vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	.dockerignore
#	.gitignore
#	CMakeLists.txt
#	Makefile
#	tests/CMakeLists.txt
This commit is contained in:
Concedo 2023-10-03 23:51:30 +08:00
commit c249f7dbc5
28 changed files with 2241 additions and 433 deletions
Download patch file Download diff file Expand all files Collapse all files

5
.gitignore vendored
View file

@ -31,6 +31,7 @@ models-mnt
/embedding
/gguf
/gguf-llama-simple
/infill
/libllama.so
/llama-bench
/main
@ -82,6 +83,10 @@ 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-llama
tests/test-tokenizer-1-bpe
/koboldcpp_default.so
/koboldcpp_failsafe.so

3
common/common.cpp
View file

@ -389,6 +389,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
params.interactive_first = true;
} else if (arg == "-ins" || arg == "--instruct") {
params.instruct = true;
} else if (arg == "--infill") {
params.infill = true;
} else if (arg == "--multiline-input") {
params.multiline_input = true;
} else if (arg == "--simple-io") {
@ -921,6 +923,7 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_to
result += piece;
}
// NOTE: the original tokenizer decodes bytes after collecting the pieces.
return result;
}

1
common/common.h
View file

@ -120,6 +120,7 @@ struct gpt_params {
bool use_mlock = false; // use mlock to keep model in memory
bool numa = false; // attempt optimizations that help on some NUMA systems
bool verbose_prompt = false; // print prompt tokens before generation
bool infill = false; // use infill mode
};
bool gpt_params_parse(int argc, char ** argv, gpt_params & params);

64
convert-falcon-hf-to-gguf.py
View file

@ -20,28 +20,6 @@ if 'NO_LOCAL_GGUF' not in os.environ:
import gguf
def bytes_to_unicode():
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
"""
Returns list of utf-8 byte and a corresponding list of unicode strings.
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a significant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
cs = bs[:]
n = 0
for b in range(2**8):
if b not in bs:
bs.append(b)
cs.append(2**8+n)
n += 1
return dict(zip(bs, (chr(n) for n in cs)))
def count_model_parts(dir_model: Path) -> int:
num_parts = 0
for filename in os.listdir(dir_model):
@ -133,50 +111,32 @@ gguf_writer.add_file_type(ftype)
print("gguf: get tokenizer metadata")
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)
scores: list[float] = []
toktypes: list[int] = []
# gpt2 tokenizer
gguf_writer.add_tokenizer_model("gpt2")
with open(tokenizer_json_file, "r", encoding="utf-8") as f:
tokenizer_json = json.load(f)
print("gguf: get gpt2 tokenizer vocab")
# The number of tokens in tokenizer.json can differ from the expected vocab size.
# This causes downstream issues with mismatched tensor sizes when running the inference
vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"])
# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
tokenizer = AutoTokenizer.from_pretrained(dir_model)
# The number of tokens in tokenizer.json can differ from the expected vocab size.
# This causes downstream issues with mismatched tensor sizes when running the inference
vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
assert max(tokenizer.vocab.values()) < vocab_size
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()}
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)
tokens.append(reverse_vocab[i])
scores.append(0.0) # dummy
toktypes.append(gguf.TokenType.NORMAL)
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)

63
convert-gptneox-hf-to-gguf.py
View file

@ -19,29 +19,6 @@ 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
def bytes_to_unicode():
"""
Returns list of utf-8 byte and a corresponding list of unicode strings.
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a significant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
cs = bs[:]
n = 0
for b in range(2**8):
if b not in bs:
bs.append(b)
cs.append(2**8+n)
n += 1
return dict(zip(bs, (chr(n) for n in cs)))
def count_model_parts(dir_model: Path) -> int:
num_parts = 0
@ -130,48 +107,32 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
print("gguf: get tokenizer metadata")
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)
scores: list[float] = []
toktypes: list[int] = []
# gpt2 tokenizer
gguf_writer.add_tokenizer_model("gpt2")
with open(tokenizer_json_file, "r", encoding="utf-8") as f:
tokenizer_json = json.load(f)
print("gguf: get gpt2 tokenizer vocab")
vocab_size = len(tokenizer_json["model"]["vocab"])
# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
tokenizer = AutoTokenizer.from_pretrained(dir_model)
# The number of tokens in tokenizer.json can differ from the expected vocab size.
# This causes downstream issues with mismatched tensor sizes when running the inference
vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
assert max(tokenizer.vocab.values()) < vocab_size
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()}
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)
tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
scores.append(0.0) # dummy
toktypes.append(gguf.TokenType.NORMAL)
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)

64
convert-starcoder-hf-to-gguf.py
View file

@ -20,28 +20,6 @@ if 'NO_LOCAL_GGUF' not in os.environ:
import gguf
def bytes_to_unicode():
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
"""
Returns list of utf-8 byte and a corresponding list of unicode strings.
The reversible bpe codes work on unicode strings.
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
This is a significant percentage of your normal, say, 32K bpe vocab.
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
And avoids mapping to whitespace/control characters the bpe code barfs on.
"""
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
cs = bs[:]
n = 0
for b in range(2**8):
if b not in bs:
bs.append(b)
cs.append(2**8+n)
n += 1
return dict(zip(bs, (chr(n) for n in cs)))
def count_model_parts(dir_model: Path) -> int:
num_parts = 0
for filename in os.listdir(dir_model):
@ -117,50 +95,32 @@ gguf_writer.add_file_type(ftype)
print("gguf: get tokenizer metadata")
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)
scores: list[float] = []
toktypes: list[int] = []
# gpt2 tokenizer
gguf_writer.add_tokenizer_model("gpt2")
with open(tokenizer_json_file, "r", encoding="utf-8") as f:
tokenizer_json = json.load(f)
print("gguf: get gpt2 tokenizer vocab")
# The number of tokens in tokenizer.json can differ from the expected vocab size.
# This causes downstream issues with mismatched tensor sizes when running the inference
vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"])
# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
tokenizer = AutoTokenizer.from_pretrained(dir_model)
# The number of tokens in tokenizer.json can differ from the expected vocab size.
# This causes downstream issues with mismatched tensor sizes when running the inference
vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
assert max(tokenizer.vocab.values()) < vocab_size
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()}
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)
tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
scores.append(0.0) # dummy
toktypes.append(gguf.TokenType.NORMAL)
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)

29
convert.py
View file

@ -41,8 +41,7 @@ if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
NDArray: TypeAlias = 'np.ndarray[Any, Any]'
ARCH=gguf.MODEL_ARCH.LLAMA
NAMES=gguf.MODEL_TENSOR_NAMES[ARCH]
ARCH = gguf.MODEL_ARCH.LLAMA
DEFAULT_CONCURRENCY = 8
#
@ -339,29 +338,15 @@ class BpeVocab:
def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
tokenizer = self.bpe_tokenizer
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")
# 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'<unk>':
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
reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
for i, _ in enumerate(tokenizer):
yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
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
yield text.encode("utf-8"), score, gguf.TokenType.CONTROL
def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
yield from self.bpe_tokens()
@ -953,7 +938,7 @@ class OutputFile:
of.close()
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
wq_type = model[gguf.TENSOR_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):
return GGMLFileType.AllF32

2
examples/finetune/convert-finetune-checkpoint-to-gguf.py
View file

@ -313,7 +313,7 @@ class ModelParams:
gguf_writer.add_feed_forward_length(self.get_n_ff())
def tensor_name(key, bid=None, suffix=".weight"):
return gguf.MODEL_TENSOR_NAMES[gguf.MODEL_ARCH.LLAMA][key].format(bid=bid) + suffix
return gguf.TENSOR_NAMES[key].format(bid=bid) + suffix
class Layer:
def __init__(self, params, lora_params, bid):

4
examples/finetune/finetune.cpp
View file

@ -332,8 +332,8 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
assert_shape_1d(layer.attention_norm, hparams.n_embd);
assert_shape_2d(layer.wq, hparams.n_embd, hparams.n_embd);
assert_shape_2d(layer.wk, hparams.n_embd, hparams.n_embd);
assert_shape_2d(layer.wv, hparams.n_embd, hparams.n_embd);
assert_shape_2d(layer.wk, hparams.n_embd, hparams.n_embd_gqa());
assert_shape_2d(layer.wv, hparams.n_embd, hparams.n_embd_gqa());
assert_shape_2d(layer.wo, hparams.n_embd, hparams.n_embd);
assert_shape_1d(layer.ffn_norm, hparams.n_embd);
assert_shape_2d(layer.w1, hparams.n_embd, hparams.n_ff);

8
examples/infill/CMakeLists.txt Normal file
View file

@ -0,0 +1,8 @@
set(TARGET infill)
add_executable(${TARGET} infill.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()

41
examples/infill/README.md Normal file
View file

@ -0,0 +1,41 @@
# llama.cpp/example/infill
This example shows how to use the infill mode with Code Llama models supporting infill mode.
Currently the 7B and 13B models support infill mode.
Infill supports most of the options available in the main example.
For further information have a look at the main README.md in llama.cpp/example/main/README.md
## Common Options
In this section, we cover the most commonly used options for running the `infill` program with the LLaMA models:
- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
- `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
- `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
- `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
## Input Prompts
The `infill` program provides several ways to interact with the LLaMA models using input prompts:
- `--in-prefix PROMPT_BEFORE_CURSOR`: Provide the prefix directly as a command-line option.
- `--in-suffix PROMPT_AFTER_CURSOR`: Provide the suffix directly as a command-line option.
- `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
## Interaction
The `infill` program offers a seamless way to interact with LLaMA models, allowing users to receive real-time infill suggestions. The interactive mode can be triggered using `--interactive`, and `--interactive-first`
### Interaction Options
- `-i, --interactive`: Run the program in interactive mode, allowing users to get real time code suggestions from model.
- `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
- `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
### Example
```bash
./infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n print(\"hell" --in-suffix "\n print(\"goodbye world\")\n "
```

769
examples/infill/infill.cpp Normal file
View file

@ -0,0 +1,769 @@
#include "common.h"
#include "console.h"
#include "llama.h"
#include "build-info.h"
#include "grammar-parser.h"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
#include <fstream>
#include <iostream>
#include <sstream>
#include <string>
#include <vector>
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
#include <signal.h>
#include <unistd.h>
#elif defined (_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#include <signal.h>
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static llama_context ** g_ctx;
static llama_model ** g_model;
static gpt_params * g_params;
static std::vector<llama_token> * g_input_tokens;
static std::ostringstream * g_output_ss;
static std::vector<llama_token> * g_output_tokens;
static bool is_interacting = false;
static void write_logfile(
const llama_context * ctx, const gpt_params & params, const llama_model * model,
const std::vector<llama_token> & input_tokens, const std::string & output,
const std::vector<llama_token> & 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: infill\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)
static void sigint_handler(int signo) {
if (signo == SIGINT) {
if (!is_interacting) {
is_interacting = true;
} else {
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);
}
}
}
#endif
int main(int argc, char ** argv) {
gpt_params params;
g_params = &params;
if (!gpt_params_parse(argc, argv, params)) {
return 1;
}
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("infill", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
console::init(params.simple_io, params.use_color);
atexit([]() { console::cleanup(); });
if (params.logits_all) {
printf("\n************\n");
printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
printf("************\n\n");
return 0;
}
if (params.embedding) {
printf("\n************\n");
printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
printf("************\n\n");
return 0;
}
if (params.n_ctx != 0 && params.n_ctx < 8) {
LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
params.n_ctx = 8;
}
if (params.instruct) {
printf("\n************\n");
printf("%s: please use the 'main' tool for instruct mode\n", __func__);
printf("************\n\n");
return 0;
}
if (!params.antiprompt.empty()) {
printf("\n************\n");
printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
printf("************\n\n");
return 0;
}
if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
printf("\n************\n");
printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
printf("************\n\n");
return 0;
}
if (params.random_prompt) {
printf("\n************\n");
printf("%s: please use the 'main' tool for random prompt mode\n", __func__);
printf("************\n\n");
return 0;
}
if (!params.path_prompt_cache.empty()) {
printf("\n************\n");
printf("%s: infill does not support prompt caching\n", __func__);
printf("************\n\n");
return 0;
}
if (params.rope_freq_base != 0.0) {
LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
}
if (params.rope_freq_scale != 0.0) {
LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
}
LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
if (params.seed == LLAMA_DEFAULT_SEED) {
params.seed = time(NULL);
}
LOG_TEE("%s: seed = %u\n", __func__, params.seed);
std::mt19937 rng(params.seed);
LOG("%s: llama backend init\n", __func__);
llama_backend_init(params.numa);
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
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);
ctx_guidance = llama_new_context_with_model(model, lparams);
}
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
LOG("n_ctx: %d\n", n_ctx);
if (n_ctx > n_ctx_train) {
LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
}
// print system information
{
LOG_TEE("\n");
LOG_TEE("%s\n", get_system_info(params).c_str());
}
const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
LOG("add_bos: %d\n", add_bos);
std::vector<llama_token> embd_inp;
std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
embd_inp = inp_pfx;
embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
embd_inp.push_back(llama_token_middle(ctx));
LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
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
std::vector<llama_token> guidance_inp;
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<llama_token> 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));
}
if ((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;
}
// number of tokens to keep when resetting context
if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) {
params.n_keep = (int)embd_inp.size();
}
LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
// enable interactive mode if interactive start is specified
if (params.interactive_first) {
params.interactive = true;
}
if (params.verbose_prompt) {
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++) {
LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
if (ctx_guidance) {
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++) {
LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
}
}
if (params.n_keep > 0) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
LOG_TEE("'\n");
}
LOG_TEE("\n");
}
if (params.interactive) {
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
struct sigaction sigint_action;
sigint_action.sa_handler = sigint_handler;
sigemptyset (&sigint_action.sa_mask);
sigint_action.sa_flags = 0;
sigaction(SIGINT, &sigint_action, NULL);
#elif defined (_WIN32)
auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
};
SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
#endif
LOG_TEE("%s: interactive mode on.\n", __func__);
if (params.input_prefix_bos) {
LOG_TEE("Input prefix with BOS\n");
}
if (!params.input_prefix.empty()) {
LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
}
if (!params.input_suffix.empty()) {
LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
}
}
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);
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;
if (!params.grammar.empty()) {
parsed_grammar = grammar_parser::parse(params.grammar.c_str());
// will be empty (default) if there are parse errors
if (parsed_grammar.rules.empty()) {
return 1;
}
LOG_TEE("%s: grammar:\n", __func__);
grammar_parser::print_grammar(stderr, parsed_grammar);
LOG_TEE("\n");
{
auto it = params.logit_bias.find(llama_token_eos(ctx));
if (it != params.logit_bias.end() && it->second == -INFINITY) {
LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
}
}
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
grammar = llama_grammar_init(
grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
}
// TODO: replace with ring-buffer
std::vector<llama_token> last_tokens(n_ctx);
std::fill(last_tokens.begin(), last_tokens.end(), 0);
LOG_TEE("\n##### Infill mode #####\n\n");
if (params.infill) {
printf("\n************\n");
printf("no need to specify '--infill', always running infill\n");
printf("************\n\n");
}
if (params.interactive) {
const char *control_message;
if (params.multiline_input) {
control_message = " - To return control to LLaMa, end your input with '\\'.\n"
" - To return control without starting a new line, end your input with '/'.\n";
} else {
control_message = " - Press Return to return control to LLaMa.\n"
" - 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";
}
LOG_TEE("== Running in interactive mode. ==\n");
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
LOG_TEE( " - Press Ctrl+C to interject at any time.\n");
#endif
LOG_TEE( "%s\n", control_message);
is_interacting = params.interactive_first;
}
bool input_echo = true;
int n_past = 0;
int n_remain = params.n_predict;
int n_consumed = 0;
int n_past_guidance = 0;
std::vector<int> input_tokens; g_input_tokens = &input_tokens;
std::vector<int> 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);
std::vector<llama_token> embd;
std::vector<llama_token> embd_guidance;
const int n_vocab = llama_n_vocab(model);
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
while (n_remain != 0 || params.interactive) {
// predict
if (!embd.empty()) {
// Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
// --prompt or --file which uses the same value.
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) {
const int skipped_tokens = (int) embd.size() - max_embd_size;
embd.resize(max_embd_size);
console::set_display(console::error);
printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
console::set_display(console::reset);
fflush(stdout);
}
// infinite text generation via context swapping
// if we run out of context:
// - take the n_keep first tokens from the original prompt (via n_past)
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
if (params.n_predict == -2) {
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 - 1;
const int n_discard = n_left/2;
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1);
llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
n_past -= n_discard;
if (ctx_guidance) {
n_past_guidance -= n_discard;
}
LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
}
// evaluate tokens in batches
// embd is typically prepared beforehand to fit within a batch, but not always
if (ctx_guidance) {
int input_size = 0;
llama_token * input_buf = NULL;
if (n_past_guidance < (int) guidance_inp.size()) {
// Guidance context should have the same data with these modifications:
//
// * Replace the initial prompt
// * Shift everything by guidance_offset
embd_guidance = guidance_inp;
if (embd.begin() + original_prompt_len < embd.end()) {
embd_guidance.insert(
embd_guidance.end(),
embd.begin() + original_prompt_len,
embd.end()
);
}
input_buf = embd_guidance.data();
input_size = embd_guidance.size();
LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
} else {
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_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
return 1;
}
n_past_guidance += n_eval;
}
}
for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
int n_eval = (int) embd.size() - i;
if (n_eval > params.n_batch) {
n_eval = params.n_batch;
}
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
return 1;
}
n_past += n_eval;
LOG("n_past = %d\n", n_past);
}
}
embd.clear();
embd_guidance.clear();
if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
const llama_token id = llama_sample_token(ctx, ctx_guidance, grammar, params, last_tokens, candidates);
last_tokens.erase(last_tokens.begin());
last_tokens.push_back(id);
LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
embd.push_back(id);
// echo this to console
input_echo = true;
// 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_tokens.erase(last_tokens.begin());
last_tokens.push_back(embd_inp[n_consumed]);
++n_consumed;
if ((int) embd.size() >= params.n_batch) {
break;
}
}
}
// display text
if (input_echo) {
for (auto id : embd) {
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);
}
// reset color to default if we there is no pending user input
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) {
// deal with eot token in infill mode
if ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
if(is_interacting && !params.interactive_first) {
// print an eot token
printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
}
fflush(stdout);
printf("\n");
console::set_display(console::user_input);
std::string buffer;
std::string line;
bool another_line=true;
// set a new prefix via stdin
do {
another_line = console::readline(line, params.multiline_input);
buffer += line;
} while (another_line);
// check if we got an empty line, if so we use the old input
if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
params.input_prefix = buffer;
}
buffer.clear();
// set a new suffix via stdin
do {
another_line = console::readline(line, params.multiline_input);
buffer += line;
} while (another_line);
// check if we got an empty line
if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
params.input_suffix = buffer;
}
buffer.clear();
// done taking input, reset color
console::set_display(console::reset);
// tokenize new prefix and suffix
std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, add_bos);
std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, add_bos);
inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(ctx));
inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(ctx));
embd_inp = inp_pfx;
embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end());
embd_inp.push_back(llama_token_middle(ctx));
embd.clear();
embd_guidance.clear();
n_remain = params.n_predict;
n_past = 0;
n_consumed = 0;
// LOG_TEE("took new input\n");
is_interacting = false;
}
// deal with end of text token in interactive mode
else if (last_tokens.back() == llama_token_eos(ctx)) {
LOG("found EOS token\n");
if (params.interactive) {
is_interacting = true;
printf("\n");
console::set_display(console::user_input);
fflush(stdout);
}
}
if (n_past > 0 && is_interacting && !params.interactive) {
LOG("waiting for user input\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());
}
std::string line;
bool another_line = true;
do {
another_line = console::readline(line, params.multiline_input);
buffer += line;
} while (another_line);
// done taking input, reset color
console::set_display(console::reset);
// Add tokens to embd only if the input buffer is non-empty
// Entering a empty line lets the user pass control back
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();
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());
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();
LOG("n_remain: %d\n", n_remain);
} else {
LOG("empty line, passing control back\n");
}
input_echo = false; // do not echo this again
}
if (n_past > 0) {
if (is_interacting) {
// reset grammar state if we're restarting generation
if (grammar != NULL) {
llama_grammar_free(grammar);
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
grammar = llama_grammar_init(
grammar_rules.data(), grammar_rules.size(),
parsed_grammar.symbol_ids.at("root"));
}
}
is_interacting = false;
}
}
// end of text token
if (!embd.empty() && embd.back() == llama_token_eos(ctx) && !params.interactive) {
break;
}
// In interactive mode, respect the maximum number of tokens and drop back to user input when reached.
// 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;
}
}
if (!params.interactive && n_remain <= 0) {
printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
fflush(stdout);
}
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);
if (grammar != NULL) {
llama_grammar_free(grammar);
}
llama_backend_free();
#ifndef LOG_DISABLE_LOGS
LOG_TEE("Log end\n");
#endif // LOG_DISABLE_LOGS
return 0;
}

