @@ DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_convers @dataclass(frozen=True) class QuantizedDataType(DataType): block_size: int - quantized_dtype: 'np.dtype[Any]' + quantized_dtype: np.dtype[Any] ggml_type: gguf.GGMLQuantizationType def quantize(self, arr: NDArray) -> NDArray: @@ -84,7 +85,7 @@ class Q8_0QuantizedDataType(QuantizedDataType): n_blocks = arr.size // self.block_size blocks = arr.reshape((n_blocks, self.block_size)) # Much faster implementation of block quantization contributed by @Cebtenzzre - def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[Tuple[Any, Any]]: + def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]: d = abs(blocks).max(axis = 1) / np.float32(127) with np.errstate(divide = 'ignore'): qs = (blocks / d[:, None]).round() @@ -98,13 +99,13 @@ DT_Q8_0 = Q8_0QuantizedDataType('Q8_0', quantized_dtype = np.dtype([('d', ' DataType: + def type_for_tensor(self, name: str, tensor: LazyTensor) -> DataType: dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self) if dt is None: raise ValueError(self) # 1D tensors are always F32. return dt if len(tensor.shape) > 1 else DT_F32 -GGML_FILE_TYPE_TO_DATA_TYPE: Dict[GGMLFileType, DataType] = { +GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = { GGMLFileType.AllF32 : DT_F32, GGMLFileType.MostlyF16 : DT_F16, GGMLFileType.MostlyQ8_0: DT_Q8_0, @@ -148,13 +149,13 @@ class Params: n_head_kv: int f_norm_eps: float - f_rope_freq_base: Optional[float] = None - f_rope_scale: Optional[float] = None + f_rope_freq_base: float | None = None + f_rope_scale: float | None = None - ftype: Optional[GGMLFileType] = None + ftype: GGMLFileType | None = None # path to the directory containing the model files - path_model: Optional['Path'] = None + path_model: Path | None = None @staticmethod def find_n_mult(n_ff: int, n_embd: int) -> int: @@ -166,7 +167,7 @@ class Params: raise Exception(f"failed to find n_mult for (n_ff={n_ff}, n_embd={n_embd}).") @staticmethod - def guessed(model: 'LazyModel') -> 'Params': + def guessed(model: LazyModel) -> Params: # try transformer naming first n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape @@ -202,7 +203,7 @@ class Params: ) @staticmethod - def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params': + def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) n_vocab = config["vocab_size"] @@ -247,7 +248,7 @@ class Params: # LLaMA v2 70B params.json # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1 @staticmethod - def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params': + def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) n_vocab = config["vocab_size"] if "vocab_size" in config else -1 @@ -291,7 +292,7 @@ class Params: ) @staticmethod - def load(model_plus: 'ModelPlus') -> 'Params': + def load(model_plus: ModelPlus) -> Params: hf_config_path = model_plus.paths[0].parent / "config.json" orig_config_path = model_plus.paths[0].parent / "params.json" @@ -314,9 +315,9 @@ class Params: # class BpeVocab: - def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None: self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read()) - added_tokens: Dict[str, int] + added_tokens: dict[str, int] if fname_added_tokens is not None: added_tokens = json.load(open(fname_added_tokens, encoding="utf-8")) else: @@ -335,9 +336,9 @@ class BpeVocab: self.fname_tokenizer = fname_tokenizer self.fname_added_tokens = fname_added_tokens - def bpe_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: tokenizer = self.bpe_tokenizer - from transformers.models.gpt2 import tokenization_gpt2 + from transformers.models.gpt2 import tokenization_gpt2 # type: ignore[import] byte_encoder = tokenization_gpt2.bytes_to_unicode() byte_decoder = {v: k for k, v in byte_encoder.items()} for i, item in enumerate(tokenizer): @@ -345,12 +346,12 @@ class BpeVocab: score: float = -i yield text, score, gguf.TokenType.USER_DEFINED - def added_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: for text in self.added_tokens_list: score = -1000.0 yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED - def all_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: yield from self.bpe_tokens() yield from self.added_tokens() @@ -359,9 +360,9 @@ class BpeVocab: class SentencePieceVocab: - def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None: self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer)) - added_tokens: Dict[str, int] + added_tokens: dict[str, int] if fname_added_tokens is not None: added_tokens = json.load(open(fname_added_tokens, encoding="utf-8")) else: @@ -380,7 +381,7 @@ class SentencePieceVocab: self.fname_tokenizer = fname_tokenizer self.fname_added_tokens = fname_added_tokens - def sentencepiece_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: tokenizer = self.sentencepiece_tokenizer for i in range(tokenizer.vocab_size()): piece = tokenizer.id_to_piece(i) @@ -404,19 +405,19 @@ class SentencePieceVocab: yield text, score, toktype - def added_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: for text in self.added_tokens_list: score = -1000.0 yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED - def all_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: yield from self.sentencepiece_tokens() yield from self.added_tokens() def __repr__(self) -> str: return f"" -Vocab = Union[BpeVocab, SentencePieceVocab] +Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab' # # data loading @@ -436,15 +437,15 @@ class Tensor(metaclass=ABCMeta): data_type: DataType @abstractmethod - def astype(self, data_type: DataType) -> 'Tensor': ... + def astype(self, data_type: DataType) -> Tensor: ... @abstractmethod - def permute(self, n_head: int, n_head_kv: int) -> 'Tensor': ... + def permute(self, n_head: int, n_head_kv: int) -> Tensor: ... @abstractmethod - def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> 'UnquantizedTensor': ... + def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor: ... @abstractmethod - def part(self, n_part: int) -> 'UnquantizedTensor': ... + def part(self, n_part: int) -> UnquantizedTensor: ... @abstractmethod - def to_ggml(self) -> 'GGMLCompatibleTensor': ... + def to_ggml(self) -> GGMLCompatibleTensor: ... def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray: @@ -465,22 +466,22 @@ class UnquantizedTensor(Tensor): self.ndarray = bf16_to_fp32(self.ndarray) return UnquantizedTensor(self.ndarray.astype(dtype)) - def to_ggml(self) -> 'UnquantizedTensor': + def to_ggml(self) -> UnquantizedTensor: return self - def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> 'UnquantizedTensor': + def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor: r = self.ndarray.shape[0] // 3 return UnquantizedTensor(permute(self.ndarray[r * n_part : r * n_part + r, ...], n_head, n_head_kv)) - def part(self, n_part: int) -> 'UnquantizedTensor': + def part(self, n_part: int) -> UnquantizedTensor: r = self.ndarray.shape[0] // 3 return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...]) - def permute(self, n_head: int, n_head_kv: int) -> 'UnquantizedTensor': + def permute(self, n_head: int, n_head_kv: int) -> UnquantizedTensor: return UnquantizedTensor(permute(self.ndarray, n_head, n_head_kv)) -def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, convert: bool = False) -> NDArray: +def load_unquantized(lazy_tensor: LazyTensor, expected_dtype: Any = None, convert: bool = False) -> NDArray: tensor = lazy_tensor.load() assert isinstance(tensor, UnquantizedTensor) @@ -496,13 +497,13 @@ def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, conv return tensor.ndarray -GGMLCompatibleTensor = Union[UnquantizedTensor] +GGMLCompatibleTensor = UnquantizedTensor @dataclass class LazyTensor: _load: Callable[[], Tensor] - shape: List[int] + shape: list[int] data_type: DataType description: str @@ -513,7 +514,7 @@ class LazyTensor: (self.data_type, ret.data_type, self.description) return ret - def astype(self, data_type: DataType) -> 'LazyTensor': + def astype(self, data_type: DataType) -> LazyTensor: self.validate_conversion_to(data_type) def load() -> Tensor: @@ -525,24 +526,24 @@ class LazyTensor: raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.') -LazyModel = Dict[str, LazyTensor] +LazyModel = dict[str, LazyTensor] @dataclass class ModelPlus: model: LazyModel - paths: List[Path] # Where this was read from. + paths: list[Path] # Where this was read from. format: Literal['ggml', 'torch', 'safetensors', 'none'] - vocab: Optional[Vocab] # For GGML models (which have vocab built in), the vocab. + vocab: Vocab | None # For GGML models (which have vocab built in), the vocab. -def merge_sharded(models: List[LazyModel]) -> LazyModel: +def merge_sharded(models: list[LazyModel]) -> LazyModel: # Original LLaMA models have each file contain one part of each tensor. # Use a dict instead of a set to preserve order. names = {name: None for model in models for name in model} def convert(name: str) -> LazyTensor: - lazy_tensors: List[LazyTensor] = [model[name] for model in models] + lazy_tensors: list[LazyTensor] = [model[name] for model in models] if len(lazy_tensors) == 1: # only one file; don't go through this procedure since there might # be quantized tensors @@ -570,7 +571,7 @@ def merge_sharded(models: List[LazyModel]) -> LazyModel: return {name: convert(name) for name in names} -def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus: +def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus: formats = set(mp.format for mp in models_plus) assert len(formats) == 1, "different formats?" format = formats.pop() @@ -674,7 +675,7 @@ class LazyUnpickler(pickle.Unpickler): def rebuild_from_type_v2(func, new_type, args, state): return func(*args) - CLASSES: Dict[Tuple[str, str], Any] = { + CLASSES: dict[tuple[str, str], Any] = { # getattr used here as a workaround for mypy not being smart enough to detrmine # the staticmethods have a __func__ attribute. ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'), @@ -707,15 +708,15 @@ def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus: def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus: header_size, = struct.unpack(' LazyTensor: + def convert(info: dict[str, Any]) -> LazyTensor: data_type = SAFETENSORS_DATA_TYPES[info['dtype']] numpy_dtype = data_type.dtype - shape: List[int] = info['shape'] + shape: list[int] = info['shape'] begin, end = info['data_offsets'] assert 0 <= begin <= end <= len(byte_buf) assert end - begin == math.prod(shape) * numpy_dtype.itemsize @@ -754,7 +755,7 @@ def lazy_load_file(path: Path) -> ModelPlus: In = TypeVar('In') Out = TypeVar('Out') -def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: Optional[int] = None, use_processpool_executor: bool = False) -> Iterable[Out]: +def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: int | None = None, use_processpool_executor: bool = False) -> Iterable[Out]: '''Parallel map, but with backpressure. If the caller doesn't call `next` fast enough, this will stop calling `func` at some point rather than letting results pile up in memory. Specifically, there is a max of one @@ -763,13 +764,13 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc yield from map(func, iterable) # Not reached. iterable = iter(iterable) - executor_class: Union[Type[ThreadPoolExecutor], Type[ProcessPoolExecutor]] + executor_class: type[ThreadPoolExecutor] | type[ProcessPoolExecutor] if use_processpool_executor: executor_class = ProcessPoolExecutor else: executor_class = ThreadPoolExecutor with executor_class(max_workers = max_workers) as executor: - futures: List[concurrent.futures.Future[Out]] = [] + futures: list[concurrent.futures.Future[Out]] = [] done = False for _ in range(concurrency): try: @@ -893,13 +894,13 @@ class OutputFile: of.close() @staticmethod - def do_item(item: Tuple[str, LazyTensor]) -> Tuple[DataType, NDArray]: + def do_item(item: tuple[str, LazyTensor]) -> tuple[DataType, NDArray]: name, lazy_tensor = item tensor = lazy_tensor.load().to_ggml() return (lazy_tensor.data_type, tensor.ndarray) @staticmethod - def maybe_do_quantize(item: Tuple[DataType, NDArray]) -> NDArray: + def maybe_do_quantize(item: tuple[DataType, NDArray]) -> NDArray: dt, arr = item if not isinstance(dt, QuantizedDataType): return arr @@ -940,7 +941,7 @@ class OutputFile: of.close() -def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFileType: +def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType: wq_type = model[NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32): @@ -960,7 +961,7 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM def convert_model_names(model: LazyModel, params: Params) -> LazyModel: tmap = gguf.TensorNameMap(ARCH, params.n_layer) - should_skip: Set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, [])) + should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, [])) tmp = model @@ -995,12 +996,12 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel: return out -def nth_multifile_path(path: Path, n: int) -> Optional[Path]: +def nth_multifile_path(path: Path, n: int) -> Path | None: '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return the nth path in the model. ''' # Support the following patterns: - patterns: List[Tuple[str, str]] = [ + patterns: list[tuple[str, str]] = [ # - x.00.pth, x.01.pth, etc. (r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'), # - x-00001-of-00002.bin, x-00002-of-00002.bin, etc. @@ -1016,11 +1017,11 @@ def nth_multifile_path(path: Path, n: int) -> Optional[Path]: return None -def find_multifile_paths(path: Path) -> List[Path]: +def find_multifile_paths(path: Path) -> list[Path]: '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return the whole list of paths in the model. ''' - ret: List[Path] = [] + ret: list[Path] = [] for i in itertools.count(): nth_path = nth_multifile_path(path, i) if nth_path is None: @@ -1051,7 +1052,7 @@ def load_some_model(path: Path) -> ModelPlus: path = files[0] paths = find_multifile_paths(path) - models_plus: List[ModelPlus] = [] + models_plus: list[ModelPlus] = [] for path in paths: print(f"Loading model file {path}") models_plus.append(lazy_load_file(path)) @@ -1060,7 +1061,7 @@ def load_some_model(path: Path) -> ModelPlus: return model_plus -def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, SentencePieceVocab]: +def load_vocab(path: Path, vocabtype: str | None) -> Vocab: # Be extra-friendly and accept either a file or a directory. Also, if it's # a directory, it might be the model directory, and tokenizer.model might # be in the parent of that. @@ -1091,7 +1092,7 @@ def load_vocab(path: Path, vocabtype: Optional[str]) -> Union[BpeVocab, Sentence raise ValueError(f"Unsupported vocabulary type {vocabtype}") -def default_outfile(model_paths: List[Path], file_type: GGMLFileType) -> Path: +def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path: namestr = { GGMLFileType.AllF32: "f32", GGMLFileType.MostlyF16: "f16", @@ -1114,7 +1115,7 @@ def do_dump_model(model_plus: ModelPlus) -> None: print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") -def main(args_in: Optional[List[str]] = None) -> None: +def main(args_in: list[str] | None = None) -> None: parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file") parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py index de3edbc99..b1bc4205b 100644 --- a/gguf-py/gguf/gguf.py +++ b/gguf-py/gguf/gguf.py @@ -1,16 +1,18 @@ #!/usr/bin/env python3 -import shutil -import sys -import struct -import tempfile -import numpy as np +from __future__ import annotations + import json import os -from pathlib import Path - +import shutil +import struct +import sys +import tempfile from enum import IntEnum, auto from io import BufferedWriter -from typing import Any, BinaryIO, Callable, IO, Dict, List, Optional, Sequence, Tuple, Union +from pathlib import Path +from typing import IO, Any, BinaryIO, Callable, Sequence + +import numpy as np # # constants @@ -103,7 +105,7 @@ class MODEL_TENSOR(IntEnum): FFN_NORM : int = auto() -MODEL_ARCH_NAMES: Dict[MODEL_ARCH, str] = { +MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.LLAMA: "llama", MODEL_ARCH.FALCON: "falcon", MODEL_ARCH.GPT2: "gpt2", @@ -112,7 +114,7 @@ MODEL_ARCH_NAMES: Dict[MODEL_ARCH, str] = { MODEL_ARCH.MPT: "mpt", } -MODEL_TENSOR_NAMES: Dict[MODEL_ARCH, Dict[MODEL_TENSOR, str]] = { +MODEL_TENSOR_NAMES: dict[MODEL_ARCH, dict[MODEL_TENSOR, str]] = { MODEL_ARCH.LLAMA: { MODEL_TENSOR.TOKEN_EMBD: "token_embd", MODEL_TENSOR.OUTPUT_NORM: "output_norm", @@ -158,7 +160,7 @@ MODEL_TENSOR_NAMES: Dict[MODEL_ARCH, Dict[MODEL_TENSOR, str]] = { } # tensors that will not be serialized -MODEL_TENSOR_SKIP: Dict[MODEL_ARCH, List[MODEL_TENSOR]] = { +MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_ARCH.LLAMA: [ MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, @@ -167,7 +169,7 @@ MODEL_TENSOR_SKIP: Dict[MODEL_ARCH, List[MODEL_TENSOR]] = { class TensorNameMap: - mappings_cfg: Dict[MODEL_TENSOR, Tuple[str, ...]] = { + mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { # Token embeddings MODEL_TENSOR.TOKEN_EMBD: ( "gpt_neox.embed_in", # gptneox @@ -203,7 +205,7 @@ class TensorNameMap: ), } - block_mappings_cfg: Dict[MODEL_TENSOR, Tuple[str, ...]] = { + block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { # Attention norm MODEL_TENSOR.ATTN_NORM: ( "gpt_neox.layers.{bid}.input_layernorm", # gptneox @@ -298,9 +300,9 @@ class TensorNameMap: ), } - mapping: Dict[str, Tuple[MODEL_TENSOR, str]] + mapping: dict[str, tuple[MODEL_TENSOR, str]] - tensor_names: Dict[MODEL_TENSOR, str] + tensor_names: dict[MODEL_TENSOR, str] def __init__(self, arch: MODEL_ARCH, n_blocks: int): mapping = self.mapping = {} @@ -321,7 +323,7 @@ class TensorNameMap: key = key.format(bid = bid) mapping[key] = (tensor, tensor_name) - def get_type_and_name(self, key: str, try_suffixes: Sequence[str]) -> Optional[Tuple[MODEL_TENSOR, str]]: + def get_type_and_name(self, key: str, try_suffixes: Sequence[str]) -> tuple[MODEL_TENSOR, str] | None: result = self.mapping.get(key) if result is not None: return result @@ -332,13 +334,13 @@ class TensorNameMap: return (result[0], result[1] + suffix) return None - def get_name(self, key: str, try_suffixes: Sequence[str]) -> Optional[str]: + def get_name(self, key: str, try_suffixes: Sequence[str]) -> str | None: result = self.get_type_and_name(key, try_suffixes = try_suffixes) if result is None: return None return result[1] - def get_type(self, key: str, try_suffixes: Sequence[str]) -> Optional[MODEL_TENSOR]: + def get_type(self, key: str, try_suffixes: Sequence[str]) -> MODEL_TENSOR | None: result = self.get_type_and_name(key, try_suffixes = try_suffixes) if result is None: return None @@ -432,10 +434,10 @@ class GGUFWriter: ti_data = b"" ti_data_count = 0 use_temp_file: bool - temp_file: Optional[tempfile.SpooledTemporaryFile[bytes]] = None - tensors: List[Tuple[np.ndarray[Any, Any], int]] + temp_file: tempfile.SpooledTemporaryFile[bytes] | None = None + tensors: list[tuple[np.ndarray[Any, Any], int]] - def __init__(self, path: Union[os.PathLike[str], str], arch: str, use_temp_file = True): + def __init__(self, path: os.PathLike[str] | str, arch: str, use_temp_file = True): self.fout = open(path, "wb") self.arch = arch self.add_architecture() @@ -531,7 +533,7 @@ class GGUFWriter: GGUFValueType.FLOAT64: " int: return ((x + n - 1) // n) * n - def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: Union[np.dtype[np.float16], np.dtype[np.float32]], tensor_nbytes: int, raw_dtype: Optional[GGMLQuantizationType] = None): + def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype[np.float16] | np.dtype[np.float32], tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None): assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now" encoded_name = name.encode("utf8") @@ -580,7 +582,7 @@ class GGUFWriter: self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment) self.ti_data_count += 1 - def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Optional[Sequence[int]] = None, raw_dtype: Optional[GGMLQuantizationType] = None): + def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None, raw_dtype: GGMLQuantizationType | None = None): if self.use_temp_file and self.temp_file is None: fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024) fp.seek(0) @@ -600,7 +602,7 @@ class GGUFWriter: if pad != 0: self.temp_file.write(bytes([0] * pad)) - def write_padding(self, fp: BinaryIO, n: int, align: Optional[int] = None): + def write_padding(self, fp: BinaryIO, n: int, align: int | None = None): pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n if pad != 0: fp.write(bytes([0] * pad)) @@ -726,13 +728,13 @@ class GGUFWriter: def add_tokenizer_model(self, model: str): self.add_string(KEY_TOKENIZER_MODEL, model) - def add_token_list(self, tokens: Union[Sequence[str], Sequence[bytes], Sequence[bytearray]]): + def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]): self.add_array(KEY_TOKENIZER_LIST, tokens) - def add_token_merges(self, merges: Union[Sequence[str], Sequence[bytes], Sequence[bytearray]]): + def add_token_merges(self, merges: Sequence[str] | Sequence[bytes] | Sequence[bytearray]): self.add_array(KEY_TOKENIZER_MERGES, merges) - def add_token_types(self, types: Union[Sequence[TokenType], Sequence[int]]): + def add_token_types(self, types: Sequence[TokenType] | Sequence[int]): self.add_array(KEY_TOKENIZER_TOKEN_TYPE, types) def add_token_scores(self, scores: Sequence[float]): @@ -756,11 +758,11 @@ class GGUFWriter: class SpecialVocab: load_merges: bool = False - merges: List[str] = [] - special_token_types: Tuple[str, ...] = tuple(('bos', 'eos', 'unk', 'sep', 'pad')) - special_token_ids: Dict[str, int] = {} + merges: list[str] = [] + special_token_types: tuple[str, ...] = ('bos', 