From 9560655409dc80771a9b19e838ff47c5c1df6483 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Andr=C3=A1s=20Salamon?= Date: Tue, 16 May 2023 16:46:34 +0100 Subject: [PATCH 01/20] define default model path once, sync path with readme (#1366) --- examples/common.h | 2 +- examples/embedding/embedding.cpp | 1 - examples/main/main.cpp | 1 - examples/perplexity/perplexity.cpp | 1 - examples/save-load-state/save-load-state.cpp | 1 - 5 files changed, 1 insertion(+), 5 deletions(-) diff --git a/examples/common.h b/examples/common.h index 717838f06..f4e07a252 100644 --- a/examples/common.h +++ b/examples/common.h @@ -45,7 +45,7 @@ struct gpt_params { float mirostat_tau = 5.00f; // target entropy float mirostat_eta = 0.10f; // learning rate - std::string model = "models/lamma-7B/ggml-model.bin"; // model path + std::string model = "models/7B/ggml-model.bin"; // model path std::string prompt = ""; std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state std::string input_prefix = ""; // string to prefix user inputs with diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index bb3fd50a9..c24f7f820 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -6,7 +6,6 @@ int main(int argc, char ** argv) { gpt_params params; - params.model = "models/llama-7B/ggml-model.bin"; if (gpt_params_parse(argc, argv, params) == false) { return 1; diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 8543414dd..fe1c847a7 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -50,7 +50,6 @@ void sigint_handler(int signo) { int main(int argc, char ** argv) { gpt_params params; - params.model = "models/llama-7B/ggml-model.bin"; if (gpt_params_parse(argc, argv, params) == false) { return 1; diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 9212dee5c..9d38626cb 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -116,7 +116,6 @@ void perplexity(llama_context * ctx, const gpt_params & params) { int main(int argc, char ** argv) { gpt_params params; - params.model = "models/llama-7B/ggml-model.bin"; params.n_batch = 512; if (gpt_params_parse(argc, argv, params) == false) { diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index ea0a984d9..355969579 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -8,7 +8,6 @@ int main(int argc, char ** argv) { gpt_params params; - params.model = "models/llama-7B/ggml-model.bin"; params.seed = 42; params.n_threads = 4; params.repeat_last_n = 64; From 42627421ece816e632e6a0d757fa75150c687f87 Mon Sep 17 00:00:00 2001 From: Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com> Date: Wed, 17 May 2023 01:36:47 +0700 Subject: [PATCH 02/20] ~7% faster Q5_1 AVX2 code (#1477) --- ggml.c | 39 ++++++++++++++++++++++++++++++--------- 1 file changed, 30 insertions(+), 9 deletions(-) diff --git a/ggml.c b/ggml.c index 4311ce7cf..dbef99312 100644 --- a/ggml.c +++ b/ggml.c @@ -543,12 +543,7 @@ static inline __m256 sum_i16_pairs_float(const __m256i x) { return _mm256_cvtepi32_ps(summed_pairs); } -// multiply int8_t, add results pairwise twice and return as float vector -static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { - // Get absolute values of x vectors - const __m256i ax = _mm256_sign_epi8(x, x); - // Sign the values of the y vectors - const __m256i sy = _mm256_sign_epi8(y, x); +static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) { #if __AVXVNNI__ const __m256i zero = _mm256_setzero_si256(); const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy); @@ -560,6 +555,21 @@ static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { #endif } +// multiply int8_t, add results pairwise twice and return as float vector +static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { +#if __AVXVNNIINT8__ + const __m256i zero = _mm256_setzero_si256(); + const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y); + return _mm256_cvtepi32_ps(summed_pairs); +#else + // Get absolute values of x vectors + const __m256i ax = _mm256_sign_epi8(x, x); + // Sign the values of the y vectors + const __m256i sy = _mm256_sign_epi8(y, x); + return mul_sum_us8_pairs_float(ax, sy); +#endif +} + static inline __m128i packNibbles( __m256i bytes ) { // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh @@ -619,6 +629,17 @@ static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) { return _mm256_cvtepi32_ps(summed_pairs); } +static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) { + const __m128i axl = _mm256_castsi256_si128(ax); + const __m128i axh = _mm256_extractf128_si256(ax, 1); + const __m128i syl = _mm256_castsi256_si128(sy); + const __m128i syh = _mm256_extractf128_si256(sy, 1); + // Perform multiplication and create 16-bit values + const __m128i dotl = _mm_maddubs_epi16(axl, syl); + const __m128i doth = _mm_maddubs_epi16(axh, syh); + return sum_i16_pairs_float(doth, dotl); +} + // multiply int8_t, add results pairwise twice and return as float vector static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { const __m128i xl = _mm256_castsi256_si128(x); @@ -2434,7 +2455,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const __m256i bx = bytes_from_nibbles_32(x[i].qs); const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs ); - const __m256 xy = mul_sum_i8_pairs_float(bx, by); + const __m256 xy = mul_sum_us8_pairs_float(bx, by); // Accumulate d0*d1*x*y #if defined(__AVX2__) @@ -2906,7 +2927,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * const __m256 dy = _mm256_broadcast_ss(&y[i].d); const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256 q = mul_sum_i8_pairs_float(bx, by); + const __m256 q = mul_sum_us8_pairs_float(bx, by); acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc); } @@ -2940,7 +2961,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * const __m256 dy = _mm256_broadcast_ss(&y[i].d); const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - const __m256 q = mul_sum_i8_pairs_float(bx, by); + const __m256 q = mul_sum_us8_pairs_float(bx, by); acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc); } From 2b2646931bd2a2eb3e21c6f3733cc0e090b2e24b Mon Sep 17 00:00:00 2001 From: Tom Jobbins <784313+TheBloke@users.noreply.github.com> Date: Tue, 16 May 2023 23:04:35 +0100 Subject: [PATCH 03/20] convert.py: Support models which are stored in a single pytorch_model.bin (#1469) * Support models in a single pytorch_model.bin * Remove spurious line with typo --- convert.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/convert.py b/convert.py index 8f4f0399e..ece5a0266 100644 --- a/convert.py +++ b/convert.py @@ -121,7 +121,6 @@ def make_tensors_list() -> List[str]: f'layers.{i}.feed_forward.w1.weight', f'layers.{i}.feed_forward.w2.weight', f'layers.{i}.feed_forward.w3.weight', - f'layers.{i}.atttention_norm.weight', f'layers.{i}.ffn_norm.weight', ] return ret @@ -1055,7 +1054,7 @@ def load_some_model(path: Path) -> ModelPlus: files = list(path.glob("model-00001-of-*.safetensors")) if not files: # Try the PyTorch patterns too, with lower priority - globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"] + globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin" ] files = [file for glob in globs for file in path.glob(glob)] if not files: # Try GGML too, but with lower priority, since if both a non-GGML From c238b5873a1ea496db03ffcfe124c9d0d83afbc6 Mon Sep 17 00:00:00 2001 From: rankaiyx Date: Wed, 17 May 2023 22:47:58 +0800 Subject: [PATCH 04/20] benchmark-matmul: Print the average of the test results (#1490) --- examples/benchmark/benchmark-matmult.cpp | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp index 7d237be02..446b8e8fb 100644 --- a/examples/benchmark/benchmark-matmult.cpp +++ b/examples/benchmark/benchmark-matmult.cpp @@ -211,6 +211,7 @@ int main(int argc, char ** argv) { printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n"); printf("=====================================================================================\n"); + double gflops_sum = 0; for (int i=0;i Date: Wed, 17 May 2023 22:12:01 +0000 Subject: [PATCH 05/20] Remove unused n_parts parameter (#1509) --- examples/common.cpp | 8 -------- examples/common.h | 1 - examples/quantize-stats/quantize-stats.cpp | 1 - examples/save-load-state/save-load-state.cpp | 1 - llama.cpp | 1 - llama.h | 1 - 6 files changed, 13 deletions(-) diff --git a/examples/common.cpp b/examples/common.cpp index 259880a7c..a6abc4977 100644 --- a/examples/common.cpp +++ b/examples/common.cpp @@ -321,12 +321,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { invalid_param = true; break; } - } else if (arg == "--n-parts") { - if (++i >= argc) { - invalid_param = true; - break; - } - params.n_parts = std::stoi(argv[i]); } else if (arg == "-h" || arg == "--help") { gpt_print_usage(argc, argv, default_params); exit(0); @@ -418,7 +412,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stderr, " --no-penalize-nl do not penalize newline token\n"); fprintf(stderr, " --memory-f32 use f32 instead of f16 for memory key+value\n"); fprintf(stderr, " --temp N temperature (default: %.1f)\n", (double)params.temp); - fprintf(stderr, " --n-parts N number of model parts (default: -1 = determine from dimensions)\n"); fprintf(stderr, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); fprintf(stderr, " --perplexity