10
examples/main-cmake-pkg/CMakeLists.txt
View file

@ -28,6 +28,16 @@ configure_file(${_common_path}/../build-info.h
target_include_directories(common PUBLIC ${LLAMA_INCLUDE_DIR}
${CMAKE_CURRENT_BINARY_DIR})
# If the common project was part of "main-cmake-pkg" the transient
# defines would automatically be attached. Because the common func-
# tionality is separate, but dependent upon the defines, it must be
# explicitly extracted from the "llama" target.
#
get_target_property(_llama_transient_defines llama
INTERFACE_COMPILE_DEFINITIONS)
target_compile_definitions(common PRIVATE "${_llama_transient_defines}")
add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp)
target_include_directories(${TARGET} PRIVATE ${_common_path})
install(TARGETS ${TARGET} RUNTIME)

10
examples/server/README.md
View file

@ -176,6 +176,16 @@ node index.js
`content`: Set the text to process.
**POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
*Options:*
`input_prefix`: Set the prefix of the code to infill.
`input_suffix`: Set the suffix of the code to infill.
It also accepts all the options of `/completion` except `stream` and `prompt`.
## More examples
### Interactive mode

206
examples/server/server.cpp
View file

@ -342,6 +342,70 @@ struct llama_server_context
return true;
}
void loadInfill()
{
auto prefix_tokens = tokenize(params.input_prefix, true); // always add BOS
auto suffix_tokens = tokenize(params.input_suffix, true); // always add BOS
prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
prefix_tokens.push_back(llama_token_middle(ctx));
auto prompt_tokens = prefix_tokens;
num_prompt_tokens = prompt_tokens.size();
if (params.n_keep < 0)
{
params.n_keep = (int)num_prompt_tokens;
}
params.n_keep = std::min(params.n_ctx - 4, params.n_keep);
// if input prompt is too big, truncate like normal
if (num_prompt_tokens >= (size_t)params.n_ctx)
{
printf("Input prompt is too big, truncating. Can only take %d tokens but got %zu\n", params.n_ctx, num_prompt_tokens);
// todo we probably want to cut from both sides
const int n_left = (params.n_ctx - params.n_keep) / 2;
std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
std::copy(prompt_tokens.end() - params.n_ctx, prompt_tokens.end(), last_n_tokens.begin());
LOG_VERBOSE("input truncated", {
{"n_ctx", params.n_ctx},
{"n_keep", params.n_keep},
{"n_left", n_left},
{"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
});
truncated = true;
prompt_tokens = new_tokens;
}
else
{
const size_t ps = num_prompt_tokens;
std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
}
// compare the evaluated prompt with the new prompt
n_past = common_part(embd, prompt_tokens);
embd = prompt_tokens;
if (n_past == num_prompt_tokens)
{
// we have to evaluate at least 1 token to generate logits.
printf("we have to evaluate at least 1 token to generate logits\n");
n_past--;
}
LOG_VERBOSE("prompt ingested", {
{"n_past", n_past},
{"cached", tokens_to_str(ctx, embd.cbegin(), embd.cbegin() + n_past)},
{"to_eval", tokens_to_str(ctx, embd.cbegin() + n_past, embd.cend())},
});
has_next_token = true;
}
void loadPrompt()
{
auto prompt_tokens = tokenize(prompt, true); // always add BOS
@ -1219,6 +1283,27 @@ static void parse_options_completion(const json &body, llama_server_context &lla
LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama));
}
static void parse_options_infill(const json &body, llama_server_context &llama)
{
if (body.count("input_prefix") != 0)
{
llama.params.input_prefix = body["input_prefix"];
}
else
{
llama.params.input_prefix = "";
}
if (body.count("input_suffix") != 0)
{
llama.params.input_suffix = body["input_suffix"];
}
else
{
llama.params.input_suffix = "";
}
parse_options_completion(body, llama);
}
static void log_server_request(const Request &req, const Response &res)
{
LOG_INFO("request", {
@ -1519,6 +1604,127 @@ int main(int argc, char **argv)
res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
} });
svr.Post("/infill", [&llama](const Request &req, Response &res)
{
auto lock = llama.lock();
llama.rewind();
llama_reset_timings(llama.ctx);
parse_options_infill(json::parse(req.body), llama);
if (!llama.loadGrammar())
{
res.status = 400;
return;
}
llama.loadInfill();
llama.beginCompletion();
const auto chunked_content_provider = [&](size_t, DataSink & sink) {
size_t sent_count = 0;
size_t sent_token_probs_index = 0;
while (llama.has_next_token) {
const completion_token_output token_with_probs = llama.doCompletion();
if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
continue;
}
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());
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);
}
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);
sent_count += to_send.size();
std::vector<completion_token_output> probs_output = {};
if (llama.params.n_probs > 0) {
const std::vector<llama_token> 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<completion_token_output>(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;
}
}
if (!llama.has_next_token) {
// Generation is done, send extra information.
const json data = format_final_response(
llama,
"",
std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index)
);
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;
}
}
}
llama_print_timings(llama.ctx);
sink.done();
return true;
};
const auto on_complete = [&](bool) {
llama.mutex.unlock();
};
lock.release();
res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
});
svr.Get("/model.json", [&llama](const Request &, Response &res)
{
const json data = format_generation_settings(llama);

2
examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py
View file

@ -364,7 +364,7 @@ class ModelParams:
gguf_writer.add_feed_forward_length(self.get_n_ff())
def tensor_name(key, bid=None):
return gguf.MODEL_TENSOR_NAMES[gguf.MODEL_ARCH.LLAMA][key].format(bid=bid) + ".weight"
return gguf.TENSOR_NAMES[key].format(bid=bid) + ".weight"
class Layer:
def __init__(self, params, bid):

90
ggml-opencl.cpp
View file

@ -1484,10 +1484,15 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const int64_t r2 = ne12 / ne02;
const int64_t r3 = ne13 / ne03;
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
@ -1506,13 +1511,22 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
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);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
int64_t pi02 = -1;
int64_t pi03 = -1;
for (int64_t i13 = 0; i13 < ne13; i13++) {
int64_t i03 = i13 / r3;
for (int64_t i12 = 0; i12 < ne12; i12++) {
int64_t i02 = i12 / r2;
// copy data to device
if (src0->backend != GGML_BACKEND_GPU) {
if (src0->backend != GGML_BACKEND_GPU && (i02 != pi02 || i03 != pi03)) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
pi02 = i02;
pi03 = i03;
}
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
CL_CHECK(clFinish(queue));
@ -1534,7 +1548,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
}
@ -1556,6 +1570,8 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb10 = src1->nb[0];
const int nb11 = src1->nb[1];
@ -1565,6 +1581,9 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const int64_t r2 = ne12 / ne02;
const int64_t r3 = ne13 / ne03;
const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
const int x_ne = ne01 * ne00;
@ -1586,32 +1605,41 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
bool src1_cont_rows = nb10 == sizeof(float);
bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
int64_t pi02 = -1;
int64_t pi03 = -1;
for (int64_t i13 = 0; i13 < ne13; i13++) {
int64_t i03 = i13 / r3;
for (int64_t i12 = 0; i12 < ne12; i12++) {
int64_t i02 = i12 / r2;
// copy src0 to device
if (src0->backend != GGML_BACKEND_GPU) {
if (src0->backend != GGML_BACKEND_GPU && (i02 != pi02 || i03 != pi03)) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
pi02 = i02;
pi03 = i03;
}
// convert src1 to fp16
// TODO: use multiple threads
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i13 * ne12 + i12);
char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
if (src1_cont_rows) {
if (src1_cont_cols) {
ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
for (int64_t i11 = 0; i11 < ne11; i11++) {
ggml_fp32_to_fp16_row((float *) (src1i + i11*nb11), tmp + i11*ne10, ne10);
}
}
}
else {
for (int64_t i01 = 0; i01 < ne11; i01++) {
for (int64_t i00 = 0; i00 < ne10; i00++) {
for (int64_t i11 = 0; i11 < ne11; i11++) {
for (int64_t i10 = 0; i10 < ne10; i10++) {
// very slow due to no inlining
tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
tmp[i11*ne10 + i10] = ggml_fp32_to_fp16(*(float *) (src1i + i11*nb11 + i10*nb10));
}
}
}
@ -1641,7 +1669,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
// copy dst to host, then convert to float
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
ggml_fp16_to_fp32_row(tmp, d, d_ne);
}
@ -1662,12 +1690,17 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_type type = src0->type;
const bool mul_mat_vec = ne11 == 1;
const int64_t r2 = ne12 / ne02;
const int64_t r3 = ne13 / ne03;
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
@ -1700,12 +1733,23 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
size_t ev_idx = 0;
std::vector<cl_event> events;
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
int64_t pi02 = -1;
int64_t pi03 = -1;
for (int64_t i13 = 0; i13 < ne13; i13++) {
int64_t i03 = i13 / r3;
for (int64_t i12 = 0; i12 < ne12; i12++) {
int64_t i02 = i12 / r2;
// copy src0 to device if necessary
if (src0->backend == GGML_BACKEND_CPU) {
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
if (i02 != pi02 || i03 != pi03) {
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
pi02 = i02;
pi03 = i03;
}
} else if (src0->backend == GGML_BACKEND_GPU) {
d_Q = (cl_mem) src0->extra;
} else {
@ -1714,7 +1758,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
// copy src1 to device
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, events.data() + ev_idx++));
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, events.data() + ev_idx++));
// compute
const size_t global = ne01 * CL_DMMV_BLOCK_SIZE;
@ -1735,7 +1779,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, NULL, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i03, i02, NULL));
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
events.emplace_back();
@ -1760,7 +1804,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
for (auto *event : events) {
clReleaseEvent(event);