'eos', 'unk', 'sep', 'pad') + special_token_ids: dict[str, int] = {} - def __init__(self, path: Path, load_merges: bool = False, special_token_types: Optional[Tuple[str, ...]] = None): + def __init__(self, path: Path, load_merges: bool = False, special_token_types: tuple[str, ...] | None = None): self.special_token_ids = {} self.load_merges = load_merges if special_token_types is not None: @@ -821,7 +823,7 @@ class SpecialVocab: print(f'gguf: Adding {len(self.merges)} merge(s).') gw.add_token_merges(self.merges) for typ, tokid in self.special_token_ids.items(): - handler: Optional[Callable[[int], None]] = getattr(gw, f'add_{typ}_token_id', None) + handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None) if handler is None: print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping') continue diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index c66b069f9..26f792b14 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.2.1" +version = "0.3.1" description = "Write ML models in GGUF for GGML" authors = ["GGML "] packages = [ diff --git a/mypy.ini b/mypy.ini new file mode 100644 index 000000000..55c168f2d --- /dev/null +++ b/mypy.ini @@ -0,0 +1,5 @@ +[mypy] +strict = true +allow_untyped_calls = true +allow_untyped_defs = true +allow_incomplete_defs = true From e8422de39e4aa2f7e50574124b060a80607e654a Mon Sep 17 00:00:00 2001 From: DannyDaemonic Date: Thu, 31 Aug 2023 04:21:45 -0700 Subject: [PATCH 108/147] @vxiiduu's fix for PrefetchVirtualMemory (#2930) Reimplement fix for `PrefetchVirtualMemory`. Co-authored-by: vxiiduu <73044267+vxiiduu@users.noreply.github.com> --- llama.cpp | 27 ++++++++++++++++----------- 1 file changed, 16 insertions(+), 11 deletions(-) diff --git a/llama.cpp b/llama.cpp index 95ee6ffe4..98a5da963 100644 --- a/llama.cpp +++ b/llama.cpp @@ -611,20 +611,25 @@ struct llama_mmap { throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str())); } - #if _WIN32_WINNT >= _WIN32_WINNT_WIN8 if (prefetch) { - // Advise the kernel to preload the mapped memory - WIN32_MEMORY_RANGE_ENTRY range; - range.VirtualAddress = addr; - range.NumberOfBytes = (SIZE_T)size; - if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) { - fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n", - llama_format_win_err(GetLastError()).c_str()); + // PrefetchVirtualMemory is only present on Windows 8 and above, so we dynamically load it + BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG); + HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll"); + + // may fail on pre-Windows 8 systems + pPrefetchVirtualMemory = reinterpret_cast (GetProcAddress(hKernel32, "PrefetchVirtualMemory")); + + if (pPrefetchVirtualMemory) { + // advise the kernel to preload the mapped memory + WIN32_MEMORY_RANGE_ENTRY range; + range.VirtualAddress = addr; + range.NumberOfBytes = (SIZE_T)size; + if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) { + fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } } } - #else - #pragma message("warning: You are building for pre-Windows 8; prefetch not supported") - #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8 } ~llama_mmap() { From aeefac4ff760acea5afe66fbfe8d7eca1937b79c Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Thu, 31 Aug 2023 16:49:24 -0600 Subject: [PATCH 109/147] scripts: Use local gguf package when running from repo (#2927) * scripts: Use local gguf when running from repo --- convert-falcon-hf-to-gguf.py | 5 ++++- convert-gptneox-hf-to-gguf.py | 5 ++++- convert-llama-ggmlv3-to-gguf.py | 6 +++++- convert.py | 6 +++++- .../convert-train-checkpoint-to-gguf.py | 5 ++++- 5 files changed, 22 insertions(+), 5 deletions(-) diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py index ec786ff67..271e58972 100755 --- a/convert-falcon-hf-to-gguf.py +++ b/convert-falcon-hf-to-gguf.py @@ -11,11 +11,14 @@ import sys from pathlib import Path from typing import Any -import gguf import numpy as np import torch from transformers import AutoTokenizer # type: ignore[import] +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) +import gguf + def bytes_to_unicode(): # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py index 852123d99..b9c8b4607 100755 --- a/convert-gptneox-hf-to-gguf.py +++ b/convert-gptneox-hf-to-gguf.py @@ -11,11 +11,14 @@ import sys from pathlib import Path from typing import Any -import gguf import numpy as np import torch from transformers import AutoTokenizer # type: ignore[import] +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) +import gguf + # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py diff --git a/convert-llama-ggmlv3-to-gguf.py b/convert-llama-ggmlv3-to-gguf.py index 3f39bc39e..08ba0c490 100755 --- a/convert-llama-ggmlv3-to-gguf.py +++ b/convert-llama-ggmlv3-to-gguf.py @@ -7,9 +7,13 @@ import struct import sys from pathlib import Path -import gguf import numpy as np +import os +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) +import gguf + # Note: Does not support GGML_QKK_64 QK_K = 256 # Items here are (block size, type size) diff --git a/convert.py b/convert.py index 9a39edb99..5cc3f6e66 100755 --- a/convert.py +++ b/convert.py @@ -25,10 +25,14 @@ from dataclasses import dataclass from pathlib import Path from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar -import gguf import numpy as np from sentencepiece import SentencePieceProcessor # type: ignore[import] +import os +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) +import gguf + if TYPE_CHECKING: from typing import TypeAlias diff --git a/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py index 01b3ee92a..a527d6153 100644 --- a/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py +++ b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py @@ -2,13 +2,16 @@ # train-text-from-scratch checkpoint --> gguf conversion import argparse -import gguf import os import struct import sys import numpy as np from pathlib import Path +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / '..' / '..' / 'gguf-py' / 'gguf')) +import gguf + # gguf constants LLM_KV_OPTIMIZER_TYPE = "optimizer.type" LLM_KV_OPTIMIZER_TYPE_ADAM = "adam" From 528134dd0267838d9c0250cf1d9621631dff09b2 Mon Sep 17 00:00:00 2001 From: slaren Date: Fri, 1 Sep 2023 01:32:09 +0200 Subject: [PATCH 110/147] remove convert-llama-7b-pth-to-gguf.py and convert-llama-hf-to-gguf.py (#2906) --- convert-llama-7b-pth-to-gguf.py | 261 ----------------------------- convert-llama-hf-to-gguf.py | 280 -------------------------------- 2 files changed, 541 deletions(-) delete mode 100755 convert-llama-7b-pth-to-gguf.py delete mode 100755 convert-llama-hf-to-gguf.py diff --git a/convert-llama-7b-pth-to-gguf.py b/convert-llama-7b-pth-to-gguf.py deleted file mode 100755 index 6574c11dd..000000000 --- a/convert-llama-7b-pth-to-gguf.py +++ /dev/null @@ -1,261 +0,0 @@ -#!/usr/bin/env python3 -# 7b pth llama --> gguf conversion -# Only models with a single datafile are supported, like 7B -# HF files required in the model dir: config.json tokenizer_config.json tokenizer.json tokenizer.model - -from __future__ import annotations - -import argparse -import json -import os -import struct -import sys -from pathlib import Path -from typing import TYPE_CHECKING, Any - -import gguf -import numpy as np -import torch -from sentencepiece import SentencePieceProcessor # type: ignore[import] - -if TYPE_CHECKING: - from typing import TypeAlias - -NDArray: TypeAlias = 'np.ndarray[Any, Any]' - - -def count_model_parts(dir_model: Path) -> int: - num_parts = 0 - for filename in os.listdir(dir_model): - if filename.startswith("consolidated."): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - return num_parts - - -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a PyTorch 7B LLaMA model to a GGML compatible file") - parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") - parser.add_argument("ftype", type=int, choices=[0, 1], help="output format - use 0 for float32, 1 for float16", default = 1) - return parser.parse_args() - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file = sys.stderr) - sys.exit(1) - -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - -print("gguf: loading model "+dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - -if hparams["architectures"][0] != "LlamaForCausalLM": - print("Model architecture not supported: " + hparams["architectures"][0]) - sys.exit() - -# get number of model parts -num_parts = count_model_parts(dir_model) - -if num_parts > 1: - print("gguf: Only models with a single datafile are supported.") - - sys.exit() - -ARCH=gguf.MODEL_ARCH.LLAMA -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) - - -print("gguf: get model metadata") - -block_count = hparams["num_hidden_layers"] -head_count = hparams["num_attention_heads"] - -if "num_key_value_heads" in hparams: - head_count_kv = hparams["num_key_value_heads"] -else: - head_count_kv = head_count - -if "_name_or_path" in hparams: - hf_repo = hparams["_name_or_path"] -else: - hf_repo = "" - -if "max_sequence_length" in hparams: - ctx_length = hparams["max_sequence_length"] -elif "max_position_embeddings" in hparams: - ctx_length = hparams["max_position_embeddings"] -else: - print("gguf: can not find ctx length parameter.") - - sys.exit() - - -gguf_writer.add_name(dir_model.name) -gguf_writer.add_source_hf_repo(hf_repo) -gguf_writer.add_tensor_data_layout("Meta AI original pth") -gguf_writer.add_context_length(ctx_length) -gguf_writer.add_embedding_length(hparams["hidden_size"]) -gguf_writer.add_block_count(block_count) -gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) -gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"]) -gguf_writer.add_head_count(head_count) -gguf_writer.add_head_count_kv(head_count_kv) -gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"]) - -if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]: - if "type" in hparams["rope_scaling"]: - if hparams["rope_scaling"]["type"] == "linear": - gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"]) - - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytes] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -tokenizer_model_file = dir_model / 'tokenizer.model' -if not tokenizer_model_file.is_file(): - print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr) - sys.exit(1) - -# vocab type sentencepiece -print("gguf: get sentencepiece tokenizer vocab and scores") - -tokenizer = SentencePieceProcessor(str(tokenizer_model_file)) - -for i in range(tokenizer.vocab_size()): - text: bytes - score: float - - piece = tokenizer.id_to_piece(i) - text = piece.encode("utf-8") - score = tokenizer.get_score(i) - - toktype = 1 # defualt to normal token type - if tokenizer.is_unknown(i): - toktype = 2 - if tokenizer.is_control(i): - toktype = 3 - - # toktype = 4 is user-defined = tokens from added_tokens.json - - if tokenizer.is_unused(i): - toktype = 5 - if tokenizer.is_byte(i): - toktype = 6 - - tokens.append(text) - scores.append(score) - toktypes.append(toktype) - -added_tokens_file = dir_model / 'added_tokens.json' -if added_tokens_file.is_file(): - with open(added_tokens_file, "r", encoding="utf-8") as f: - addtokens_json = json.load(f) - - print("gguf: get added tokens") - - for key in addtokens_json: - tokens.append( key.encode("utf-8") ) - scores.append(-1000.0) - toktypes.append(4) # user-defined token type - -gguf_writer.add_tokenizer_model("llama") -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_scores(scores) -gguf_writer.add_token_types(toktypes) - -special_vocab = gguf.SpecialVocab(dir_model) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH,block_count) - -# tensor info -print("gguf: get tensor metadata") - -part_names = (f"consolidated.{n:02}.pth" for n in range(0, num_parts)) - -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") - - for name in model_part.keys(): - data = model_part[name] - - # we don't need these - if name == "rope.freqs": - continue - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - data = data.squeeze().numpy() - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: - data = data.astype(np.float16) - - print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - - gguf_writer.add_tensor(new_name, data) - - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") diff --git a/convert-llama-hf-to-gguf.py b/convert-llama-hf-to-gguf.py deleted file mode 100755 index c453c83c3..000000000 --- a/convert-llama-hf-to-gguf.py +++ /dev/null @@ -1,280 +0,0 @@ -#!/usr/bin/env python3 -# HF llama --> gguf conversion - -from __future__ import annotations - -import argparse -import json -import os -import struct -import sys -from pathlib import Path -from typing import TYPE_CHECKING, Any - -import gguf -import numpy as np -import torch -from sentencepiece import SentencePieceProcessor # type: ignore[import] - -if TYPE_CHECKING: - from typing import TypeAlias - -NDArray: TypeAlias = 'np.ndarray[Any, Any]' - -# reverse HF permute back to original pth layout -# https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py - - -def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray: - if n_kv_head is not None and n_head != n_kv_head: - n_head //= n_kv_head - - return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) - .swapaxes(1, 2) - .reshape(weights.shape)) - - -def count_model_parts(dir_model: str) -> int: - num_parts = 0 - - for filename in os.listdir(dir_model): - if filename.startswith("pytorch_model-"): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - - return num_parts - - -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file") - parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") - parser.add_argument("ftype", type=int, choices=[0, 1], help="output format - use 0 for float32, 1 for float16", default = 1) - return parser.parse_args() - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file = sys.stderr) - sys.exit(1) - -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - -print("gguf: loading model "+dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - -if hparams["architectures"][0] != "LlamaForCausalLM": - print("Model architecture not supported: " + hparams["architectures"][0]) - - sys.exit() - -# get number of model parts -num_parts = count_model_parts(dir_model) - -ARCH=gguf.MODEL_ARCH.LLAMA -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) - -print("gguf: get model metadata") - -block_count = hparams["num_hidden_layers"] -head_count = hparams["num_attention_heads"] - -if "num_key_value_heads" in hparams: - head_count_kv = hparams["num_key_value_heads"] -else: - head_count_kv = head_count - -if "_name_or_path" in hparams: - hf_repo = hparams["_name_or_path"] -else: - hf_repo = "" - -if "max_sequence_length" in hparams: - ctx_length = hparams["max_sequence_length"] -elif "max_position_embeddings" in hparams: - ctx_length = hparams["max_position_embeddings"] -else: - print("gguf: can not find ctx length parameter.") - - sys.exit() - - -gguf_writer.add_name(dir_model.name) -gguf_writer.add_source_hf_repo(hf_repo) -gguf_writer.add_tensor_data_layout("Meta AI original pth") -gguf_writer.add_context_length(ctx_length) -gguf_writer.add_embedding_length(hparams["hidden_size"]) -gguf_writer.add_block_count(block_count) -gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) -gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"]) -gguf_writer.add_head_count(head_count) -gguf_writer.add_head_count_kv(head_count_kv) -gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"]) - -if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]: - if "type" in hparams["rope_scaling"]: - if hparams["rope_scaling"]["type"] == "linear": - gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"]) - - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytes] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -tokenizer_model_file = dir_model / 'tokenizer.model' -if not tokenizer_model_file.is_file(): - print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr) - sys.exit(1) - -# vocab type sentencepiece -print("gguf: get sentencepiece tokenizer vocab, scores and token types") - -tokenizer = SentencePieceProcessor(str(tokenizer_model_file)) - -for i in range(tokenizer.vocab_size()): - text: bytes - score: float - - piece = tokenizer.id_to_piece(i) - text = piece.encode("utf-8") - score = tokenizer.get_score(i) - - toktype = 1 # defualt to normal token type - if tokenizer.is_unknown(i): - toktype = 2 - if tokenizer.is_control(i): - toktype = 3 - - # toktype = 4 is user-defined = tokens from added_tokens.json - - if tokenizer.is_unused(i): - toktype = 5 - if tokenizer.is_byte(i): - toktype = 6 - - tokens.append(text) - scores.append(score) - toktypes.append(toktype) - -added_tokens_file = dir_model / 'added_tokens.json' -if added_tokens_file.is_file(): - with open(added_tokens_file, "r", encoding="utf-8") as f: - addtokens_json = json.load(f) - - print("gguf: get added tokens") - - for key in addtokens_json: - tokens.append( key.encode("utf-8") ) - scores.append(-1000.0) - toktypes.append(4) # user-defined token type - - -gguf_writer.add_tokenizer_model("llama") -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_scores(scores) -gguf_writer.add_token_types(toktypes) - -special_vocab = gguf.SpecialVocab(dir_model) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH,block_count) - -# tensor info -print("gguf: get tensor metadata") - -if num_parts == 0: - part_names = iter(("pytorch_model.bin",)) -else: - part_names = ( - f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) - ) - -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") - - for name in model_part.keys(): - data = model_part[name] - - # we don't need these - if name.endswith(".rotary_emb.inv_freq"): - continue - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - data = data.squeeze().numpy() - - # reverse permute these - if name.endswith(".q_proj.weight"): - data = reverse_hf_permute(data, head_count) - if name.endswith(".k_proj.weight"): - data = reverse_hf_permute(data, head_count, head_count_kv) - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: - data = data.astype(np.float16) - - print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - - gguf_writer.add_tensor(new_name, data) - - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") From bce1fef328941499dc0acb76cc7fd7ac90449c2f Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Thu, 31 Aug 2023 22:13:51 -0400 Subject: [PATCH 111/147] convert : fix another python 3.8 issue (#2949) --- convert.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert.py b/convert.py index 5cc3f6e66..6c89b5ecc 100755 --- a/convert.py +++ b/convert.py @@ -530,7 +530,7 @@ class LazyTensor: raise ValueError(f'Cannot validate conversion from {self.data_type} to {data_type}.') -LazyModel = dict[str, LazyTensor] +LazyModel: TypeAlias = 'dict[str, LazyTensor]' @dataclass From e8d91589258f9204397a7ac5f9b3c857835c98f8 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Fri, 1 Sep 2023 11:15:57 +0300 Subject: [PATCH 112/147] metal: somewhat faster f16 x f32 matrix multiply kernel (#2951) * Somewhat faster f16 x f32 matrix multiply kernel * Better use 32 thread groups for f16 x f32 --------- Co-authored-by: Iwan Kawrakow --- ggml-metal.m | 2 +- ggml-metal.metal | 38 ++++++++++++++++++++++++++++---------- 2 files changed, 29 insertions(+), 11 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index e929c4b07..8c3c64f53 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -840,7 +840,7 @@ void ggml_metal_graph_compute( switch (src0t) { case GGML_TYPE_F16: { - nth0 = 64; + nth0 = 32; nth1 = 1; [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32]; } break; diff --git a/ggml-metal.metal b/ggml-metal.metal index 82e1a0c7a..02db5323e 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -528,24 +528,42 @@ kernel void kernel_mul_mat_f16_f32( device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); - sum[tpitg.x] = 0.0f; + uint ith = tpitg.x; + uint nth = tptg.x; - for (int i = tpitg.x; i < ne00; i += tptg.x) { - sum[tpitg.x] += (float) x[i] * (float) y[i]; + sum[ith] = 0.0f; + + for (int i = ith; i < ne00; i += nth) { + sum[ith] += (float) x[i] * (float) y[i]; } // accumulate the sum from all threads in the threadgroup threadgroup_barrier(mem_flags::mem_threadgroup); - for (uint i = tptg.x/2; i > 0; i /= 2) { - if (tpitg.x < i) { - sum[tpitg.x] += sum[tpitg.x + i]; - } - threadgroup_barrier(mem_flags::mem_threadgroup); + if (ith%4 == 0) { + for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i]; } - - if (tpitg.x == 0) { + threadgroup_barrier(mem_flags::mem_threadgroup); + if (ith%16 == 0) { + for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i]; + } + threadgroup_barrier(mem_flags::mem_threadgroup); + if (ith == 0) { + for (int i = 16; i < nth; i += 16) sum[0] += sum[i]; dst[im*ne1*ne0 + r1*ne0 + r0] = sum[0]; } + + // Original implementation. Left behind commented out for now + //threadgroup_barrier(mem_flags::mem_threadgroup); + //for (uint i = tptg.x/2; i > 0; i /= 2) { + // if (tpitg.x < i) { + // sum[tpitg.x] += sum[tpitg.x + i]; + // } + // threadgroup_barrier(mem_flags::mem_threadgroup); + //} + // + //if (tpitg.x == 0) { + // dst[im*ne1*ne0 + r1*ne0 + r0] = sum[0]; + //} } kernel void kernel_alibi_f32( From 18705a30ef3d6a89e1d7c6cb8cfe8633f760cb53 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Fri, 1 Sep 2023 05:03:49 -0400 Subject: [PATCH 113/147] llama2c : fix segfault and alloc-dealloc-mismatch (#2913) * llama2c : fix segfault if vocab is not found * llama2c : fix mismatch between new[] and delete * llama2c : fix basename on Windows * llama2c : use a destructor to prevent memory leaks --- .../convert-llama2c-to-ggml.cpp | 43 ++++++++++--------- 1 file changed, 23 insertions(+), 20 