compute perplexity over the prompt\n"); fprintf(stderr, " --keep number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); @@ -473,7 +466,6 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) { auto lparams = llama_context_default_params(); lparams.n_ctx = params.n_ctx; - lparams.n_parts = params.n_parts; lparams.n_gpu_layers = params.n_gpu_layers; lparams.seed = params.seed; lparams.f16_kv = params.memory_f16; diff --git a/examples/common.h b/examples/common.h index f4e07a252..2ad20ba50 100644 --- a/examples/common.h +++ b/examples/common.h @@ -24,7 +24,6 @@ struct gpt_params { int32_t seed = -1; // RNG seed int32_t n_threads = get_num_physical_cores(); int32_t n_predict = -1; // new tokens to predict - int32_t n_parts = -1; // amount of model parts (-1 = determine from model dimensions) 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 diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index 9a2aa7c64..085fdde3c 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -321,7 +321,6 @@ int main(int argc, char ** argv) { auto lparams = llama_context_default_params(); lparams.n_ctx = 256; - lparams.n_parts = 1; lparams.seed = 1; lparams.f16_kv = false; lparams.use_mlock = false; diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index 355969579..91f04b6c7 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -26,7 +26,6 @@ int main(int argc, char ** argv) { auto lparams = llama_context_default_params(); lparams.n_ctx = params.n_ctx; - lparams.n_parts = params.n_parts; lparams.seed = params.seed; lparams.f16_kv = params.memory_f16; lparams.use_mmap = params.use_mmap; diff --git a/llama.cpp b/llama.cpp index 98f49abd7..6e19064fc 100644 --- a/llama.cpp +++ b/llama.cpp @@ -812,7 +812,6 @@ static bool kv_cache_init( struct llama_context_params llama_context_default_params() { struct llama_context_params result = { /*.n_ctx =*/ 512, - /*.n_parts =*/ -1, /*.gpu_layers =*/ 0, /*.seed =*/ -1, /*.f16_kv =*/ false, diff --git a/llama.h b/llama.h index 21cba8cf6..f955fa23d 100644 --- a/llama.h +++ b/llama.h @@ -55,7 +55,6 @@ extern "C" { struct llama_context_params { int n_ctx; // text context - int n_parts; // -1 for default int n_gpu_layers; // number of layers to store in VRAM int seed; // RNG seed, -1 for random From ee9654138ab0ae5f138f4abddf56ca234ea3c352 Mon Sep 17 00:00:00 2001 From: DannyDaemonic Date: Thu, 18 May 2023 10:30:40 -0700 Subject: [PATCH 06/20] Fixes #1511 lambda issue for w64devkit (mingw) (#1513) * Fix for w64devkit and mingw --- examples/main/main.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/main/main.cpp b/examples/main/main.cpp index fe1c847a7..18673ed2e 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -241,7 +241,7 @@ int main(int argc, char ** argv) { sigint_action.sa_flags = 0; sigaction(SIGINT, &sigint_action, NULL); #elif defined (_WIN32) - auto console_ctrl_handler = [](DWORD ctrl_type) -> BOOL { + auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL { return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false; }; SetConsoleCtrlHandler(static_cast(console_ctrl_handler), true); From 5ea43392731040b454c293123839b90e159cbb99 Mon Sep 17 00:00:00 2001 From: Erik Scholz Date: Thu, 18 May 2023 19:31:01 +0200 Subject: [PATCH 07/20] make kv_f16 the default for api users (#1517) --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 6e19064fc..1f9d37844 100644 --- a/llama.cpp +++ b/llama.cpp @@ -814,7 +814,7 @@ struct llama_context_params llama_context_default_params() { /*.n_ctx =*/ 512, /*.gpu_layers =*/ 0, /*.seed =*/ -1, - /*.f16_kv =*/ false, + /*.f16_kv =*/ true, /*.logits_all =*/ false, /*.vocab_only =*/ false, /*.use_mmap =*/ true, From 4b7e245adf63db675c3daab4a9bfddd451ef4097 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 19 May 2023 20:14:51 +0300 Subject: [PATCH 08/20] minor : fix compile warnings --- examples/common.cpp | 4 ++-- examples/common.h | 6 +++--- llama.cpp | 2 +- tests/test-sampling.cpp | 10 ++++++---- 4 files changed, 12 insertions(+), 10 deletions(-) diff --git a/examples/common.cpp b/examples/common.cpp index a6abc4977..a4fea4af4 100644 --- a/examples/common.cpp +++ b/examples/common.cpp @@ -749,7 +749,7 @@ bool console_readline(console_state & con_st, std::string & line) { break; } - if (input_char == WEOF || input_char == 0x04 /* Ctrl+D*/) { + if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) { end_of_stream = true; break; } @@ -764,7 +764,7 @@ bool console_readline(console_state & con_st, std::string & line) { char32_t code = getchar32(); if (code == '[' || code == 0x1B) { // Discard the rest of the escape sequence - while ((code = getchar32()) != WEOF) { + while ((code = getchar32()) != (char32_t) WEOF) { if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') { break; } diff --git a/examples/common.h b/examples/common.h index 2ad20ba50..2b66382a6 100644 --- a/examples/common.h +++ b/examples/common.h @@ -44,15 +44,15 @@ struct gpt_params { float mirostat_tau = 5.00f; // target entropy float mirostat_eta = 0.10f; // learning rate - std::string model = "models/7B/ggml-model.bin"; // model path - std::string prompt = ""; + std::string model = "models/7B/ggml-model.bin"; // model path + std::string prompt = ""; std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state std::string input_prefix = ""; // string to prefix user inputs with std::string input_suffix = ""; // string to suffix user inputs with std::vector antiprompt; // string upon seeing which more user input is prompted std::string lora_adapter = ""; // lora adapter path - std::string lora_base = ""; // base model path for the lora adapter + std::string lora_base = ""; // base model path for the lora adapter bool memory_f16 = true; // use f16 instead of f32 for memory kv bool random_prompt = false; // do not randomize prompt if none provided diff --git a/llama.cpp b/llama.cpp index 1f9d37844..1802d2319 100644 --- a/llama.cpp +++ b/llama.cpp @@ -941,7 +941,7 @@ static void llama_model_load_internal( size_t ctx_size; size_t mmapped_size; ml->calc_sizes(&ctx_size, &mmapped_size); - fprintf(stderr, "%s: ggml ctx size = %6.2f KB\n", __func__, ctx_size/1024.0); + fprintf(stderr, "%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/1024.0/1024.0); // print memory requirements { diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index 9174c1e37..ebfc17c18 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -1,14 +1,16 @@ -#include "llama.h" #include "ggml.h" -#include -#include +#include "llama.h" + +#ifdef NDEBUG +#undef NDEBUG +#endif + #include #include #include #include #include - void dump(const llama_token_data_array * candidates) { for (size_t i = 0; i < candidates->size; i++) { printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit); From 79e3efb0e97b65b6cc72cd9ee970fa8189ad79a4 Mon Sep 17 00:00:00 2001 From: David Kennedy Date: Fri, 19 May 2023 13:16:30 -0400 Subject: [PATCH 09/20] readme : adds WizardLM to the list of supported models (#1485) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 1d84a5e6d..6a67765aa 100644 --- a/README.md +++ b/README.md @@ -80,6 +80,7 @@ as the main playground for developing new features for the [ggml](https://github - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/) - [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy) - [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b) +- [X] [WizardLM](https://github.com/nlpxucan/WizardLM) **Bindings:** From 7694b52b9a206b93d59139c3c7c9b55da0f5aa59 Mon Sep 17 00:00:00 2001 From: Jason McCartney Date: Fri, 19 May 2023 10:24:59 -0700 Subject: [PATCH 10/20] main : make reverse prompt option act as a stop token in non-interactive mode (#1032) * Make reverse prompt option act as a stop token in non-interactive scenarios * Making requested review changes * Update gpt_params_parse and fix a merge error * Revert "Update gpt_params_parse and fix a merge error" This reverts commit 2bb2ff1748513591ad45b175a75ed1d8089d84c8. * Update gpt_params_parse and fix a merge error take 2 --- examples/common.cpp | 6 +++--- examples/main/main.cpp | 26 ++++++++++++++++++-------- 2 files changed, 21 insertions(+), 11 deletions(-) diff --git a/examples/common.cpp b/examples/common.cpp index a4fea4af4..e89df537e 100644 --- a/examples/common.cpp +++ b/examples/common.cpp @@ -351,7 +351,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { } if (params.prompt_cache_all && (params.interactive || params.interactive_first || - params.instruct || params.antiprompt.size())) { + params.instruct)) { fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n"); gpt_print_usage(argc, argv, default_params); exit(1); @@ -373,8 +373,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stderr, " -ins, --instruct run in instruction mode (use with Alpaca models)\n"); fprintf(stderr, " --multiline-input allows you to write or paste multiple lines without ending each in '\\'\n"); fprintf(stderr, " -r PROMPT, --reverse-prompt PROMPT\n"); - fprintf(stderr, " run in interactive mode and poll user input upon seeing PROMPT (can be\n"); - fprintf(stderr, " specified more than once for multiple prompts).\n"); + fprintf(stderr, " halt generation at PROMPT, return control in interactive mode\n"); + fprintf(stderr, " (can be specified more than once for multiple prompts).