380
gguf-py/gguf/gguf.py
View file

@ -85,10 +85,12 @@ class MODEL_ARCH(IntEnum):
GPTNEOX : int = auto()
MPT : int = auto()
STARCODER : int = auto()
BERT : int = auto()
class MODEL_TENSOR(IntEnum):
TOKEN_EMBD : int = auto()
TOKEN_TYPES : int = auto()
POS_EMBD : int = auto()
OUTPUT : int = auto()
OUTPUT_NORM : int = auto()
@ -116,78 +118,138 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.GPTNEOX: "gptneox",
MODEL_ARCH.MPT: "mpt",
MODEL_ARCH.STARCODER: "starcoder",
MODEL_ARCH.BERT: "bert",
}
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",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ROPE_FREQS: "rope_freqs",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.ATTN_Q: "blk.{bid}.attn_q",
MODEL_TENSOR.ATTN_K: "blk.{bid}.attn_k",
MODEL_TENSOR.ATTN_V: "blk.{bid}.attn_v",
MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output",
MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
},
MODEL_ARCH.GPTNEOX: {
MODEL_TENSOR.TOKEN_EMBD: "token_embd",
MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv",
MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
},
MODEL_ARCH.FALCON: {
MODEL_TENSOR.TOKEN_EMBD: "token_embd",
MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2",
MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv",
MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
},
MODEL_ARCH.BAICHUAN: {
MODEL_TENSOR.TOKEN_EMBD: "token_embd",
MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ROPE_FREQS: "rope_freqs",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.ATTN_Q: "blk.{bid}.attn_q",
MODEL_TENSOR.ATTN_K: "blk.{bid}.attn_k",
MODEL_TENSOR.ATTN_V: "blk.{bid}.attn_v",
MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output",
MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
},
MODEL_ARCH.STARCODER: {
MODEL_TENSOR.TOKEN_EMBD: "token_embd",
MODEL_TENSOR.POS_EMBD: "position_embd",
MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv",
MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
},
MODEL_ARCH.GPT2: {
TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.TOKEN_EMBD: "token_embd",
MODEL_TENSOR.TOKEN_TYPES: "token_types",
MODEL_TENSOR.POS_EMBD: "position_embd",
MODEL_TENSOR.OUTPUT_NORM: "output_norm",
MODEL_TENSOR.OUTPUT: "output",
MODEL_TENSOR.ROPE_FREQS: "rope_freqs",
MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm",
MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2",
MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv",
MODEL_TENSOR.ATTN_Q: "blk.{bid}.attn_q",
MODEL_TENSOR.ATTN_K: "blk.{bid}.attn_k",
MODEL_TENSOR.ATTN_V: "blk.{bid}.attn_v",
MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output",
MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd",
MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm",
MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate",
MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down",
MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up",
}
MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_ARCH.LLAMA: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GPTNEOX: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.FALCON: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_NORM_2,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.BAICHUAN: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FREQS,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_ROT_EMBD,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.STARCODER: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.POS_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.BERT: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.TOKEN_TYPES,
MODEL_TENSOR.POS_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.MPT: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GPTJ: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GPT2: [
# TODO
},
],
# TODO
}
@ -208,31 +270,40 @@ 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
"gpt_neox.embed_in", # gptneox
"transformer.wte", # gpt2 gpt-j mpt
"transformer.word_embeddings", # falcon
"model.embed_tokens", # llama-hf
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert
),
# Token type embeddings
MODEL_TENSOR.TOKEN_TYPES: (
"embeddings.token_type_embeddings", # bert
),
# Position embeddings
MODEL_TENSOR.POS_EMBD: (
"transformer.wpe", # gpt2
"transformer.wpe", # gpt2
"embeddings.position_embeddings", # bert
),
# Output
MODEL_TENSOR.OUTPUT: (
"embed_out", # gptneox
"lm_head", # gpt2 mpt falcon llama-hf baichuan
"output", # llama-pth
"embed_out", # gptneox
"lm_head", # gpt2 gpt-j mpt falcon llama-hf baichuan
"output", # llama-pth
),
# Output norm
MODEL_TENSOR.OUTPUT_NORM: (
"gpt_neox.final_layer_norm", # gptneox
"transformer.ln_f", # gpt2 falcon
"model.norm", # llama-hf baichuan
"norm", # llama-pth
"gpt_neox.final_layer_norm", # gptneox
"transformer.ln_f", # gpt2 gpt-j falcon
"model.norm", # llama-hf baichuan
"norm", # llama-pth
"embeddings.LayerNorm", # bert
"transformer.norm_f", # mpt
),
# Rope frequencies
@ -244,13 +315,14 @@ class TensorNameMap:
block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
# Attention norm
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
"gpt_neox.layers.{bid}.input_layernorm", # gptneox
"transformer.h.{bid}.ln_1", # gpt2 gpt-j
"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
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
),
# Attention norm 2
@ -260,38 +332,46 @@ class TensorNameMap:
# Attention query-key-value
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
"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
MODEL_TENSOR.ATTN_Q: (
"model.layers.{bid}.self_attn.q_proj", # llama-hf
"layers.{bid}.attention.wq", # llama-pth
"model.layers.{bid}.self_attn.q_proj", # llama-hf
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.h.{bid}.attn.q_proj", # gpt-j
),
# Attention key
MODEL_TENSOR.ATTN_K: (
"model.layers.{bid}.self_attn.k_proj", # llama-hf
"layers.{bid}.attention.wk", # llama-pth
"model.layers.{bid}.self_attn.k_proj", # llama-hf
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.h.{bid}.attn.k_proj", # gpt-j
),
# Attention value
MODEL_TENSOR.ATTN_V: (
"model.layers.{bid}.self_attn.v_proj", # llama-hf
"layers.{bid}.attention.wv", # llama-pth
"model.layers.{bid}.self_attn.v_proj", # llama-hf
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.h.{bid}.attn.v_proj", # gpt-j
),
# Attention output
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
"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
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.h.{bid}.attn.out_proj", # gpt-j
),
# Rotary embeddings
@ -302,21 +382,24 @@ class TensorNameMap:
# Feed-forward norm
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
"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
"encoder.layer.{bid}.output.LayerNorm", # bert
),
# Feed-forward up
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
"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
"encoder.layer.{bid}.intermediate.dense", # bert
"transformer.h.{bid}.mlp.fc_in", # gpt-j
),
# Feed-forward gate
@ -327,39 +410,37 @@ class TensorNameMap:
# Feed-forward down
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
"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
"encoder.layer.{bid}.output.dense", # bert
"transformer.h.{bid}.mlp.fc_out", # gpt-j
),
}
mapping: dict[str, tuple[MODEL_TENSOR, str]]
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]
self.mapping = {}
for tensor, keys in self.mappings_cfg.items():
tensor_name = tensor_names.get(tensor)
if tensor_name is None:
if tensor not in MODEL_TENSORS[arch]:
continue
mapping[tensor_name] = (tensor, tensor_name)
tensor_name = TENSOR_NAMES[tensor]
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
mapping[key] = (tensor, tensor_name)
self.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:
if tensor not in MODEL_TENSORS[arch]:
continue
tensor_name = tensor_name.format(bid = bid)
mapping[tensor_name] = (tensor, tensor_name)
tensor_name = TENSOR_NAMES[tensor].format(bid = bid)
self.mapping[tensor_name] = (tensor, tensor_name)
for key in keys:
key = key.format(bid = bid)
mapping[key] = (tensor, tensor_name)
self.mapping[key] = (tensor, tensor_name)
def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None:
result = self.mapping.get(key)
@ -800,22 +881,25 @@ class SpecialVocab:
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: tuple[str, ...] | None = None):
def __init__(
self, path: str | os.PathLike[str], 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:
self.special_token_types = special_token_types
self.load(path)
self._load(Path(path))
def load(self, path: Path):
if not self.try_load_from_tokenizer_json(path):
self.try_load_from_config_json(path)
def _load(self, path: Path) -> None:
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:
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:
with open(tokenizer_file, encoding = 'utf-8') as f:
tokenizer = json.load(f)
if self.load_merges:
merges = tokenizer.get('model', {}).get('merges')
@ -825,7 +909,7 @@ class SpecialVocab:
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:
with open(tokenizer_config_file, encoding = 'utf-8') as f:
tokenizer_config = json.load(f)
for typ in self.special_token_types:
entry = tokenizer_config.get(f'{typ}_token')
@ -844,11 +928,11 @@ class SpecialVocab:
break
return True
def try_load_from_config_json(self, path: Path) -> bool:
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:
with open(config_file, 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')
@ -856,7 +940,7 @@ class SpecialVocab:
self.special_token_ids[typ] = maybe_token_id
return True
def add_to_gguf(self, gw: GGUFWriter):
def add_to_gguf(self, gw: GGUFWriter) -> None:
if len(self.merges) > 0:
print(f'gguf: Adding {len(self.merges)} merge(s).')
gw.add_token_merges(self.merges)
@ -868,8 +952,8 @@ class SpecialVocab:
print(f'gguf: Setting special token type {typ} to {tokid}')
handler(tokid)
def __repr__(self):
return f'<SpecialVocab with {len(self.merges)} merges and special tokens {self.special_token_ids if self.special_token_ids else "unset"}>'
def __repr__(self) -> str:
return f'<SpecialVocab with {len(self.merges)} merges and special tokens {self.special_token_ids or "unset"}>'
# Example usage:

2
gguf-py/pyproject.toml
View file

@ -1,6 +1,6 @@
[tool.poetry]
name = "gguf"
version = "0.3.3"
version = "0.4.0"
description = "Write ML models in GGUF for GGML"
authors = ["GGML <ggml@ggml.ai>"]
packages = [

301
llama.cpp
View file

@ -1,6 +1,8 @@
#define LLAMA_API_INTERNAL
#include "llama.h"
#include "unicode.h"
#include "ggml.h"
#include "ggml-alloc.h"
@ -1081,6 +1083,10 @@ struct llama_vocab {
id special_pad_id = -1;
id linefeed_id = 13;
id special_prefix_id = 32007;
id special_middle_id = 32009;
id special_suffix_id = 32008;
id special_eot_id = 32010;
int find_bpe_rank(std::string token_left, std::string token_right) const {
replace_all(token_left, " ", "\u0120");
@ -1982,6 +1988,7 @@ static void llm_load_vocab(
for (int i = 0; i < n_merges; i++) {
const std::string word = gguf_get_arr_str(ctx, merges_keyidx, i);
GGML_ASSERT(codepoints_from_utf8(word).size() > 0);
std::string first;
std::string second;
@ -2016,6 +2023,7 @@ static void llm_load_vocab(
for (uint32_t i = 0; i < n_vocab; i++) {
std::string word = gguf_get_arr_str(ctx, token_idx, i);
GGML_ASSERT(codepoints_from_utf8(word).size() > 0);
vocab.token_to_id[word] = i;
@ -2024,12 +2032,13 @@ static void llm_load_vocab(
token_data.score = scores ? scores[i] : 0.0f;
token_data.type = toktypes ? (llama_token_type) toktypes[i] : LLAMA_TOKEN_TYPE_NORMAL;
}
GGML_ASSERT(vocab.id_to_token.size() == vocab.token_to_id.size());
// determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
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];
vocab.linefeed_id = llama_tokenize_internal(vocab, "\u010A", false)[0];
}
// special tokens
@ -4238,18 +4247,41 @@ static bool llama_is_byte_token(const llama_vocab & vocab, llama_token id) {
return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_BYTE;
}
static uint8_t llama_token_to_byte(const llama_vocab & vocab, llama_token id) {
static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id) {
return vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_USER_DEFINED;
}
static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
GGML_ASSERT(llama_is_byte_token(vocab, id));
const auto& token_data = vocab.id_to_token.at(id);
auto buf = token_data.text.substr(3, 2);
return strtol(buf.c_str(), NULL, 16);
switch (llama_vocab_get_type(vocab)) {
case LLAMA_VOCAB_TYPE_SPM: {
auto buf = token_data.text.substr(3, 2);
return strtol(buf.c_str(), NULL, 16);
}
case LLAMA_VOCAB_TYPE_BPE: {
GGML_ASSERT(false);
return unicode_to_bytes_bpe(token_data.text);
}
default:
GGML_ASSERT(false);
}
}
static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
char buf[7];
int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch);
GGML_ASSERT(0 <= result && result < 7);
return vocab.token_to_id.at(buf);
switch (llama_vocab_get_type(vocab)) {
case LLAMA_VOCAB_TYPE_SPM: {
char buf[7];
int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch);
GGML_ASSERT(0 <= result && result < 7);
return vocab.token_to_id.at(buf);
}
case LLAMA_VOCAB_TYPE_BPE: {
return vocab.token_to_id.at(bytes_to_unicode_bpe(ch));
}
default:
GGML_ASSERT(false);
}
}
static void llama_escape_whitespace(std::string & text) {
@ -4529,15 +4561,9 @@ 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()) {
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);
throw std::runtime_error("ERROR: byte not found in vocab");
}
output.push_back((*token_multibyte).second);
}
} else {
output.push_back((*token).second);
@ -4574,23 +4600,144 @@ private:
work_queue.push(bigram);
}
// probably not 100% correct
static std::vector<std::string> bpe_gpt2_preprocess(const std::string & text) {
std::vector<std::string> words;
std::vector<std::string> bpe_gpt2_preprocess(const std::string & text) {
std::vector<std::string> bpe_words;
std::vector<std::string> bpe_encoded_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::string token = "";
// GPT2 system regex: 's|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+
bool collecting_numeric = false;
bool collecting_letter = false;
bool collecting_special = false;
bool collecting_whitespace_lookahead = false;
bool collecting = false;
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());
std::vector<std::string> text_utf;
text_utf.reserve(text.size());
bpe_words.reserve(text.size());
bpe_encoded_words.reserve(text.size());
auto cps = codepoints_from_utf8(text);
for (size_t i = 0; i < cps.size(); ++i)
text_utf.emplace_back(codepoint_to_utf8(cps[i]));
for (int i = 0; i < (int)text_utf.size(); i++) {
const std::string & utf_char = text_utf[i];
bool split_condition = false;
// const char* text_pos = raw_text_p + utf_char.seq_offset_bytes;
int bytes_remain = text_utf.size() - i;
// forward backward lookups
const std::string & utf_char_next = (i + 1 < (int)text_utf.size()) ? text_utf[i + 1] : "";
const std::string & utf_char_next_next = (i + 2 < (int)text_utf.size()) ? text_utf[i + 2] : "";
// handling contractions
if (!split_condition && bytes_remain >= 2) {
// 's|'t|'m|'d
if (utf_char == "\'" && (utf_char_next == "s" || utf_char_next == "t" || utf_char_next == "m" || utf_char_next == "d")) {
split_condition = true;
}
if (split_condition) {
if (token.size()) {
bpe_words.emplace_back(token); // push previous content as token
}
token = utf_char + utf_char_next;
bpe_words.emplace_back(token);
token = "";
i++;
continue;
}
}
if (!split_condition && bytes_remain >= 3) {
// 're|'ve|'ll
if (utf_char == "\'" && (
(utf_char_next == "r" || utf_char_next_next == "e") ||
(utf_char_next == "v" || utf_char_next_next == "e") ||
(utf_char_next == "l" || utf_char_next_next == "l"))
) {
split_condition = true;
}
if (split_condition) {
// current token + next token can be defined
if (token.size()) {
bpe_words.emplace_back(token); // push previous content as token
}
token = utf_char + utf_char_next + utf_char_next_next;
bpe_words.emplace_back(token); // the contraction
token = "";
i += 2;
continue;
}
}
if (!split_condition && !collecting) {
if (codepoint_type(utf_char) == CODEPOINT_TYPE_LETTER || (!token.size() && utf_char == " " && codepoint_type(utf_char_next) == CODEPOINT_TYPE_LETTER)) {
collecting_letter = true;
collecting = true;
}
else if (codepoint_type(utf_char) == CODEPOINT_TYPE_DIGIT || (!token.size() && utf_char == " " && codepoint_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
collecting_numeric = true;
collecting = true;
}
else if (
((codepoint_type(utf_char) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char) != CODEPOINT_TYPE_DIGIT) && (codepoint_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) ||
(!token.size() && utf_char == " " && codepoint_type(utf_char_next) != CODEPOINT_TYPE_LETTER && codepoint_type(utf_char_next) != CODEPOINT_TYPE_DIGIT && codepoint_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE)
) {
collecting_special = true;
collecting = true;
}
else if (codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE && codepoint_type(utf_char_next) == CODEPOINT_TYPE_WHITESPACE) {
collecting_whitespace_lookahead = true;
collecting = true;
}
else if (codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE) {
split_condition = true;
}
}
else if (!split_condition && collecting) {
if (collecting_letter && codepoint_type(utf_char) != CODEPOINT_TYPE_LETTER) {
split_condition = true;
}
else if (collecting_numeric && codepoint_type(utf_char) != CODEPOINT_TYPE_DIGIT) {
split_condition = true;
}
else if (collecting_special && (codepoint_type(utf_char) == CODEPOINT_TYPE_LETTER || codepoint_type(utf_char) == CODEPOINT_TYPE_DIGIT || codepoint_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) {
split_condition = true;
}
else if (collecting_whitespace_lookahead && codepoint_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE) {
split_condition = true;
}
}
if (utf_char_next == "") {
split_condition = true; // final
token += utf_char;
}
if (split_condition) {
if (token.size()) {
bpe_words.emplace_back(token);
}
token = utf_char;
collecting = false;
collecting_letter = false;
collecting_numeric = false;
collecting_special = false;
collecting_whitespace_lookahead = false;
}
else {
token += utf_char;
}
}
return words;
for (std::string & word : bpe_words) {
std::string encoded_token = "";
for (char & c : word) {
encoded_token += bytes_to_unicode_bpe(c);
}
bpe_encoded_words.emplace_back(encoded_token);
}
return bpe_encoded_words;
}
const llama_vocab & vocab;
@ -6765,13 +6912,14 @@ struct llama_context * llama_new_context_with_model(
#ifdef GGML_USE_METAL
if (model->n_gpu_layers > 0) {
ggml_metal_log_set_callback(llama_log_callback_default, NULL);
ctx->ctx_metal = ggml_metal_init(1);
if (!ctx->ctx_metal) {
LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__);
llama_free(ctx);
return NULL;
}
ggml_metal_log_set_callback(llama_log_callback_default, NULL);
//ggml_metal_graph_find_concurrency(ctx->ctx_metal, gf, false);
//ggml_allocr_set_parse_seq(ctx->alloc, ggml_metal_get_concur_list(ctx->ctx_metal), ggml_metal_if_optimized(ctx->ctx_metal));
}
@ -7498,6 +7646,22 @@ llama_token llama_token_eos(const struct llama_context * ctx) {
llama_token llama_token_nl(const struct llama_context * ctx) {
return ctx->model.vocab.linefeed_id;
}
llama_token llama_token_prefix(const struct llama_context * ctx) {
return ctx->model.vocab.special_prefix_id;
}
llama_token llama_token_middle(const struct llama_context * ctx) {
return ctx->model.vocab.special_middle_id;
}
llama_token llama_token_suffix(const struct llama_context * ctx) {
return ctx->model.vocab.special_suffix_id;
}
llama_token llama_token_eot(const struct llama_context * ctx) {
return ctx->model.vocab.special_eot_id;
}
int llama_tokenize(
const struct llama_model * model,
@ -7520,35 +7684,66 @@ int llama_tokenize(
return res.size();
}
static std::string llama_decode_text(const std::string & text) {
std::string decoded_text;
auto unicode_sequences = codepoints_from_utf8(text);
for (auto& unicode_sequence : unicode_sequences) {
decoded_text += unicode_to_bytes_bpe(codepoint_to_utf8(unicode_sequence));
}
return decoded_text;
}
// does not write null-terminator to buf
int llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int length) {
if (0 <= token && token < llama_n_vocab(model)) {
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) {
switch (llama_vocab_get_type(model->vocab)) {
case LLAMA_VOCAB_TYPE_SPM: {
if (llama_is_normal_token(model->vocab, token)) {
std::string result = model->vocab.id_to_token[token].text;
llama_unescape_whitespace(result);
if (length < (int) result.length()) {
return -result.length();
}
memcpy(buf, result.c_str(), result.length());
return result.length();
} else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT
if (length < 3) {
return -3;
}
memcpy(buf, "\xe2\x96\x85", 3);
return 3;
} else if (llama_is_control_token(model->vocab, token)) {
;
} else if (llama_is_byte_token(model->vocab, token)) {
if (length < 1) {
return -1;
}
buf[0] = llama_token_to_byte(model->vocab, token);
return 1;
} else {
GGML_ASSERT(false);
}
if (length < (int) result.length()) {
return -result.length();
break;
}
case LLAMA_VOCAB_TYPE_BPE: {
if (llama_is_normal_token(model->vocab, token)) {
std::string result = model->vocab.id_to_token[token].text;
result = llama_decode_text(result);
if (length < (int) result.length()) {
return -result.length();
}
memcpy(buf, result.c_str(), result.length());
return result.length();
} else if (llama_is_control_token(model->vocab, token)) {
;
} else {
GGML_ASSERT(false);
}
memcpy(buf, result.c_str(), result.length());
return result.length();
} else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT
if (length < 3) {
return -3;
}
buf[0] = '\xe2';
buf[1] = '\x96';
buf[2] = '\x85';
return 3;
} else if (llama_is_control_token(model->vocab, token)) {
// do nothing
} else if (llama_is_byte_token(model->vocab, token)) {
if (length < 1) {
return -1;
}
buf[0] = llama_token_to_byte(model->vocab, token);
return 1;
break;
}
default:
GGML_ASSERT(false);
}
}
return 0;