deletions(-) diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index e9e070b1f..0b03c9d2b 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -75,7 +75,7 @@ typedef struct { int seq_len; // max sequence length } Config; -typedef struct { +struct TransformerWeights { // token embedding table float* token_embedding_table; // (vocab_size, dim) // weights for rmsnorms @@ -97,7 +97,22 @@ typedef struct { // float* freq_cis_imag; // (seq_len, dim/2) // (optional) classifier weights for the logits, on the last layer float* wcls; -} TransformerWeights; + + ~TransformerWeights() { + delete[] token_embedding_table; + delete[] rms_att_weight; + delete[] rms_ffn_weight; + delete[] wq; + delete[] wk; + delete[] wv; + delete[] wo; + delete[] w1; + delete[] w2; + delete[] w3; + delete[] rms_final_weight; + delete[] wcls; + } +}; void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) { // we calloc instead of malloc to keep valgrind happy @@ -173,21 +188,6 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shar return 0; } -void free_weights(TransformerWeights* w) { - delete w->token_embedding_table; - delete w->rms_att_weight; - delete w->rms_ffn_weight; - delete w->wq; - delete w->wk; - delete w->wv; - delete w->wo; - delete w->w1; - delete w->w2; - delete w->w3; - delete w->rms_final_weight; - if (w->wcls) delete w->wcls; -} - void print_sample_weights(TransformerWeights *w){ printf("----- Quick print of first of the weight vales of all the variables\n"); printf("%f\n", w->token_embedding_table[0]); @@ -596,6 +596,10 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) // assume llama2.c vocabulary printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename); llama_file file(filename, "rb"); + if (!file.fp) { + fprintf(stderr, "error: %s: %s\n", strerror(errno), filename); + exit(1); + } const int n_vocab = config->vocab_size; /* uint32_t max_token_length = */ file.read_u32(); // unused vocab->id_to_token.resize(n_vocab); @@ -898,7 +902,7 @@ bool params_parse(int argc, char ** argv, struct train_params * params) { } std::string basename(const std::string &path) { - size_t pos = path.find_last_of("/"); + size_t pos = path.find_last_of("/\\"); if (pos == std::string::npos) { return path; } @@ -911,7 +915,7 @@ int main(int argc, char ** argv) { return 1; } Config config; - TransformerWeights weights; + TransformerWeights weights = {}; { FILE *file = fopen(params.fn_llama2c_model, "rb"); if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; } @@ -953,6 +957,5 @@ int main(int argc, char ** argv) { printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model); ggml_free(model.ctx); - free_weights(&weights); return 0; } From 4dcd47d71df8ca4edcc31302744bd93f0c31298e Mon Sep 17 00:00:00 2001 From: staviq Date: Fri, 1 Sep 2023 11:07:06 +0200 Subject: [PATCH 114/147] logs : fix mingw-like builds (fixes #2898) (#2911) * fix mingw-like builds * formatting * make LOG_COMPAT easier to override and extend * simplify win detection * fix for #2940 --- Makefile | 10 +++++----- common/log.h | 20 ++++++++++---------- 2 files changed, 15 insertions(+), 15 deletions(-) diff --git a/Makefile b/Makefile index b750540fe..b56df3d8a 100644 --- a/Makefile +++ b/Makefile @@ -79,6 +79,11 @@ ifdef LLAMA_SERVER_VERBOSE CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE) endif +ifdef LLAMA_DISABLE_LOGS + CFLAGS += -DLOG_DISABLE_LOGS + CXXFLAGS += -DLOG_DISABLE_LOGS +endif # LLAMA_DISABLE_LOGS + # warnings CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \ -Wmissing-prototypes -Werror=implicit-int @@ -343,11 +348,6 @@ k_quants.o: k_quants.c k_quants.h $(CC) $(CFLAGS) -c $< -o $@ endif # LLAMA_NO_K_QUANTS -ifdef LLAMA_DISABLE_LOGS - CFLAGS += -DLOG_DISABLE_LOGS - CXXFLAGS += -DLOG_DISABLE_LOGS -endif # LLAMA_DISABLE_LOGS - # # Print build information # diff --git a/common/log.h b/common/log.h index c1364187d..bf9fafd68 100644 --- a/common/log.h +++ b/common/log.h @@ -154,7 +154,7 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base // #include "log.h" // #ifndef LOG_NO_TIMESTAMPS - #ifndef _WIN32 + #ifndef _MSC_VER #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] " #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast>(std::chrono::system_clock::now().time_since_epoch())).count() #else @@ -167,7 +167,7 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base #endif #ifdef LOG_TEE_TIMESTAMPS - #ifndef _WIN32 + #ifndef _MSC_VER #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] " #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast>(std::chrono::system_clock::now().time_since_epoch())).count() #else @@ -187,7 +187,7 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base // #include "log.h" // #ifndef LOG_NO_FILE_LINE_FUNCTION - #ifndef _WIN32 + #ifndef _MSC_VER #define LOG_FLF_FMT "[%24s:%5d][%24s] " #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__ #else @@ -200,7 +200,7 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base #endif #ifdef LOG_TEE_FILE_LINE_FUNCTION - #ifndef _WIN32 + #ifndef _MSC_VER #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] " #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__ #else @@ -224,7 +224,7 @@ enum LogTriState // INTERNAL, DO NOT USE // USE LOG() INSTEAD // -#ifndef _WIN32 +#ifndef _MSC_VER #define LOG_IMPL(str, ...) \ { \ if (LOG_TARGET != nullptr) \ @@ -247,7 +247,7 @@ enum LogTriState // INTERNAL, DO NOT USE // USE LOG_TEE() INSTEAD // -#ifndef _WIN32 +#ifndef _MSC_VER #define LOG_TEE_IMPL(str, ...) \ { \ if (LOG_TARGET != nullptr) \ @@ -284,7 +284,7 @@ enum LogTriState // Main LOG macro. // behaves like printf, and supports arguments the exact same way. // -#ifndef _WIN32 +#ifndef _MSC_VER #define LOG(...) LOG_IMPL(__VA_ARGS__, "") #else #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "") @@ -298,14 +298,14 @@ enum LogTriState // Secondary target can be changed just like LOG_TARGET // by defining LOG_TEE_TARGET // -#ifndef _WIN32 +#ifndef _MSC_VER #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "") #else #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "") #endif // LOG macro variants with auto endline. -#ifndef _WIN32 +#ifndef _MSC_VER #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n") #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n") #else @@ -461,7 +461,7 @@ inline void log_test() LOG("13 Hello World this time in yet new file?\n") log_set_target(log_filename_generator("llama_autonamed", "log")); LOG("14 Hello World in log with generated filename!\n") -#ifdef _WIN32 +#ifdef _MSC_VER LOG_TEE("15 Hello msvc TEE without arguments\n") LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test") LOG_TEELN("17 Hello msvc TEELN without arguments\n") From 13268c533177a4dc76bce0b465645d74f0d51d55 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 1 Sep 2023 13:42:41 +0300 Subject: [PATCH 115/147] metal : slight speed-up for add and mul kernels (#2917) --- ggml-metal.m | 20 ++++++++++++++++---- ggml-metal.metal | 32 ++++++++++++++++---------------- 2 files changed, 32 insertions(+), 20 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 8c3c64f53..4267db9be 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -680,6 +680,12 @@ void ggml_metal_graph_compute( } break; case GGML_OP_ADD: { + GGML_ASSERT(ggml_is_contiguous(src0)); + + // utilize float4 + GGML_ASSERT(ne00 % 4 == 0); + const int64_t nb = ne00/4; + if (ggml_nelements(src1) == ne10) { // src1 is a row [encoder setComputePipelineState:ctx->pipeline_add_row]; @@ -689,14 +695,20 @@ void ggml_metal_graph_compute( [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&nb length:sizeof(nb) atIndex:3]; - const int64_t n = ggml_nelements(dst); + const int64_t n = ggml_nelements(dst)/4; [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; case GGML_OP_MUL: { + GGML_ASSERT(ggml_is_contiguous(src0)); + + // utilize float4 + GGML_ASSERT(ne00 % 4 == 0); + const int64_t nb = ne00/4; + if (ggml_nelements(src1) == ne10) { // src1 is a row [encoder setComputePipelineState:ctx->pipeline_mul_row]; @@ -706,9 +718,9 @@ void ggml_metal_graph_compute( [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&nb length:sizeof(nb) atIndex:3]; - const int64_t n = ggml_nelements(dst); + const int64_t n = ggml_nelements(dst)/4; [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; diff --git a/ggml-metal.metal b/ggml-metal.metal index 02db5323e..8cdf0b9d2 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -25,9 +25,9 @@ typedef struct { } block_q8_0; kernel void kernel_add( - device const float * src0, - device const float * src1, - device float * dst, + device const float4 * src0, + device const float4 * src1, + device float4 * dst, uint tpig[[thread_position_in_grid]]) { dst[tpig] = src0[tpig] + src1[tpig]; } @@ -35,18 +35,18 @@ kernel void kernel_add( // assumption: src1 is a row // broadcast src1 into src0 kernel void kernel_add_row( - device const float * src0, - device const float * src1, - device float * dst, - constant int64_t & ne00, + device const float4 * src0, + device const float4 * src1, + device float4 * dst, + constant int64_t & nb, uint tpig[[thread_position_in_grid]]) { - dst[tpig] = src0[tpig] + src1[tpig % ne00]; + dst[tpig] = src0[tpig] + src1[tpig % nb]; } kernel void kernel_mul( - device const float * src0, - device const float * src1, - device float * dst, + device const float4 * src0, + device const float4 * src1, + device float4 * dst, uint tpig[[thread_position_in_grid]]) { dst[tpig] = src0[tpig] * src1[tpig]; } @@ -54,12 +54,12 @@ kernel void kernel_mul( // assumption: src1 is a row // broadcast src1 into src0 kernel void kernel_mul_row( - device const float * src0, - device const float * src1, - device float * dst, - constant int64_t & ne00, + device const float4 * src0, + device const float4 * src1, + device float4 * dst, + constant int64_t & nb, uint tpig[[thread_position_in_grid]]) { - dst[tpig] = src0[tpig] * src1[tpig % ne00]; + dst[tpig] = src0[tpig] * src1[tpig % nb]; } kernel void kernel_scale( From 5aec2cfaac386eb09aebb75b805860828f00de91 Mon Sep 17 00:00:00 2001 From: Tameem <113388789+AhmadTameem@users.noreply.github.com> Date: Fri, 1 Sep 2023 18:27:40 +0500 Subject: [PATCH 116/147] ggml : add RISC-V vector intrinsics support (#2929) * added support for RISCV CFLAGS & native compile + cross compile options * Add RISC-V Vector Intrinsics Support Added RVV intrinsics for following ggml_vec_dot_q4_0_q8_0 ggml_vec_dot_q4_1_q8_1 ggml_vec_dot_q5_0_q8_0 ggml_vec_dot_q5_1_q8_1 ggml_vec_dot_q8_0_q8_0 Co-authored-by: Sharafat Signed-off-by: Ahmad Tameem --------- Signed-off-by: Ahmad Tameem Co-authored-by: moiz.hussain Co-authored-by: Sharafat --- Makefile | 13 ++++ ggml.c | 227 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 240 insertions(+) diff --git a/Makefile b/Makefile index b56df3d8a..8f73297f4 100644 --- a/Makefile +++ b/Makefile @@ -35,6 +35,11 @@ ifndef UNAME_M UNAME_M := $(shell uname -m) endif +ifdef RISCV_CROSS_COMPILE +CC := riscv64-unknown-linux-gnu-gcc +CXX := riscv64-unknown-linux-gnu-g++ +endif + CCV := $(shell $(CC) --version | head -n 1) CXXV := $(shell $(CXX) --version | head -n 1) @@ -150,6 +155,9 @@ endif # Architecture specific # TODO: probably these flags need to be tweaked on some architectures # feel free to update the Makefile for your architecture and send a pull request or issue + +ifndef RISCV + ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64)) # Use all CPU extensions that are available: CFLAGS += -march=native -mtune=native @@ -198,6 +206,11 @@ ifneq ($(filter ppc64%,$(UNAME_M)),) endif endif +else + CFLAGS += -march=rv64gcv -mabi=lp64d + CXXFLAGS += -march=rv64gcv -mabi=lp64d +endif + ifndef LLAMA_NO_K_QUANTS CFLAGS += -DGGML_USE_K_QUANTS CXXFLAGS += -DGGML_USE_K_QUANTS diff --git a/ggml.c b/ggml.c index 46ce4a581..cf3955f7f 100644 --- a/ggml.c +++ b/ggml.c @@ -301,6 +301,10 @@ typedef double ggml_float; #endif #endif +#ifdef __riscv_v_intrinsic +#include +#endif + #ifdef __F16C__ #ifdef _MSC_VER @@ -2677,6 +2681,41 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * } *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3); +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + + size_t vl = __riscv_vsetvl_e8m1(qk/2); + + for (int i = 0; i < nb; i++) { + vuint8m1_t tx = __riscv_vle8_v_u8m1(x[i].qs, vl); + + vint8m1_t y0 = __riscv_vle8_v_i8m1(y[i].qs, vl); + vint8m1_t y1 = __riscv_vle8_v_i8m1(y[i].qs+16, vl); + + vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl); + vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl); + + vint8m1_t x_ai = __riscv_vreinterpret_v_u8m1_i8m1(x_a); + vint8m1_t x_li = __riscv_vreinterpret_v_u8m1_i8m1(x_l); + + vint8m1_t v0 = __riscv_vsub_vx_i8m1(x_ai, 8, vl); + vint8m1_t v1 = __riscv_vsub_vx_i8m1(x_li, 8, vl); + + vint16m2_t vec_mul1 = __riscv_vwmul_vv_i16m2(v0, y0, vl); + vint16m2_t vec_mul2 = __riscv_vwmul_vv_i16m2(v1, y1, vl); + + vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); + + vint32m1_t vs1 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul1, vec_zero, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(vs1); + sumi += __riscv_vmv_x_s_i32m1_i32(vs2); + + sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d); + } + + *s = sumf; #else // scalar float sumf = 0.0; @@ -2803,6 +2842,38 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * } *s = hsum_float_8(acc) + summs; +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + + size_t vl = __riscv_vsetvl_e8m1(qk/2); + + for (int i = 0; i < nb; i++) { + vuint8m1_t tx = __riscv_vle8_v_u8m1(x[i].qs, vl); + + vint8m1_t y0 = __riscv_vle8_v_i8m1(y[i].qs, vl); + vint8m1_t y1 = __riscv_vle8_v_i8m1(y[i].qs+16, vl); + + vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl); + vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl); + + vint8m1_t v0 = __riscv_vreinterpret_v_u8m1_i8m1(x_a); + vint8m1_t v1 = __riscv_vreinterpret_v_u8m1_i8m1(x_l); + + vint16m2_t vec_mul1 = __riscv_vwmul_vv_i16m2(v0, y0, vl); + vint16m2_t vec_mul2 = __riscv_vwmul_vv_i16m2(v1, y1, vl); + + vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); + + vint32m1_t vs1 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul1, vec_zero, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(vs1); + sumi += __riscv_vmv_x_s_i32m1_i32(vs2); + + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + } + + *s = sumf; #else // scalar float sumf = 0.0; @@ -3037,6 +3108,76 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * } *s = hsum_float_8(acc); +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + + uint32_t qh; + + // These temp values are for masking and shift operations + uint32_t temp_1[16] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}; + uint32_t temp_2[16] = {0x1, 0x2, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80, + 0x100, 0x200, 0x400, 0x800, 0x1000, 0x2000, 0x4000, 0x8000}; + + size_t vl = __riscv_vsetvl_e8m1(qk/2); + + for (int i = 0; i < nb; i++) { + memcpy(&qh, x[i].qh, sizeof(uint32_t)); + + // temporary registers + vuint32m4_t vt_1 = __riscv_vle32_v_u32m4(temp_2, vl); + vuint32m4_t vt_2 = __riscv_vle32_v_u32m4(temp_1, vl); + vuint32m4_t vt_3 = __riscv_vsll_vx_u32m4(vt_1, 16, vl); + vuint32m4_t vt_4 = __riscv_vadd_vx_u32m4(vt_2, 12, vl); + + // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + vuint32m4_t xha_0 = __riscv_vand_vx_u32m4(vt_1, qh, vl); + vuint32m4_t xhr_0 = __riscv_vsrl_vv_u32m4(xha_0, vt_2, vl); + vuint32m4_t xhl_0 = __riscv_vsll_vx_u32m4(xhr_0, 4, vl); + + // ((qh & (1u << (j + 16))) >> (j + 12)); + vuint32m4_t xha_1 = __riscv_vand_vx_u32m4(vt_3, qh, vl); + vuint32m4_t xhl_1 = __riscv_vsrl_vv_u32m4(xha_1, vt_4, vl); + + // narrowing + vuint16m2_t xhc_0 = __riscv_vncvt_x_x_w_u16m2(xhl_0, vl); + vuint8m1_t xh_0 = __riscv_vncvt_x_x_w_u8m1(xhc_0, vl); + + vuint16m2_t xhc_1 = __riscv_vncvt_x_x_w_u16m2(xhl_1, vl); + vuint8m1_t xh_1 = __riscv_vncvt_x_x_w_u8m1(xhc_1, vl); + + // load + vuint8m1_t tx = __riscv_vle8_v_u8m1(x[i].qs, vl); + + vint8m1_t y0 = __riscv_vle8_v_i8m1(y[i].qs, vl); + vint8m1_t y1 = __riscv_vle8_v_i8m1(y[i].qs+16, vl); + + vuint8m1_t x_at = __riscv_vand_vx_u8m1(tx, 0x0F, vl); + vuint8m1_t x_lt = __riscv_vsrl_vx_u8m1(tx, 0x04, vl); + + vuint8m1_t x_a = __riscv_vor_vv_u8m1(x_at, xh_0, vl); + vuint8m1_t x_l = __riscv_vor_vv_u8m1(x_lt, xh_1, vl); + + vint8m1_t x_ai = __riscv_vreinterpret_v_u8m1_i8m1(x_a); + vint8m1_t x_li = __riscv_vreinterpret_v_u8m1_i8m1(x_l); + + vint8m1_t v0 = __riscv_vsub_vx_i8m1(x_ai, 16, vl); + vint8m1_t v1 = __riscv_vsub_vx_i8m1(x_li, 16, vl); + + vint16m2_t vec_mul1 = __riscv_vwmul_vv_i16m2(v0, y0, vl); + vint16m2_t vec_mul2 = __riscv_vwmul_vv_i16m2(v1, y1, vl); + + vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); + + vint32m1_t vs1 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul1, vec_zero, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(vs1); + sumi += __riscv_vmv_x_s_i32m1_i32(vs2); + + sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi; + } + + *s = sumf; #else // scalar float sumf = 0.0; @@ -3293,6 +3434,72 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * } *s = hsum_float_8(acc) + summs; +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + + uint32_t qh; + + // These temp values are for shift operations + uint32_t temp_1[16] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}; + + size_t vl = __riscv_vsetvl_e8m1(qk/2); + + for (int i = 0; i < nb; i++) { + memcpy(&qh, x[i].qh, sizeof(uint32_t)); + + // temporary registers + vuint32m4_t vt_1 = __riscv_vle32_v_u32m4(temp_1, vl); + vuint32m4_t vt_2 = __riscv_vadd_vx_u32m4(vt_1, 12, vl); + + // load qh + vuint32m4_t vqh = __riscv_vmv_v_x_u32m4(qh, vl); + + // ((qh >> (j + 0)) << 4) & 0x10; + vuint32m4_t xhr_0 = __riscv_vsrl_vv_u32m4(vqh, vt_1, vl); + vuint32m4_t xhl_0 = __riscv_vsll_vx_u32m4(xhr_0, 4, vl); + vuint32m4_t xha_0 = __riscv_vand_vx_u32m4(xhl_0, 0x10, vl); + + // ((qh >> (j + 12)) ) & 0x10; + vuint32m4_t xhr_1 = __riscv_vsrl_vv_u32m4(vqh, vt_2, vl); + vuint32m4_t xha_1 = __riscv_vand_vx_u32m4(xhr_1, 0x10, vl); + + // narrowing + vuint16m2_t xhc_0 = __riscv_vncvt_x_x_w_u16m2(xha_0, vl); + vuint8m1_t xh_0 = __riscv_vncvt_x_x_w_u8m1(xhc_0, vl); + + vuint16m2_t xhc_1 = __riscv_vncvt_x_x_w_u16m2(xha_1, vl); + vuint8m1_t xh_1 = __riscv_vncvt_x_x_w_u8m1(xhc_1, vl); + + // load + vuint8m1_t tx = __riscv_vle8_v_u8m1(x[i].qs, vl); + + vint8m1_t y0 = __riscv_vle8_v_i8m1(y[i].qs, vl); + vint8m1_t y1 = __riscv_vle8_v_i8m1(y[i].qs+16, vl); + + vuint8m1_t x_at = __riscv_vand_vx_u8m1(tx, 0x0F, vl); + vuint8m1_t x_lt = __riscv_vsrl_vx_u8m1(tx, 0x04, vl); + + vuint8m1_t x_a = __riscv_vor_vv_u8m1(x_at, xh_0, vl); + vuint8m1_t x_l = __riscv_vor_vv_u8m1(x_lt, xh_1, vl); + + vint8m1_t v0 = __riscv_vreinterpret_v_u8m1_i8m1(x_a); + vint8m1_t v1 = __riscv_vreinterpret_v_u8m1_i8m1(x_l); + + vint16m2_t vec_mul1 = __riscv_vwmul_vv_i16m2(v0, y0, vl); + vint16m2_t vec_mul2 = __riscv_vwmul_vv_i16m2(v1, y1, vl); + + vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); + + vint32m1_t vs1 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul1, vec_zero, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(vs1); + sumi += __riscv_vmv_x_s_i32m1_i32(vs2); + + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + } + + *s = sumf; #else // scalar float sumf = 0.0; @@ -3404,6 +3611,26 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * } *s = hsum_float_8(acc); +#elif defined(__riscv_v_intrinsic) + float sumf = 0.0; + size_t vl = __riscv_vsetvl_e8m1(qk); + + for (int i = 0; i < nb; i++) { + // load elements + vint8m1_t bx = __riscv_vle8_v_i8m1(x[i].qs, vl); + vint8m1_t by = __riscv_vle8_v_i8m1(y[i].qs, vl); + + vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx, by, vl); + + vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl); + vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl); + + int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum); + + sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)); + } + + *s = sumf; #else // scalar float sumf = 0.0; From d8d6977f48f1fa402ade38ad32c5b5fb1358d059 Mon Sep 17 00:00:00 2001 From: Ben Siraphob Date: Fri, 1 Sep 2023 09:32:14 -0400 Subject: [PATCH 117/147] examples : add C grammar (#2357) --- grammars/c.gbnf | 42 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 42 insertions(+) create mode 100644 grammars/c.gbnf diff --git a/grammars/c.gbnf b/grammars/c.gbnf new file mode 100644 index 000000000..4a0331dd2 --- /dev/null +++ b/grammars/c.gbnf @@ -0,0 +1,42 @@ +root ::= (declaration)* + +declaration ::= dataType identifier "(" parameter? ")" "{" statement* "}" + +dataType ::= "int" ws | "float" ws | "char" ws +identifier ::= [a-zA-Z_] [a-zA-Z_0-9]* + +parameter ::= dataType identifier + +statement ::= + ( dataType identifier ws "=" ws expression ";" ) | + ( identifier ws "=" ws expression ";" ) | + ( identifier ws "(" argList? ")" ";" ) | + ( "return" ws expression ";" ) | + ( "while" "(" condition ")" "{" statement* "}" ) | + ( "for" "(" forInit ";" ws condition ";" ws forUpdate ")" "{" statement* "}" ) | + ( "if" "(" condition ")" "{" statement* "}" ("else" "{" statement* "}")? ) | + ( singleLineComment ) | + ( multiLineComment ) + +forInit ::= dataType identifier ws "=" ws expression | identifier ws "=" ws expression +forUpdate ::= identifier ws "=" ws expression + +condition ::= expression relationOperator expression +relationOperator ::= ("<=" | "<" | "==" | "!