\n"); fprintf(stderr, " --color colorise output to distinguish prompt and user input from generations\n"); fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for < 0)\n"); fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads); diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 18673ed2e..4d886f8de 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -208,8 +208,8 @@ int main(int argc, char ** argv) { params.antiprompt.push_back("### Instruction:\n\n"); } - // enable interactive mode if reverse prompt or interactive start is specified - if (params.antiprompt.size() != 0 || params.interactive_first) { + // enable interactive mode if interactive start is specified + if (params.interactive_first) { params.interactive = true; } @@ -305,7 +305,7 @@ int main(int argc, char ** argv) { std::vector embd; - while (n_remain != 0 || params.interactive) { + while ((n_remain != 0 && !is_antiprompt) || params.interactive) { // predict if (embd.size() > 0) { // infinite text generation via context swapping @@ -503,9 +503,8 @@ int main(int argc, char ** argv) { console_set_color(con_st, CONSOLE_COLOR_DEFAULT); } - // in interactive mode, and not currently processing queued inputs; - // check if we should prompt the user for more - if (params.interactive && (int) embd_inp.size() <= n_consumed) { + // if not currently processing queued inputs; + if ((int) embd_inp.size() <= n_consumed) { // check for reverse prompt if (params.antiprompt.size()) { @@ -516,10 +515,21 @@ int main(int argc, char ** argv) { is_antiprompt = false; // Check if each of the reverse prompts appears at the end of the output. + // If we're not running interactively, the reverse prompt might be tokenized with some following characters + // so we'll compensate for that by widening the search window a bit. for (std::string & antiprompt : params.antiprompt) { - if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) { - is_interacting = true; + size_t extra_padding = params.interactive ? 0 : 2; + size_t search_start_pos = last_output.length() > static_cast(antiprompt.length() + extra_padding) + ? last_output.length() - static_cast(antiprompt.length() + extra_padding) + : 0; + + if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) { + if (params.interactive) { + is_interacting = true; + console_set_color(con_st, CONSOLE_COLOR_USER_INPUT); + } is_antiprompt = true; + fflush(stdout); break; } } From 943e6081cc939df7584f8f0ab7057a39c2ef3271 Mon Sep 17 00:00:00 2001 From: Evan Jones Date: Fri, 19 May 2023 13:39:51 -0400 Subject: [PATCH 11/20] examples : add persistent chat (#1495) * examples : add persistent chat * examples : fix whitespace --------- Co-authored-by: Georgi Gerganov --- examples/chat-persistent.sh | 151 ++++++++++++++++++++++++++++++++++++ 1 file changed, 151 insertions(+) create mode 100755 examples/chat-persistent.sh diff --git a/examples/chat-persistent.sh b/examples/chat-persistent.sh new file mode 100755 index 000000000..b32284b49 --- /dev/null +++ b/examples/chat-persistent.sh @@ -0,0 +1,151 @@ +#!/bin/bash + +set -euo pipefail + +cd "$(dirname "$0")/.." || exit + +if [[ -z "${PROMPT_CACHE_FILE+x}" || -z "${CHAT_SAVE_DIR+x}" ]]; then + echo >&2 "error: PROMPT_CACHE_FILE and CHAT_SAVE_DIR must be provided" + exit 1 +fi + +MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}" +PROMPT_TEMPLATE="${PROMPT_TEMPLATE:-./prompts/chat.txt}" +USER_NAME="${USER_NAME:-User}" +AI_NAME="${AI_NAME:-ChatLLaMa}" +DATE_TIME="$(date +%H:%M)" +DATE_YEAR="$(date +%Y)" + +LOG="${CHAT_SAVE_DIR}/main.log" +LOG_BG="${CHAT_SAVE_DIR}/main-bg.log" +CUR_PROMPT_FILE="${CHAT_SAVE_DIR}/current-prompt.txt" +CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin" +NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt" +NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin" + +SESSION_SIZE_MSG_PATTERN='main: session file matches \d+ / \d+' +SAMPLE_TIME_MSG_PATTERN='sample time =\s+\d+.\d+ ms /\s+\d+' +SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d" + +CTX_SIZE=2048 +CTX_ROTATE_POINT=$((CTX_SIZE * 3 / 5)) # REVIEW +OPTS=(--model "$MODEL" --ctx_size "$CTX_SIZE" --repeat_last_n 256 "$@") + +# An unbuffered `tail -c+N` +skip_bytes() { + LANG=C IFS= read -r -n "$1" -d '' c + while LANG=C IFS= read -r -n 1 -d '' c; do + printf '%s' "$c" + done +} + +mkdir -p "$CHAT_SAVE_DIR" +echo >"$LOG" +trap "tail -n100 ${LOG}" EXIT + +if [[ ! -e "$CUR_PROMPT_FILE" ]]; then + sed -e "s/\[\[USER_NAME\]\]/${USER_NAME}/g" \ + -e "s/\[\[AI_NAME\]\]/${AI_NAME}/g" \ + -e "s/\[\[DATE_TIME\]\]/${DATE_TIME}/g" \ + -e "s/\[\[DATE_YEAR\]\]/${DATE_YEAR}/g" \ + "$PROMPT_TEMPLATE" >"$CUR_PROMPT_FILE" +fi + +if [[ ! -e "$NEXT_PROMPT_FILE" ]]; then + sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE" +fi + +if [[ "$(tail -c4 "$NEXT_PROMPT_FILE")" != "..." ]]; then + echo '...' >>"$NEXT_PROMPT_FILE" +fi + +if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then + echo 'Prompt cache does not exist, building...' + # Default batch_size to 8 here for better user feedback during initial prompt processing + ./main 2>>"$LOG" \ + --batch_size 8 \ + "${OPTS[@]}" \ + --prompt-cache "$PROMPT_CACHE_FILE" \ + --file "$CUR_PROMPT_FILE" \ + --n_predict 1 + echo + echo 'Done!' +fi + +if [[ ! -e "$CUR_PROMPT_CACHE" ]]; then + cp "$PROMPT_CACHE_FILE" "$CUR_PROMPT_CACHE" +fi +if [[ ! -e "$NEXT_PROMPT_CACHE" ]]; then + cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE" +fi + +printf '%s ' "$(< "$CUR_PROMPT_FILE")" +n_tokens=0 + +while read -e line; do + # Limit generation to remaining context, with a buffer and estimating 2 chars/token for input + n_predict=$((CTX_SIZE - n_tokens - ${#line} / 2 - 32)) + + # Swap prompts when we're about to run out of context + if ((n_predict <= 0)); then + wait # for background main (below) to finish with next prompt + mv "$NEXT_PROMPT_FILE" "$CUR_PROMPT_FILE" + mv "$NEXT_PROMPT_CACHE" "$CUR_PROMPT_CACHE" + + sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE" + echo '...' >>"$NEXT_PROMPT_FILE" + cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE" + + n_tokens=0 + n_predict=$((CTX_SIZE / 2)) + fi + + echo " ${line}" >>"$CUR_PROMPT_FILE" + if ((n_tokens > CTX_ROTATE_POINT)); then + echo " ${line}" >>"$NEXT_PROMPT_FILE" + fi + + n_prompt_len_pre=$(($(wc -c <"$CUR_PROMPT_FILE"))) + + printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE" + + ./main 2>>"$LOG" "${OPTS[@]}" \ + --prompt-cache "$CUR_PROMPT_CACHE" \ + --prompt-cache-all \ + --file "$CUR_PROMPT_FILE" \ + --reverse-prompt "${USER_NAME}:" \ + --n_predict "$n_predict" | + skip_bytes 1 | # skip BOS token added by ./main + tee "$CUR_PROMPT_FILE.tmp" | # save prompt + generation to tmp file + skip_bytes "$n_prompt_len_pre" # print generation + + mv "$CUR_PROMPT_FILE.tmp" "$CUR_PROMPT_FILE" + + # if we hit n_predict instead of reverse-prompt, we need to add the prompt + if [[ "$(tail -n1 "$CUR_PROMPT_FILE")" != "${USER_NAME}:" ]]; then + printf '\n%s:' "$USER_NAME" + printf '\n%s:' "$USER_NAME" >> "$CUR_PROMPT_FILE" + fi + + printf ' ' + + # HACK get num tokens from debug message + # TODO get both messages in one go + if ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" || + ! sample_time_msg="$( tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then + echo >&2 "Couldn't get number of tokens from ./main output!" + exit 1 + fi + + n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg"))) + + if ((n_tokens > CTX_ROTATE_POINT)); then + tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE" + fi + + # Update cache for next prompt in background, ideally during user input + ./main >>"$LOG_BG" 2>&1 "${OPTS[@]}" \ + --prompt-cache "$NEXT_PROMPT_CACHE" \ + --file "$NEXT_PROMPT_FILE" \ + --n_predict 1 & +done From 6986c7835adc13ba3f9d933b95671bb1f3984dc6 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 19 May 2023 21:17:28 +0300 Subject: [PATCH 12/20] tests : add missing header --- tests/test-sampling.