5
llama.h
View file

@ -490,6 +490,11 @@ extern "C" {
LLAMA_API llama_token llama_token_bos(const struct llama_context * ctx); // beginning-of-sentence
LLAMA_API llama_token llama_token_eos(const struct llama_context * ctx); // end-of-sentence
LLAMA_API llama_token llama_token_nl (const struct llama_context * ctx); // next-line
// codellama infill tokens
LLAMA_API llama_token llama_token_prefix(const struct llama_context * ctx); // Beginning of infill prefix
LLAMA_API llama_token llama_token_middle(const struct llama_context * ctx); // Beginning of infill middle
LLAMA_API llama_token llama_token_suffix(const struct llama_context * ctx); // Beginning of infill suffix
LLAMA_API llama_token llama_token_eot (const struct llama_context * ctx); // End of infill middle
//
// Tokenization

BIN
models/ggml-vocab-aquila.gguf Normal file
View file

Binary file not shown.

BIN
models/ggml-vocab-falcon.gguf Normal file
View file

Binary file not shown.

2
scripts/LlamaConfig.cmake.in
View file

@ -56,11 +56,13 @@ find_library(llama_LIBRARY llama
HINTS ${LLAMA_LIB_DIR})
set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
set(_llama_transient_defines "@LLAMA_TRANSIENT_DEFINES@")
add_library(llama UNKNOWN IMPORTED)
set_target_properties(llama
PROPERTIES
INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
INTERFACE_COMPILE_DEFINITIONS "${_llama_transient_defines}"
IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
IMPORTED_LOCATION "${llama_LIBRARY}"
INTERFACE_COMPILE_FEATURES cxx_std_11

9
tests/test-tokenizer-0-falcon.cpp
View file

@ -1,5 +1,6 @@
#include "llama.h"
#include "common.h"
#include "console.h"
#include <cstdio>
#include <string>
@ -85,12 +86,18 @@ int main(int argc, char **argv) {
}
if (llama_vocab_type(model) != LLAMA_VOCAB_TYPE_BPE) {
fprintf(stderr, "%s : error: vocab type is not SPM\n", __func__);
fprintf(stderr, "%s : error: vocab type is not BPE\n", __func__);
llama_free_model(model);
llama_free(ctx);
return 2;
}
#ifdef _WIN32
// We need this for unicode console support
console::init(false, false);
atexit([]() { console::cleanup(); });
#endif
bool success = true;
for (const auto & test_kv : k_tests()) {

113
tests/test-tokenizer-1-bpe.cpp Normal file
View file

@ -0,0 +1,113 @@
#include "llama.h"
#include "common.h"
#include "unicode.h"
#include "console.h"
#include <cassert>
#include <cstdio>
#include <cstring>
#include <string>
#include <codecvt>
#include <map>
#include <vector>
#include <locale>
int main(int argc, char **argv) {
if (argc < 2) {
fprintf(stderr, "Usage: %s <vocab-file>\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 mparams = llama_model_default_params();
mparams.vocab_only = true;
model = llama_load_model_from_file(fname.c_str(), mparams);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
return 1;
}
auto cparams = llama_context_default_params();
ctx = llama_new_context_with_model(model, cparams);
if (ctx == NULL) {
fprintf(stderr, "%s: error: failed to load vocab '%s'\n", __func__, fname.c_str());
llama_free_model(model);
return 1;
}
}
GGML_ASSERT(llama_vocab_type(model) == LLAMA_VOCAB_TYPE_BPE);
#ifdef _WIN32
// We need this for unicode console support
console::init(false, false);
atexit([]() { console::cleanup(); });
#endif
const int n_vocab = llama_n_vocab(model);
for (int i = 0; i < n_vocab; ++i) {
std::string str = llama_detokenize_bpe(ctx, std::vector<int>(1, i));
try {
auto cps = codepoints_from_utf8(str);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (check != str) {
fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",
__func__, i, str.c_str(), str.length(), check.c_str(), check.length());
return 2;
}
}
catch (const std::invalid_argument &) {
fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
}
}
for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
// NOTE: these exceptions seem to be necessary, because the GPT2 tokenizer doesn't want to interfere with some ASCII control characters
if ((cp < 0x03 || cp > 0x05) && cp != 0x0b && cp != 0x11 && (cp < 0x13 || cp > 0x17) && cp != 0x19 && (cp < 0x1c || cp > 0x1e) && (cp < 0xd800 || cp > 0xdfff)) {
std::string str = " " + codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (str != check) {
fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
__func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
return 3;
}
}
}
// TODO: why doesn't this work for the full range of Unicodes?
// for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
for (uint32_t cp = 0x10000; cp < 0x00080000; ++cp) {
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (str != check) {
fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
__func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
return 4;
}
}
llama_free_model(model);
llama_free(ctx);
llama_backend_free();
return 0;
}

29
tests/test-tokenizer-1-llama.cpp
View file

@ -1,5 +1,6 @@
#include "llama.h"
#include "common.h"
#include "unicode.h"
#include "console.h"
#include <cassert>
@ -11,30 +12,6 @@
#include <vector>
#include <locale>
typedef int codepoint;
static std::string codepoint_to_utf8(codepoint cp) {
std::string result;
if (0x00 <= cp && cp <= 0x7f) {
result.push_back(cp);
} else if (0x80 <= cp && cp <= 0x7ff) {
result.push_back(0xc0 | ((cp >> 6) & 0x1f));
result.push_back(0x80 | (cp & 0x3f));
} else if (0x800 <= cp && cp <= 0xffff) {
result.push_back(0xe0 | ((cp >> 12) & 0x0f));
result.push_back(0x80 | ((cp >> 6) & 0x3f));
result.push_back(0x80 | (cp & 0x3f));
} else if (0x10000 <= cp && cp <= 0x10ffff) {
result.push_back(0xf0 | ((cp >> 18) & 0x07));
result.push_back(0x80 | ((cp >> 12) & 0x3f));
result.push_back(0x80 | ((cp >> 6) & 0x3f));
result.push_back(0x80 | (cp & 0x3f));
} else {
throw std::invalid_argument("invalid codepoint");
}
return result;
}
int main(int argc, char **argv) {
if (argc < 2) {
fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);
@ -95,7 +72,7 @@ int main(int argc, char **argv) {
}
}
for (codepoint cp = 0x0000; cp < 0xffff; ++cp) {
for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
if (cp < 0xd800 || cp > 0xdfff) {
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
@ -107,7 +84,7 @@ int main(int argc, char **argv) {
}
}
}
for (codepoint cp = 0x10000; cp < 0x0010ffff; ++cp) {
for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_spm(ctx, tokens);