=" | ">=" | ">") + +expression ::= term (("+" | "-") term)* +term ::= factor(("*" | "/") factor)* + +factor ::= identifier | number | unaryTerm | funcCall | parenExpression +unaryTerm ::= "-" factor +funcCall ::= identifier "(" argList? ")" +parenExpression ::= "(" ws expression ws ")" + +argList ::= expression ("," ws expression)* + +number ::= [0-9]+ + +singleLineComment ::= "//" [^\n]* "\n" +multiLineComment ::= "/*" ( [^*] | ("*" [^/]) )* "*/" + +ws ::= ([ \t\n]+) From ef156499721c67748cde01a5436cb6f0648bb4b4 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Fri, 1 Sep 2023 09:34:50 -0400 Subject: [PATCH 118/147] build : fix most gcc and clang warnings (#2861) * fix most gcc and clang warnings * baby-llama : remove commented opt_params_adam * fix some MinGW warnings * fix more MinGW warnings --- CMakeLists.txt | 5 +++++ Makefile | 7 ++++++- common/common.cpp | 6 ++++-- common/console.cpp | 1 + examples/baby-llama/baby-llama.cpp | 5 ----- examples/beam-search/beam-search.cpp | 8 +++++--- examples/server/server.cpp | 8 +++++--- k_quants.c | 8 ++------ llama.cpp | 4 ++-- 9 files changed, 30 insertions(+), 22 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index d6c1b3b33..1b7cce9f1 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -403,6 +403,7 @@ if (LLAMA_ALL_WARNINGS) -Wpointer-arith -Wmissing-prototypes -Werror=implicit-int + -Wno-unused-function ) set(cxx_flags -Wall @@ -412,6 +413,10 @@ if (LLAMA_ALL_WARNINGS) -Wno-unused-function -Wno-multichar ) + if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU") + # g++ only + set(cxx_flags ${cxx_flags} -Wno-format-truncation) + endif() else() # todo : msvc endif() diff --git a/Makefile b/Makefile index 8f73297f4..ef1eef6ac 100644 --- a/Makefile +++ b/Makefile @@ -91,9 +91,14 @@ endif # LLAMA_DISABLE_LOGS # warnings CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \ - -Wmissing-prototypes -Werror=implicit-int + -Wmissing-prototypes -Werror=implicit-int -Wno-unused-function CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar +ifeq '' '$(findstring clang++,$(CXX))' + # g++ only + CXXFLAGS += -Wno-format-truncation +endif + # OS specific # TODO: support Windows ifeq ($(UNAME_S),Linux) diff --git a/common/common.cpp b/common/common.cpp index ed09fc27d..41fc59ced 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -24,7 +24,9 @@ #if defined(_WIN32) #define WIN32_LEAN_AND_MEAN -#define NOMINMAX +#ifndef NOMINMAX +# define NOMINMAX +#endif #include #include #include @@ -1027,7 +1029,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str()); fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n"); fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false"); - fprintf(stream, "hellaswag_tasks: %ld # default: 400\n", params.hellaswag_tasks); + fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks); const auto logit_bias_eos = params.logit_bias.find(llama_token_eos(lctx)); const bool ignore_eos = logit_bias_eos != params.logit_bias.end() && logit_bias_eos->second == -INFINITY; diff --git a/common/console.cpp b/common/console.cpp index 8efa2a674..23545e5be 100644 --- a/common/console.cpp +++ b/common/console.cpp @@ -235,6 +235,7 @@ namespace console { int estimateWidth(char32_t codepoint) { #if defined(_WIN32) + (void)codepoint; return 1; #else return wcwidth(codepoint); diff --git a/examples/baby-llama/baby-llama.cpp b/examples/baby-llama/baby-llama.cpp index 6fa55b319..a99ece9a6 100644 --- a/examples/baby-llama/baby-llama.cpp +++ b/examples/baby-llama/baby-llama.cpp @@ -1617,15 +1617,10 @@ int main(int argc, char ** argv) { float error_before_opt = ggml_get_f32_1d(e, 0); - struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM); struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS); - opt_params_adam.print_forward_graph = false; - opt_params_adam.print_backward_graph = false; opt_params_lbfgs.print_forward_graph = false; opt_params_lbfgs.print_backward_graph = false; - opt_params_adam.adam.n_iter = 16; opt_params_lbfgs.lbfgs.n_iter = 16; - // ggml_opt(ctx0, opt_params_adam, e); ggml_opt(ctx0, opt_params_lbfgs, e); // ggml_build_forward_expand(&gf, e); diff --git a/examples/beam-search/beam-search.cpp b/examples/beam-search/beam-search.cpp index 42c7c7254..4d021434b 100644 --- a/examples/beam-search/beam-search.cpp +++ b/examples/beam-search/beam-search.cpp @@ -22,7 +22,9 @@ #include #elif defined (_WIN32) #define WIN32_LEAN_AND_MEAN -#define NOMINMAX +#ifndef NOMINMAX +# define NOMINMAX +#endif #include #include #endif @@ -73,7 +75,7 @@ void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_stat assert(0u < beams_state.n_beams); const llama_token * tokens = beams_state.beam_views[0].tokens; std::copy(tokens, tokens + n, callback_data.response.end() - n); - printf("%lu", n); + printf("%zu", n); } fflush(stdout); #if 1 // DEBUG: print current beams for this iteration @@ -145,7 +147,7 @@ int main(int argc, char ** argv) if (tokens_list.size() > max_tokens_list_size) { - fprintf( stderr , "%s: error: prompt too long (%lu tokens, max %lu)\n" , + fprintf( stderr , "%s: error: prompt too long (%zu tokens, max %zu)\n" , __func__ , tokens_list.size() , max_tokens_list_size ); return 1; } diff --git a/examples/server/server.cpp b/examples/server/server.cpp index b485a5ead..09eac2ec2 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -17,6 +17,8 @@ #include "completion.js.hpp" #include "json-schema-to-grammar.mjs.hpp" +#include + #ifndef SERVER_VERBOSE #define SERVER_VERBOSE 1 #endif @@ -1038,7 +1040,7 @@ static json format_timings(llama_server_context &llama) { const auto timings = llama_get_timings(llama.ctx); - assert(timings.n_eval == llama.num_tokens_predicted); + assert(timings.n_eval == ptrdiff_t(llama.num_tokens_predicted)); return json{ {"prompt_n", timings.n_p_eval}, @@ -1239,7 +1241,7 @@ void beam_search_callback(void * callback_data, llama_beams_state beams_state) { const llama_token * tokens = beams_state.beam_views[0].tokens; const auto map = [](llama_token tok) { return completion_token_output{{},tok}; }; std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map); - printf("%lu", n); + printf("%zu", n); } fflush(stdout); #if 0 // DEBUG: print current beams for this iteration @@ -1548,7 +1550,7 @@ int main(int argc, char **argv) svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep) { - const auto * fmt = "500 Internal Server Error\n%s"; + const char fmt[] = "500 Internal Server Error\n%s"; char buf[BUFSIZ]; try { std::rethrow_exception(std::move(ep)); diff --git a/k_quants.c b/k_quants.c index 3a9b1dafd..3deeaedf7 100644 --- a/k_quants.c +++ b/k_quants.c @@ -183,13 +183,9 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t int ntry, float alpha) { float min = x[0]; float max = x[0]; - float sum_x = 0; - float sum_x2 = 0; for (int i = 1; i < n; ++i) { if (x[i] < min) min = x[i]; if (x[i] > max) max = x[i]; - sum_x += x[i]; - sum_x2 += x[i]*x[i]; } if (max == min) { for (int i = 0; i < n; ++i) L[i] = 0; @@ -2060,7 +2056,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri __m256 acc = _mm256_setzero_ps(); - uint32_t *aux; + const uint32_t *aux; for (int i = 0; i < nb; ++i) { @@ -2070,7 +2066,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const int8_t * restrict q8 = y[i].qs; // Set up scales - aux = (uint32_t *)x[i].scales; + aux = (const uint32_t *)x[i].scales; __m128i scales128 = _mm_set_epi32( ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4), ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4), diff --git a/llama.cpp b/llama.cpp index 98a5da963..5ca119238 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3600,7 +3600,7 @@ static void llama_grammar_advance_stack( std::vector> & new_stacks) { if (stack.empty()) { - new_stacks.push_back(stack); + new_stacks.emplace_back(stack); return; } @@ -3637,7 +3637,7 @@ static void llama_grammar_advance_stack( } case LLAMA_GRETYPE_CHAR: case LLAMA_GRETYPE_CHAR_NOT: - new_stacks.push_back(stack); + new_stacks.emplace_back(stack); break; default: // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range From 49bb9cbe0f598bc43be539b0df8eafb2130cfad3 Mon Sep 17 00:00:00 2001 From: Konstantin Herud Date: Fri, 1 Sep 2023 15:36:14 +0200 Subject: [PATCH 119/147] docs : add java-llama.cpp to README.md (#2935) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index d727b0554..4e6a0957d 100644 --- a/README.md +++ b/README.md @@ -114,6 +114,7 @@ as the main playground for developing new features for the [ggml](https://github - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s) - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj) - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn) +- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp) **UI:** From ee8654bcd0146708988a703e54406d5b553712ea Mon Sep 17 00:00:00 2001 From: m3ndax Date: Fri, 1 Sep 2023 15:47:27 +0200 Subject: [PATCH 120/147] minor : add const qualifiers (#2853) * made the methods const # Conflicts: # examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp * made method const * Update convert-llama2c-to-ggml.cpp removed write_raw and write_u32 * llama2c : remove misleading const --------- Co-authored-by: Georgi Gerganov --- examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp | 2 +- llama.cpp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index 0b03c9d2b..0ee7adc52 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -637,7 +637,7 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) } } -void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * karpathy_weights){ +void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) { int ct; switch (gg_weights->n_dims){ case 1: diff --git a/llama.cpp b/llama.cpp index 5ca119238..23b251caf 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4393,7 +4393,7 @@ struct llama_logit_info { } return min_heap; } - float probability_from_logit(float logit) { + float probability_from_logit(float logit) const { return normalizer * std::exp(logit - max_l); } }; From 6c9c23429bf4e4fcaaddbebadc4638558430a7f2 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Fri, 1 Sep 2023 09:53:14 -0400 Subject: [PATCH 121/147] make : use unaligned vector moves on MinGW (#2945) Fixes #2922 --- Makefile | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/Makefile b/Makefile index ef1eef6ac..23f050c0d 100644 --- a/Makefile +++ b/Makefile @@ -177,6 +177,14 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64)) #CXXFLAGS += -mssse3 endif +# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves. +# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412 +# https://github.com/ggerganov/llama.cpp/issues/2922 +ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))' + CFLAGS += -Xassembler -muse-unaligned-vector-move + CXXFLAGS += -Xassembler -muse-unaligned-vector-move +endif + ifneq ($(filter aarch64%,$(UNAME_M)),) # Apple M1, M2, etc. # Raspberry Pi 3, 4, Zero 2 (64-bit) From 0d5893668625456c94bbadfddc53fc69cd51c223 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 1 Sep 2023 17:00:40 +0300 Subject: [PATCH 122/147] llama2c : rename function --- .../convert-llama2c-to-ggml.cpp | 28 +++++++++---------- 1 file changed, 14 insertions(+), 14 deletions(-) diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp index 0ee7adc52..9e856c21a 100644 --- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp +++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp @@ -637,7 +637,7 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) } } -void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) { +void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) { int ct; switch (gg_weights->n_dims){ case 1: @@ -674,13 +674,13 @@ void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, const float } void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) { - // stuff AK weights into GG weights one by one. + // convert AK weights into GG weights one by one. // w->token_embedding_table -> model->tok_embeddings // float* -> struct ggml_tensor - stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table); - stuff_karpathy_weights_into_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table); + convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table); + convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table); - stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight); + convert_weights_ak_to_gg(model->norm, w->rms_final_weight); //print_row(model->norm, 0); // for rms-att-weight @@ -690,18 +690,18 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod for (uint32_t i = 0; i < model->hparams.n_layer; ++i){ auto & layer = model->layers[i]; // 1d - stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]); - stuff_karpathy_weights_into_gg(layer.ffn_norm , &w->rms_ffn_weight[i*row_length]); + convert_weights_ak_to_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]); + convert_weights_ak_to_gg(layer.ffn_norm , &w->rms_ffn_weight[i*row_length]); // from 3d matrix layer x dim x dim to 2d matrix dim x dim - stuff_karpathy_weights_into_gg(layer.wq , &w->wq[i*row_length*row_length]); - stuff_karpathy_weights_into_gg(layer.wk , &w->wk[i*row_length*row_length]); - stuff_karpathy_weights_into_gg(layer.wv , &w->wv[i*row_length*row_length]); - stuff_karpathy_weights_into_gg(layer.wo , &w->wo[i*row_length*row_length]); + convert_weights_ak_to_gg(layer.wq , &w->wq[i*row_length*row_length]); + convert_weights_ak_to_gg(layer.wk , &w->wk[i*row_length*row_length]); + convert_weights_ak_to_gg(layer.wv , &w->wv[i*row_length*row_length]); + convert_weights_ak_to_gg(layer.wo , &w->wo[i*row_length*row_length]); - stuff_karpathy_weights_into_gg(layer.w1 , &w->w1[i*row_length*n_ff]); - stuff_karpathy_weights_into_gg(layer.w2 , &w->w2[i*n_ff*row_length]); - stuff_karpathy_weights_into_gg(layer.w3 , &w->w3[i*row_length*n_ff]); + convert_weights_ak_to_gg(layer.w1 , &w->w1[i*row_length*n_ff]); + convert_weights_ak_to_gg(layer.w2 , &w->w2[i*n_ff*row_length]); + convert_weights_ak_to_gg(layer.w3 , &w->w3[i*row_length*n_ff]); } struct gguf_context * ctx = gguf_init_empty(); From 5d6f19f16b2173afe2d5c6aee2f5c9fc31038eba Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Fri, 1 Sep 2023 08:02:48 -0600 Subject: [PATCH 123/147] Allow quantize to only copy tensors, some other improvements (#2931) * Allow quantize tool to only copy tensors to allow repackaging models. * Slightly better logic when requantizing. * Change help message to go to `stdout`. --- examples/quantize/quantize.cpp | 24 +++++++++++++++++++----- llama.cpp | 25 +++++++++++++++++-------- llama.h | 1 + 3 files changed, 37 insertions(+), 13 deletions(-) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index df9a214fc..c174be069 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -35,6 +35,8 @@ static const std::vector QUANT_OPTIONS = { { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", }, { "F16", LLAMA_FTYPE_MOSTLY_F16, "13.00G @ 7B", }, { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, + // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching. + { "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", }, }; @@ -71,12 +73,17 @@ bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std: // ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads] // void usage(const char * executable) { - fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable); - fprintf(stderr, " --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n"); - fprintf(stderr, " --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n"); - fprintf(stderr, "\nAllowed quantization types:\n"); + printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable); + printf(" --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n"); + printf(" --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n"); + printf("\nAllowed quantization types:\n"); for (auto & it : QUANT_OPTIONS) { - printf(" %2d or %-6s : %s\n", it.ftype, it.name.c_str(), it.desc.c_str()); + if (it.name != "COPY") { + printf(" %2d or ", it.ftype); + } else { + printf(" "); + } + printf("%-6s : %s\n", it.name.c_str(), it.desc.c_str()); } exit(1); } @@ -121,6 +128,9 @@ int main(int argc, char ** argv) { // export as [inp path]/ggml-model-[ftype].gguf fname_out = fpath + "ggml-model-" + ftype_str + ".gguf"; arg_idx++; + if (ftype_str == "COPY") { + params.only_copy = true; + } } else { fname_out = argv[arg_idx]; @@ -133,6 +143,10 @@ int main(int argc, char ** argv) { if (!try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) { fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]); return 1; + } else { + if (ftype_str == "COPY") { + params.only_copy = true; + } } arg_idx++; } diff --git a/llama.cpp b/llama.cpp index 23b251caf..3114d3311 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4683,6 +4683,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s llm_load_arch(*ml, model); llm_load_hparams(*ml, model, 0, 0, 0); + if (params->only_copy) { + ftype = model.ftype; + } + const size_t align = GGUF_DEFAULT_ALIGNMENT; struct gguf_context * ctx_out = gguf_init_empty(); @@ -4769,18 +4773,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // quantize only 2D tensors quantize &= (tensor->n_dims == 2); quantize &= params->quantize_output_tensor || name != "output.weight"; - quantize &= quantized_type != tensor->type; + quantize &= !params->only_copy; enum ggml_type new_type; void * new_data; size_t new_size; - if (!quantize) { - new_type = tensor->type; - new_data = tensor->data; - new_size = ggml_nbytes(tensor); - LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0); - } else { + if (quantize) { new_type = quantized_type; #ifdef GGML_USE_K_QUANTS // TODO: avoid hardcoded tensor names - use the TN_* constants @@ -4879,7 +4878,16 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } } #endif - + // If we've decided to quantize to the same type the tensor is already + // in then there's nothing to do. + quantize = tensor->type != new_type; + } + if (!quantize) { + new_type = tensor->type; + new_data = tensor->data; + new_size = ggml_nbytes(tensor); + LLAMA_LOG_INFO("size = %8.3f MB\n", ggml_nbytes(tensor)/1024.0/1024.0); + } else { const size_t nelements = ggml_nelements(tensor); float * f32_data; @@ -5310,6 +5318,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() { /*.ftype =*/ LLAMA_FTYPE_MOSTLY_Q5_1, /*.allow_requantize =*/ false, /*.quantize_output_tensor =*/ true, + /*.only_copy =*/ false, }; return result; diff --git a/llama.h b/llama.h index 6e5e1df63..422f28527 100644 --- a/llama.h +++ b/llama.h @@ -164,6 +164,7 @@ extern "C" { enum llama_ftype ftype; // quantize to this llama_ftype bool allow_requantize; // allow quantizing non-f32/f16 tensors bool quantize_output_tensor; // quantize output.weight + bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored } llama_model_quantize_params; // grammar types From 69fdbb9abc8907dd2a9ffdd840cba92d678a660a Mon Sep 17 00:00:00 2001 From: ZHAOKAI WANG Date: Fri, 1 Sep 2023 22:06:44 +0800 Subject: [PATCH 124/147] readme : quick start command fix (#2908) * quick start command fix * quick start win command fix --- examples/main/README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/examples/main/README.md b/examples/main/README.md index d555afdcc..2773fe976 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -34,7 +34,7 @@ For an interactive experience, try this command: #### Unix-based systems (Linux, macOS, etc.): ```bash -./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " \ +./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -p \ 'User: Hi AI: Hello. I am an AI chatbot. Would you like to talk? User: Sure! @@ -45,7 +45,7 @@ User:' #### Windows: ```powershell -main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -e --prompt "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:" +main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -e -p "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:" ``` The following command generates "infinite" text from a starting prompt (you can use `Ctrl-C` to stop it): From f04d0028444bc9b3d4225fba47e19d4c3aeb3741 Mon Sep 17 00:00:00 2001 From: Engininja2 <139037756+Engininja2@users.noreply.github.com> Date: Fri, 1 Sep 2023 15:33:19 -0600 Subject: [PATCH 125/147] cuda : vsubss4 for older versions of ROCm/clang (#2942) --- ggml-cuda.cu | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 5fd625630..8357f32f7 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -81,12 +81,29 @@ #if defined(GGML_USE_HIPBLAS) #define __CUDA_ARCH__ 1300 +#ifndef __has_builtin + #define __has_builtin(x) 0 +#endif + typedef int8_t int8x4_t __attribute__((ext_vector_type(4))); static __device__ __forceinline__ int __vsubss4(const int a, const int b) { const int8x4_t va = reinterpret_cast(a); const int8x4_t vb = reinterpret_cast(b); +#if __has_builtin(__builtin_elementwise_sub_sat) const int8x4_t c = __builtin_elementwise_sub_sat(va, vb); return reinterpret_cast(c); +#else + int8x4_t c; + int16_t tmp; +#pragma unroll + for (int i = 0; i < 4; i++) { + tmp = va[i] - vb[i]; + if(tmp > std::numeric_limits::max()) tmp = std::numeric_limits::max(); + if(tmp < std::numeric_limits::min()) tmp = std::numeric_limits::min(); + c[i] = tmp; + } + return reinterpret_cast(c); +#endif // __has_builtin(__builtin_elementwise_sub_sat) } static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { From 571083f508266c4eb5cb5457d836df5dd3c173ce Mon Sep 17 00:00:00 2001 From: Jhen-Jie Hong Date: Sat, 2 Sep 2023 08:31:46 +0800 Subject: [PATCH 126/147] server : avoid aniprompt in probabilities of final response (#2849) --- examples/server/server.cpp | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 09eac2ec2..94def943b 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1379,7 +1379,13 @@ int main(int argc, char **argv) } } - const json data = format_final_response(llama, llama.generated_text, llama.generated_token_probs); + auto probs = llama.generated_token_probs; + if (llama.params.n_probs > 0 && llama.stopped_word) { + const std::vector stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false); + probs = std::vector(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size()); + } + + const json data = format_final_response(llama, llama.generated_text, probs); llama_print_timings(llama.ctx); @@ -1456,7 +1462,11 @@ int main(int argc, char **argv) if (!llama.has_next_token) { // Generation is done, send extra information. - const json data = format_final_response(llama, "", llama.generated_token_probs); + const json data = format_final_response( + llama, + "", + std::vector(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index) + ); const std::string str = "data: " + From 21f3d1be867b4d7be07c26f5da6e4bc69bcf4d27 Mon Sep 17 00:00:00 2001 From: Jhen-Jie Hong Date: Sat, 2 Sep 2023 20:23:45 +0800 Subject: [PATCH 127/147] k-quants : fix build on armv7 (android only) (#2920) * k-quants : fix build on armv7 * ggml : cleanup unused arm32 specific impl * k-quants : avoid some unused vzero / mzero define * ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm --- ggml-alloc.c | 7 +++++++ ggml.c | 46 ---------------------------------------------- k_quants.c | 40 +++++++++++++++++++++++++++++++++++----- 3 files changed, 42 insertions(+), 51 deletions(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index f07a4a217..459f121ca 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -284,7 +284,14 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) // address and size of the buffer when measuring // it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers static void * const MEASURE_BASE_ADDR = (void *) 0x1000; +#if defined(__ARM_NEON) && !defined(__aarch64__) +// 32-bit +// TODO: Use for 32-bit x86 as well +static const size_t MEASURE_MAX_SIZE = (1ULL<<32) - 1; // 4 GB +#else +// 64-bit static const size_t MEASURE_MAX_SIZE = 1ULL<<40; // 1 TB +#endif struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) { struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */); diff --git a/ggml.c b/ggml.c index cf3955f7f..38b1155c1 100644 --- a/ggml.c +++ b/ggml.c @@ -817,46 +817,6 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 #if !defined(__aarch64__) -inline static uint16_t vaddvq_u8(uint8x16_t v) { - return - (uint16_t)vgetq_lane_u8(v, 0) + (uint16_t)vgetq_lane_u8(v, 1) + - (uint16_t)vgetq_lane_u8(v, 2) + (uint16_t)vgetq_lane_u8(v, 3) + - (uint16_t)vgetq_lane_u8(v, 4) + (uint16_t)vgetq_lane_u8(v, 5) + - (uint16_t)vgetq_lane_u8(v, 6) + (uint16_t)vgetq_lane_u8(v, 7) + - (uint16_t)vgetq_lane_u8(v, 8) + (uint16_t)vgetq_lane_u8(v, 9) + - (uint16_t)vgetq_lane_u8(v, 10) + (uint16_t)vgetq_lane_u8(v, 11) + - (uint16_t)vgetq_lane_u8(v, 12) + (uint16_t)vgetq_lane_u8(v, 13) + - (uint16_t)vgetq_lane_u8(v, 14) + (uint16_t)vgetq_lane_u8(v, 15); -} - -inline static int16_t vaddvq_s8(int8x16_t v) { - return - (int16_t)vgetq_lane_s8(v, 0) + (int16_t)vgetq_lane_s8(v, 1) + - (int16_t)vgetq_lane_s8(v, 2) + (int16_t)vgetq_lane_s8(v, 3) + - (int16_t)vgetq_lane_s8(v, 4) + (int16_t)vgetq_lane_s8(v, 5) + - (int16_t)vgetq_lane_s8(v, 6) + (int16_t)vgetq_lane_s8(v, 7) + - (int16_t)vgetq_lane_s8(v, 8) + (int16_t)vgetq_lane_s8(v, 9) + - (int16_t)vgetq_lane_s8(v, 10) + (int16_t)vgetq_lane_s8(v, 11) + - (int16_t)vgetq_lane_s8(v, 12) + (int16_t)vgetq_lane_s8(v, 13) + - (int16_t)vgetq_lane_s8(v, 14) + (int16_t)vgetq_lane_s8(v, 15); -} - -inline static int32_t vaddvq_s16(int16x8_t v) { - return - (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + - (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + - (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + - (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); -} - -inline static uint32_t vaddvq_u16(uint16x8_t v) { - return - (uint32_t)vgetq_lane_u16(v, 0) + (uint32_t)vgetq_lane_u16(v, 1) + - (uint32_t)vgetq_lane_u16(v, 2) + (uint32_t)vgetq_lane_u16(v, 3) + - (uint32_t)vgetq_lane_u16(v, 4) + (uint32_t)vgetq_lane_u16(v, 5) + - (uint32_t)vgetq_lane_u16(v, 6) + (uint32_t)vgetq_lane_u16(v, 7); -} - inline static int32_t vaddvq_s32(int32x4_t v) { return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); } @@ -865,12 +825,6 @@ inline static float vaddvq_f32(float32x4_t v) { return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3); } -inline static float vminvq_f32(float32x4_t v) { - return - MIN(MIN(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)), - MIN(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); -} - inline static float vmaxvq_f32(float32x4_t v) { return MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)), diff --git a/k_quants.c b/k_quants.c index 3deeaedf7..4accd2480 100644 --- a/k_quants.c +++ b/k_quants.c @@ -13,6 +13,26 @@ // #include +#if !defined(__aarch64__) +inline static int32_t vaddvq_s16(int16x8_t v) { + return + (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + + (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + + (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + + (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); +} + +inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) { + int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a)); + int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b)); + return vcombine_s16(a0, b0); +} + +inline static int32_t vaddvq_s32(int32x4_t v) { + return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); +} +#endif + #else #ifdef __wasm_simd128__ @@ -1302,7 +1322,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t m3 = vdupq_n_u8(0x3); const uint8x16_t m4 = vdupq_n_u8(0xF); +#if defined(__ARM_FEATURE_DOTPROD) const int32x4_t vzero = vdupq_n_s32(0); +#endif int8x16x2_t q2bytes; uint8_t aux[16]; @@ -1608,7 +1630,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri #ifdef __ARM_NEON const uint8x16_t m3 = vdupq_n_u8(0x3); +#if defined(__ARM_FEATURE_DOTPROD) const int32x4_t vzero = vdupq_n_s32(0); +#endif int8x16x4_t q2bytes; @@ -2592,8 +2616,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict q4 = x[i].qs; const int8_t * restrict q8 = y[i].qs; - //int32x4_t isum = mzero; - int32_t sumi1 = 0; int32_t sumi2 = 0; @@ -3092,9 +3114,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri #ifdef __ARM_NEON const uint8x16_t m4b = vdupq_n_u8(0xf); - const int32x4_t mzero = vdupq_n_s32(0); const uint8x16_t mone = vdupq_n_u8(1); const uint8x16_t mtwo = vdupq_n_u8(2); +#if defined(__ARM_FEATURE_DOTPROD) + const int32x4_t mzero = vdupq_n_s32(0); +#endif int8x16x4_t q5bytes; @@ -3437,8 +3461,10 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri #ifdef __ARM_NEON const uint8x16_t m4b = vdupq_n_u8(0xf); - const int32x4_t mzero = vdupq_n_s32(0); const uint8x16_t mh = vdupq_n_u8(16); +#if defined(__ARM_FEATURE_DOTPROD) + const int32x4_t mzero = vdupq_n_s32(0); +#endif int8x16x4_t q5bytes; uint8x16x4_t q5h; @@ -3656,7 +3682,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri float sum = 0; const uint8x16_t m4b = vdupq_n_u8(0xF); +#if defined(__ARM_FEATURE_DOTPROD) const int32x4_t vzero = vdupq_n_s32(0); +#endif //const int8x16_t m32s = vdupq_n_s8(32); const uint8x16_t mone = vdupq_n_u8(3); @@ -4045,8 +4073,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri float sum = 0; const uint8x16_t m4b = vdupq_n_u8(0xF); - const int32x4_t vzero = vdupq_n_s32(0); const int8x16_t m32s = vdupq_n_s8(32); +#if defined(__ARM_FEATURE_DOTPROD) + const int32x4_t vzero = vdupq_n_s32(0); +#endif const uint8x16_t mone = vdupq_n_u8(3); From 8b56b4f2c396eae1f4417e5a859557fed989e0ee Mon Sep 17 00:00:00 2001 From: Karsten Weiss Date: Sat, 2 Sep 2023 14:29:09 +0200 Subject: [PATCH 128/147] metal : show all Metal device instances in the system (#2952) * ggml_metal_init: Show all Metal device instances in the system Also show the default Metal device that was picked. * Update ggml-metal.m --------- Co-authored-by: Georgi Gerganov --- ggml-metal.m | 18 ++++++++++++++++-- 1 file changed, 16 insertions(+), 2 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 4267db9be..88e7e1356 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -116,10 +116,24 @@ static NSString * const msl_library_source = @"see metal.metal"; struct ggml_metal_context * ggml_metal_init(int n_cb) { metal_printf("%s: allocating\n", __func__); - struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context)); + // Show all the Metal device instances in the system + NSArray * devices = MTLCopyAllDevices(); + id device; + NSString * s; + for (device in devices) { + s = [device name]; + metal_printf("%s: found device: %s\n", __func__, [s UTF8String]); + } + // Pick and show default Metal device + device = MTLCreateSystemDefaultDevice(); + s = [device name]; + metal_printf("%s: picking default device: %s\n", __func__, [s UTF8String]); + + // Configure context + struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context)); + ctx->device = device; ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS); - ctx->device = MTLCreateSystemDefaultDevice(); ctx->queue = [ctx->device newCommandQueue]; ctx->n_buffers = 0; ctx->concur_list_len = 0; From 52315a421674ff64305dbf082f69e4ec77f0a3f3 Mon Sep 17 00:00:00 2001 From: bandoti <141645996+bandoti@users.noreply.github.com> Date: Sat, 2 Sep 2023 09:53:18 -0300 Subject: [PATCH 129/147] readme : update clblast instructions (#2903) * Update Windows CLBlast instructions * Update Windows CLBlast instructions * Remove trailing whitespace --- README.md | 40 +++++++++++++++++++++++++++++++++++----- 1 file changed, 35 insertions(+), 5 deletions(-) diff --git a/README.md b/README.md index 4e6a0957d..5eda5f006 100644 --- a/README.md +++ b/README.md @@ -464,6 +464,8 @@ Building the program with BLAS support may lead to some performance improvements You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK). - For Ubuntu or Debian, the packages `opencl-headers`, `ocl-icd` may be needed. + - For Windows, a pre-built SDK is available on the [OpenCL Releases](https://github.com/KhronosGroup/OpenCL-SDK/releases) page. + -
    Installing the OpenCL SDK from source @@ -481,10 +483,27 @@ Building the program with BLAS support may lead to some performance improvements ```
    - Installing CLBlast: it may be found in your operating system's packages. + ##### Installing CLBlast + + Pre-built CLBlast binaries may be found on the [CLBlast Releases](https://github.com/CNugteren/CLBlast/releases) page. For Unix variants, it may also be found in your operating system's packages. + + Alternatively, they may be built from source. -
    - If not, then installing from source: + Windows: + + ```cmd + set OPENCL_SDK_ROOT="C:/OpenCL-SDK-v2023.04.17-Win-x64" + git clone https://github.com/CNugteren/CLBlast.git + mkdir CLBlast\build + cd CLBlast\build + cmake .. -DBUILD_SHARED_LIBS=OFF -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT=%OPENCL_SDK_ROOT% -G "Visual Studio 17 2022" -A x64 + cmake --build . --config Release + cmake --install . --prefix C:/CLBlast + ``` + + -
    + Unix: ```sh git clone https://github.com/CNugteren/CLBlast.git @@ -498,21 +517,32 @@ Building the program with BLAS support may lead to some performance improvements Where `/some/path` is where the built library will be installed (default is `/usr/local`).
    - Building: + ##### Building Llama with CLBlast - Build with make: ```sh make LLAMA_CLBLAST=1 ``` - - CMake: + - CMake (Unix): ```sh mkdir build cd build cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_dir=/some/path cmake --build . --config Release ``` + - CMake (Windows): + ```cmd + set CL_BLAST_CMAKE_PKG="C:/CLBlast/lib/cmake/CLBlast" + git clone https://github.com/ggerganov/llama.cpp + cd llama.cpp + mkdir build + cd build + cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=%CL_BLAST_CMAKE_PKG% -G "Visual Studio 17 2022" -A x64 + cmake --build . --config Release + cmake --install . --prefix C:/LlamaCPP + ``` - Running: + ##### Running Llama with CLBlast The CLBlast build supports `--gpu-layers|-ngl` like the CUDA version does. From 3358c381f6251bf6e65855e1c93bfaa9ec82ddb3 Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Sat, 2 Sep 2023 11:53:55 -0600 Subject: [PATCH 130/147] logging: Fix creating empty file even when disabled (#2966) * logging: Fix creating empty file even when disabled * Minor formatting fix Co-authored-by: staviq --------- Co-authored-by: staviq --- common/log.h | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/common/log.h b/common/log.h index bf9fafd68..0b9b01052 100644 --- a/common/log.h +++ b/common/log.h @@ -341,14 +341,14 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri } } + if (_disabled) + { + // Log is disabled + return nullptr; + } + if (_initialized) { - if (_disabled) - { - // Log is disabled - return nullptr; - } - // with fallback in case something went wrong return logfile ? logfile : stderr; } From bc054af97ac68a4b726e972cb283eb9565253ed5 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Sun, 3 Sep 2023 01:26:59 -0400 Subject: [PATCH 131/147] make : support overriding CFLAGS/CXXFLAGS/CPPFLAGS/LDFLAGS (#2886) * make : remove unused -DGGML_BIG_ENDIAN * make : put preprocessor stuff in CPPFLAGS * make : pass Raspberry Pi arch flags to g++ as well * make : support overriding CFLAGS/CXXFLAGS/CPPFLAGS/LDFLAGS * make : fix inverted conditional --- Makefile | 150 ++++++++++++++++++++++++------------------------------- 1 file changed, 66 insertions(+), 84 deletions(-) diff --git a/Makefile b/Makefile index 23f050c0d..e214970f8 100644 --- a/Makefile +++ b/Makefile @@ -67,21 +67,21 @@ OPT = -Ofast else OPT = -O3 endif -CFLAGS = -I. $(OPT) -std=c11 -fPIC -CXXFLAGS = -I. -I./common $(OPT) -std=c++11 -fPIC -LDFLAGS = +MK_CPPFLAGS = -I. -Icommon +MK_CFLAGS = $(CPPFLAGS) $(OPT) -std=c11 -fPIC +MK_CXXFLAGS = $(CPPFLAGS) $(OPT) -std=c++11 -fPIC +MK_LDFLAGS = ifdef LLAMA_DEBUG - CFLAGS += -O0 -g - CXXFLAGS += -O0 -g - LDFLAGS += -g + MK_CFLAGS += -O0 -g + MK_CXXFLAGS += -O0 -g + MK_LDFLAGS += -g else - CFLAGS += -DNDEBUG - CXXFLAGS += -DNDEBUG + MK_CPPFLAGS += -DNDEBUG endif ifdef LLAMA_SERVER_VERBOSE - CXXFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE) + MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE) endif ifdef LLAMA_DISABLE_LOGS @@ -90,9 +90,9 @@ ifdef LLAMA_DISABLE_LOGS endif # LLAMA_DISABLE_LOGS # warnings -CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \ - -Wmissing-prototypes -Werror=implicit-int -Wno-unused-function -CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar +MK_CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith \ + -Wmissing-prototypes -Werror=implicit-int -Wno-unused-function +MK_CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar ifeq '' '$(findstring clang++,$(CXX))' # g++ only @@ -101,29 +101,9 @@ endif # OS specific # TODO: support Windows -ifeq ($(UNAME_S),Linux) - CFLAGS += -pthread - CXXFLAGS += -pthread -endif -ifeq ($(UNAME_S),Darwin) - CFLAGS += -pthread - CXXFLAGS += -pthread -endif -ifeq ($(UNAME_S),FreeBSD) - CFLAGS += -pthread - CXXFLAGS += -pthread -endif -ifeq ($(UNAME_S),NetBSD) - CFLAGS += -pthread - CXXFLAGS += -pthread -endif -ifeq ($(UNAME_S),OpenBSD) - CFLAGS += -pthread - CXXFLAGS += -pthread -endif -ifeq ($(UNAME_S),Haiku) - CFLAGS += -pthread - CXXFLAGS += -pthread +ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)' + MK_CFLAGS += -pthread + MK_CXXFLAGS += -pthread endif # detect Windows @@ -149,12 +129,11 @@ ifeq ($(_WIN32),1) endif ifdef LLAMA_GPROF - CFLAGS += -pg - CXXFLAGS += -pg + MK_CFLAGS += -pg + MK_CXXFLAGS += -pg endif ifdef LLAMA_PERF - CFLAGS += -DGGML_PERF - CXXFLAGS += -DGGML_PERF + MK_CPPFLAGS += -DGGML_PERF endif # Architecture specific @@ -165,16 +144,16 @@ ifndef RISCV ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64)) # Use all CPU extensions that are available: - CFLAGS += -march=native -mtune=native - CXXFLAGS += -march=native -mtune=native + MK_CFLAGS += -march=native -mtune=native + MK_CXXFLAGS += -march=native -mtune=native # Usage AVX-only - #CFLAGS += -mfma -mf16c -mavx - #CXXFLAGS += -mfma -mf16c -mavx + #MK_CFLAGS += -mfma -mf16c -mavx + #MK_CXXFLAGS += -mfma -mf16c -mavx # Usage SSSE3-only (Not is SSE3!) - #CFLAGS += -mssse3 - #CXXFLAGS += -mssse3 + #MK_CFLAGS += -mssse3 + #MK_CXXFLAGS += -mssse3 endif # The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves. @@ -188,34 +167,33 @@ endif ifneq ($(filter aarch64%,$(UNAME_M)),) # Apple M1, M2, etc. # Raspberry Pi 3, 4, Zero 2 (64-bit) - CFLAGS += -mcpu=native - CXXFLAGS += -mcpu=native + MK_CFLAGS += -mcpu=native + MK_CXXFLAGS += -mcpu=native endif ifneq ($(filter armv6%,$(UNAME_M)),) # Raspberry Pi 1, Zero - CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access + MK_CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access + MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access endif ifneq ($(filter armv7%,$(UNAME_M)),) # Raspberry Pi 2 - CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations + MK_CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations + MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations endif ifneq ($(filter armv8%,$(UNAME_M)),) # Raspberry Pi 3, 4, Zero 2 (32-bit) - CFLAGS += -mfp16-format=ieee -mno-unaligned-access + MK_CFLAGS += -mfp16-format=ieee -mno-unaligned-access + MK_CXXFLAGS += -mfp16-format=ieee -mno-unaligned-access endif ifneq ($(filter ppc64%,$(UNAME_M)),) POWER9_M := $(shell grep "POWER9" /proc/cpuinfo) ifneq (,$(findstring POWER9,$(POWER9_M))) - CFLAGS += -mcpu=power9 - CXXFLAGS += -mcpu=power9 - endif - # Require c++23's std::byteswap for big-endian support. - ifeq ($(UNAME_M),ppc64) - CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN + MK_CFLAGS += -mcpu=power9 + MK_CXXFLAGS += -mcpu=power9 endif endif @@ -225,12 +203,10 @@ else endif ifndef LLAMA_NO_K_QUANTS - CFLAGS += -DGGML_USE_K_QUANTS - CXXFLAGS += -DGGML_USE_K_QUANTS + MK_CPPFLAGS += -DGGML_USE_K_QUANTS OBJS += k_quants.o ifdef LLAMA_QKK_64 - CFLAGS += -DGGML_QKK_64 - CXXFLAGS += -DGGML_QKK_64 + MK_CPPFLAGS += -DGGML_QKK_64 endif endif @@ -238,31 +214,32 @@ ifndef LLAMA_NO_ACCELERATE # Mac M1 - include Accelerate framework. # `-framework Accelerate` works on Mac Intel as well, with negliable performance boost (as of the predict time). ifeq ($(UNAME_S),Darwin) - CFLAGS += -DGGML_USE_ACCELERATE - LDFLAGS += -framework Accelerate + MK_CPPFLAGS += -DGGML_USE_ACCELERATE + MK_LDFLAGS += -framework Accelerate endif endif # LLAMA_NO_ACCELERATE ifdef LLAMA_MPI - CFLAGS += -DGGML_USE_MPI -Wno-cast-qual - CXXFLAGS += -DGGML_USE_MPI -Wno-cast-qual + MK_CPPFLAGS += -DGGML_USE_MPI + MK_CFLAGS += -Wno-cast-qual + MK_CXXFLAGS += -Wno-cast-qual OBJS += ggml-mpi.o endif # LLAMA_MPI ifdef LLAMA_OPENBLAS - CFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags openblas) - LDFLAGS += $(shell pkg-config --libs openblas) + MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas) + MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas) + MK_LDFLAGS += $(shell pkg-config --libs openblas) endif # LLAMA_OPENBLAS ifdef LLAMA_BLIS - CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis - LDFLAGS += -lblis -L/usr/local/lib + MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis + MK_LDFLAGS += -lblis -L/usr/local/lib endif # LLAMA_BLIS ifdef LLAMA_CUBLAS - CFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include - CXXFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include - LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib + MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include + MK_LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib OBJS += ggml-cuda.o NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math ifdef LLAMA_CUDA_NVCC @@ -313,14 +290,15 @@ endif # LLAMA_CUBLAS ifdef LLAMA_CLBLAST - CFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL) - CXXFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL) + MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL) + MK_CFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL) + MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL) # Mac provides OpenCL as a framework ifeq ($(UNAME_S),Darwin) - LDFLAGS += -lclblast -framework OpenCL + MK_LDFLAGS += -lclblast -framework OpenCL else - LDFLAGS += $(shell pkg-config --libs clblast OpenCL) + MK_LDFLAGS += $(shell pkg-config --libs clblast OpenCL) endif OBJS += ggml-opencl.o @@ -335,10 +313,9 @@ ifdef LLAMA_HIPBLAS LLAMA_CUDA_DMMV_X ?= 32 LLAMA_CUDA_MMV_Y ?= 1 LLAMA_CUDA_KQUANTS_ITER ?