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index ebfc17c18..0e675127f 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -5,6 +5,7 @@ #undef NDEBUG #endif +#include #include #include #include From 2d5db48371052087a83974abda3767d1aedec598 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 19 May 2023 22:17:18 +0300 Subject: [PATCH 13/20] ggml : use F16 instead of F32 in Q4_0, Q4_1, Q8_0 (#1508) * ggml : use F16 instead of F32 in Q4_0, Q4_1 and Q8_0 * llama : bump LLAMA_FILE_VERSION to 3 * cuda : update Q4 and Q8 dequantize kernels * ggml : fix AVX dot products * readme : update performance table + hot topics --- README.md | 21 +++---- ggml-cuda.cu | 14 ++--- ggml.c | 154 +++++++++++++++++++++++++-------------------------- ggml.h | 2 +- llama.cpp | 18 +++++- llama.h | 2 +- 6 files changed, 109 insertions(+), 102 deletions(-) diff --git a/README.md b/README.md index 6a67765aa..762f4aa03 100644 --- a/README.md +++ b/README.md @@ -9,6 +9,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ **Hot topics:** +- Quantization formats `Q4` and `Q8` have changed again (19 May) - [(info)](https://github.com/ggerganov/llama.cpp/pull/1508) - Quantization formats `Q4` and `Q5` have changed - requantize any old models [(info)](https://github.com/ggerganov/llama.cpp/pull/1405) - [Roadmap May 2023](https://github.com/ggerganov/llama.cpp/discussions/1220) @@ -334,16 +335,16 @@ Several quantization methods are supported. They differ in the resulting model d | Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 | |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:| -| 7B | perplexity | 5.9066 | 6.1565 | 6.0910 | 5.9862 | 5.9481 | 5.9069 | -| 7B | file size | 13.0G | 4.0G | 4.8G | 4.4G | 4.8G | 7.1G | -| 7B | ms/tok @ 4th | 128 | 50 | 54 | 75 | 83 | 75 | -| 7B | ms/tok @ 8th | 123 | 44 | 52 | 53 | 58 | 72 | -| 7B | bits/weight | 16.0 | 5.0 | 6.0 | 5.5 | 6.0 | 9.0 | -| 13B | perplexity | 5.2543 | 5.3860 | 5.3607 | 5.2856 | 5.2706 | 5.2548 | -| 13B | file size | 25.0G | 7.6G | 9.1G | 8.4G | 9.1G | 14G | -| 13B | ms/tok @ 4th | 239 | 93 | 101 | 150 | 164 | 141 | -| 13B | ms/tok @ 8th | 240 | 81 | 96 | 96 | 104 | 136 | -| 13B | bits/weight | 16.0 | 5.0 | 6.0 | 5.5 | 6.0 | 9.0 | +| 7B | perplexity | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 | +| 7B | file size | 13.0G | 3.5G | 3.9G | 4.3G | 4.7G | 6.7G | +| 7B | ms/tok @ 4th | 127 | 55 | 54 | 76 | 83 | 72 | +| 7B | ms/tok @ 8th | 122 | 43 | 45 | 52 | 56 | 67 | +| 7B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 | +| 13B | perplexity | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 | +| 13B | file size | 25.0G | 6.8G | 7.6G | 8.3G | 9.1G | 13G | +| 13B | ms/tok @ 4th | - | 103 | 105 | 148 | 160 | 131 | +| 13B | ms/tok @ 8th | - | 73 | 82 | 98 | 105 | 128 | +| 13B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 | ### Perplexity (measuring model quality) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index f2630ec8e..688bcf799 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -42,19 +42,19 @@ typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y, #define QK4_0 32 #define QR4_0 2 typedef struct { - float d; // delta + half d; // delta uint8_t qs[QK4_0 / 2]; // nibbles / quants } block_q4_0; -static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding"); +static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding"); #define QK4_1 32 #define QR4_1 2 typedef struct { - float d; // delta - float m; // min + half d; // delta + half m; // min uint8_t qs[QK4_1 / 2]; // nibbles / quants } block_q4_1; -static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); +static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); #define QK5_0 32 #define QR5_0 2 @@ -78,10 +78,10 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + #define QK8_0 32 #define QR8_0 1 typedef struct { - float d; // delta + half d; // delta int8_t qs[QK8_0]; // quants } block_q8_0; -static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding"); +static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding"); #define CUDA_DEQUANTIZE_BLOCK_SIZE 256 #define CUDA_DMMV_BLOCK_SIZE 32 // dmmv = dequantize_mul_mat_vec diff --git a/ggml.c b/ggml.c index dbef99312..1cb89636a 100644 --- a/ggml.c +++ b/ggml.c @@ -769,18 +769,18 @@ int32x4_t vcvtnq_s32_f32(float32x4_t v) { #define QK4_0 32 typedef struct { - float d; // delta + ggml_fp16_t d; // delta uint8_t qs[QK4_0 / 2]; // nibbles / quants } block_q4_0; -static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding"); +static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding"); #define QK4_1 32 typedef struct { - float d; // delta - float m; // min + ggml_fp16_t d; // delta + ggml_fp16_t m; // min uint8_t qs[QK4_1 / 2]; // nibbles / quants } block_q4_1; -static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding"); +static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding"); #define QK5_0 32 typedef struct { @@ -801,16 +801,16 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + #define QK8_0 32 typedef struct { - float d; // delta - int8_t qs[QK8_0]; // quants + ggml_fp16_t d; // delta + int8_t qs[QK8_0]; // quants } block_q8_0; -static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding"); +static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding"); #define QK8_1 32 typedef struct { - float d; // delta - float s; // d * sum(qs[i]) - int8_t qs[QK8_1]; // quants + float d; // delta + float s; // d * sum(qs[i]) + int8_t qs[QK8_1]; // quants } block_q8_1; static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); @@ -837,7 +837,7 @@ static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * r const float d = max / -8; const float id = d ? 1.0f/d : 0.0f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); for (int j = 0; j < qk/2; ++j) { const float x0 = x[i*qk + 0 + j]*id; @@ -877,8 +877,8 @@ static void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * r const float d = (max - min) / ((1 << 4) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = d; - y[i].m = min; + y[i].d = GGML_FP32_TO_FP16(d); + y[i].m = GGML_FP32_TO_FP16(min); for (int j = 0; j < qk/2; ++j) { const float x0 = (x[i*qk + 0 + j] - min)*id; @@ -1009,7 +1009,7 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); for (int j = 0; j < QK8_0; ++j) { const float x0 = x[i*QK8_0 + j]*id; @@ -1044,7 +1044,7 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); for (int j = 0; j < 8; j++) { const float32x4_t v = vmulq_n_f32(srcv[j], id); @@ -1079,7 +1079,7 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int // Quantize these floats const float d = maxScalar / 127.f; - y[i].d = d; + y[i].d = GGML_FP32_TO_FP16(d); const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; const __m256 mul = _mm256_set1_ps( id ); @@ -1178,7 +1178,7 @@ static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * r sum += y[i].qs[QK8_1/2 + j]; } - y[i].s = d * sum; + y[i].s = sum*d; } } @@ -1330,7 +1330,7 @@ static void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict const int nb = k / qk; for (int i = 0; i < nb; i++) { - const float d = x[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d); for (int j = 0; j < qk/2; ++j) { const int x0 = (x[i].qs[j] & 0x0F) - 8; @@ -1350,8 +1350,8 @@ static void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict const int nb = k / qk; for (int i = 0; i < nb; i++) { - const float d = x[i].d; - const float m = x[i].m; + const float d = GGML_FP16_TO_FP32(x[i].d); + const float m = GGML_FP16_TO_FP32(x[i].m); for (int j = 0; j < qk/2; ++j) { const int x0 = (x[i].qs[j] & 0x0F); @@ -1426,7 +1426,7 @@ static void dequantize_row_q8_0(const void * restrict vx, float * restrict y, in const block_q8_0 * restrict x = vx; for (int i = 0; i < nb; i++) { - const float d = x[i].d; + const float d = GGML_FP16_TO_FP32(x[i].d); for (int j = 0; j < qk; ++j) { y[i*qk + j] = x[i].qs[j]*d; @@ -1690,8 +1690,9 @@ quantize_fns_t ggml_internal_get_quantize_fn(size_t i) { static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) { float tmp[8]; - for (int i = 0; i < 8; i++) + for (int i = 0; i < 8; i++) { tmp[i] = GGML_FP16_TO_FP32(x[i]); + } return _mm256_loadu_ps(tmp); } @@ -2111,8 +2112,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * const block_q8_0 * restrict y0 = &y[i + 0]; const block_q8_0 * restrict y1 = &y[i + 1]; - const uint8x16_t m4b = vdupq_n_u8(0x0F); - const int8x16_t s8b = vdupq_n_s8(0x8); + const uint8x16_t m4b = vdupq_n_u8(0x0F); + const int8x16_t s8b = vdupq_n_s8(0x8); const uint8x16_t v0_0 = vld1q_u8(x0->qs); const uint8x16_t v0_1 = vld1q_u8(x1->qs); @@ -2140,8 +2141,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); #else const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l)); const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l)); @@ -2158,8 +2159,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); #endif } @@ -2171,7 +2172,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * // Main loop for (int i = 0; i < nb; ++i) { /* Compute combined scale for the block */ - const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); + const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); __m256i bx = bytes_from_nibbles_32(x[i].qs); @@ -2195,7 +2196,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * // Main loop for (int i = 0; i < nb; ++i) { // Compute combined scale for the block - const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); + const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); const __m128i lowMask = _mm_set1_epi8(0xF); const __m128i off = _mm_set1_epi8(8); @@ -2237,7 +2238,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0); // Compute combined scale for the block 0 and 1 - const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[0].d ), _mm_set1_ps( y[0].d ) ); + const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[0].d) * GGML_FP16_TO_FP32(y[0].d) ); const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs); @@ -2255,7 +2256,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * _mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0); // Compute combined scale for the block 2 and 3 - const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[1].d ), _mm_set1_ps( y[1].d ) ); + const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[1].d) * GGML_FP16_TO_FP32(y[1].d) ); const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs); @@ -2288,7 +2289,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0); // Compute combined