462
unicode.h Normal file
View file

@ -0,0 +1,462 @@
#pragma once
#include <cassert>
#include <stdexcept>
#include <vector>
#include <unordered_map>
static const std::vector<std::pair<uint32_t, uint32_t>> digit_ranges = {
{0x30, 0x39}, {0xB2, 0xB3}, {0xB9, 0xB9}, {0x660, 0x669}, {0x6F0, 0x6F9}, {0x7C0, 0x7C9}, {0x966, 0x96F}, {0x9E6, 0x9EF}, {0xA66, 0xA6F}, {0xAE6, 0xAEF}, {0xB66, 0xB6F}, {0xBE6, 0xBEF}, {0xC66, 0xC6F},
{0xCE6, 0xCEF}, {0xD66, 0xD6F}, {0xDE6, 0xDEF}, {0xE50, 0xE59}, {0xED0, 0xED9}, {0xF20, 0xF29}, {0x1040, 0x1049}, {0x1090, 0x1099}, {0x1369, 0x1371}, {0x17E0, 0x17E9}, {0x1810, 0x1819}, {0x1946, 0x194F},
{0x19D0, 0x19DA}, {0x1A80, 0x1A89}, {0x1A90, 0x1A99}, {0x1B50, 0x1B59}, {0x1BB0, 0x1BB9}, {0x1C40, 0x1C49}, {0x1C50, 0x1C59}, {0x2070, 0x2070}, {0x2074, 0x2079}, {0x2080, 0x2089}, {0x2460, 0x2468},
{0x2474, 0x247C}, {0x2488, 0x2490}, {0x24EA, 0x24EA}, {0x24F5, 0x24FD}, {0x24FF, 0x24FF}, {0x2776, 0x277E}, {0x2780, 0x2788}, {0x278A, 0x2792}, {0xA620, 0xA629}, {0xA8D0, 0xA8D9}, {0xA900, 0xA909},
{0xA9D0, 0xA9D9}, {0xA9F0, 0xA9F9}, {0xAA50, 0xAA59}, {0xABF0, 0xABF9}, {0xFF10, 0xFF19}, {0x104A0, 0x104A9}, {0x10A40, 0x10A43}, {0x10D30, 0x10D39}, {0x10E60, 0x10E68}, {0x11052, 0x1105A},
{0x11066, 0x1106F}, {0x110F0, 0x110F9}, {0x11136, 0x1113F}, {0x111D0, 0x111D9}, {0x112F0, 0x112F9}, {0x11450, 0x11459}, {0x114D0, 0x114D9}, {0x11650, 0x11659}, {0x116C0, 0x116C9}, {0x11730, 0x11739},
{0x118E0, 0x118E9}, {0x11950, 0x11959}, {0x11C50, 0x11C59}, {0x11D50, 0x11D59}, {0x11DA0, 0x11DA9}, {0x16A60, 0x16A69}, {0x16B50, 0x16B59}, {0x1D7CE, 0x1D7FF}, {0x1E140, 0x1E149}, {0x1E2F0, 0x1E2F9},
{0x1E950, 0x1E959}, {0x1F100, 0x1F10A}, {0x1FBF0, 0x1FBF9},
};
static const std::vector<std::pair<uint32_t, uint32_t>> letter_ranges = {
{0x41, 0x5A}, {0x61, 0x7A}, {0xAA, 0xAA}, {0xB5, 0xB5}, {0xBA, 0xBA}, {0xC0, 0xD6}, {0xD8, 0xF6}, {0xF8, 0x2C1}, {0x2C6, 0x2D1}, {0x2E0, 0x2E4}, {0x2EC, 0x2EC}, {0x2EE, 0x2EE}, {0x370, 0x374},
{0x376, 0x377}, {0x37A, 0x37D}, {0x37F, 0x37F}, {0x386, 0x386}, {0x388, 0x38A}, {0x38C, 0x38C}, {0x38E, 0x3A1}, {0x3A3, 0x3F5}, {0x3F7, 0x481}, {0x48A, 0x52F}, {0x531, 0x556}, {0x559, 0x559},
{0x560, 0x588}, {0x5D0, 0x5EA}, {0x5EF, 0x5F2}, {0x620, 0x64A}, {0x66E, 0x66F}, {0x671, 0x6D3}, {0x6D5, 0x6D5}, {0x6E5, 0x6E6}, {0x6EE, 0x6EF}, {0x6FA, 0x6FC}, {0x6FF, 0x6FF}, {0x710, 0x710},
{0x712, 0x72F}, {0x74D, 0x7A5}, {0x7B1, 0x7B1}, {0x7CA, 0x7EA}, {0x7F4, 0x7F5}, {0x7FA, 0x7FA}, {0x800, 0x815}, {0x81A, 0x81A}, {0x824, 0x824}, {0x828, 0x828}, {0x840, 0x858}, {0x860, 0x86A},
{0x8A0, 0x8B4}, {0x8B6, 0x8C7}, {0x904, 0x939}, {0x93D, 0x93D}, {0x950, 0x950}, {0x958, 0x961}, {0x971, 0x980}, {0x985, 0x98C}, {0x98F, 0x990}, {0x993, 0x9A8}, {0x9AA, 0x9B0}, {0x9B2, 0x9B2},
{0x9B6, 0x9B9}, {0x9BD, 0x9BD}, {0x9CE, 0x9CE}, {0x9DC, 0x9DD}, {0x9DF, 0x9E1}, {0x9F0, 0x9F1}, {0x9FC, 0x9FC}, {0xA05, 0xA0A}, {0xA0F, 0xA10}, {0xA13, 0xA28}, {0xA2A, 0xA30}, {0xA32, 0xA33},
{0xA35, 0xA36}, {0xA38, 0xA39}, {0xA59, 0xA5C}, {0xA5E, 0xA5E}, {0xA72, 0xA74}, {0xA85, 0xA8D}, {0xA8F, 0xA91}, {0xA93, 0xAA8}, {0xAAA, 0xAB0}, {0xAB2, 0xAB3}, {0xAB5, 0xAB9}, {0xABD, 0xABD},
{0xAD0, 0xAD0}, {0xAE0, 0xAE1}, {0xAF9, 0xAF9}, {0xB05, 0xB0C}, {0xB0F, 0xB10}, {0xB13, 0xB28}, {0xB2A, 0xB30}, {0xB32, 0xB33}, {0xB35, 0xB39}, {0xB3D, 0xB3D}, {0xB5C, 0xB5D}, {0xB5F, 0xB61},
{0xB71, 0xB71}, {0xB83, 0xB83}, {0xB85, 0xB8A}, {0xB8E, 0xB90}, {0xB92, 0xB95}, {0xB99, 0xB9A}, {0xB9C, 0xB9C}, {0xB9E, 0xB9F}, {0xBA3, 0xBA4}, {0xBA8, 0xBAA}, {0xBAE, 0xBB9}, {0xBD0, 0xBD0},
{0xC05, 0xC0C}, {0xC0E, 0xC10}, {0xC12, 0xC28}, {0xC2A, 0xC39}, {0xC3D, 0xC3D}, {0xC58, 0xC5A}, {0xC60, 0xC61}, {0xC80, 0xC80}, {0xC85, 0xC8C}, {0xC8E, 0xC90}, {0xC92, 0xCA8}, {0xCAA, 0xCB3},
{0xCB5, 0xCB9}, {0xCBD, 0xCBD}, {0xCDE, 0xCDE}, {0xCE0, 0xCE1}, {0xCF1, 0xCF2}, {0xD04, 0xD0C}, {0xD0E, 0xD10}, {0xD12, 0xD3A}, {0xD3D, 0xD3D}, {0xD4E, 0xD4E}, {0xD54, 0xD56}, {0xD5F, 0xD61},
{0xD7A, 0xD7F}, {0xD85, 0xD96}, {0xD9A, 0xDB1}, {0xDB3, 0xDBB}, {0xDBD, 0xDBD}, {0xDC0, 0xDC6}, {0xE01, 0xE30}, {0xE32, 0xE33}, {0xE40, 0xE46}, {0xE81, 0xE82}, {0xE84, 0xE84}, {0xE86, 0xE8A},
{0xE8C, 0xEA3}, {0xEA5, 0xEA5}, {0xEA7, 0xEB0}, {0xEB2, 0xEB3}, {0xEBD, 0xEBD}, {0xEC0, 0xEC4}, {0xEC6, 0xEC6}, {0xEDC, 0xEDF}, {0xF00, 0xF00}, {0xF40, 0xF47}, {0xF49, 0xF6C}, {0xF88, 0xF8C},
{0x1000, 0x102A}, {0x103F, 0x103F}, {0x1050, 0x1055}, {0x105A, 0x105D}, {0x1061, 0x1061}, {0x1065, 0x1066}, {0x106E, 0x1070}, {0x1075, 0x1081}, {0x108E, 0x108E}, {0x10A0, 0x10C5}, {0x10C7, 0x10C7},
{0x10CD, 0x10CD}, {0x10D0, 0x10FA}, {0x10FC, 0x1248}, {0x124A, 0x124D}, {0x1250, 0x1256}, {0x1258, 0x1258}, {0x125A, 0x125D}, {0x1260, 0x1288}, {0x128A, 0x128D}, {0x1290, 0x12B0}, {0x12B2, 0x12B5},
{0x12B8, 0x12BE}, {0x12C0, 0x12C0}, {0x12C2, 0x12C5}, {0x12C8, 0x12D6}, {0x12D8, 0x1310}, {0x1312, 0x1315}, {0x1318, 0x135A}, {0x1380, 0x138F}, {0x13A0, 0x13F5}, {0x13F8, 0x13FD}, {0x1401, 0x166C},
{0x166F, 0x167F}, {0x1681, 0x169A}, {0x16A0, 0x16EA}, {0x16F1, 0x16F8}, {0x1700, 0x170C}, {0x170E, 0x1711}, {0x1720, 0x1731}, {0x1740, 0x1751}, {0x1760, 0x176C}, {0x176E, 0x1770}, {0x1780, 0x17B3},
{0x17D7, 0x17D7}, {0x17DC, 0x17DC}, {0x1820, 0x1878}, {0x1880, 0x1884}, {0x1887, 0x18A8}, {0x18AA, 0x18AA}, {0x18B0, 0x18F5}, {0x1900, 0x191E}, {0x1950, 0x196D}, {0x1970, 0x1974}, {0x1980, 0x19AB},
{0x19B0, 0x19C9}, {0x1A00, 0x1A16}, {0x1A20, 0x1A54}, {0x1AA7, 0x1AA7}, {0x1B05, 0x1B33}, {0x1B45, 0x1B4B}, {0x1B83, 0x1BA0}, {0x1BAE, 0x1BAF}, {0x1BBA, 0x1BE5}, {0x1C00, 0x1C23}, {0x1C4D, 0x1C4F},
{0x1C5A, 0x1C7D}, {0x1C80, 0x1C88}, {0x1C90, 0x1CBA}, {0x1CBD, 0x1CBF}, {0x1CE9, 0x1CEC}, {0x1CEE, 0x1CF3}, {0x1CF5, 0x1CF6}, {0x1CFA, 0x1CFA}, {0x1D00, 0x1DBF}, {0x1E00, 0x1F15}, {0x1F18, 0x1F1D},
{0x1F20, 0x1F45}, {0x1F48, 0x1F4D}, {0x1F50, 0x1F57}, {0x1F59, 0x1F59}, {0x1F5B, 0x1F5B}, {0x1F5D, 0x1F5D}, {0x1F5F, 0x1F7D}, {0x1F80, 0x1FB4}, {0x1FB6, 0x1FBC}, {0x1FBE, 0x1FBE}, {0x1FC2, 0x1FC4},
{0x1FC6, 0x1FCC}, {0x1FD0, 0x1FD3}, {0x1FD6, 0x1FDB}, {0x1FE0, 0x1FEC}, {0x1FF2, 0x1FF4}, {0x1FF6, 0x1FFC}, {0x2071, 0x2071}, {0x207F, 0x207F}, {0x2090, 0x209C}, {0x2102, 0x2102}, {0x2107, 0x2107},
{0x210A, 0x2113}, {0x2115, 0x2115}, {0x2119, 0x211D}, {0x2124, 0x2124}, {0x2126, 0x2126}, {0x2128, 0x2128}, {0x212A, 0x212D}, {0x212F, 0x2139}, {0x213C, 0x213F}, {0x2145, 0x2149}, {0x214E, 0x214E},
{0x2183, 0x2184}, {0x2C00, 0x2C2E}, {0x2C30, 0x2C5E}, {0x2C60, 0x2CE4}, {0x2CEB, 0x2CEE}, {0x2CF2, 0x2CF3}, {0x2D00, 0x2D25}, {0x2D27, 0x2D27}, {0x2D2D, 0x2D2D}, {0x2D30, 0x2D67}, {0x2D6F, 0x2D6F},
{0x2D80, 0x2D96}, {0x2DA0, 0x2DA6}, {0x2DA8, 0x2DAE}, {0x2DB0, 0x2DB6}, {0x2DB8, 0x2DBE}, {0x2DC0, 0x2DC6}, {0x2DC8, 0x2DCE}, {0x2DD0, 0x2DD6}, {0x2DD8, 0x2DDE}, {0x2E2F, 0x2E2F}, {0x3005, 0x3006},
{0x3031, 0x3035}, {0x303B, 0x303C}, {0x3041, 0x3096}, {0x309D, 0x309F}, {0x30A1, 0x30FA}, {0x30FC, 0x30FF}, {0x3105, 0x312F}, {0x3131, 0x318E}, {0x31A0, 0x31BF}, {0x31F0, 0x31FF}, {0x3400, 0x4DBF},
{0x4E00, 0x9FFC}, {0xA000, 0xA48C}, {0xA4D0, 0xA4FD}, {0xA500, 0xA60C}, {0xA610, 0xA61F}, {0xA62A, 0xA62B}, {0xA640, 0xA66E}, {0xA67F, 0xA69D}, {0xA6A0, 0xA6E5}, {0xA717, 0xA71F}, {0xA722, 0xA788},
{0xA78B, 0xA7BF}, {0xA7C2, 0xA7CA}, {0xA7F5, 0xA801}, {0xA803, 0xA805}, {0xA807, 0xA80A}, {0xA80C, 0xA822}, {0xA840, 0xA873}, {0xA882, 0xA8B3}, {0xA8F2, 0xA8F7}, {0xA8FB, 0xA8FB}, {0xA8FD, 0xA8FE},
{0xA90A, 0xA925}, {0xA930, 0xA946}, {0xA960, 0xA97C}, {0xA984, 0xA9B2}, {0xA9CF, 0xA9CF}, {0xA9E0, 0xA9E4}, {0xA9E6, 0xA9EF}, {0xA9FA, 0xA9FE}, {0xAA00, 0xAA28}, {0xAA40, 0xAA42}, {0xAA44, 0xAA4B},
{0xAA60, 0xAA76}, {0xAA7A, 0xAA7A}, {0xAA7E, 0xAAAF}, {0xAAB1, 0xAAB1}, {0xAAB5, 0xAAB6}, {0xAAB9, 0xAABD}, {0xAAC0, 0xAAC0}, {0xAAC2, 0xAAC2}, {0xAADB, 0xAADD}, {0xAAE0, 0xAAEA}, {0xAAF2, 0xAAF4},
{0xAB01, 0xAB06}, {0xAB09, 0xAB0E}, {0xAB11, 0xAB16}, {0xAB20, 0xAB26}, {0xAB28, 0xAB2E}, {0xAB30, 0xAB5A}, {0xAB5C, 0xAB69}, {0xAB70, 0xABE2}, {0xAC00, 0xD7A3}, {0xD7B0, 0xD7C6}, {0xD7CB, 0xD7FB},
{0xF900, 0xFA6D}, {0xFA70, 0xFAD9}, {0xFB00, 0xFB06}, {0xFB13, 0xFB17}, {0xFB1D, 0xFB1D}, {0xFB1F, 0xFB28}, {0xFB2A, 0xFB36}, {0xFB38, 0xFB3C}, {0xFB3E, 0xFB3E}, {0xFB40, 0xFB41}, {0xFB43, 0xFB44},
{0xFB46, 0xFBB1}, {0xFBD3, 0xFD3D}, {0xFD50, 0xFD8F}, {0xFD92, 0xFDC7}, {0xFDF0, 0xFDFB}, {0xFE70, 0xFE74}, {0xFE76, 0xFEFC}, {0xFF21, 0xFF3A}, {0xFF41, 0xFF5A}, {0xFF66, 0xFFBE}, {0xFFC2, 0xFFC7},
{0xFFCA, 0xFFCF}, {0xFFD2, 0xFFD7}, {0xFFDA, 0xFFDC}, {0x10000, 0x1000B}, {0x1000D, 0x10026}, {0x10028, 0x1003A}, {0x1003C, 0x1003D}, {0x1003F, 0x1004D}, {0x10050, 0x1005D}, {0x10080, 0x100FA},
{0x10280, 0x1029C}, {0x102A0, 0x102D0}, {0x10300, 0x1031F}, {0x1032D, 0x10340}, {0x10342, 0x10349}, {0x10350, 0x10375}, {0x10380, 0x1039D}, {0x103A0, 0x103C3}, {0x103C8, 0x103CF}, {0x10400, 0x1049D},
{0x104B0, 0x104D3}, {0x104D8, 0x104FB}, {0x10500, 0x10527}, {0x10530, 0x10563}, {0x10600, 0x10736}, {0x10740, 0x10755}, {0x10760, 0x10767}, {0x10800, 0x10805}, {0x10808, 0x10808}, {0x1080A, 0x10835},