= 2 - CFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS - CXXFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS - LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib - LDFLAGS += -lhipblas -lamdhip64 -lrocblas + MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS + MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib + MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas HIPFLAGS += $(addprefix --offload-arch=,$(GPU_TARGETS)) HIPFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X) HIPFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y) @@ -353,10 +330,9 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h endif # LLAMA_HIPBLAS ifdef LLAMA_METAL - CFLAGS += -DGGML_USE_METAL #-DGGML_METAL_NDEBUG - CXXFLAGS += -DGGML_USE_METAL - LDFLAGS += -framework Foundation -framework Metal -framework MetalKit - OBJS += ggml-metal.o + MK_CPPFLAGS += -DGGML_USE_METAL #-DGGML_METAL_NDEBUG + MK_LDFLAGS += -framework Foundation -framework Metal -framework MetalKit + OBJS += ggml-metal.o endif # LLAMA_METAL ifdef LLAMA_METAL @@ -369,11 +345,17 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h $(CC) $(CFLAGS) -c $< -o $@ endif # LLAMA_MPI -ifdef LLAMA_NO_K_QUANTS +ifndef LLAMA_NO_K_QUANTS k_quants.o: k_quants.c k_quants.h $(CC) $(CFLAGS) -c $< -o $@ endif # LLAMA_NO_K_QUANTS +# combine build flags with cmdline overrides +override CPPFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) +override CFLAGS := $(MK_CFLAGS) $(CFLAGS) +override CXXFLAGS := $(MK_CXXFLAGS) $(CXXFLAGS) +override LDFLAGS := $(MK_LDFLAGS) $(LDFLAGS) + # # Print build information # From 2753415afdaf22a18c49608bd9d93cfffc05d435 Mon Sep 17 00:00:00 2001 From: kchro3 <62481661+kchro3@users.noreply.github.com> Date: Sat, 2 Sep 2023 22:27:25 -0700 Subject: [PATCH 132/147] swift : add missing c file to Package.swift (#2978) --- Package.swift | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Package.swift b/Package.swift index 73d027c70..2c07710cd 100644 --- a/Package.swift +++ b/Package.swift @@ -12,7 +12,7 @@ let package = Package( name: "llama", path: ".", exclude: ["ggml-metal.metal"], - sources: ["ggml.c", "llama.cpp"], + sources: ["ggml.c", "llama.cpp", "ggml-alloc.c"], publicHeadersPath: "spm-headers", cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")], linkerSettings: [ From c42f0ec6b344e14bd81c8612ab1445b3ff77358b Mon Sep 17 00:00:00 2001 From: momonga <115213907+mmnga@users.noreply.github.com> Date: Sun, 3 Sep 2023 14:36:28 +0900 Subject: [PATCH 133/147] examples : fix gpt-neox (#2943) Co-authored-by: mmnga --- examples/gptneox-wip/gptneox-main.cpp | 13 ++++---- llama.cpp | 46 +++++++++++++++++++++++++-- 2 files changed, 51 insertions(+), 8 deletions(-) diff --git a/examples/gptneox-wip/gptneox-main.cpp b/examples/gptneox-wip/gptneox-main.cpp index 04af50245..6291523f2 100644 --- a/examples/gptneox-wip/gptneox-main.cpp +++ b/examples/gptneox-wip/gptneox-main.cpp @@ -660,9 +660,10 @@ bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt2 ggml_tensor * gpt_neox_ff( const gpt_neox_block &block, ggml_context * ctx0, - ggml_tensor * inp) { + ggml_tensor * inp, + const gpt_neox_hparams &hparams) { - ggml_tensor * cur = ggml_norm(ctx0, inp); + ggml_tensor * cur = ggml_norm(ctx0, inp, hparams.norm_eps); cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, block.ln_2_g, cur), cur), ggml_repeat(ctx0, block.ln_2_b, cur)); cur = ggml_mul_mat(ctx0, block.c_mlp_fc_w, cur); @@ -753,7 +754,7 @@ bool gpt_neox_eval( // self-attention { { - cur = ggml_norm(ctx0, inpL); + cur = ggml_norm(ctx0, inpL, hparams.norm_eps); cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, model.blocks[il].ln_1_g, cur), cur), @@ -844,7 +845,7 @@ bool gpt_neox_eval( if (hparams.par_res == 0) { struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL); - cur = gpt_neox_ff(model.blocks[il], ctx0, inpFF); + cur = gpt_neox_ff(model.blocks[il], ctx0, inpFF, hparams); // input for next layer inpL = ggml_add(ctx0, cur, inpFF); @@ -853,7 +854,7 @@ bool gpt_neox_eval( // this is independent of the self-attention result, so it could be done in parallel to the self-attention // note here we pass inpL instead of cur - cur = gpt_neox_ff(model.blocks[il], ctx0, inpL); + cur = gpt_neox_ff(model.blocks[il], ctx0, inpL, hparams); // layer input + FF cur = ggml_add(ctx0, cur, inpFF); @@ -867,7 +868,7 @@ bool gpt_neox_eval( // norm { - inpL = ggml_norm(ctx0, inpL); + inpL = ggml_norm(ctx0, inpL, hparams.norm_eps); // inpL = ln_f_g*inpL + ln_f_b inpL = ggml_add(ctx0, diff --git a/llama.cpp b/llama.cpp index 3114d3311..2b0cf30f6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -325,6 +325,44 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_GPT2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, + }, + { + LLM_ARCH_GPTJ, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, + }, + { + LLM_ARCH_GPTNEOX, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_MPT, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, + }, + { + LLM_ARCH_UNKNOWN, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + }, + }, }; static llm_arch llm_arch_from_string(const std::string & name) { @@ -1605,9 +1643,13 @@ static void llm_load_hparams( GGUF_GET_KEY(ctx, hparams.n_rot, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ROPE_DIMENSION_COUNT)); - if (hparams.n_rot != hparams.n_embd / hparams.n_head) { - throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head)); + if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) { + if (hparams.n_rot != hparams.n_embd / hparams.n_head) { + throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head)); + } } + // gpt-neox n_rot = rotary_pct * (n_embd / n_head) + // gpt-j n_rot = rotary_dim } // arch-specific KVs From 340af42f09a80e32f4998857b4f0543e41124525 Mon Sep 17 00:00:00 2001 From: Ido S Date: Sun, 3 Sep 2023 08:50:51 +0300 Subject: [PATCH 134/147] docs : add `catai` to `README.md` (#2967) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 5eda5f006..0cfd94db4 100644 --- a/README.md +++ b/README.md @@ -120,6 +120,7 @@ as the main playground for developing new features for the [ggml](https://github - [nat/openplayground](https://github.com/nat/openplayground) - [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) +- [withcatai/catai](https://github.com/withcatai/catai) --- From cff7b0bf07cb46e1ad4fd199f6bdeb538925c8c4 Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Sat, 2 Sep 2023 23:52:13 -0600 Subject: [PATCH 135/147] convert.py : BPE fixes (#2938) * convert.py: BPE fixes? * Remove unnecessary conditional in addl token error handling --- convert.py | 32 ++++++++++++++++++++++++++++---- 1 file changed, 28 insertions(+), 4 deletions(-) diff --git a/convert.py b/convert.py index 6c89b5ecc..5a7483b43 100755 --- a/convert.py +++ b/convert.py @@ -323,15 +323,27 @@ class BpeVocab: self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read()) added_tokens: dict[str, int] if fname_added_tokens is not None: + # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab. added_tokens = json.load(open(fname_added_tokens, encoding="utf-8")) else: - added_tokens = {} + # Fall back to trying to find the added tokens in tokenizer.json + tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json' + if not tokenizer_json_file.is_file(): + added_tokens = {} + else: + tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8")) + added_tokens = dict( + (item['content'], item['id']) + for item in tokenizer_json.get('added_tokens', []) + # Added tokens here can be duplicates of the main vocabulary. + if item['content'] not in self.bpe_tokenizer ) vocab_size: int = len(self.bpe_tokenizer) expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) actual_ids = sorted(added_tokens.values()) if expected_ids != actual_ids: - raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}") + expected_end_id = vocab_size + len(actual_ids) - 1 + raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}") items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) self.added_tokens_list = [text for (text, idx) in items] @@ -345,10 +357,22 @@ class BpeVocab: from transformers.models.gpt2 import tokenization_gpt2 # type: ignore[import] byte_encoder = tokenization_gpt2.bytes_to_unicode() byte_decoder = {v: k for k, v in byte_encoder.items()} + score = 0.0 for i, item in enumerate(tokenizer): text: bytes = item.encode("utf-8") - score: float = -i - yield text, score, gguf.TokenType.USER_DEFINED + # FIXME: These shouldn't be hardcoded, but it's probably better than the current behavior? + if i <= 258 and text.startswith(b'<') and text.endswith(b'>'): + if i == 0 and text == b'': + toktype = gguf.TokenType.UNKNOWN + elif i == 1 or i == 2: + toktype = gguf.TokenType.CONTROL + elif i >= 3 and text.startswith(b'<0x'): + toktype = gguf.TokenType.BYTE + else: + toktype = gguf.TokenType.NORMAL + else: + toktype = gguf.TokenType.NORMAL + yield text, score, toktype def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: for text in self.added_tokens_list: From 6a31a3bd9806c85ed08266f6ab65181da0f30d03 Mon Sep 17 00:00:00 2001 From: kchro3 <62481661+kchro3@users.noreply.github.com> Date: Sat, 2 Sep 2023 23:21:05 -0700 Subject: [PATCH 136/147] swift : add support for k-quants (#2983) --- Package.swift | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-) diff --git a/Package.swift b/Package.swift index 2c07710cd..96f52c4f0 100644 --- a/Package.swift +++ b/Package.swift @@ -12,9 +12,18 @@ let package = Package( name: "llama", path: ".", exclude: ["ggml-metal.metal"], - sources: ["ggml.c", "llama.cpp", "ggml-alloc.c"], + sources: [ + "ggml.c", + "llama.cpp", + "ggml-alloc.c", + "k_quants.c" + ], publicHeadersPath: "spm-headers", - cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")], + cSettings: [ + .unsafeFlags(["-Wno-shorten-64-to-32"]), + .define("GGML_USE_K_QUANTS"), + .define("GGML_USE_ACCELERATE") + ], linkerSettings: [ .linkedFramework("Accelerate") ] From ca82cf7bac0c91d03e3d320b3a865dd006f854ac Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sun, 3 Sep 2023 11:06:22 +0300 Subject: [PATCH 137/147] metal : more optimizations (#2959) * Very minor speedup via simd-group synchronization in f16 x f32 * Another very minor speedup on metal * Quite significant PP speedup on metal * Another attempt * Minor * Massive improvement for TG for fp16 * ~4-5% improvement for Q8_0 TG on metal --------- Co-authored-by: Iwan Kawrakow Co-authored-by: Georgi Gerganov --- ggml-metal.m | 22 +++-- ggml-metal.metal | 226 ++++++++++++++++++++++++++++++----------------- 2 files changed, 163 insertions(+), 85 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 88e7e1356..d0d23442e 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -76,6 +76,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(rms_norm); GGML_METAL_DECL_KERNEL(norm); GGML_METAL_DECL_KERNEL(mul_mat_f16_f32); + GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row); GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32); GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32); GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32); @@ -219,6 +220,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(rms_norm); GGML_METAL_ADD_KERNEL(norm); GGML_METAL_ADD_KERNEL(mul_mat_f16_f32); + GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row); GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32); GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32); GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32); @@ -284,6 +286,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(rms_norm); GGML_METAL_DEL_KERNEL(norm); GGML_METAL_DEL_KERNEL(mul_mat_f16_f32); + GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row); GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32); GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32); GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32); @@ -868,7 +871,11 @@ void ggml_metal_graph_compute( { nth0 = 32; nth1 = 1; - [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32]; + if (ne11 * ne12 < 4) { + [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32_1row]; + } else { + [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32]; + } } break; case GGML_TYPE_Q4_0: { @@ -920,8 +927,8 @@ void ggml_metal_graph_compute( GGML_ASSERT(ne02 == 1); GGML_ASSERT(ne12 == 1); - nth0 = 2; - nth1 = 32; + nth0 = 4; //1; + nth1 = 8; //32; [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32]; } break; case GGML_TYPE_Q5_K: @@ -969,9 +976,12 @@ void ggml_metal_graph_compute( [encoder setBytes:&gqa length:sizeof(gqa) atIndex:17]; if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 || - src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) { + src0t == GGML_TYPE_Q2_K) {// || src0t == GGML_TYPE_Q4_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } + else if (src0t == GGML_TYPE_Q4_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } else if (src0t == GGML_TYPE_Q3_K) { #ifdef GGML_QKK_64 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; @@ -985,8 +995,8 @@ void ggml_metal_graph_compute( else if (src0t == GGML_TYPE_Q6_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else { - [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + int64_t ny = (ne11 + 3)/4; + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } } } break; diff --git a/ggml-metal.metal b/ggml-metal.metal index 8cdf0b9d2..3fa311b40 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -133,19 +133,24 @@ kernel void kernel_soft_max( threadgroup_barrier(mem_flags::mem_threadgroup); } - // broadcast - if (tpitg[0] == 0) { - buf[0] = buf[0]; - } + //// broadcast - not needed. There is a threadgroup barrier above in the last iteration of + // the loop, and when that is done, buf[0] has the correct (synchronized) value + //if (tpitg[0] == 0) { + // buf[0] = buf[0]; + //} - threadgroup_barrier(mem_flags::mem_threadgroup); + //threadgroup_barrier(mem_flags::mem_threadgroup); const float max = buf[0]; // parallel sum buf[tpitg[0]] = 0.0f; for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) { - buf[tpitg[0]] += exp(psrc0[i00] - max); + const float exp_psrc0 = exp(psrc0[i00] - max); + buf[tpitg[0]] += exp_psrc0; + // Remember the result of exp here. exp is expensive, so we really do not + // whish to compute it twice. + pdst[i00] = exp_psrc0; } // reduce @@ -157,17 +162,18 @@ kernel void kernel_soft_max( threadgroup_barrier(mem_flags::mem_threadgroup); } - // broadcast - if (tpitg[0] == 0) { - buf[0] = buf[0]; - } + // broadcast - not needed, see above + //// broadcast + //if (tpitg[0] == 0) { + // buf[0] = buf[0]; + //} - threadgroup_barrier(mem_flags::mem_threadgroup); + //threadgroup_barrier(mem_flags::mem_threadgroup); const float sum = buf[0]; for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) { - pdst[i00] = exp(psrc0[i00] - max) / sum; + pdst[i00] /= sum; } } @@ -214,25 +220,27 @@ kernel void kernel_norm( } threadgroup_barrier(mem_flags::mem_threadgroup); } - // broadcast - if (tpitg == 0) { - sum[0] /= ne00; - } - threadgroup_barrier(mem_flags::mem_threadgroup); + //// broadcast + //if (tpitg == 0) { + // sum[0] /= ne00; + //} + //threadgroup_barrier(mem_flags::mem_threadgroup); const float mean = sum[0]; - // recenter + // recenter and VARIANCE device float * y = dst + tgpig*ne00; - for (int i00 = tpitg; i00 < ne00; i00 += ntg) { - y[i00] = x[i00] - mean; - } - - // VARIANCE - // parallel sum sum[tpitg] = 0.0f; for (int i00 = tpitg; i00 < ne00; i00 += ntg) { + y[i00] = x[i00] - mean; sum[tpitg] += y[i00] * y[i00]; } + + //// VARIANCE + //// parallel sum + //sum[tpitg] = 0.0f; + //for (int i00 = tpitg; i00 < ne00; i00 += ntg) { + // sum[tpitg] += y[i00] * y[i00]; + //} // reduce threadgroup_barrier(mem_flags::mem_threadgroup); for (uint i = ntg/2; i > 0; i /= 2) { @@ -241,11 +249,11 @@ kernel void kernel_norm( } threadgroup_barrier(mem_flags::mem_threadgroup); } - // broadcast - if (tpitg == 0) { - sum[0] /= ne00; - } - threadgroup_barrier(mem_flags::mem_threadgroup); + //// broadcast + //if (tpitg == 0) { + // sum[0] /= ne00; + //} + //threadgroup_barrier(mem_flags::mem_threadgroup); const float variance = sum[0]; const float scale = 1.0f/sqrt(variance + eps); @@ -435,6 +443,8 @@ kernel void kernel_mul_mat_q4_1_f32( mul_vec_q_n_f32(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg); } +#define NB_Q8_0 8 + kernel void kernel_mul_mat_q8_0_f32( device const void * src0, device const float * src1, @@ -463,30 +473,30 @@ kernel void kernel_mul_mat_q8_0_f32( device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0; device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1; - float yl[16]; + float yl[NB_Q8_0]; float sumf[nr]={0.f}; - const int ix = tiisg/2; - const int il = tiisg%2; + const int ix = tiisg/4; + const int il = tiisg%4; - device const float * yb = y + ix * QK8_0 + 16*il; + device const float * yb = y + ix * QK8_0 + NB_Q8_0*il; - // each thread in a SIMD group deals with half a block. - for (int ib = ix; ib < nb; ib += nw/2) { - for (int i = 0; i < 16; ++i) { + // each thread in a SIMD group deals with NB_Q8_0 quants at a time + for (int ib = ix; ib < nb; ib += nw/4) { + for (int i = 0; i < NB_Q8_0; ++i) { yl[i] = yb[i]; } for (int row = 0; row < nr; row++) { - device const int8_t * qs = x[ib+row*nb].qs + 16*il; + device const int8_t * qs = x[ib+row*nb].qs + NB_Q8_0*il; float sumq = 0.f; - for (int iq = 0; iq < 16; ++iq) { + for (int iq = 0; iq < NB_Q8_0; ++iq) { sumq += qs[iq] * yl[iq]; } sumf[row] += sumq*x[ib+row*nb].d; } - yb += QK8_0 * 16; + yb += NB_Q8_0 * nw; } for (int row = 0; row < nr; ++row) { @@ -497,6 +507,60 @@ kernel void kernel_mul_mat_q8_0_f32( } } +kernel void kernel_mul_mat_f16_f32_1row( + device const char * src0, + device const char * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiisg[[thread_index_in_simdgroup]]) { + + const int64_t r0 = tgpig.x; + const int64_t r1 = tgpig.y; + const int64_t im = tgpig.z; + + device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); + device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); + + float sumf = 0; + if (ne00 < 128) { + for (int i = tiisg; i < ne00; i += 32) { + sumf += (float) x[i] * (float) y[i]; + } + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } else { + device const half4 * x4 = (device const half4 *) x; + device const float4 * y4 = (device const float4 *) y; + for (int i = tiisg; i < ne00/4; i += 32) { + for (int k = 0; k < 4; ++k) sumf += (float)x4[i][k] * y4[i][k]; + } + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) sumf += (float) x[i] * y[i]; + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + +} + +#define N_F16_F32 4 + kernel void kernel_mul_mat_f16_f32( device const char * src0, device const char * src1, @@ -515,55 +579,58 @@ kernel void kernel_mul_mat_f16_f32( constant uint64_t & nb12, constant int64_t & ne0, constant int64_t & ne1, - threadgroup float * sum [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], - uint3 tpig[[thread_position_in_grid]], - uint3 tpitg[[thread_position_in_threadgroup]], - uint3 tptg[[threads_per_threadgroup]]) { + uint tiisg[[thread_index_in_simdgroup]]) { const int64_t r0 = tgpig.x; - const int64_t r1 = tgpig.y; + const int64_t rb = N_F16_F32*tgpig.y; const int64_t im = tgpig.z; - device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); - device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); + device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); - uint ith = tpitg.x; - uint nth = tptg.x; + if (ne00 < 128) { + for (int row = 0; row < N_F16_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } - sum[ith] = 0.0f; + device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); - for (int i = ith; i < ne00; i += nth) { - sum[ith] += (float) x[i] * (float) y[i]; + float sumf = 0; + for (int i = tiisg; i < ne00; i += 32) { + sumf += (float) x[i] * (float) y[i]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } else { + device const half4 * x4 = (device const half4 *)x; + for (int row = 0; row < N_F16_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); + device const float4 * y4 = (device const float4 *) y; + + float sumf = 0; + for (int i = tiisg; i < ne00/4; i += 32) { + for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) sumf += (float) x[i] * y[i]; + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } } - // accumulate the sum from all threads in the threadgroup - threadgroup_barrier(mem_flags::mem_threadgroup); - if (ith%4 == 0) { - for (int i = 1; i < 4; ++i) sum[ith] += sum[ith + i]; - } - threadgroup_barrier(mem_flags::mem_threadgroup); - if (ith%16 == 0) { - for (int i = 4; i < 16; i += 4) sum[ith] += sum[ith + i]; - } - threadgroup_barrier(mem_flags::mem_threadgroup); - if (ith == 0) { - for (int i = 16; i < nth; i += 16) sum[0] += sum[i]; - dst[im*ne1*ne0 + r1*ne0 + r0] = sum[0]; - } - - // Original implementation. Left behind commented out for now - //threadgroup_barrier(mem_flags::mem_threadgroup); - //for (uint i = tptg.x/2; i > 0; i /= 2) { - // if (tpitg.x < i) { - // sum[tpitg.x] += sum[tpitg.x + i]; - // } - // threadgroup_barrier(mem_flags::mem_threadgroup); - //} - // - //if (tpitg.x == 0) { - // dst[im*ne1*ne0 + r1*ne0 + r0] = sum[0]; - //} } kernel void kernel_alibi_f32( @@ -1262,7 +1329,8 @@ kernel void kernel_mul_mat_q4_K_f32( const int r0 = tgpig.x; const int r1 = tgpig.y; const int r2 = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = r0 * N_DST; const int ib_row = first_row * nb; const uint offset0 = r2/gqa*(nb*ne0); device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0; From 6460f758dbd472653296044d36bed8c4554988f5 Mon Sep 17 00:00:00 2001 From: Wentai Zhang Date: Sun, 3 Sep 2023 16:46:44 +0800 Subject: [PATCH 138/147] opencl : fix a bug in ggml_cl_pool_malloc() for ggml_cl_mul_mat_f32() (#2955) Co-authored-by: Wentai Zhang --- ggml-opencl.