scale for the block 0 and 1 - const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[i].d ), _mm_set1_ps( y[i].d ) ); + const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) ); const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs); @@ -2306,7 +2307,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * _mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0); // Compute combined scale for the block 2 and 3 - const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[i + 1].d ), _mm_set1_ps( y[i + 1].d ) ); + const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i + 1].d) * GGML_FP16_TO_FP32(y[i + 1].d) ); const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs); @@ -2354,7 +2355,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } - sumf += (x[i].d*y[i].d)*sumi; + sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d); } *s = sumf; @@ -2384,7 +2385,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const block_q8_1 * restrict y0 = &y[i + 0]; const block_q8_1 * restrict y1 = &y[i + 1]; - summs += x0->m * y0->s + x1->m * y1->s; + summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s; const uint8x16_t m4b = vdupq_n_u8(0x0F); @@ -2408,8 +2409,8 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d); #else const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l)); const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l)); @@ -2426,8 +2427,8 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d); #endif } @@ -2440,13 +2441,13 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * // Main loop for (int i = 0; i < nb; ++i) { - const float * d0 = &x[i].d; - const float * d1 = &y[i].d; + const float d0 = GGML_FP16_TO_FP32(x[i].d); + const float d1 = y[i].d; - summs += x[i].m * y[i].s; + summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; - const __m256 d0v = _mm256_broadcast_ss( d0 ); - const __m256 d1v = _mm256_broadcast_ss( d1 ); + const __m256 d0v = _mm256_set1_ps( d0 ); + const __m256 d1v = _mm256_set1_ps( d1 ); // Compute combined scales const __m256 d0d1 = _mm256_mul_ps( d0v, d1v ); @@ -2480,7 +2481,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } - sumf += (x[i].d*y[i].d)*sumi + x[i].m*y[i].s; + sumf += (GGML_FP16_TO_FP32(x[i]).d*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; } *s = sumf; @@ -2556,16 +2557,13 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - const float x0d = GGML_FP16_TO_FP32(x0->d); - const float x1d = GGML_FP16_TO_FP32(x1->d); - #if defined(__ARM_FEATURE_DOTPROD) sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), - vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d); + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), - vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d); + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); #else const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); @@ -2582,8 +2580,8 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); #endif } @@ -2658,7 +2656,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * // Main loop for (int i = 0; i < nb; i++) { /* Compute combined scale for the block */ - const __m256 d = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)), _mm256_broadcast_ss(&y[i].d)); + const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); __m256i bx = bytes_from_nibbles_32(x[i].qs); __m256i bxhi = bytes_from_bits_32(x[i].qh); @@ -2682,7 +2680,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * // Main loop for (int i = 0; i < nb; i++) { /* Compute combined scale for the block */ - const __m256 d = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)), _mm256_broadcast_ss(&y[i].d)); + const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); __m256i bx = bytes_from_nibbles_32(x[i].qs); const __m256i bxhi = bytes_from_bits_32(x[i].qh); @@ -2725,7 +2723,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); } - sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi; + sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi; } *s = sumf; @@ -2807,16 +2805,13 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - const float x0d = GGML_FP16_TO_FP32(x0->d); - const float x1d = GGML_FP16_TO_FP32(x1->d); - #if defined(__ARM_FEATURE_DOTPROD) sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), - vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d); + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), - vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d); + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d); #else const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); @@ -2833,8 +2828,8 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d); #endif } @@ -2894,15 +2889,14 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); - const float x0d = GGML_FP16_TO_FP32(x0->d); - // dot product - sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4( - wasm_i32x4_add( - wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), - wasm_i32x4_dot_i16x8(v0lfh, v1lh)), - wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), - wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d))); + sumv = wasm_f32x4_add(sumv, + wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add( + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), + wasm_i32x4_dot_i16x8(v0lfh, v1lh)), + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), + wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), + wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d)); } *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + @@ -2924,7 +2918,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10)); bx = _mm256_or_si256(bx, bxhi); - const __m256 dy = _mm256_broadcast_ss(&y[i].d); + const __m256 dy = _mm256_set1_ps(y[i].d); const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); const __m256 q = mul_sum_us8_pairs_float(bx, by); @@ -2958,7 +2952,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * bxh = _mm_or_si128(bxh, bxhih); bx = _mm256_set_m128i(bxh, bxl); - const __m256 dy = _mm256_broadcast_ss(&y[i].d); + const __m256 dy = _mm256_set1_ps(y[i].d); const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); const __m256 q = mul_sum_us8_pairs_float(bx, by); @@ -3028,11 +3022,11 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * #if defined(__ARM_FEATURE_DOTPROD) sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), - vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), x0->d*y0->d); + vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), - vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), x1->d*y1->d); + vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); #else const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0)); @@ -3050,8 +3044,8 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1)); const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3)); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), x0->d*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), x1->d*y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); #endif } @@ -3063,7 +3057,7 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * // Main loop for (int i = 0; i < nb; ++i) { // Compute combined scale for the block - const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); + const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d)); __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs); __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); @@ -3089,7 +3083,7 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * sumi += x[i].qs[j]*y[i].qs[j]; } - sumf += (x[i].d*y[i].d)*sumi; + sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)); } *s = sumf; diff --git a/ggml.h b/ggml.h index 255541d02..dce5ca1e7 100644 --- a/ggml.h +++ b/ggml.h @@ -190,7 +190,7 @@ #define GGML_FILE_MAGIC 0x67676d6c // "ggml" #define GGML_FILE_VERSION 1 -#define GGML_QNT_VERSION 1 // bump this on quantization format changes +#define GGML_QNT_VERSION 2 // bump this on quantization format changes #define GGML_QNT_VERSION_FACTOR 1000 // do not change this #define GGML_MAX_DIMS 4 diff --git a/llama.cpp b/llama.cpp index 1802d2319..6ebe85d0f 100644 --- a/llama.cpp +++ b/llama.cpp @@ -406,6 +406,7 @@ enum llama_file_version { LLAMA_FILE_VERSION_GGMF_V1, // added version field and scores in vocab LLAMA_FILE_VERSION_GGJT_V1, // added padding LLAMA_FILE_VERSION_GGJT_V2, // changed quantization format + LLAMA_FILE_VERSION_GGJT_V3, // changed Q4 and Q8 quantization format }; struct llama_file_loader { @@ -438,6 +439,8 @@ struct llama_file_loader { file_version = LLAMA_FILE_VERSION_GGJT_V1; } else if (magic == 'ggjt' && version == 2) { file_version = LLAMA_FILE_VERSION_GGJT_V2; + } else if (magic == 'ggjt' && version == 3) { + file_version = LLAMA_FILE_VERSION_GGJT_V3; } else { throw format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?", magic, version); @@ -844,7 +847,8 @@ static const char *llama_file_version_name(llama_file_version version) { case LLAMA_FILE_VERSION_GGML: return "'ggml' (old version with low tokenizer quality and no mmap support)"; case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (old version with no mmap support)"; case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (pre #1405)"; - case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (latest)"; + case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (pre #1508)"; + case LLAMA_FILE_VERSION_GGJT_V3: return "ggjt v3 (latest)"; } return "unknown"; @@ -924,11 +928,19 @@ static void llama_model_load_internal( fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type)); } - if (file_version != LLAMA_FILE_VERSION_GGJT_V2) { + if (file_version < LLAMA_FILE_VERSION_GGJT_V2) { if (hparams.ftype != LLAMA_FTYPE_ALL_F32 && hparams.ftype != LLAMA_FTYPE_MOSTLY_F16 && hparams.ftype != LLAMA_FTYPE_MOSTLY_Q8_0) { - throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1305)"); + throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1405)"); + } + } + + if (file_version < LLAMA_FILE_VERSION_GGJT_V3) { + if (hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || + hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_1 || + hparams.ftype == LLAMA_FTYPE_MOSTLY_Q8_0) { + throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1508)"); } } diff --git a/llama.h b/llama.h index f955fa23d..fd3f21e5f 100644 --- a/llama.h +++ b/llama.h @@ -19,7 +19,7 @@ # define LLAMA_API #endif -#define LLAMA_FILE_VERSION 2 +#define LLAMA_FILE_VERSION 3 #define LLAMA_FILE_MAGIC 'ggjt' #define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml' #define LLAMA_SESSION_MAGIC 'ggsn' From 4fd3e29297e3246a7be291932c115636fadb0f52 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 20 May 2023 10:13:19 +0300 Subject: [PATCH 14/20] ggml : fix scalar implementation of Q4_1 dot --- ggml.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml.c b/ggml.c index 1cb89636a..101cb733b 100644 --- a/ggml.c +++ b/ggml.c @@ -2481,7 +2481,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } - sumf += (GGML_FP16_TO_FP32(x[i]).d*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; } *s = sumf; From 8a203f9fa1b24e010be8f35ebbbd6786293684cb Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 20 May 2023 10:14:31 +0300 Subject: [PATCH 15/20] llama : fix compile warnings in llama_set_state_data() --- llama.cpp | 4 ++-- llama.h | 2 +- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/llama.cpp b/llama.cpp index 6ebe85d0f..68e3bec13 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2618,8 +2618,8 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { } // Sets the state reading from the specified source address -size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { - const uint8_t * inp = src; +size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { + uint8_t * inp = src; // set rng { diff --git a/llama.h b/llama.h index fd3f21e5f..8623e08ce 100644 --- a/llama.h +++ b/llama.h @@ -138,7 +138,7 @@ extern "C" { // Set the state reading from the specified address // Returns the number of bytes read - LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src); + LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src); // Save/load session file LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out); From 503db28849d8641d66244385e7e9649608a2e4d0 Mon Sep 17 00:00:00 2001 From: Maxime <672982+maximegmd@users.noreply.github.com> Date: Sat, 20 May 2023 09:22:37 +0200 Subject: [PATCH 16/20] llama : fix name shadowing and C4146 (#1526) * Fix name shadowing and C4146 * Fix if macros not using defined when required * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Update llama-util.h Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * Code style Co-authored-by: Georgi Gerganov --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Georgi Gerganov --- ggml.c | 4 ++-- llama-util.h | 40 ++++++++++++++++++++-------------------- llama.cpp | 7 ++++--- 3 files changed, 26 insertions(+), 25 deletions(-) diff --git a/ggml.c b/ggml.c index 101cb733b..939ab4d62 100644 --- a/ggml.c +++ b/ggml.c @@ -512,7 +512,7 @@ static inline int hsum_i32_4(const __m128i a) { return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32)); } -#if __AVX2__ || __AVX512F__ +#if defined(__AVX2__) || defined(__AVX512F__) // spread 32 bits to 32 bytes { 0x00, 0xFF } static inline __m256i bytes_from_bits_32(const uint8_t * x) { uint32_t x32; @@ -688,7 +688,7 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 #endif // __AVX__ || __AVX2__ || __AVX512F__ #endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) -#if __ARM_NEON +#if defined(__ARM_NEON) #if !defined(__aarch64__) diff --git a/llama-util.h b/llama-util.h index 88ec28dca..a79c5dadd 100644 --- a/llama-util.h +++ b/llama-util.h @@ -101,12 +101,12 @@ struct llama_file { LLAMA_ASSERT(ret == 0); // same } - void read_raw(void * ptr, size_t size) { - if (size == 0) { + void read_raw(void * ptr, size_t len) const { + if (len == 0) { return; } errno = 0; - std::size_t ret = std::fread(ptr, size, 1, fp); + std::size_t ret = std::fread(ptr, len, 1, fp); if (ferror(fp)) { throw std::runtime_error(format("read error: %s", strerror(errno))); } @@ -127,12 +127,12 @@ struct llama_file { return std::string(chars.data(), len); } - void write_raw(const void * ptr, size_t size) { - if (size == 0) { + void write_raw(const void * ptr, size_t len) const { + if (len == 0) { return; } errno = 0; - size_t ret = std::fwrite(ptr, size, 1, fp); + size_t ret = std::fwrite(ptr, len, 1, fp); if (ret != 1) { throw std::runtime_error(format("write error: %s", strerror(errno))); } @@ -267,9 +267,9 @@ struct llama_mlock { } } - void init(void * addr) { - LLAMA_ASSERT(this->addr == NULL && this->size == 0); - this->addr = addr; + void init(void * ptr) { + LLAMA_ASSERT(addr == NULL && size == 0); + addr = ptr; } void grow_to(size_t target_size) { @@ -340,14 +340,14 @@ struct llama_mlock { return (size_t) si.dwPageSize; } - bool raw_lock(void * addr, size_t size) { + bool raw_lock(void * ptr, size_t len) { for (int tries = 1; ; tries++) { - if (VirtualLock(addr, size)) { + if (VirtualLock(ptr, len)) { return true; } if (tries == 2) { fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n", - size, this->size, llama_format_win_err(GetLastError()).c_str()); + len, size, llama_format_win_err(GetLastError()).c_str()); return false; } @@ -363,7 +363,7 @@ struct llama_mlock { // is equal to the number of pages in its minimum working set minus // a small overhead." // Hopefully a megabyte is enough overhead: - size_t increment = size + 1048576; + size_t increment = len + 1048576; // The minimum must be <= the maximum, so we need to increase both: min_ws_size += increment; max_ws_size += increment; @@ -375,8 +375,8 @@ struct llama_mlock { } } - void raw_unlock(void * addr, size_t size) { - if (!VirtualUnlock(addr, size)) { + void raw_unlock(void * ptr, size_t len) { + if (!VirtualUnlock(ptr, len)) { fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n", llama_format_win_err(GetLastError()).c_str()); } @@ -388,12 +388,12 @@ struct llama_mlock { return (size_t) 65536; } - bool raw_lock(const void * addr, size_t size) { + bool raw_lock(const void * addr, size_t len) { fprintf(stderr, "warning: mlock not supported on this system\n"); return false; } - void raw_unlock(const void * addr, size_t size) {} + void raw_unlock(const void * addr, size_t len) {} #endif }; @@ -404,10 +404,10 @@ struct llama_buffer { llama_buffer() = default; - void resize(size_t size) { + void resize(size_t len) { delete[] addr; - addr = new uint8_t[size]; - this->size = size; + addr = new uint8_t[len]; + size = len; } ~llama_buffer() { diff --git a/llama.cpp b/llama.cpp index 68e3bec13..dd449592a 100644 --- a/llama.cpp +++ b/llama.cpp @@ -45,6 +45,7 @@ enum e_model { MODEL_65B, }; + static const size_t MB = 1024*1024; // computed for n_ctx == 2048 @@ -110,7 +111,7 @@ struct llama_hparams { enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16; bool operator!=(const llama_hparams & other) const { - return memcmp(this, &other, sizeof(llama_hparams)); + return static_cast(memcmp(this, &other, sizeof(llama_hparams))); } }; @@ -502,7 +503,7 @@ struct llama_file_loader { if (file_version >= LLAMA_FILE_VERSION_GGJT_V1) { // skip to the next multiple of 32 bytes - file.seek(-file.tell() & 31, SEEK_CUR); + file.seek(-static_cast(file.tell()) & 31, SEEK_CUR); } shard.file_idx = file_idx; shard.file_off = file.tell(); @@ -577,7 +578,7 @@ struct llama_file_saver { file.write_u32(new_type); file.write_raw(tensor.ne.data(), sizeof(tensor.ne[0]) * tensor.ne.size()); file.write_raw(tensor.name.data(), tensor.name.size()); - file.seek(-file.tell() & 31, SEEK_CUR); + file.seek(-static_cast(file.tell()) & 31, SEEK_CUR); LLAMA_ASSERT(new_size == llama_calc_tensor_size(tensor.ne, new_type)); file.write_raw(new_data, new_size); } From d2c59b8ba498ab01e65203dde6fe95236d20f6e7 Mon Sep 17 00:00:00 2001 From: DannyDaemonic Date: Sat, 20 May 2023 00:40:02 -0700 Subject: [PATCH 17/20] Fix for mingw (#1462) --- examples/common.cpp | 32 ++++++++++++++++++++++++++++++++ 1 file changed, 32 insertions(+) diff --git a/examples/common.cpp b/examples/common.cpp index e89df537e..1308f8410 100644 --- a/examples/common.cpp +++ b/examples/common.cpp @@ -578,6 +578,37 @@ void console_set_color(console_state & con_st, console_color_t color) { } char32_t getchar32() { +#if defined(_WIN32) + HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE); + wchar_t high_surrogate = 0; + + while (true) { + INPUT_RECORD record; + DWORD count; + if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) { + return WEOF; + } + + if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) { + wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar; + if (wc == 0) { + continue; + } + + if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate + high_surrogate = wc; + continue; + } else if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate + if (high_surrogate != 0) { // Check if we have a high surrogate + return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000; + } + } + + high_surrogate = 0; // Reset the high surrogate + return static_cast(wc); + } + } +#else wchar_t wc = getwchar(); if (static_cast(wc) == WEOF) { return WEOF; @@ -596,6 +627,7 @@ char32_t getchar32() { #endif return static_cast(wc); +#endif } void pop_cursor(console_state & con_st) { From ec2e10c4443209da56b431b24dd0845b60e757fb Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 20 May 2023 11:06:11 +0300 Subject: [PATCH 18/20] llama : add llama_init_backend() API (close #1527) --- examples/benchmark/benchmark-matmult.cpp | 3 ++- examples/embedding/embedding.cpp | 2 ++ examples/main/main.cpp | 3 +-- examples/perplexity/perplexity.cpp | 2 ++ examples/quantize/quantize.cpp | 21 ++++++---------- llama.cpp | 15 ++++++++++++ llama.h | 31 +++++++++++++++--------- 7 files changed, 48 insertions(+), 29 deletions(-) diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp index 446b8e8fb..9f9ed9db0 100644 --- a/examples/benchmark/benchmark-matmult.cpp +++ b/examples/benchmark/benchmark-matmult.cpp @@ -1,6 +1,7 @@ -#include #include "ggml.h" #include "build-info.h" + +#include #include #include #include diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index c24f7f820..03603b10f 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -31,6 +31,8 @@ int main(int argc, char ** argv) { params.prompt = gpt_random_prompt(rng); } + llama_init_backend(); + llama_context * ctx; // load the model diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 4d886f8de..47b418d97 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -96,8 +96,7 @@ int main(int argc, char ** argv) { params.prompt = gpt_random_prompt(rng); } -// params.prompt = R"(// this function checks if the number n is prime -//bool is_prime(int n) {)"; + llama_init_backend(); llama_context * ctx; g_ctx = &ctx; diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 9d38626cb..e19c6825f 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -143,6 +143,8 @@ int main(int argc, char ** argv) { params.prompt = gpt_random_prompt(rng); } + llama_init_backend(); + llama_context * ctx; // load the model and apply lora adapter, if any diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 115d8fb1b..769dd36a4 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -1,7 +1,7 @@ -#include "ggml.h" -#include "llama.h" #include "build-info.h" +#include "llama.h" + #include #include #include @@ -42,8 +42,6 @@ bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::st // ./quantize models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads] // int main(int argc, char ** argv) { - ggml_time_init(); - if (argc < 3) { fprintf(stderr, "usage: %s model-f32.bin [model-quant.bin] type [nthreads]\n", argv[0]); for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) { @@ -52,12 +50,7 @@ int main(int argc, char ** argv) { return 1; } - // needed to initialize f16 tables - { - struct ggml_init_params params = { 0, NULL, false }; - struct ggml_context * ctx = ggml_init(params); - ggml_free(ctx); - } + llama_init_backend(); // parse command line arguments const std::string fname_inp = argv[1]; @@ -116,25 +109,25 @@ int main(int argc, char ** argv) { } fprintf(stderr, "\n"); - const int64_t t_main_start_us = ggml_time_us(); + const int64_t t_main_start_us = llama_time_us(); int64_t t_quantize_us = 0; // load the model { - const int64_t t_start_us = ggml_time_us(); + const int64_t t_start_us = llama_time_us(); if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) { fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str()); return 1; } - t_quantize_us = ggml_time_us() - t_start_us; + t_quantize_us = llama_time_us() - t_start_us; } // report timing { - const int64_t t_main_end_us = ggml_time_us(); + const int64_t t_main_end_us = llama_time_us(); printf("\n"); printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0); diff --git a/llama.cpp b/llama.cpp index dd449592a..5e6980b48 100644 --- a/llama.cpp +++ b/llama.cpp @@ -839,6 +839,21 @@ bool llama_mlock_supported() { return llama_mlock::SUPPORTED; } +void llama_init_backend() { + ggml_time_init(); + + // needed to initialize f16 tables + { + struct ggml_init_params params = { 0, NULL, false }; + struct ggml_context * ctx = ggml_init(params); + ggml_free(ctx); + } +} + +int64_t llama_time_us() { + return ggml_time_us(); +} + // // model loading // diff --git a/llama.h b/llama.h index 8623e08ce..0a63d034b 100644 --- a/llama.h +++ b/llama.h @@ -40,9 +40,9 @@ extern "C" { typedef int llama_token; typedef struct llama_token_data { - llama_token id; // token id - float logit; // log-odds of the token - float p; // probability of the token + llama_token id; // token id + float logit; // log-odds of the token + float p; // probability of the token } llama_token_data; typedef struct llama_token_data_array { @@ -73,16 +73,16 @@ extern "C" { // model file types enum llama_ftype { - LLAMA_FTYPE_ALL_F32 = 0, - LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors + LLAMA_FTYPE_ALL_F32 = 0, + LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 - // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed - // LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed - LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors + // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed + // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed + LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors }; LLAMA_API struct llama_context_params llama_context_default_params(); @@ -90,6 +90,13 @@ extern "C" { LLAMA_API bool llama_mmap_supported(); LLAMA_API bool llama_mlock_supported(); + // TODO: not great API - very likely to change + // Initialize the llama + ggml backend + // Call once at the start of the program + LLAMA_API void llama_init_backend(); + + LLAMA_API int64_t llama_time_us(); + // Various functions for loading a ggml llama model. // Allocate (almost) all memory needed for the model. // Return NULL on failure From 07e9ace0f9da424d82e75df969642522880feb92 Mon Sep 17 00:00:00 2001 From: Zenix Date: Sat, 20 May 2023 18:02:48 +0900 Subject: [PATCH 19/20] feature : add blis and other BLAS implementation support (#1502) * feature: add blis support * feature: allow all BLA_VENDOR to be assigned in cmake arguments. align with whisper.cpp pr 927 * fix: version detection for BLA_SIZEOF_INTEGER, recover min version of cmake * Fix typo in INTEGER Co-authored-by: Georgi Gerganov --------- Co-authored-by: Georgi Gerganov --- BLIS.md | 67 ++++++++++++++++++++++++++++++++++++++++++++++++++ CMakeLists.txt | 37 +++++++++++----------------- Makefile | 4 +++ README.md | 19 ++++++++++++-- 4 files changed, 103 insertions(+), 24 deletions(-) create mode 100644 BLIS.md diff --git a/BLIS.md b/BLIS.md new file mode 100644 index 000000000..9b3c30605 --- /dev/null +++ b/BLIS.md @@ -0,0 +1,67 @@ +BLIS Installation Manual +------------------------ + +BLIS is a portable software framework for high-performance BLAS-like dense linear algebra libraries. It has received awards and recognition, including the 2023 James H. Wilkinson Prize for Numerical Software and the 2020 SIAM Activity Group on Supercomputing Best Paper Prize. BLIS provides a new BLAS-like API and a compatibility layer for traditional BLAS routine calls. It offers features such as object-based API, typed API, BLAS and CBLAS compatibility layers. + +Project URL: https://github.com/flame/blis + +### Prepare: + +Compile BLIS: + +```bash +git clone https://github.com/flame/blis +cd blis +./configure --enable-cblas -t openmp,pthreads auto +# will install to /usr/local/ by default. +make -j +``` + +Install BLIS: + +```bash +sudo make install +``` + +We recommend using openmp since it's easier to modify the cores been used. + +### llama.cpp compilation + +Makefile: + +```bash +make LLAMA_BLIS=1 -j +# make LLAMA_BLIS=1 benchmark-matmult +``` + +CMake: + +```bash +mkdir build +cd build +cmake -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=FLAME .. +make -j +``` + +### llama.cpp execution + +According to the BLIS documentation, we could set the following +environment variables to modify the behavior of openmp: + +``` +export GOMP_GPU_AFFINITY="0-19" +export BLIS_NUM_THREADS=14 +``` + +And then run the binaries as normal. + + +### Intel specific issue + +Some might get the error message saying that `libimf.so` cannot be found. +Please follow this [stackoverflow page](https://stackoverflow.com/questions/70687930/intel-oneapi-2022-libimf-so-no-such-file-or-directory-during-openmpi-compila). + +### Reference: + +1. https://github.com/flame/blis#getting-started +2. https://github.com/flame/blis/blob/master/docs/Multithreading.md diff --git