{0x10837, 0x10838}, {0x1083C, 0x1083C}, {0x1083F, 0x10855}, {0x10860, 0x10876}, {0x10880, 0x1089E}, {0x108E0, 0x108F2}, {0x108F4, 0x108F5}, {0x10900, 0x10915}, {0x10920, 0x10939}, {0x10980, 0x109B7},
{0x109BE, 0x109BF}, {0x10A00, 0x10A00}, {0x10A10, 0x10A13}, {0x10A15, 0x10A17}, {0x10A19, 0x10A35}, {0x10A60, 0x10A7C}, {0x10A80, 0x10A9C}, {0x10AC0, 0x10AC7}, {0x10AC9, 0x10AE4}, {0x10B00, 0x10B35},
{0x10B40, 0x10B55}, {0x10B60, 0x10B72}, {0x10B80, 0x10B91}, {0x10C00, 0x10C48}, {0x10C80, 0x10CB2}, {0x10CC0, 0x10CF2}, {0x10D00, 0x10D23}, {0x10E80, 0x10EA9}, {0x10EB0, 0x10EB1}, {0x10F00, 0x10F1C},
{0x10F27, 0x10F27}, {0x10F30, 0x10F45}, {0x10FB0, 0x10FC4}, {0x10FE0, 0x10FF6}, {0x11003, 0x11037}, {0x11083, 0x110AF}, {0x110D0, 0x110E8}, {0x11103, 0x11126}, {0x11144, 0x11144}, {0x11147, 0x11147},
{0x11150, 0x11172}, {0x11176, 0x11176}, {0x11183, 0x111B2}, {0x111C1, 0x111C4}, {0x111DA, 0x111DA}, {0x111DC, 0x111DC}, {0x11200, 0x11211}, {0x11213, 0x1122B}, {0x11280, 0x11286}, {0x11288, 0x11288},
{0x1128A, 0x1128D}, {0x1128F, 0x1129D}, {0x1129F, 0x112A8}, {0x112B0, 0x112DE}, {0x11305, 0x1130C}, {0x1130F, 0x11310}, {0x11313, 0x11328}, {0x1132A, 0x11330}, {0x11332, 0x11333}, {0x11335, 0x11339},
{0x1133D, 0x1133D}, {0x11350, 0x11350}, {0x1135D, 0x11361}, {0x11400, 0x11434}, {0x11447, 0x1144A}, {0x1145F, 0x11461}, {0x11480, 0x114AF}, {0x114C4, 0x114C5}, {0x114C7, 0x114C7}, {0x11580, 0x115AE},
{0x115D8, 0x115DB}, {0x11600, 0x1162F}, {0x11644, 0x11644}, {0x11680, 0x116AA}, {0x116B8, 0x116B8}, {0x11700, 0x1171A}, {0x11800, 0x1182B}, {0x118A0, 0x118DF}, {0x118FF, 0x11906}, {0x11909, 0x11909},
{0x1190C, 0x11913}, {0x11915, 0x11916}, {0x11918, 0x1192F}, {0x1193F, 0x1193F}, {0x11941, 0x11941}, {0x119A0, 0x119A7}, {0x119AA, 0x119D0}, {0x119E1, 0x119E1}, {0x119E3, 0x119E3}, {0x11A00, 0x11A00},
{0x11A0B, 0x11A32}, {0x11A3A, 0x11A3A}, {0x11A50, 0x11A50}, {0x11A5C, 0x11A89}, {0x11A9D, 0x11A9D}, {0x11AC0, 0x11AF8}, {0x11C00, 0x11C08}, {0x11C0A, 0x11C2E}, {0x11C40, 0x11C40}, {0x11C72, 0x11C8F},
{0x11D00, 0x11D06}, {0x11D08, 0x11D09}, {0x11D0B, 0x11D30}, {0x11D46, 0x11D46}, {0x11D60, 0x11D65}, {0x11D67, 0x11D68}, {0x11D6A, 0x11D89}, {0x11D98, 0x11D98}, {0x11EE0, 0x11EF2}, {0x11FB0, 0x11FB0},
{0x12000, 0x12399}, {0x12480, 0x12543}, {0x13000, 0x1342E}, {0x14400, 0x14646}, {0x16800, 0x16A38}, {0x16A40, 0x16A5E}, {0x16AD0, 0x16AED}, {0x16B00, 0x16B2F}, {0x16B40, 0x16B43}, {0x16B63, 0x16B77},
{0x16B7D, 0x16B8F}, {0x16E40, 0x16E7F}, {0x16F00, 0x16F4A}, {0x16F50, 0x16F50}, {0x16F93, 0x16F9F}, {0x16FE0, 0x16FE1}, {0x16FE3, 0x16FE3}, {0x17000, 0x187F7}, {0x18800, 0x18CD5}, {0x18D00, 0x18D08},
{0x1B000, 0x1B11E}, {0x1B150, 0x1B152}, {0x1B164, 0x1B167}, {0x1B170, 0x1B2FB}, {0x1BC00, 0x1BC6A}, {0x1BC70, 0x1BC7C}, {0x1BC80, 0x1BC88}, {0x1BC90, 0x1BC99}, {0x1D400, 0x1D454}, {0x1D456, 0x1D49C},
{0x1D49E, 0x1D49F}, {0x1D4A2, 0x1D4A2}, {0x1D4A5, 0x1D4A6}, {0x1D4A9, 0x1D4AC}, {0x1D4AE, 0x1D4B9}, {0x1D4BB, 0x1D4BB}, {0x1D4BD, 0x1D4C3}, {0x1D4C5, 0x1D505}, {0x1D507, 0x1D50A}, {0x1D50D, 0x1D514},
{0x1D516, 0x1D51C}, {0x1D51E, 0x1D539}, {0x1D53B, 0x1D53E}, {0x1D540, 0x1D544}, {0x1D546, 0x1D546}, {0x1D54A, 0x1D550}, {0x1D552, 0x1D6A5}, {0x1D6A8, 0x1D6C0}, {0x1D6C2, 0x1D6DA}, {0x1D6DC, 0x1D6FA},
{0x1D6FC, 0x1D714}, {0x1D716, 0x1D734}, {0x1D736, 0x1D74E}, {0x1D750, 0x1D76E}, {0x1D770, 0x1D788}, {0x1D78A, 0x1D7A8}, {0x1D7AA, 0x1D7C2}, {0x1D7C4, 0x1D7CB}, {0x1E100, 0x1E12C}, {0x1E137, 0x1E13D},
{0x1E14E, 0x1E14E}, {0x1E2C0, 0x1E2EB}, {0x1E800, 0x1E8C4}, {0x1E900, 0x1E943}, {0x1E94B, 0x1E94B}, {0x1EE00, 0x1EE03}, {0x1EE05, 0x1EE1F}, {0x1EE21, 0x1EE22}, {0x1EE24, 0x1EE24}, {0x1EE27, 0x1EE27},
{0x1EE29, 0x1EE32}, {0x1EE34, 0x1EE37}, {0x1EE39, 0x1EE39}, {0x1EE3B, 0x1EE3B}, {0x1EE42, 0x1EE42}, {0x1EE47, 0x1EE47}, {0x1EE49, 0x1EE49}, {0x1EE4B, 0x1EE4B}, {0x1EE4D, 0x1EE4F}, {0x1EE51, 0x1EE52},
{0x1EE54, 0x1EE54}, {0x1EE57, 0x1EE57}, {0x1EE59, 0x1EE59}, {0x1EE5B, 0x1EE5B}, {0x1EE5D, 0x1EE5D}, {0x1EE5F, 0x1EE5F}, {0x1EE61, 0x1EE62}, {0x1EE64, 0x1EE64}, {0x1EE67, 0x1EE6A}, {0x1EE6C, 0x1EE72},
{0x1EE74, 0x1EE77}, {0x1EE79, 0x1EE7C}, {0x1EE7E, 0x1EE7E}, {0x1EE80, 0x1EE89}, {0x1EE8B, 0x1EE9B}, {0x1EEA1, 0x1EEA3}, {0x1EEA5, 0x1EEA9}, {0x1EEAB, 0x1EEBB}, {0x20000, 0x2A6DD}, {0x2A700, 0x2B734},
{0x2B740, 0x2B81D}, {0x2B820, 0x2CEA1}, {0x2CEB0, 0x2EBE0}, {0x2F800, 0x2FA1D}, {0x30000, 0x3134A},
};
static const std::vector<std::pair<uint32_t, uint32_t>> whitespace_ranges = {
{0x9, 0xD}, {0x1C, 0x20}, {0x85, 0x85}, {0xA0, 0xA0}, {0x1680, 0x1680}, {0x2000, 0x200A}, {0x2028, 0x2029}, {0x202F, 0x202F}, {0x205F, 0x205F}, {0x3000, 0x3000},
};
static const std::vector<std::pair<uint32_t, uint32_t>> accent_mark_ranges = {
{0x300, 0x36F}, {0x483, 0x489}, {0x591, 0x5BD}, {0x5BF, 0x5BF}, {0x5C1, 0x5C2}, {0x5C4, 0x5C5}, {0x5C7, 0x5C7}, {0x610, 0x61A}, {0x64B, 0x65F}, {0x670, 0x670}, {0x6D6, 0x6DC}, {0x6DF, 0x6E4},
{0x6E7, 0x6E8}, {0x6EA, 0x6ED}, {0x711, 0x711}, {0x730, 0x74A}, {0x7A6, 0x7B0}, {0x7EB, 0x7F3}, {0x7FD, 0x7FD}, {0x816, 0x819}, {0x81B, 0x823}, {0x825, 0x827}, {0x829, 0x82D}, {0x859, 0x85B},
{0x8D3, 0x8E1}, {0x8E3, 0x903}, {0x93A, 0x93C}, {0x93E, 0x94F}, {0x951, 0x957}, {0x962, 0x963}, {0x981, 0x983}, {0x9BC, 0x9BC}, {0x9BE, 0x9C4}, {0x9C7, 0x9C8}, {0x9CB, 0x9CD}, {0x9D7, 0x9D7},
{0x9E2, 0x9E3}, {0x9FE, 0x9FE}, {0xA01, 0xA03}, {0xA3C, 0xA3C}, {0xA3E, 0xA42}, {0xA47, 0xA48}, {0xA4B, 0xA4D}, {0xA51, 0xA51}, {0xA70, 0xA71}, {0xA75, 0xA75}, {0xA81, 0xA83}, {0xABC, 0xABC},
{0xABE, 0xAC5}, {0xAC7, 0xAC9}, {0xACB, 0xACD}, {0xAE2, 0xAE3}, {0xAFA, 0xAFF}, {0xB01, 0xB03}, {0xB3C, 0xB3C}, {0xB3E, 0xB44}, {0xB47, 0xB48}, {0xB4B, 0xB4D}, {0xB55, 0xB57}, {0xB62, 0xB63},
{0xB82, 0xB82}, {0xBBE, 0xBC2}, {0xBC6, 0xBC8}, {0xBCA, 0xBCD}, {0xBD7, 0xBD7}, {0xC00, 0xC04}, {0xC3E, 0xC44}, {0xC46, 0xC48}, {0xC4A, 0xC4D}, {0xC55, 0xC56}, {0xC62, 0xC63}, {0xC81, 0xC83},
{0xCBC, 0xCBC}, {0xCBE, 0xCC4}, {0xCC6, 0xCC8}, {0xCCA, 0xCCD}, {0xCD5, 0xCD6}, {0xCE2, 0xCE3}, {0xD00, 0xD03}, {0xD3B, 0xD3C}, {0xD3E, 0xD44}, {0xD46, 0xD48}, {0xD4A, 0xD4D}, {0xD57, 0xD57},
{0xD62, 0xD63}, {0xD81, 0xD83}, {0xDCA, 0xDCA}, {0xDCF, 0xDD4}, {0xDD6, 0xDD6}, {0xDD8, 0xDDF}, {0xDF2, 0xDF3}, {0xE31, 0xE31}, {0xE34, 0xE3A}, {0xE47, 0xE4E}, {0xEB1, 0xEB1}, {0xEB4, 0xEBC},
{0xEC8, 0xECD}, {0xF18, 0xF19}, {0xF35, 0xF35}, {0xF37, 0xF37}, {0xF39, 0xF39}, {0xF3E, 0xF3F}, {0xF71, 0xF84}, {0xF86, 0xF87}, {0xF8D, 0xF97}, {0xF99, 0xFBC}, {0xFC6, 0xFC6}, {0x102B, 0x103E},
{0x1056, 0x1059}, {0x105E, 0x1060}, {0x1062, 0x1064}, {0x1067, 0x106D}, {0x1071, 0x1074}, {0x1082, 0x108D}, {0x108F, 0x108F}, {0x109A, 0x109D}, {0x135D, 0x135F}, {0x1712, 0x1714}, {0x1732, 0x1734},
{0x1752, 0x1753}, {0x1772, 0x1773}, {0x17B4, 0x17D3}, {0x17DD, 0x17DD}, {0x180B, 0x180D}, {0x1885, 0x1886}, {0x18A9, 0x18A9}, {0x1920, 0x192B}, {0x1930, 0x193B}, {0x1A17, 0x1A1B}, {0x1A55, 0x1A5E},
{0x1A60, 0x1A7C}, {0x1A7F, 0x1A7F}, {0x1AB0, 0x1AC0}, {0x1B00, 0x1B04}, {0x1B34, 0x1B44}, {0x1B6B, 0x1B73}, {0x1B80, 0x1B82}, {0x1BA1, 0x1BAD}, {0x1BE6, 0x1BF3}, {0x1C24, 0x1C37}, {0x1CD0, 0x1CD2},
{0x1CD4, 0x1CE8}, {0x1CED, 0x1CED}, {0x1CF4, 0x1CF4}, {0x1CF7, 0x1CF9}, {0x1DC0, 0x1DF9}, {0x1DFB, 0x1DFF}, {0x20D0, 0x20F0}, {0x2CEF, 0x2CF1}, {0x2D7F, 0x2D7F}, {0x2DE0, 0x2DFF}, {0x302A, 0x302F},
{0x3099, 0x309A}, {0xA66F, 0xA672}, {0xA674, 0xA67D}, {0xA69E, 0xA69F}, {0xA6F0, 0xA6F1}, {0xA802, 0xA802}, {0xA806, 0xA806}, {0xA80B, 0xA80B}, {0xA823, 0xA827}, {0xA82C, 0xA82C}, {0xA880, 0xA881},
{0xA8B4, 0xA8C5}, {0xA8E0, 0xA8F1}, {0xA8FF, 0xA8FF}, {0xA926, 0xA92D}, {0xA947, 0xA953}, {0xA980, 0xA983}, {0xA9B3, 0xA9C0}, {0xA9E5, 0xA9E5}, {0xAA29, 0xAA36}, {0xAA43, 0xAA43}, {0xAA4C, 0xAA4D},
{0xAA7B, 0xAA7D}, {0xAAB0, 0xAAB0}, {0xAAB2, 0xAAB4}, {0xAAB7, 0xAAB8}, {0xAABE, 0xAABF}, {0xAAC1, 0xAAC1}, {0xAAEB, 0xAAEF}, {0xAAF5, 0xAAF6}, {0xABE3, 0xABEA}, {0xABEC, 0xABED}, {0xFB1E, 0xFB1E},
{0xFE00, 0xFE0F}, {0xFE20, 0xFE2F}, {0x101FD, 0x101FD}, {0x102E0, 0x102E0}, {0x10376, 0x1037A}, {0x10A01, 0x10A03}, {0x10A05, 0x10A06}, {0x10A0C, 0x10A0F}, {0x10A38, 0x10A3A}, {0x10A3F, 0x10A3F},
{0x10AE5, 0x10AE6}, {0x10D24, 0x10D27}, {0x10EAB, 0x10EAC}, {0x10F46, 0x10F50}, {0x11000, 0x11002}, {0x11038, 0x11046}, {0x1107F, 0x11082}, {0x110B0, 0x110BA}, {0x11100, 0x11102}, {0x11127, 0x11134},
{0x11145, 0x11146}, {0x11173, 0x11173}, {0x11180, 0x11182}, {0x111B3, 0x111C0}, {0x111C9, 0x111CC}, {0x111CE, 0x111CF}, {0x1122C, 0x11237}, {0x1123E, 0x1123E}, {0x112DF, 0x112EA}, {0x11300, 0x11303},
{0x1133B, 0x1133C}, {0x1133E, 0x11344}, {0x11347, 0x11348}, {0x1134B, 0x1134D}, {0x11357, 0x11357}, {0x11362, 0x11363}, {0x11366, 0x1136C}, {0x11370, 0x11374}, {0x11435, 0x11446}, {0x1145E, 0x1145E},