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml-opencl.cpp b/ggml-opencl.cpp index eb214a836..3d50a7f08 100644 --- a/ggml-opencl.cpp +++ b/ggml-opencl.cpp @@ -1493,7 +1493,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr if (src0->backend == GGML_BACKEND_GPU) { // NOLINT d_X = (cl_mem) src0->data; } else { - d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size); + d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size); } cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size); cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size); From afc43d5f82588d2ed71ea104e8262f5e5da13980 Mon Sep 17 00:00:00 2001 From: Alon Date: Sun, 3 Sep 2023 11:48:49 +0300 Subject: [PATCH 139/147] cov : add Code Coverage and codecov.io integration (#2928) * update .gitignore * makefile: add coverage support (lcov, gcovr) * add code-coverage workflow * update code coverage workflow * wun on ubuntu 20.04 * use gcc-8 * check why the job hang * add env vars * add LLAMA_CODE_COVERAGE=1 again * - add CODECOV_TOKEN - add missing make lcov-report * install lcov * update make file -pb flag * remove unused GGML_NITER from workflows * wrap coverage output files in COV_TARGETS --- .github/workflows/build.yml | 1 - .github/workflows/code-coverage.yml | 36 +++++++++++++++++++++++++++++ .gitignore | 7 ++++++ Makefile | 22 +++++++++++++++++- 4 files changed, 64 insertions(+), 2 deletions(-) create mode 100644 .github/workflows/code-coverage.yml diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 20fd8c2b5..9d0a6c222 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -18,7 +18,6 @@ on: env: BRANCH_NAME: ${{ github.head_ref || github.ref_name }} GGML_NLOOP: 3 - GGML_NITER: 1 GGML_N_THREADS: 1 jobs: diff --git a/.github/workflows/code-coverage.yml b/.github/workflows/code-coverage.yml new file mode 100644 index 000000000..392db8a08 --- /dev/null +++ b/.github/workflows/code-coverage.yml @@ -0,0 +1,36 @@ +name: Code Coverage +on: [push, pull_request] + +env: + GGML_NLOOP: 3 + GGML_N_THREADS: 1 + +jobs: + run: + runs-on: ubuntu-20.04 + steps: + - name: Checkout + uses: actions/checkout@v3 + + - name: Dependencies + run: | + sudo apt-get update + sudo apt-get install build-essential gcc-8 lcov + + - name: Build + run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests + + - name: Run tests + run: CC=gcc-8 make test + + - name: Generate coverage report + run: | + make coverage + make lcov-report + + - name: Upload coverage to Codecov + uses: codecov/codecov-action@v3 + env: + CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }} + with: + files: lcov-report/coverage.info diff --git a/.gitignore b/.gitignore index 8b5f45a2d..f9244fadc 100644 --- a/.gitignore +++ b/.gitignore @@ -6,6 +6,10 @@ *.exe *.dll *.log +*.gcov +*.gcno +*.gcda +*.dot .DS_Store .build/ .cache/ @@ -17,6 +21,9 @@ .vs/ .vscode/ +lcov-report/ +gcovr-report/ + build*/ out/ tmp/ diff --git a/Makefile b/Makefile index e214970f8..c042bf0e5 100644 --- a/Makefile +++ b/Makefile @@ -4,6 +4,9 @@ BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-tex # Binaries only useful for tests TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1 +# Code coverage output files +COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report + default: $(BUILD_TARGETS) test: @@ -23,6 +26,18 @@ test: all: $(BUILD_TARGETS) $(TEST_TARGETS) +coverage: ## Run code coverage + gcov -pb tests/*.cpp + +lcov-report: coverage ## Generate lcov report + mkdir -p lcov-report + lcov --capture --directory . --output-file lcov-report/coverage.info + genhtml lcov-report/coverage.info --output-directory lcov-report + +gcovr-report: coverage ## Generate gcovr report + mkdir -p gcovr-report + gcovr --root . --html --html-details --output gcovr-report/coverage.html + ifndef UNAME_S UNAME_S := $(shell uname -s) endif @@ -84,6 +99,11 @@ ifdef LLAMA_SERVER_VERBOSE MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE) endif + +ifdef LLAMA_CODE_COVERAGE + CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase '' +endif + ifdef LLAMA_DISABLE_LOGS CFLAGS += -DLOG_DISABLE_LOGS CXXFLAGS += -DLOG_DISABLE_LOGS @@ -399,7 +419,7 @@ libllama.so: llama.o ggml.o $(OBJS) $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) clean: - rm -vf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h $(BUILD_TARGETS) $(TEST_TARGETS) + rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS) # # Examples From d9151e6f570eb20bfd54427bd8a337d9b1a08018 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 3 Sep 2023 12:40:56 +0300 Subject: [PATCH 140/147] metal : revert 6af0bab until we fix it This restores the generated text to be the same as before #2959 --- ggml-metal.metal | 74 ++++++++++++------------------------------------ 1 file changed, 18 insertions(+), 56 deletions(-) diff --git a/ggml-metal.metal b/ggml-metal.metal index 3fa311b40..1d324e466 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -536,27 +536,14 @@ kernel void kernel_mul_mat_f16_f32_1row( device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); float sumf = 0; - if (ne00 < 128) { - for (int i = tiisg; i < ne00; i += 32) { - sumf += (float) x[i] * (float) y[i]; - } - float all_sum = simd_sum(sumf); - if (tiisg == 0) { - dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; - } - } else { - device const half4 * x4 = (device const half4 *) x; - device const float4 * y4 = (device const float4 *) y; - for (int i = tiisg; i < ne00/4; i += 32) { - for (int k = 0; k < 4; ++k) sumf += (float)x4[i][k] * y4[i][k]; - } - float all_sum = simd_sum(sumf); - if (tiisg == 0) { - for (int i = 4*(ne00/4); i < ne00; ++i) sumf += (float) x[i] * y[i]; - dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; - } + for (int i = tiisg; i < ne00; i += 32) { + sumf += (float) x[i] * (float) y[i]; } + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } } #define N_F16_F32 4 @@ -588,49 +575,24 @@ kernel void kernel_mul_mat_f16_f32( device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); - if (ne00 < 128) { - for (int row = 0; row < N_F16_F32; ++row) { - int r1 = rb + row; - if (r1 >= ne11) { - break; - } - - device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); - - float sumf = 0; - for (int i = tiisg; i < ne00; i += 32) { - sumf += (float) x[i] * (float) y[i]; - } - - float all_sum = simd_sum(sumf); - if (tiisg == 0) { - dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; - } + for (int row = 0; row < N_F16_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; } - } else { - device const half4 * x4 = (device const half4 *)x; - for (int row = 0; row < N_F16_F32; ++row) { - int r1 = rb + row; - if (r1 >= ne11) { - break; - } - device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); - device const float4 * y4 = (device const float4 *) y; + device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); - float sumf = 0; - for (int i = tiisg; i < ne00/4; i += 32) { - for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k]; - } + float sumf = 0; + for (int i = tiisg; i < ne00; i += 32) { + sumf += (float) x[i] * (float) y[i]; + } - float all_sum = simd_sum(sumf); - if (tiisg == 0) { - for (int i = 4*(ne00/4); i < ne00; ++i) sumf += (float) x[i] * y[i]; - dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; - } + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; } } - } kernel void kernel_alibi_f32( From 37301347767d555d0a66c043ce4ef6ead8e61c55 Mon Sep 17 00:00:00 2001 From: opparco Date: Sun, 3 Sep 2023 19:18:09 +0900 Subject: [PATCH 141/147] llama : fix bpe tokenize from byte (#2889) --- llama.cpp | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) diff --git a/llama.cpp b/llama.cpp index 2b0cf30f6..c97c1462f 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3366,9 +3366,15 @@ struct llm_tokenizer_bpe { std::string byte_str(1, *j); auto token_multibyte = vocab.token_to_id.find(byte_str); if (token_multibyte == vocab.token_to_id.end()) { - fprintf(stderr,"ERROR: byte not found in vocab: '%s'\n", byte_str.c_str()); + try { + llama_token token_byte = llama_byte_to_token(vocab, *j); + output.push_back(token_byte); + } catch (const std::out_of_range & err) { + fprintf(stderr,"ERROR: byte not found in vocab: '%s'\n", byte_str.c_str()); + } + } else { + output.push_back((*token_multibyte).second); } - output.push_back((*token_multibyte).second); } } else { output.push_back((*token).second); From 73a12a6344d5da4d8e2eba5d12221b8bc6895931 Mon Sep 17 00:00:00 2001 From: Alon Date: Sun, 3 Sep 2023 13:19:01 +0300 Subject: [PATCH 142/147] cov : disable comment in PRs (#2989) --- codecov.yml | 14 ++++++++++++++ 1 file changed, 14 insertions(+) create mode 100644 codecov.yml diff --git a/codecov.yml b/codecov.yml new file mode 100644 index 000000000..a301c5b2c --- /dev/null +++ b/codecov.yml @@ -0,0 +1,14 @@ +comment: off + +coverage: + status: + project: + default: + target: auto + threshold: 0 + base: auto + patch: + default: + target: auto + threshold: 0 + base: auto From b7f2aa9e512c3be2e863d877cbb1056d7c4a03f8 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 3 Sep 2023 13:23:33 +0300 Subject: [PATCH 143/147] metal : restore 363f0bf and fix reduce in F16_F32 kernels (#2986) --- ggml-metal.metal | 76 ++++++++++++++++++++++++++++++++++++------------ 1 file changed, 57 insertions(+), 19 deletions(-) diff --git a/ggml-metal.metal b/ggml-metal.metal index 1d324e466..119fcbeb6 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -536,14 +536,27 @@ kernel void kernel_mul_mat_f16_f32_1row( device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); float sumf = 0; - for (int i = tiisg; i < ne00; i += 32) { - sumf += (float) x[i] * (float) y[i]; + if (ne00 < 128) { + for (int i = tiisg; i < ne00; i += 32) { + sumf += (float) x[i] * (float) y[i]; + } + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } else { + device const half4 * x4 = (device const half4 *) x; + device const float4 * y4 = (device const float4 *) y; + for (int i = tiisg; i < ne00/4; i += 32) { + for (int k = 0; k < 4; ++k) sumf += (float)x4[i][k] * y4[i][k]; + } + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i]; + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } } - float all_sum = simd_sum(sumf); - if (tiisg == 0) { - dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; - } } #define N_F16_F32 4 @@ -570,29 +583,54 @@ kernel void kernel_mul_mat_f16_f32( uint tiisg[[thread_index_in_simdgroup]]) { const int64_t r0 = tgpig.x; - const int64_t rb = N_F16_F32*tgpig.y; + const int64_t rb = tgpig.y*N_F16_F32; const int64_t im = tgpig.z; device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); - for (int row = 0; row < N_F16_F32; ++row) { - int r1 = rb + row; - if (r1 >= ne11) { - break; + if (ne00 < 128) { + for (int row = 0; row < N_F16_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); + + float sumf = 0; + for (int i = tiisg; i < ne00; i += 32) { + sumf += (float) x[i] * (float) y[i]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } } + } else { + device const half4 * x4 = (device const half4 *)x; + for (int row = 0; row < N_F16_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } - device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); + device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12); + device const float4 * y4 = (device const float4 *) y; - float sumf = 0; - for (int i = tiisg; i < ne00; i += 32) { - sumf += (float) x[i] * (float) y[i]; - } + float sumf = 0; + for (int i = tiisg; i < ne00/4; i += 32) { + for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k]; + } - float all_sum = simd_sum(sumf); - if (tiisg == 0) { - dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i]; + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } } } + } kernel void kernel_alibi_f32( From 6519e9c99cffbad19b31bcba86df48c500628c09 Mon Sep 17 00:00:00 2001 From: Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com> Date: Sun, 3 Sep 2023 04:38:43 -0600 Subject: [PATCH 144/147] gguf(python): Fix special vocab handling when id < 0 (#2984) --- gguf-py/gguf/gguf.py | 4 ++-- gguf-py/pyproject.toml | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py index b1bc4205b..d377cd56d 100644 --- a/gguf-py/gguf/gguf.py +++ b/gguf-py/gguf/gguf.py @@ -801,7 +801,7 @@ class SpecialVocab: else: continue for maybe_token_id in (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content): - if isinstance(maybe_token_id, int): + if isinstance(maybe_token_id, int) and maybe_token_id >= 0: self.special_token_ids[typ] = maybe_token_id break return True @@ -814,7 +814,7 @@ class SpecialVocab: config = json.load(f) for typ in self.special_token_types: maybe_token_id = config.get(f'{typ}_token_id') - if isinstance(maybe_token_id, int): + if isinstance(maybe_token_id, int) and maybe_token_id >= 0: self.special_token_ids[typ] = maybe_token_id return True diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index 26f792b14..8da60de1b 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.3.1" +version = "0.3.2" description = "Write ML models in GGUF for GGML" authors = ["GGML "] packages = [ From 8f429fa5111901f9646cf998643ac5310846d487 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 3 Sep 2023 13:42:56 +0300 Subject: [PATCH 145/147] perplexity : fix ETA by warming up the model with an empty run --- common/common.cpp | 8 ++++++++ examples/main/main.cpp | 8 -------- 2 files changed, 8 insertions(+), 8 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 41fc59ced..a1c3dc780 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -752,6 +752,14 @@ std::tuple llama_init_from_gpt_par params.logit_bias[llama_token_eos(lctx)] = -INFINITY; } + { + LOG("warming up the model with an empty run\n"); + + const std::vector tmp = { llama_token_bos(lctx), }; + llama_eval(lctx, tmp.data(), tmp.size(), 0, params.n_threads); + llama_reset_timings(lctx); + } + return std::make_tuple(model, lctx); } diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 7117db4b0..db98312ca 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -492,14 +492,6 @@ int main(int argc, char ** argv) { std::vector embd; std::vector embd_guidance; - { - LOG("warming up the model with an empty run\n"); - - const std::vector tmp = { llama_token_bos(ctx), }; - llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads); - llama_reset_timings(ctx); - } - while ((n_remain != 0 && !is_antiprompt) || params.interactive) { // predict if (embd.size() > 0) { From 47068e517004d90f13c16352bb3b4cafd53a00cd Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 3 Sep 2023 15:12:08 +0300 Subject: [PATCH 146/147] speculative : PoC for speeding-up inference via speculative sampling (#2926) * speculative : initial example * speculative : print encoding speed * speculative : add --draft CLI arg --- common/common.cpp | 140 ++++++++++++++++ common/common.h | 36 +++++ examples/CMakeLists.txt | 1 + examples/main/main.cpp | 136 +++------------- examples/speculative/CMakeLists.txt | 8 + examples/speculative/speculative.cpp | 234 +++++++++++++++++++++++++++ 6 files changed, 440 insertions(+), 115 deletions(-) create mode 100644 examples/speculative/CMakeLists.txt create mode 100644 examples/speculative/speculative.cpp diff --git a/common/common.cpp b/common/common.cpp index a1c3dc780..313821375 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -305,6 +305,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.n_keep = std::stoi(argv[i]); + } else if (arg == "--draft") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.n_draft = std::stoi(argv[i]); } else if (arg == "--chunks") { if (++i >= argc) { invalid_param = true; @@ -317,6 +323,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.model = argv[i]; + } else if (arg == "-md" || arg == "--model-draft") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.model_draft = argv[i]; } else if (arg == "-a" || arg == "--alias") { if (++i >= argc) { invalid_param = true; @@ -638,6 +650,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stdout, " --hellaswag compute HellaSwag score over random tasks from datafile supplied with -f\n"); fprintf(stdout, " --hellaswag-tasks N number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks); fprintf(stdout, " --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); + fprintf(stdout, " --draft N number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft); fprintf(stdout, " --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks); if (llama_mlock_supported()) { fprintf(stdout, " --mlock force system to keep model in RAM rather than swapping or compressing\n"); @@ -669,6 +682,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stdout, " --lora-base FNAME optional model to use as a base for the layers modified by the LoRA adapter\n"); fprintf(stdout, " -m FNAME, --model FNAME\n"); fprintf(stdout, " model path (default: %s)\n", params.model.c_str()); + fprintf(stdout, " -md FNAME, --model-draft FNAME\n"); + fprintf(stdout, " draft model for speculative decoding (default: %s)\n", params.model.c_str()); fprintf(stdout, " -ld LOGDIR, --logdir LOGDIR\n"); fprintf(stdout, " path under which to save YAML logs (no logging if unset)\n"); fprintf(stdout, "\n"); @@ -832,6 +847,130 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector & last_tokens, + std::vector & candidates, + int idx) { + const int n_ctx = llama_n_ctx(ctx); + const int n_vocab = llama_n_vocab(ctx); + + const float temp = params.temp; + const int32_t top_k = params.top_k <= 0 ? n_vocab : params.top_k; + const float top_p = params.top_p; + const float tfs_z = params.tfs_z; + const float typical_p = params.typical_p; + const int32_t repeat_last_n = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n; + const float repeat_penalty = params.repeat_penalty; + const float alpha_presence = params.presence_penalty; + const float alpha_frequency = params.frequency_penalty; + const int mirostat = params.mirostat; + const float mirostat_tau = params.mirostat_tau; + const float mirostat_eta = params.mirostat_eta; + const bool penalize_nl = params.penalize_nl; + + llama_token id = 0; + + float * logits = llama_get_logits(ctx) + idx * n_vocab; + + // Apply params.logit_bias map + for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) { + logits[it->first] += it->second; + } + + candidates.clear(); + for (llama_token token_id = 0; token_id < n_vocab; token_id++) { + candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); + } + + llama_token_data_array cur_p = { candidates.data(), candidates.size(), false }; + + if (ctx_guidance) { + llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale); + } + + // apply penalties + if (!last_tokens.empty()) { + const float nl_logit = logits[llama_token_nl(ctx)]; + const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx); + + llama_sample_repetition_penalty(ctx, &cur_p, + last_tokens.data() + last_tokens.size() - last_n_repeat, + last_n_repeat, repeat_penalty); + llama_sample_frequency_and_presence_penalties(ctx, &cur_p, + last_tokens.data() + last_tokens.size() - last_n_repeat, + last_n_repeat, alpha_frequency, alpha_presence); + + if (!penalize_nl) { + for (size_t idx = 0; idx < cur_p.size; idx++) { + if (cur_p.data[idx].id == llama_token_nl(ctx)) { + cur_p.data[idx].logit = nl_logit; + break; + } + } + } + } + + if (grammar != NULL) { + llama_sample_grammar(ctx, &cur_p, grammar); + } + + if (temp <= 0) { + // Greedy sampling + id = llama_sample_token_greedy(ctx, &cur_p); + } else { + if (mirostat == 1) { + static float mirostat_mu = 2.0f * mirostat_tau; + const int mirostat_m = 100; + llama_sample_temperature(ctx, &cur_p, temp); + id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu); + } else if (mirostat == 2) { + static float mirostat_mu = 2.0f * mirostat_tau; + llama_sample_temperature(ctx, &cur_p, temp); + id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu); + } else { + // Temperature sampling + llama_sample_top_k (ctx, &cur_p, top_k, 1); + llama_sample_tail_free (ctx, &cur_p, tfs_z, 1); + llama_sample_typical (ctx, &cur_p, typical_p, 1); + llama_sample_top_p (ctx, &cur_p, top_p, 1); + llama_sample_temperature(ctx, &cur_p, temp); + + { + const int n_top = 10; + LOG("top %d candidates:\n", n_top); + + for (int i = 0; i < n_top; i++) { + const llama_token id = cur_p.data[i].id; + LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p); + } + } + + id = llama_sample_token(ctx, &cur_p); + + LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str()); + } + } + // printf("`%d`", candidates_p.size); + + if (grammar != NULL) { + llama_grammar_accept_token(ctx, grammar, id); + } + + return id; +} + +// +// YAML utils +// + // returns true if successful, false otherwise bool create_directory_with_parents(const std::string & path) { #ifdef _WIN32 @@ -1070,6 +1209,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "mirostat_lr: %f # default: 0.1\n", params.mirostat_eta); fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false"); fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str()); + fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str()); fprintf(stream, "mtest: %s # default: false\n", params.mem_test ? "true" : "false"); fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false"); fprintf(stream, "n_gpu_layers: %d # default: 0\n", params.n_gpu_layers); diff --git a/common/common.h b/common/common.h index 5a379688e..105fb09e4 100644 --- a/common/common.h +++ b/common/common.h @@ -32,6 +32,7 @@ struct gpt_params { int32_t n_ctx = 512; // context size int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS) int32_t n_keep = 0; // number of tokens to keep from initial prompt + int32_t n_draft = 16; // number of tokens to draft during speculative decoding int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited) int32_t n_gpu_layers = 0; // number of layers to store in VRAM int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors @@ -63,6 +64,7 @@ struct gpt_params { float cfg_scale = 1.f; // How strong is guidance std::string model = "models/7B/ggml-model-f16.gguf"; // model path + std::string model_draft = ""; // draft model for speculative decoding std::string model_alias = "unknown"; // model alias std::string prompt = ""; std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state @@ -156,6 +158,40 @@ std::string llama_detokenize_bpe( llama_context * ctx, const std::vector & tokens); +// +// Sampling utils +// + +// this is a common sampling function used across the examples for convenience +// it can serve as a starting point for implementing your own sampling function +// +// required: +// - ctx: context to use for sampling +// - params: sampling parameters +// +// optional: +// - ctx_guidance: context to use for classifier-free guidance, ignore if NULL +// - grammar: grammar to use for sampling, ignore if NULL +// - last_tokens: needed for repetition penalty, ignore if empty +// - idx: sample from llama_get_logits(ctx) + idx * n_vocab +// +// returns: +// - token: sampled token +// - candidates: vector of candidate tokens +// +llama_token llama_sample_token( + struct llama_context * ctx, + struct llama_context * ctx_guidance, + struct llama_grammar * grammar, + const struct gpt_params & params, + const std::vector & last_tokens, + std::vector & candidates, + int idx = 0); + +// +// YAML utils +// + bool create_directory_with_parents(const std::string & path); void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector & data); void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector & data); diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 6e65eb087..884c42764 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -23,6 +23,7 @@ else() add_subdirectory(train-text-from-scratch) add_subdirectory(convert-llama2c-to-ggml) add_subdirectory(simple) + add_subdirectory(speculative) add_subdirectory(embd-input) add_subdirectory(llama-bench) add_subdirectory(beam-search) diff --git a/examples/main/main.cpp b/examples/main/main.cpp index db98312ca..922b9a980 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -116,7 +116,7 @@ int main(int argc, char ** argv) { #ifndef LOG_DISABLE_LOGS log_set_target(log_filename_generator("main", "log")); LOG_TEE("Log start\n"); - log_dump_cmdline(argc,argv); + log_dump_cmdline(argc, argv); #endif // LOG_DISABLE_LOGS // TODO: Dump params ? @@ -425,8 +425,9 @@ int main(int argc, char ** argv) { LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep); LOG_TEE("\n\n"); + struct llama_grammar * grammar = NULL; grammar_parser::parse_state parsed_grammar; - llama_grammar * grammar = NULL; + if (!params.grammar.empty()) { parsed_grammar = grammar_parser::parse(params.grammar.c_str()); // will be empty (default) if there are parse errors @@ -450,8 +451,8 @@ int main(int argc, char ** argv) { } // TODO: replace with ring-buffer - std::vector last_n_tokens(n_ctx); - std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0); + std::vector last_tokens(n_ctx); + std::fill(last_tokens.begin(), last_tokens.end(), 0); if (params.interactive) { const char *control_message; @@ -492,6 +493,11 @@ int main(int argc, char ** argv) { std::vector embd; std::vector embd_guidance; + const int n_vocab = llama_n_vocab(ctx); + + std::vector candidates; + candidates.reserve(n_vocab); + while ((n_remain != 0 && !is_antiprompt) || params.interactive) { // predict if (embd.size() > 0) { @@ -529,8 +535,8 @@ int main(int argc, char ** argv) { LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance); - // insert n_left/2 tokens at the start of embd from last_n_tokens - embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size()); + // insert n_left/2 tokens at the start of embd from last_tokens + embd.insert(embd.begin(), last_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_tokens.end() - embd.size()); LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd)); @@ -629,20 +635,6 @@ int main(int argc, char ** argv) { embd_guidance.clear(); if ((int) embd_inp.size() <= n_consumed && !is_interacting) { - const float temp = params.temp; - const int32_t top_k = params.top_k <= 0 ? llama_n_vocab(ctx) : params.top_k; - const float top_p = params.top_p; - const float tfs_z = params.tfs_z; - const float typical_p = params.typical_p; - const int32_t repeat_last_n = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n; - const float repeat_penalty = params.repeat_penalty; - const float alpha_presence = params.presence_penalty; - const float alpha_frequency = params.frequency_penalty; - const int mirostat = params.mirostat; - const float mirostat_tau = params.mirostat_tau; - const float mirostat_eta = params.mirostat_eta; - const bool penalize_nl = params.penalize_nl; - // optionally save the session on first sample (for faster prompt loading next time) if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) { need_to_save_session = false; @@ -651,98 +643,12 @@ int main(int argc, char ** argv) { LOG("saved session to %s\n", path_session.c_str()); } - llama_token id = 0; + const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates); - { - auto logits = llama_get_logits(ctx); - auto n_vocab = llama_n_vocab(ctx); + last_tokens.erase(last_tokens.begin()); + last_tokens.push_back(id); - // Apply params.logit_bias map - for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) { - logits[it->first] += it->second; - } - - std::vector candidates; - candidates.reserve(n_vocab); - for (llama_token token_id = 0; token_id < n_vocab; token_id++) { - candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); - } - - llama_token_data_array cur_p = { candidates.data(), candidates.size(), false }; - - if (ctx_guidance) { - llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale); - } - - // Apply penalties - float nl_logit = logits[llama_token_nl(ctx)]; - auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx); - llama_sample_repetition_penalty(ctx, &cur_p, - last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, - last_n_repeat, repeat_penalty); - llama_sample_frequency_and_presence_penalties(ctx, &cur_p, - last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, - last_n_repeat, alpha_frequency, alpha_presence); - if (!penalize_nl) { - for (size_t idx = 0; idx < cur_p.size; idx++) { - if (cur_p.data[idx].id == llama_token_nl(ctx)) { - cur_p.data[idx].logit = nl_logit; - break; - } - } - } - - if (grammar != NULL) { - llama_sample_grammar(ctx, &cur_p, grammar); - } - - if (temp <= 0) { - // Greedy sampling - id = llama_sample_token_greedy(ctx, &cur_p); - } else { - if (mirostat == 1) { - static float mirostat_mu = 2.0f * mirostat_tau; - const int mirostat_m = 100; - llama_sample_temperature(ctx, &cur_p, temp); - id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu); - } else if (mirostat == 2) { - static float mirostat_mu = 2.0f * mirostat_tau; - llama_sample_temperature(ctx, &cur_p, temp); - id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &mirostat_mu); - } else { - // Temperature sampling - llama_sample_top_k (ctx, &cur_p, top_k, 1); - llama_sample_tail_free (ctx, &cur_p, tfs_z, 1); - llama_sample_typical (ctx, &cur_p, typical_p, 1); - llama_sample_top_p (ctx, &cur_p, top_p, 1); - llama_sample_temperature(ctx, &cur_p, temp); - - { - const int n_top = 10; - LOG("top %d candidates:\n", n_top); - - for (int i = 0; i < n_top; i++) { - const llama_token id = cur_p.data[i].id; - LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p); - } - } - - id = llama_sample_token(ctx, &cur_p); - - LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str()); - } - } - // printf("`%d`", candidates_p.size); - - if (grammar != NULL) { - llama_grammar_accept_token(ctx, grammar, id); - } - - last_n_tokens.erase(last_n_tokens.begin()); - last_n_tokens.push_back(id); - - LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_n_tokens)); - } + LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens)); embd.push_back(id); @@ -758,8 +664,8 @@ int main(int argc, char ** argv) { LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed); while ((int) embd_inp.size() > n_consumed) { embd.push_back(embd_inp[n_consumed]); - last_n_tokens.erase(last_n_tokens.begin()); - last_n_tokens.push_back(embd_inp[n_consumed]); + last_tokens.erase(last_tokens.begin()); + last_tokens.push_back(embd_inp[n_consumed]); ++n_consumed; if ((int) embd.size() >= params.n_batch) { break; @@ -792,7 +698,7 @@ int main(int argc, char ** argv) { // check for reverse prompt if (params.antiprompt.size()) { std::string last_output; - for (auto id : last_n_tokens) { + for (auto id : last_tokens) { last_output += llama_token_to_piece(ctx, id); } @@ -823,7 +729,7 @@ int main(int argc, char ** argv) { } // deal with end of text token in interactive mode - if (last_n_tokens.back() == llama_token_eos(ctx)) { + if (last_tokens.back() == llama_token_eos(ctx)) { LOG("found EOS token\n"); if (params.interactive) { @@ -925,7 +831,7 @@ int main(int argc, char ** argv) { if (grammar != NULL) { llama_grammar_free(grammar); - std::vector grammar_rules( parsed_grammar.c_rules()); + std::vector grammar_rules(parsed_grammar.c_rules()); grammar = llama_grammar_init( grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root")); diff --git a/examples/speculative/CMakeLists.txt b/examples/speculative/CMakeLists.txt new file mode 100644 index 000000000..6c5c9456e --- /dev/null +++ b/examples/speculative/CMakeLists.txt @@ -0,0 +1,8 @@ +set(TARGET speculative) +add_executable(${TARGET} speculative.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) +if(TARGET BUILD_INFO) + add_dependencies(${TARGET} BUILD_INFO) +endif() diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp new file mode 100644 index 000000000..f0400c13f --- /dev/null +++ b/examples/speculative/speculative.cpp @@ -0,0 +1,234 @@ +#ifndef _GNU_SOURCE +#define _GNU_SOURCE +#endif + +#include "build-info.h" + +#include "common.h" +#include "llama.h" + +#include +#include +#include +#include + +int main(int argc, char ** argv) { + gpt_params params; + + if (gpt_params_parse(argc, argv, params) == false) { + return 1; + } + + if (params.model_draft.empty()) { + fprintf(stderr, "%s: error: --model-draft is required\n", __func__); + return 1; + } + +#ifndef LOG_DISABLE_LOGS + log_set_target(log_filename_generator("speculative", "log")); + LOG_TEE("Log start\n"); + log_dump_cmdline(argc, argv); +#endif // LOG_DISABLE_LOGS + + // init llama.cpp + llama_backend_init(params.numa); + + llama_model * model_tgt = NULL; + llama_model * model_dft = NULL; + + llama_context * ctx_tgt = NULL; + llama_context * ctx_dft = NULL; + + // load the target model + params.perplexity = true; // HACK: enable logits_all = true + std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params); + + // load the draft model + params.model = params.model_draft; + std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params); + + // tokenize the prompt + std::vector inp; + inp = ::llama_tokenize(ctx_tgt, params.prompt, true); + + const int max_context_size = llama_n_ctx(ctx_tgt); + const int max_tokens_list_size = max_context_size - 4; + + if ((int) inp.size() > max_tokens_list_size) { + fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size); + return 1; + } + + fprintf(stderr, "\n\n"); + + for (auto id : inp) { + fprintf(stderr, "%s", llama_token_to_piece(ctx_tgt, id).c_str()); + } + + fflush(stderr); + + const int n_input = inp.size(); + + const auto t_enc_start = ggml_time_us(); + + // eval the prompt with both models + llama_eval(ctx_tgt, inp.data(), int(inp.size() - 1), 0, params.n_threads); + llama_eval(ctx_tgt, &inp.back(), 1, inp.size() - 1, params.n_threads); + llama_eval(ctx_dft, inp.data(), int(inp.size()), 0, params.n_threads); + + const auto t_enc_end = ggml_time_us(); + + // the 2 models should have the same vocab + const int n_ctx = llama_n_ctx(ctx_tgt); + const int n_vocab = llama_n_vocab(ctx_tgt); + //GGML_ASSERT(n_vocab == llama_n_vocab(ctx_dft)); + + // how many tokens to draft each time + const int n_draft = params.n_draft; + + int n_predict = 0; + int n_drafted = 0; + int n_accept = 0; + + int n_past_tgt = inp.size(); + int n_past_dft = inp.size(); + + std::vector drafted; + + std::vector last_tokens(n_ctx); + std::fill(last_tokens.begin(), last_tokens.end(), 0); + + for (auto & id : inp) { + last_tokens.erase(last_tokens.begin()); + last_tokens.push_back(id); + } + + std::vector candidates; + candidates.reserve(n_vocab); + + // used to determine end of generation + bool has_eos = false; + + const auto t_dec_start = ggml_time_us(); + + while (true) { + LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted)); + + // sample from the drafted tokens if any + int i_dft = 0; + while (true) { + const llama_token id = llama_sample_token(ctx_tgt, NULL, NULL, params, last_tokens, candidates, i_dft); + + last_tokens.erase(last_tokens.begin()); + last_tokens.push_back(id); + + //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, last_tokens)); + + const std::string token_str = llama_token_to_piece(ctx_tgt, id); + printf("%s", token_str.c_str()); + fflush(stdout); + + if (id == llama_token_eos(ctx_tgt)) { + has_eos = true; + } + + ++n_predict; + + if (i_dft < (int) drafted.size() && id == drafted[i_dft]) { + LOG("drafted token %d accepted\n", id); + ++n_accept; + ++n_past_tgt; + ++n_past_dft; + ++i_dft; + + continue; + } + + // the drafted token was rejected or we are out of drafted tokens + llama_eval(ctx_dft, &id, 1, n_past_dft, params.n_threads); + ++n_past_dft; + + drafted.clear(); + drafted.push_back(id); + + break; + } + + if (n_predict > params.n_predict || has_eos) { + break; + } + + // sample n_draft tokens from the draft model picking the best token + int n_past_cur = n_past_dft; + for (int i = 0; i < n_draft; ++i) { + float * logits = llama_get_logits(ctx_dft); + + candidates.clear(); + for (llama_token token_id = 0; token_id < n_vocab; token_id++) { + candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); + } + + llama_token_data_array cur_p = { candidates.data(), candidates.size(), false }; + + // computes softmax and sorts the candidates + llama_sample_softmax(ctx_dft, &cur_p); + + for (int i = 0; i < 3; ++i) { + LOG(" - draft candidate %d: %d (%.3f)\n", i, cur_p.data[i].id, cur_p.data[i].p); + } + + // too low probability, stop drafting + if (cur_p.data[0].p < 2*cur_p.data[1].p) { + break; + } + + drafted.push_back(cur_p.data[0].id); + ++n_drafted; + + if (i < n_draft - 1) { + // evaluate the drafted token on the draft model + llama_eval(ctx_dft, &drafted.back(), 1, n_past_cur, params.n_threads); + ++n_past_cur; + } + } + + // evaluate the target model on the drafted tokens + llama_eval(ctx_tgt, drafted.data(), drafted.size(), n_past_tgt, params.n_threads); + ++n_past_tgt; + + drafted.erase(drafted.begin()); + } + + auto t_dec_end = ggml_time_us(); + + LOG_TEE("\n\n"); + + LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f)); + LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f)); + + // TODO: make sure these numbers are computed correctly + LOG_TEE("\n"); + LOG_TEE("n_draft = %d\n", n_draft); + LOG_TEE("n_predict = %d\n", n_predict); + LOG_TEE("n_drafted = %d\n", n_drafted); + LOG_TEE("n_accept = %d\n", n_accept); + LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted); + + LOG_TEE("\ndraft:\n"); + llama_print_timings(ctx_dft); + + LOG_TEE("\ntarget:\n"); + llama_print_timings(ctx_tgt); + + llama_free(ctx_tgt); + llama_free_model(model_tgt); + + llama_free(ctx_dft); + llama_free_model(model_dft); + + llama_backend_free(); + + fprintf(stderr, "\n\n"); + + return 0; +} From cf9b08485c4c2d4d945c6e74fe20f273a38b6104 Mon Sep 17 00:00:00 2001 From: slaren Date: Sun, 3 Sep 2023 20:34:09 +0200 Subject: [PATCH 147/147] ggml-alloc : use virtual memory for measurement (#2973) * ggml-alloc : use virtual memory for measurement * compatibility fixes for MAP_ANONYMOUS * fallback to fixed address for systems without virtual memory --- ggml-alloc.c | 123 ++++++++++++++++++++++++++++++++++++++++----------- 1 file changed, 97 insertions(+), 26 deletions(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index 459f121ca..c1939a4b7 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -1,3 +1,8 @@ +// defines MAP_ANONYMOUS +#ifndef _GNU_SOURCE +#define _GNU_SOURCE +#endif + #include "ggml-alloc.h" #include "ggml.h" #include @@ -6,6 +11,26 @@ #include #include +#ifdef __has_include + #if __has_include() + #include + #if defined(_POSIX_MAPPED_FILES) + #include + #include + #endif + #endif +#endif + +#if defined(_WIN32) + #define WIN32_LEAN_AND_MEAN + #ifndef NOMINMAX + #define NOMINMAX + #endif + #include + #include +#endif + + #define UNUSED(x) (void)(x) #define MAX(a, b) ((a) > (b) ? (a) : (b)) #define GGML_MAX_CONCUR (2*GGML_MAX_NODES) @@ -99,19 +124,24 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens } #endif - -static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { return ggml_nbytes(tensor); UNUSED(alloc); } +// check if a tensor is allocated by this buffer +static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) { + void * ptr = tensor->data; + return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size; +} + void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { #ifdef GGML_ALLOCATOR_DEBUG GGML_ASSERT(ggml_is_view(tensor) == false); // views generally get data pointer from one of their sources GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated #endif - size_t size = ggml_allocator_get_alloc_size(alloc, tensor); + size_t size = ggml_allocr_get_alloc_size(alloc, tensor); size = aligned_offset(NULL, size, alloc->alignment); AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size); @@ -177,17 +207,17 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) } // this is a very naive implementation, but for our case the number of free blocks should be very small -static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { void * ptr = tensor->data; - if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) { + if (ggml_allocr_is_own(alloc, tensor) == false) { // the tensor was not allocated in this buffer // this can happen because the graph allocator will try to free weights and other tensors from different buffers // the easiest way to deal with this is just to ignore it return; } - size_t size = ggml_allocator_get_alloc_size(alloc, tensor); + size_t size = ggml_allocr_get_alloc_size(alloc, tensor); size = aligned_offset(NULL, size, alloc->alignment); AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks); @@ -281,24 +311,64 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) return alloc; } -// address and size of the buffer when measuring -// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers -static void * const MEASURE_BASE_ADDR = (void *) 0x1000; -#if defined(__ARM_NEON) && !defined(__aarch64__) -// 32-bit -// TODO: Use for 32-bit x86 as well -static const size_t MEASURE_MAX_SIZE = (1ULL<<32) - 1; // 4 GB +// OS specific functions to allocate and free uncommitted virtual memory +static void * alloc_vmem(size_t size) { +#if defined(_WIN32) + return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS); +#elif defined(_POSIX_MAPPED_FILES) + return mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0); #else -// 64-bit -static const size_t MEASURE_MAX_SIZE = 1ULL<<40; // 1 TB + // use a fixed address for other platforms + uintptr_t base_addr = (uintptr_t)-size - 0x100; + return (void *)base_addr; #endif +} + +static void free_vmem(void * base_addr, size_t size) { +#if defined(_WIN32) + VirtualFree(base_addr, 0, MEM_RELEASE); + UNUSED(size); +#elif defined(_POSIX_MAPPED_FILES) + munmap(base_addr, size); +#else + // nothing to do + UNUSED(base_addr); + UNUSED(size); +#endif +} + +// allocate uncommitted virtual memory to measure the size of the graph +static void alloc_measure_vmem(void ** base_addr, size_t * size) { + // 1TB for 64-bit, 1GB for 32-bit + *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<40; + do { + *base_addr = alloc_vmem(*size); + if (*base_addr != NULL) { + AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr); + return; + } + // try again with half the size + *size /= 2; + } while (*size > 0); + + GGML_ASSERT(!"failed to allocate virtual memory for measure buffer"); +} + +static void free_measure_vmem(void * base_addr, size_t size) { + free_vmem(base_addr, size); +} struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) { struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */); + void * base_addr; + size_t size; + + alloc_measure_vmem(&base_addr, &size); + *alloc = (struct ggml_allocr){ - /*.data = */ MEASURE_BASE_ADDR, - /*.size = */ MEASURE_MAX_SIZE, + /*.data = */ base_addr, + /*.size = */ size, /*.alignment = */ alignment, /*.n_free_blocks = */ 0, /*.free_blocks = */ {{0}}, @@ -318,6 +388,9 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) { } void ggml_allocr_free(struct ggml_allocr * alloc) { + if (alloc->measure) { + free_measure_vmem(alloc->data, alloc->size); + } free(alloc); } @@ -387,8 +460,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) } // if the node's data is external, then we cannot re-use it - if ((char *) parent->data < (char *) alloc->data || - (char *) parent->data >= ((char *) alloc->data + alloc->size)) { + if (ggml_allocr_is_own(alloc, parent) == false) { AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data); continue; } @@ -422,7 +494,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) } } -static size_t ggml_allocator_alloc_graph_tensors_n( +static size_t ggml_allocr_alloc_graph_tensors_n( struct ggml_allocr * alloc, struct ggml_cgraph ** graphs, int n_graphs, struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) { @@ -500,11 +572,10 @@ static size_t ggml_allocator_alloc_graph_tensors_n( AT_PRINTF("\n"); } - // update parents // update immediately if there is no parse_seq // update only at barriers if there is parse_seq - if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] == -1) { + if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) { int update_start = alloc->parse_seq_len ? last_barrier_pos : ind; int update_end = alloc->parse_seq_len ? ind : ind + 1; for (int i = update_start; i < update_end; i++) { @@ -528,12 +599,12 @@ static size_t ggml_allocator_alloc_graph_tensors_n( view_src_hn->n_views -= 1; AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views); if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) { - ggml_allocator_free_tensor(alloc, view_src); + ggml_allocr_free_tensor(alloc, view_src); } } else { if (parent->data != node->data) { - ggml_allocator_free_tensor(alloc, parent); + ggml_allocr_free_tensor(alloc, parent); } } } @@ -550,7 +621,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n( for (int i = 0; outputs[g][i] != NULL; i++) { struct ggml_tensor * output = outputs[g][i]; AT_PRINTF("output: %s\n", output->name); - ggml_allocator_free_tensor(alloc, output); + ggml_allocr_free_tensor(alloc, output); } } } @@ -559,5 +630,5 @@ static size_t ggml_allocator_alloc_graph_tensors_n( } size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) { - return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL); + return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL); }