a/CMakeLists.txt b/CMakeLists.txt index 48e3238df..0876ab90a 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -65,7 +65,8 @@ endif() # 3rd party libs option(LLAMA_ACCELERATE "llama: enable Accelerate framework" ON) -option(LLAMA_OPENBLAS "llama: use OpenBLAS" OFF) +option(LLAMA_BLAS "llama: use BLAS" OFF) +option(LLAMA_BLAS_VENDOR "llama: BLA_VENDOR from https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors" Generic) option(LLAMA_CUBLAS "llama: use cuBLAS" OFF) option(LLAMA_CLBLAST "llama: use CLBlast" OFF) @@ -145,36 +146,28 @@ if (APPLE AND LLAMA_ACCELERATE) endif() endif() -if (LLAMA_OPENBLAS) +if (LLAMA_BLAS) if (LLAMA_STATIC) set(BLA_STATIC ON) endif() - - set(BLA_VENDOR OpenBLAS) + if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22) + set(BLA_SIZEOF_INTEGER 8) + endif() + set(BLA_VENDOR ${LLAMA_BLAS_VENDOR}) find_package(BLAS) if (BLAS_FOUND) - message(STATUS "OpenBLAS found") + message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}") + add_compile_options(${BLAS_LINKER_FLAGS}) add_compile_definitions(GGML_USE_OPENBLAS) - add_link_options(${BLAS_LIBRARIES}) - set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} openblas) + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES}) - # find header file - set(OPENBLAS_INCLUDE_SEARCH_PATHS - /usr/include - /usr/include/openblas - /usr/include/openblas-base - /usr/local/include - /usr/local/include/openblas - /usr/local/include/openblas-base - /opt/OpenBLAS/include - $ENV{OpenBLAS_HOME} - $ENV{OpenBLAS_HOME}/include - ) - find_path(OPENBLAS_INC NAMES cblas.h PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS}) - add_compile_options(-I${OPENBLAS_INC}) + message("${BLAS_LIBRARIES}") + include_directories(${BLAS_INCLUDE_DIRS}) else() - message(WARNING "OpenBLAS not found") + message(WARNING "BLAS not found, please refer to " + "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors" + " to set correct LLAMA_BLAS_VENDOR") endif() endif() diff --git a/Makefile b/Makefile index f9ec8797a..cefa0b4a5 100644 --- a/Makefile +++ b/Makefile @@ -122,6 +122,10 @@ ifdef LLAMA_OPENBLAS LDFLAGS += -lopenblas endif endif +ifdef LLAMA_BLIS + CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis + LDFLAGS += -lblis -L/usr/local/lib +endif 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 diff --git a/README.md b/README.md index 762f4aa03..102cde43f 100644 --- a/README.md +++ b/README.md @@ -56,7 +56,7 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant - Mixed F16 / F32 precision - 4-bit, 5-bit and 8-bit integer quantization support - Runs on the CPU -- OpenBLAS 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 The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022). @@ -274,10 +274,25 @@ Building the program with BLAS support may lead to some performance improvements ```bash mkdir build cd build - cmake .. -DLLAMA_OPENBLAS=ON + cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS cmake --build . --config Release ``` +- BLIS + + Check [BLIS.md](BLIS.md) for more information. + +- Intel MKL + + By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. You may also specify it by: + + ```bash + mkdir build + cd build + cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx + cmake --build . -config Release + ``` + - cuBLAS This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). From ea600071cb005267e9e8f2629c1e406dd5fde083 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 20 May 2023 12:03:48 +0300 Subject: [PATCH 20/20] Revert "feature : add blis and other BLAS implementation support (#1502)" This reverts commit 07e9ace0f9da424d82e75df969642522880feb92. --- BLIS.md | 67 -------------------------------------------------- CMakeLists.txt | 37 +++++++++++++++++----------- Makefile | 4 --- README.md | 19 ++------------ 4 files changed, 24 insertions(+), 103 deletions(-) delete mode 100644 BLIS.md diff --git a/BLIS.md b/BLIS.md deleted file mode 100644 index 9b3c30605..000000000 --- a/BLIS.md +++ /dev/null @@ -1,67 +0,0 @@ -BLIS Installation Manual ------------------------- - -BLIS is a portable software framework for high-performance BLAS-like dense linear algebra libraries. It has received awards and recognition, including the 2023 James H. Wilkinson Prize for Numerical Software and the 2020 SIAM Activity Group on Supercomputing Best Paper Prize. BLIS provides a new BLAS-like API and a compatibility layer for traditional BLAS routine calls. It offers features such as object-based API, typed API, BLAS and CBLAS compatibility layers. - -Project URL: https://github.com/flame/blis - -### Prepare: - -Compile BLIS: - -```bash -git clone https://github.com/flame/blis -cd blis -./configure --enable-cblas -t openmp,pthreads auto -# will install to /usr/local/ by default. -make -j -``` - -Install BLIS: - -```bash -sudo make install -``` - -We recommend using openmp since it's easier to modify the cores been used. - -### llama.cpp compilation - -Makefile: - -```bash -make LLAMA_BLIS=1 -j -# make LLAMA_BLIS=1 benchmark-matmult -``` - -CMake: - -```bash -mkdir build -cd build -cmake -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=FLAME .. -make -j -``` - -### llama.cpp execution - -According to the BLIS documentation, we could set the following -environment variables to modify the behavior of openmp: - -``` -export GOMP_GPU_AFFINITY="0-19" -export BLIS_NUM_THREADS=14 -``` - -And then run the binaries as normal. - - -### Intel specific issue - -Some might get the error message saying that `libimf.so` cannot be found. -Please follow this [stackoverflow page](https://stackoverflow.com/questions/70687930/intel-oneapi-2022-libimf-so-no-such-file-or-directory-during-openmpi-compila). - -### Reference: - -1. https://github.com/flame/blis#getting-started -2. https://github.com/flame/blis/blob/master/docs/Multithreading.md diff --git a/CMakeLists.txt b/CMakeLists.txt index 0876ab90a..48e3238df 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -65,8 +65,7 @@ endif() # 3rd party libs option(LLAMA_ACCELERATE "llama: enable Accelerate framework" ON) -option(LLAMA_BLAS "llama: use BLAS" OFF) -option(LLAMA_BLAS_VENDOR "llama: BLA_VENDOR from https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors" Generic) +option(LLAMA_OPENBLAS "llama: use OpenBLAS" OFF) option(LLAMA_CUBLAS "llama: use cuBLAS" OFF) option(LLAMA_CLBLAST "llama: use CLBlast" OFF) @@ -146,28 +145,36 @@ if (APPLE AND LLAMA_ACCELERATE) endif() endif() -if (LLAMA_BLAS) +if (LLAMA_OPENBLAS) if (LLAMA_STATIC) set(BLA_STATIC ON) endif() - if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22) - set(BLA_SIZEOF_INTEGER 8) - endif() - set(BLA_VENDOR ${LLAMA_BLAS_VENDOR}) + + set(BLA_VENDOR OpenBLAS) find_package(BLAS) if (BLAS_FOUND) - message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}") + message(STATUS "OpenBLAS found") - add_compile_options(${BLAS_LINKER_FLAGS}) add_compile_definitions(GGML_USE_OPENBLAS) - set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES}) + add_link_options(${BLAS_LIBRARIES}) + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} openblas) - message("${BLAS_LIBRARIES}") - include_directories(${BLAS_INCLUDE_DIRS}) + # find header file + set(OPENBLAS_INCLUDE_SEARCH_PATHS + /usr/include + /usr/include/openblas + /usr/include/openblas-base + /usr/local/include + /usr/local/include/openblas + /usr/local/include/openblas-base + /opt/OpenBLAS/include + $ENV{OpenBLAS_HOME} + $ENV{OpenBLAS_HOME}/include + ) + find_path(OPENBLAS_INC NAMES cblas.h PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS}) + add_compile_options(-I${OPENBLAS_INC}) else() - message(WARNING "BLAS not found, please refer to " - "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors" - " to set correct LLAMA_BLAS_VENDOR") + message(WARNING "OpenBLAS not found") endif() endif() diff --git a/Makefile b/Makefile index cefa0b4a5..f9ec8797a 100644 --- a/Makefile +++ b/Makefile @@ -122,10 +122,6 @@ ifdef LLAMA_OPENBLAS LDFLAGS += -lopenblas endif endif -ifdef LLAMA_BLIS - CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis - LDFLAGS += -lblis -L/usr/local/lib -endif 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 diff --git a/README.md b/README.md index 102cde43f..762f4aa03 100644 --- a/README.md +++ b/README.md @@ -56,7 +56,7 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant - Mixed F16 / F32 precision - 4-bit, 5-bit and 8-bit integer quantization support - Runs on the CPU -- 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 +- OpenBLAS support - cuBLAS and CLBlast support The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022). @@ -274,25 +274,10 @@ Building the program with BLAS support may lead to some performance improvements ```bash mkdir build cd build - cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS + cmake .. -DLLAMA_OPENBLAS=ON cmake --build . --config Release ``` -- BLIS - - Check [BLIS.md](BLIS.md) for more information. - -- Intel MKL - - By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. You may also specify it by: - - ```bash - mkdir build - cd build - cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx - cmake --build . -config Release - ``` - - cuBLAS This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).