{0x114B0, 0x114C3}, {0x115AF, 0x115B5}, {0x115B8, 0x115C0}, {0x115DC, 0x115DD}, {0x11630, 0x11640}, {0x116AB, 0x116B7}, {0x1171D, 0x1172B}, {0x1182C, 0x1183A}, {0x11930, 0x11935}, {0x11937, 0x11938},
{0x1193B, 0x1193E}, {0x11940, 0x11940}, {0x11942, 0x11943}, {0x119D1, 0x119D7}, {0x119DA, 0x119E0}, {0x119E4, 0x119E4}, {0x11A01, 0x11A0A}, {0x11A33, 0x11A39}, {0x11A3B, 0x11A3E}, {0x11A47, 0x11A47},
{0x11A51, 0x11A5B}, {0x11A8A, 0x11A99}, {0x11C2F, 0x11C36}, {0x11C38, 0x11C3F}, {0x11C92, 0x11CA7}, {0x11CA9, 0x11CB6}, {0x11D31, 0x11D36}, {0x11D3A, 0x11D3A}, {0x11D3C, 0x11D3D}, {0x11D3F, 0x11D45},
{0x11D47, 0x11D47}, {0x11D8A, 0x11D8E}, {0x11D90, 0x11D91}, {0x11D93, 0x11D97}, {0x11EF3, 0x11EF6}, {0x16AF0, 0x16AF4}, {0x16B30, 0x16B36}, {0x16F4F, 0x16F4F}, {0x16F51, 0x16F87}, {0x16F8F, 0x16F92},
{0x16FE4, 0x16FE4}, {0x16FF0, 0x16FF1}, {0x1BC9D, 0x1BC9E}, {0x1D165, 0x1D169}, {0x1D16D, 0x1D172}, {0x1D17B, 0x1D182}, {0x1D185, 0x1D18B}, {0x1D1AA, 0x1D1AD}, {0x1D242, 0x1D244}, {0x1DA00, 0x1DA36},
{0x1DA3B, 0x1DA6C}, {0x1DA75, 0x1DA75}, {0x1DA84, 0x1DA84}, {0x1DA9B, 0x1DA9F}, {0x1DAA1, 0x1DAAF}, {0x1E000, 0x1E006}, {0x1E008, 0x1E018}, {0x1E01B, 0x1E021}, {0x1E023, 0x1E024}, {0x1E026, 0x1E02A},
{0x1E130, 0x1E136}, {0x1E2EC, 0x1E2EF}, {0x1E8D0, 0x1E8D6}, {0x1E944, 0x1E94A}, {0xE0100, 0xE01EF},
};
static const std::vector<std::pair<uint32_t, uint32_t>> punctuation_ranges = {
{0x21, 0x23}, {0x25, 0x2A}, {0x2C, 0x2F}, {0x3A, 0x3B}, {0x3F, 0x40}, {0x5B, 0x5D}, {0x5F, 0x5F}, {0x7B, 0x7B}, {0x7D, 0x7D}, {0xA1, 0xA1}, {0xA7, 0xA7}, {0xAB, 0xAB}, {0xB6, 0xB7}, {0xBB, 0xBB},
{0xBF, 0xBF}, {0x37E, 0x37E}, {0x387, 0x387}, {0x55A, 0x55F}, {0x589, 0x58A}, {0x5BE, 0x5BE}, {0x5C0, 0x5C0}, {0x5C3, 0x5C3}, {0x5C6, 0x5C6}, {0x5F3, 0x5F4}, {0x609, 0x60A}, {0x60C, 0x60D},
{0x61B, 0x61B}, {0x61E, 0x61F}, {0x66A, 0x66D}, {0x6D4, 0x6D4}, {0x700, 0x70D}, {0x7F7, 0x7F9}, {0x830, 0x83E}, {0x85E, 0x85E}, {0x964, 0x965}, {0x970, 0x970}, {0x9FD, 0x9FD}, {0xA76, 0xA76},
{0xAF0, 0xAF0}, {0xC77, 0xC77}, {0xC84, 0xC84}, {0xDF4, 0xDF4}, {0xE4F, 0xE4F}, {0xE5A, 0xE5B}, {0xF04, 0xF12}, {0xF14, 0xF14}, {0xF3A, 0xF3D}, {0xF85, 0xF85}, {0xFD0, 0xFD4}, {0xFD9, 0xFDA},
{0x104A, 0x104F}, {0x10FB, 0x10FB}, {0x1360, 0x1368}, {0x1400, 0x1400}, {0x166E, 0x166E}, {0x169B, 0x169C}, {0x16EB, 0x16ED}, {0x1735, 0x1736}, {0x17D4, 0x17D6}, {0x17D8, 0x17DA}, {0x1800, 0x180A},
{0x1944, 0x1945}, {0x1A1E, 0x1A1F}, {0x1AA0, 0x1AA6}, {0x1AA8, 0x1AAD}, {0x1B5A, 0x1B60}, {0x1BFC, 0x1BFF}, {0x1C3B, 0x1C3F}, {0x1C7E, 0x1C7F}, {0x1CC0, 0x1CC7}, {0x1CD3, 0x1CD3}, {0x2010, 0x2027},
{0x2030, 0x2043}, {0x2045, 0x2051}, {0x2053, 0x205E}, {0x207D, 0x207E}, {0x208D, 0x208E}, {0x2308, 0x230B}, {0x2329, 0x232A}, {0x2768, 0x2775}, {0x27C5, 0x27C6}, {0x27E6, 0x27EF}, {0x2983, 0x2998},
{0x29D8, 0x29DB}, {0x29FC, 0x29FD}, {0x2CF9, 0x2CFC}, {0x2CFE, 0x2CFF}, {0x2D70, 0x2D70}, {0x2E00, 0x2E2E}, {0x2E30, 0x2E4F}, {0x2E52, 0x2E52}, {0x3001, 0x3003}, {0x3008, 0x3011}, {0x3014, 0x301F},
{0x3030, 0x3030}, {0x303D, 0x303D}, {0x30A0, 0x30A0}, {0x30FB, 0x30FB}, {0xA4FE, 0xA4FF}, {0xA60D, 0xA60F}, {0xA673, 0xA673}, {0xA67E, 0xA67E}, {0xA6F2, 0xA6F7}, {0xA874, 0xA877}, {0xA8CE, 0xA8CF},
{0xA8F8, 0xA8FA}, {0xA8FC, 0xA8FC}, {0xA92E, 0xA92F}, {0xA95F, 0xA95F}, {0xA9C1, 0xA9CD}, {0xA9DE, 0xA9DF}, {0xAA5C, 0xAA5F}, {0xAADE, 0xAADF}, {0xAAF0, 0xAAF1}, {0xABEB, 0xABEB}, {0xFD3E, 0xFD3F},
{0xFE10, 0xFE19}, {0xFE30, 0xFE52}, {0xFE54, 0xFE61}, {0xFE63, 0xFE63}, {0xFE68, 0xFE68}, {0xFE6A, 0xFE6B}, {0xFF01, 0xFF03}, {0xFF05, 0xFF0A}, {0xFF0C, 0xFF0F}, {0xFF1A, 0xFF1B}, {0xFF1F, 0xFF20},
{0xFF3B, 0xFF3D}, {0xFF3F, 0xFF3F}, {0xFF5B, 0xFF5B}, {0xFF5D, 0xFF5D}, {0xFF5F, 0xFF65}, {0x10100, 0x10102}, {0x1039F, 0x1039F}, {0x103D0, 0x103D0}, {0x1056F, 0x1056F}, {0x10857, 0x10857},
{0x1091F, 0x1091F}, {0x1093F, 0x1093F}, {0x10A50, 0x10A58}, {0x10A7F, 0x10A7F}, {0x10AF0, 0x10AF6}, {0x10B39, 0x10B3F}, {0x10B99, 0x10B9C}, {0x10EAD, 0x10EAD}, {0x10F55, 0x10F59}, {0x11047, 0x1104D},
{0x110BB, 0x110BC}, {0x110BE, 0x110C1}, {0x11140, 0x11143}, {0x11174, 0x11175}, {0x111C5, 0x111C8}, {0x111CD, 0x111CD}, {0x111DB, 0x111DB}, {0x111DD, 0x111DF}, {0x11238, 0x1123D}, {0x112A9, 0x112A9},
{0x1144B, 0x1144F}, {0x1145A, 0x1145B}, {0x1145D, 0x1145D}, {0x114C6, 0x114C6}, {0x115C1, 0x115D7}, {0x11641, 0x11643}, {0x11660, 0x1166C}, {0x1173C, 0x1173E}, {0x1183B, 0x1183B}, {0x11944, 0x11946},
{0x119E2, 0x119E2}, {0x11A3F, 0x11A46}, {0x11A9A, 0x11A9C}, {0x11A9E, 0x11AA2}, {0x11C41, 0x11C45}, {0x11C70, 0x11C71}, {0x11EF7, 0x11EF8}, {0x11FFF, 0x11FFF}, {0x12470, 0x12474}, {0x16A6E, 0x16A6F},
{0x16AF5, 0x16AF5}, {0x16B37, 0x16B3B}, {0x16B44, 0x16B44}, {0x16E97, 0x16E9A}, {0x16FE2, 0x16FE2}, {0x1BC9F, 0x1BC9F}, {0x1DA87, 0x1DA8B}, {0x1E95E, 0x1E95F},
};
static const std::vector<std::pair<uint32_t, uint32_t>> symbol_ranges = {
{0x24, 0x24}, {0x2B, 0x2B}, {0x3C, 0x3E}, {0x5E, 0x5E}, {0x60, 0x60}, {0x7C, 0x7C}, {0x7E, 0x7E}, {0xA2, 0xA6}, {0xA8, 0xA9}, {0xAC, 0xAC}, {0xAE, 0xB1}, {0xB4, 0xB4}, {0xB8, 0xB8}, {0xD7, 0xD7},
{0xF7, 0xF7}, {0x2C2, 0x2C5}, {0x2D2, 0x2DF}, {0x2E5, 0x2EB}, {0x2ED, 0x2ED}, {0x2EF, 0x2FF}, {0x375, 0x375}, {0x384, 0x385}, {0x3F6, 0x3F6}, {0x482, 0x482}, {0x58D, 0x58F}, {0x606, 0x608},
{0x60B, 0x60B}, {0x60E, 0x60F}, {0x6DE, 0x6DE}, {0x6E9, 0x6E9}, {0x6FD, 0x6FE}, {0x7F6, 0x7F6}, {0x7FE, 0x7FF}, {0x9F2, 0x9F3}, {0x9FA, 0x9FB}, {0xAF1, 0xAF1}, {0xB70, 0xB70}, {0xBF3, 0xBFA},
{0xC7F, 0xC7F}, {0xD4F, 0xD4F}, {0xD79, 0xD79}, {0xE3F, 0xE3F}, {0xF01, 0xF03}, {0xF13, 0xF13}, {0xF15, 0xF17}, {0xF1A, 0xF1F}, {0xF34, 0xF34}, {0xF36, 0xF36}, {0xF38, 0xF38}, {0xFBE, 0xFC5},
{0xFC7, 0xFCC}, {0xFCE, 0xFCF}, {0xFD5, 0xFD8}, {0x109E, 0x109F}, {0x1390, 0x1399}, {0x166D, 0x166D}, {0x17DB, 0x17DB}, {0x1940, 0x1940}, {0x19DE, 0x19FF}, {0x1B61, 0x1B6A}, {0x1B74, 0x1B7C},
{0x1FBD, 0x1FBD}, {0x1FBF, 0x1FC1}, {0x1FCD, 0x1FCF}, {0x1FDD, 0x1FDF}, {0x1FED, 0x1FEF}, {0x1FFD, 0x1FFE}, {0x2044, 0x2044}, {0x2052, 0x2052}, {0x207A, 0x207C}, {0x208A, 0x208C}, {0x20A0, 0x20BF},
{0x2100, 0x2101}, {0x2103, 0x2106}, {0x2108, 0x2109}, {0x2114, 0x2114}, {0x2116, 0x2118}, {0x211E, 0x2123}, {0x2125, 0x2125}, {0x2127, 0x2127}, {0x2129, 0x2129}, {0x212E, 0x212E}, {0x213A, 0x213B},
{0x2140, 0x2144}, {0x214A, 0x214D}, {0x214F, 0x214F}, {0x218A, 0x218B}, {0x2190, 0x2307}, {0x230C, 0x2328}, {0x232B, 0x2426}, {0x2440, 0x244A}, {0x249C, 0x24E9}, {0x2500, 0x2767}, {0x2794, 0x27C4},
{0x27C7, 0x27E5}, {0x27F0, 0x2982}, {0x2999, 0x29D7}, {0x29DC, 0x29FB}, {0x29FE, 0x2B73}, {0x2B76, 0x2B95}, {0x2B97, 0x2BFF}, {0x2CE5, 0x2CEA}, {0x2E50, 0x2E51}, {0x2E80, 0x2E99}, {0x2E9B, 0x2EF3},
{0x2F00, 0x2FD5}, {0x2FF0, 0x2FFB}, {0x3004, 0x3004}, {0x3012, 0x3013}, {0x3020, 0x3020}, {0x3036, 0x3037}, {0x303E, 0x303F}, {0x309B, 0x309C}, {0x3190, 0x3191}, {0x3196, 0x319F}, {0x31C0, 0x31E3},
{0x3200, 0x321E}, {0x322A, 0x3247}, {0x3250, 0x3250}, {0x3260, 0x327F}, {0x328A, 0x32B0}, {0x32C0, 0x33FF}, {0x4DC0, 0x4DFF}, {0xA490, 0xA4C6}, {0xA700, 0xA716}, {0xA720, 0xA721}, {0xA789, 0xA78A},
{0xA828, 0xA82B}, {0xA836, 0xA839}, {0xAA77, 0xAA79}, {0xAB5B, 0xAB5B}, {0xAB6A, 0xAB6B}, {0xFB29, 0xFB29}, {0xFBB2, 0xFBC1}, {0xFDFC, 0xFDFD}, {0xFE62, 0xFE62}, {0xFE64, 0xFE66}, {0xFE69, 0xFE69},
{0xFF04, 0xFF04}, {0xFF0B, 0xFF0B}, {0xFF1C, 0xFF1E}, {0xFF3E, 0xFF3E}, {0xFF40, 0xFF40}, {0xFF5C, 0xFF5C}, {0xFF5E, 0xFF5E}, {0xFFE0, 0xFFE6}, {0xFFE8, 0xFFEE}, {0xFFFC, 0xFFFD}, {0x10137, 0x1013F},
{0x10179, 0x10189}, {0x1018C, 0x1018E}, {0x10190, 0x1019C}, {0x101A0, 0x101A0}, {0x101D0, 0x101FC}, {0x10877, 0x10878}, {0x10AC8, 0x10AC8}, {0x1173F, 0x1173F}, {0x11FD5, 0x11FF1}, {0x16B3C, 0x16B3F},
{0x16B45, 0x16B45}, {0x1BC9C, 0x1BC9C}, {0x1D000, 0x1D0F5}, {0x1D100, 0x1D126}, {0x1D129, 0x1D164}, {0x1D16A, 0x1D16C}, {0x1D183, 0x1D184}, {0x1D18C, 0x1D1A9}, {0x1D1AE, 0x1D1E8}, {0x1D200, 0x1D241},
{0x1D245, 0x1D245}, {0x1D300, 0x1D356}, {0x1D6C1, 0x1D6C1}, {0x1D6DB, 0x1D6DB}, {0x1D6FB, 0x1D6FB}, {0x1D715, 0x1D715}, {0x1D735, 0x1D735}, {0x1D74F, 0x1D74F}, {0x1D76F, 0x1D76F}, {0x1D789, 0x1D789},
{0x1D7A9, 0x1D7A9}, {0x1D7C3, 0x1D7C3}, {0x1D800, 0x1D9FF}, {0x1DA37, 0x1DA3A}, {0x1DA6D, 0x1DA74}, {0x1DA76, 0x1DA83}, {0x1DA85, 0x1DA86}, {0x1E14F, 0x1E14F}, {0x1E2FF, 0x1E2FF}, {0x1ECAC, 0x1ECAC},
{0x1ECB0, 0x1ECB0}, {0x1ED2E, 0x1ED2E}, {0x1EEF0, 0x1EEF1}, {0x1F000, 0x1F02B}, {0x1F030, 0x1F093}, {0x1F0A0, 0x1F0AE}, {0x1F0B1, 0x1F0BF}, {0x1F0C1, 0x1F0CF}, {0x1F0D1, 0x1F0F5}, {0x1F10D, 0x1F1AD},
{0x1F1E6, 0x1F202}, {0x1F210, 0x1F23B}, {0x1F240, 0x1F248}, {0x1F250, 0x1F251}, {0x1F260, 0x1F265}, {0x1F300, 0x1F6D7}, {0x1F6E0, 0x1F6EC}, {0x1F6F0, 0x1F6FC}, {0x1F700, 0x1F773}, {0x1F780, 0x1F7D8},
{0x1F7E0, 0x1F7EB}, {0x1F800, 0x1F80B}, {0x1F810, 0x1F847}, {0x1F850, 0x1F859}, {0x1F860, 0x1F887}, {0x1F890, 0x1F8AD}, {0x1F8B0, 0x1F8B1}, {0x1F900, 0x1F978}, {0x1F97A, 0x1F9CB}, {0x1F9CD, 0x1FA53},
{0x1FA60, 0x1FA6D}, {0x1FA70, 0x1FA74}, {0x1FA78, 0x1FA7A}, {0x1FA80, 0x1FA86}, {0x1FA90, 0x1FAA8}, {0x1FAB0, 0x1FAB6}, {0x1FAC0, 0x1FAC2}, {0x1FAD0, 0x1FAD6}, {0x1FB00, 0x1FB92}, {0x1FB94, 0x1FBCA},
};
static const std::vector<std::pair<uint32_t, uint32_t>> control_ranges = {
{0x0, 0x8}, {0xE, 0x1B}, {0x7F, 0x84}, {0x86, 0x9F}, {0xAD, 0xAD}, {0x378, 0x379}, {0x380, 0x383}, {0x38B, 0x38B}, {0x38D, 0x38D}, {0x3A2, 0x3A2}, {0x530, 0x530}, {0x557, 0x558}, {0x58B, 0x58C},
{0x590, 0x590}, {0x5C8, 0x5CF}, {0x5EB, 0x5EE}, {0x5F5, 0x605}, {0x61C, 0x61D}, {0x6DD, 0x6DD}, {0x70E, 0x70F}, {0x74B, 0x74C}, {0x7B2, 0x7BF}, {0x7FB, 0x7FC}, {0x82E, 0x82F}, {0x83F, 0x83F},
{0x85C, 0x85D}, {0x85F, 0x85F}, {0x86B, 0x89F}, {0x8B5, 0x8B5}, {0x8C8, 0x8D2}, {0x8E2, 0x8E2}, {0x984, 0x984}, {0x98D, 0x98E}, {0x991, 0x992}, {0x9A9, 0x9A9}, {0x9B1, 0x9B1}, {0x9B3, 0x9B5},
{0x9BA, 0x9BB}, {0x9C5, 0x9C6}, {0x9C9, 0x9CA}, {0x9CF, 0x9D6}, {0x9D8, 0x9DB}, {0x9DE, 0x9DE}, {0x9E4, 0x9E5}, {0x9FF, 0xA00}, {0xA04, 0xA04}, {0xA0B, 0xA0E}, {0xA11, 0xA12}, {0xA29, 0xA29},
{0xA31, 0xA31}, {0xA34, 0xA34}, {0xA37, 0xA37}, {0xA3A, 0xA3B}, {0xA3D, 0xA3D}, {0xA43, 0xA46}, {0xA49, 0xA4A}, {0xA4E, 0xA50}, {0xA52, 0xA58}, {0xA5D, 0xA5D}, {0xA5F, 0xA65}, {0xA77, 0xA80},
{0xA84, 0xA84}, {0xA8E, 0xA8E}, {0xA92, 0xA92}, {0xAA9, 0xAA9}, {0xAB1, 0xAB1}, {0xAB4, 0xAB4}, {0xABA, 0xABB}, {0xAC6, 0xAC6}, {0xACA, 0xACA}, {0xACE, 0xACF}, {0xAD1, 0xADF}, {0xAE4, 0xAE5},
{0xAF2, 0xAF8}, {0xB00, 0xB00}, {0xB04, 0xB04}, {0xB0D, 0xB0E}, {0xB11, 0xB12}, {0xB29, 0xB29}, {0xB31, 0xB31}, {0xB34, 0xB34}, {0xB3A, 0xB3B}, {0xB45, 0xB46}, {0xB49, 0xB4A}, {0xB4E, 0xB54},
{0xB58, 0xB5B}, {0xB5E, 0xB5E}, {0xB64, 0xB65}, {0xB78, 0xB81}, {0xB84, 0xB84}, {0xB8B, 0xB8D}, {0xB91, 0xB91}, {0xB96, 0xB98}, {0xB9B, 0xB9B}, {0xB9D, 0xB9D}, {0xBA0, 0xBA2}, {0xBA5, 0xBA7},
{0xBAB, 0xBAD}, {0xBBA, 0xBBD}, {0xBC3, 0xBC5}, {0xBC9, 0xBC9}, {0xBCE, 0xBCF}, {0xBD1, 0xBD6}, {0xBD8, 0xBE5}, {0xBFB, 0xBFF}, {0xC0D, 0xC0D}, {0xC11, 0xC11}, {0xC29, 0xC29}, {0xC3A, 0xC3C},
{0xC45, 0xC45}, {0xC49, 0xC49}, {0xC4E, 0xC54}, {0xC57, 0xC57}, {0xC5B, 0xC5F}, {0xC64, 0xC65}, {0xC70, 0xC76}, {0xC8D, 0xC8D}, {0xC91, 0xC91}, {0xCA9, 0xCA9}, {0xCB4, 0xCB4}, {0xCBA, 0xCBB},
{0xCC5, 0xCC5}, {0xCC9, 0xCC9}, {0xCCE, 0xCD4}, {0xCD7, 0xCDD}, {0xCDF, 0xCDF}, {0xCE4, 0xCE5}, {0xCF0, 0xCF0}, {0xCF3, 0xCFF}, {0xD0D, 0xD0D}, {0xD11, 0xD11}, {0xD45, 0xD45}, {0xD49, 0xD49},
{0xD50, 0xD53}, {0xD64, 0xD65}, {0xD80, 0xD80}, {0xD84, 0xD84}, {0xD97, 0xD99}, {0xDB2, 0xDB2}, {0xDBC, 0xDBC}, {0xDBE, 0xDBF}, {0xDC7, 0xDC9}, {0xDCB, 0xDCE}, {0xDD5, 0xDD5}, {0xDD7, 0xDD7},
{0xDE0, 0xDE5}, {0xDF0, 0xDF1}, {0xDF5, 0xE00}, {0xE3B, 0xE3E}, {0xE5C, 0xE80}, {0xE83, 0xE83}, {0xE85, 0xE85}, {0xE8B, 0xE8B}, {0xEA4, 0xEA4}, {0xEA6, 0xEA6}, {0xEBE, 0xEBF}, {0xEC5, 0xEC5},
{0xEC7, 0xEC7}, {0xECE, 0xECF}, {0xEDA, 0xEDB}, {0xEE0, 0xEFF}, {0xF48, 0xF48}, {0xF6D, 0xF70}, {0xF98, 0xF98}, {0xFBD, 0xFBD}, {0xFCD, 0xFCD}, {0xFDB, 0xFFF}, {0x10C6, 0x10C6}, {0x10C8, 0x10CC},
{0x10CE, 0x10CF}, {0x1249, 0x1249}, {0x124E, 0x124F}, {0x1257, 0x1257}, {0x1259, 0x1259}, {0x125E, 0x125F}, {0x1289, 0x1289}, {0x128E, 0x128F}, {0x12B1, 0x12B1}, {0x12B6, 0x12B7}, {0x12BF, 0x12BF},
{0x12C1, 0x12C1}, {0x12C6, 0x12C7}, {0x12D7, 0x12D7}, {0x1311, 0x1311}, {0x1316, 0x1317}, {0x135B, 0x135C}, {0x137D, 0x137F}, {0x139A, 0x139F}, {0x13F6, 0x13F7}, {0x13FE, 0x13FF}, {0x169D, 0x169F},
{0x16F9, 0x16FF}, {0x170D, 0x170D}, {0x1715, 0x171F}, {0x1737, 0x173F}, {0x1754, 0x175F}, {0x176D, 0x176D}, {0x1771, 0x1771}, {0x1774, 0x177F}, {0x17DE, 0x17DF}, {0x17EA, 0x17EF}, {0x17FA, 0x17FF},
{0x180E, 0x180F}, {0x181A, 0x181F}, {0x1879, 0x187F}, {0x18AB, 0x18AF}, {0x18F6, 0x18FF}, {0x191F, 0x191F}, {0x192C, 0x192F}, {0x193C, 0x193F}, {0x1941, 0x1943}, {0x196E, 0x196F}, {0x1975, 0x197F},
{0x19AC, 0x19AF}, {0x19CA, 0x19CF}, {0x19DB, 0x19DD}, {0x1A1C, 0x1A1D}, {0x1A5F, 0x1A5F}, {0x1A7D, 0x1A7E}, {0x1A8A, 0x1A8F}, {0x1A9A, 0x1A9F}, {0x1AAE, 0x1AAF}, {0x1AC1, 0x1AFF}, {0x1B4C, 0x1B4F},
{0x1B7D, 0x1B7F}, {0x1BF4, 0x1BFB}, {0x1C38, 0x1C3A}, {0x1C4A, 0x1C4C}, {0x1C89, 0x1C8F}, {0x1CBB, 0x1CBC}, {0x1CC8, 0x1CCF}, {0x1CFB, 0x1CFF}, {0x1DFA, 0x1DFA}, {0x1F16, 0x1F17}, {0x1F1E, 0x1F1F},
{0x1F46, 0x1F47}, {0x1F4E, 0x1F4F}, {0x1F58, 0x1F58}, {0x1F5A, 0x1F5A}, {0x1F5C, 0x1F5C}, {0x1F5E, 0x1F5E}, {0x1F7E, 0x1F7F}, {0x1FB5, 0x1FB5}, {0x1FC5, 0x1FC5}, {0x1FD4, 0x1FD5}, {0x1FDC, 0x1FDC},
{0x1FF0, 0x1FF1}, {0x1FF5, 0x1FF5}, {0x1FFF, 0x1FFF}, {0x200B, 0x200F}, {0x202A, 0x202E}, {0x2060, 0x206F}, {0x2072, 0x2073}, {0x208F, 0x208F}, {0x209D, 0x209F}, {0x20C0, 0x20CF}, {0x20F1, 0x20FF},
{0x218C, 0x218F}, {0x2427, 0x243F}, {0x244B, 0x245F}, {0x2B74, 0x2B75}, {0x2B96, 0x2B96}, {0x2C2F, 0x2C2F}, {0x2C5F, 0x2C5F}, {0x2CF4, 0x2CF8}, {0x2D26, 0x2D26}, {0x2D28, 0x2D2C}, {0x2D2E, 0x2D2F},
{0x2D68, 0x2D6E}, {0x2D71, 0x2D7E}, {0x2D97, 0x2D9F}, {0x2DA7, 0x2DA7}, {0x2DAF, 0x2DAF}, {0x2DB7, 0x2DB7}, {0x2DBF, 0x2DBF}, {0x2DC7, 0x2DC7}, {0x2DCF, 0x2DCF}, {0x2DD7, 0x2DD7}, {0x2DDF, 0x2DDF},
{0x2E53, 0x2E7F}, {0x2E9A, 0x2E9A}, {0x2EF4, 0x2EFF}, {0x2FD6, 0x2FEF}, {0x2FFC, 0x2FFF}, {0x3040, 0x3040}, {0x3097, 0x3098}, {0x3100, 0x3104}, {0x3130, 0x3130}, {0x318F, 0x318F}, {0x31E4, 0x31EF},
{0x321F, 0x321F}, {0x9FFD, 0x9FFF}, {0xA48D, 0xA48F}, {0xA4C7, 0xA4CF}, {0xA62C, 0xA63F}, {0xA6F8, 0xA6FF}, {0xA7C0, 0xA7C1}, {0xA7CB, 0xA7F4}, {0xA82D, 0xA82F}, {0xA83A, 0xA83F}, {0xA878, 0xA87F},
{0xA8C6, 0xA8CD}, {0xA8DA, 0xA8DF}, {0xA954, 0xA95E}, {0xA97D, 0xA97F}, {0xA9CE, 0xA9CE}, {0xA9DA, 0xA9DD}, {0xA9FF, 0xA9FF}, {0xAA37, 0xAA3F}, {0xAA4E, 0xAA4F}, {0xAA5A, 0xAA5B}, {0xAAC3, 0xAADA},
{0xAAF7, 0xAB00}, {0xAB07, 0xAB08}, {0xAB0F, 0xAB10}, {0xAB17, 0xAB1F}, {0xAB27, 0xAB27}, {0xAB2F, 0xAB2F}, {0xAB6C, 0xAB6F}, {0xABEE, 0xABEF}, {0xABFA, 0xABFF}, {0xD7A4, 0xD7AF}, {0xD7C7, 0xD7CA},
{0xD7FC, 0xF8FF}, {0xFA6E, 0xFA6F}, {0xFADA, 0xFAFF}, {0xFB07, 0xFB12}, {0xFB18, 0xFB1C}, {0xFB37, 0xFB37}, {0xFB3D, 0xFB3D}, {0xFB3F, 0xFB3F}, {0xFB42, 0xFB42}, {0xFB45, 0xFB45}, {0xFBC2, 0xFBD2},
{0xFD40, 0xFD4F}, {0xFD90, 0xFD91}, {0xFDC8, 0xFDEF}, {0xFDFE, 0xFDFF}, {0xFE1A, 0xFE1F}, {0xFE53, 0xFE53}, {0xFE67, 0xFE67}, {0xFE6C, 0xFE6F}, {0xFE75, 0xFE75}, {0xFEFD, 0xFF00}, {0xFFBF, 0xFFC1},
{0xFFC8, 0xFFC9}, {0xFFD0, 0xFFD1}, {0xFFD8, 0xFFD9}, {0xFFDD, 0xFFDF}, {0xFFE7, 0xFFE7}, {0xFFEF, 0xFFFB}, {0xFFFE, 0xFFFF}, {0x1000C, 0x1000C}, {0x10027, 0x10027}, {0x1003B, 0x1003B},
{0x1003E, 0x1003E}, {0x1004E, 0x1004F}, {0x1005E, 0x1007F}, {0x100FB, 0x100FF}, {0x10103, 0x10106}, {0x10134, 0x10136}, {0x1018F, 0x1018F}, {0x1019D, 0x1019F}, {0x101A1, 0x101CF}, {0x101FE, 0x1027F},
{0x1029D, 0x1029F}, {0x102D1, 0x102DF}, {0x102FC, 0x102FF}, {0x10324, 0x1032C}, {0x1034B, 0x1034F}, {0x1037B, 0x1037F}, {0x1039E, 0x1039E}, {0x103C4, 0x103C7}, {0x103D6, 0x103FF}, {0x1049E, 0x1049F},
{0x104AA, 0x104AF}, {0x104D4, 0x104D7}, {0x104FC, 0x104FF}, {0x10528, 0x1052F}, {0x10564, 0x1056E}, {0x10570, 0x105FF}, {0x10737, 0x1073F}, {0x10756, 0x1075F}, {0x10768, 0x107FF}, {0x10806, 0x10807},
{0x10809, 0x10809}, {0x10836, 0x10836}, {0x10839, 0x1083B}, {0x1083D, 0x1083E}, {0x10856, 0x10856}, {0x1089F, 0x108A6}, {0x108B0, 0x108DF}, {0x108F3, 0x108F3}, {0x108F6, 0x108FA}, {0x1091C, 0x1091E},
{0x1093A, 0x1093E}, {0x10940, 0x1097F}, {0x109B8, 0x109BB}, {0x109D0, 0x109D1}, {0x10A04, 0x10A04}, {0x10A07, 0x10A0B}, {0x10A14, 0x10A14}, {0x10A18, 0x10A18}, {0x10A36, 0x10A37}, {0x10A3B, 0x10A3E},
{0x10A49, 0x10A4F}, {0x10A59, 0x10A5F}, {0x10AA0, 0x10ABF}, {0x10AE7, 0x10AEA}, {0x10AF7, 0x10AFF}, {0x10B36, 0x10B38}, {0x10B56, 0x10B57}, {0x10B73, 0x10B77}, {0x10B92, 0x10B98}, {0x10B9D, 0x10BA8},
{0x10BB0, 0x10BFF}, {0x10C49, 0x10C7F}, {0x10CB3, 0x10CBF}, {0x10CF3, 0x10CF9}, {0x10D28, 0x10D2F}, {0x10D3A, 0x10E5F}, {0x10E7F, 0x10E7F}, {0x10EAA, 0x10EAA}, {0x10EAE, 0x10EAF}, {0x10EB2, 0x10EFF},
{0x10F28, 0x10F2F}, {0x10F5A, 0x10FAF}, {0x10FCC, 0x10FDF}, {0x10FF7, 0x10FFF}, {0x1104E, 0x11051}, {0x11070, 0x1107E}, {0x110BD, 0x110BD}, {0x110C2, 0x110CF}, {0x110E9, 0x110EF}, {0x110FA, 0x110FF},
{0x11135, 0x11135}, {0x11148, 0x1114F}, {0x11177, 0x1117F}, {0x111E0, 0x111E0}, {0x111F5, 0x111FF}, {0x11212, 0x11212}, {0x1123F, 0x1127F}, {0x11287, 0x11287}, {0x11289, 0x11289}, {0x1128E, 0x1128E},
{0x1129E, 0x1129E}, {0x112AA, 0x112AF}, {0x112EB, 0x112EF}, {0x112FA, 0x112FF}, {0x11304, 0x11304}, {0x1130D, 0x1130E}, {0x11311, 0x11312}, {0x11329, 0x11329}, {0x11331, 0x11331}, {0x11334, 0x11334},
{0x1133A, 0x1133A}, {0x11345, 0x11346}, {0x11349, 0x1134A}, {0x1134E, 0x1134F}, {0x11351, 0x11356}, {0x11358, 0x1135C}, {0x11364, 0x11365}, {0x1136D, 0x1136F}, {0x11375, 0x113FF}, {0x1145C, 0x1145C},
{0x11462, 0x1147F}, {0x114C8, 0x114CF}, {0x114DA, 0x1157F}, {0x115B6, 0x115B7}, {0x115DE, 0x115FF}, {0x11645, 0x1164F}, {0x1165A, 0x1165F}, {0x1166D, 0x1167F}, {0x116B9, 0x116BF}, {0x116CA, 0x116FF},
{0x1171B, 0x1171C}, {0x1172C, 0x1172F}, {0x11740, 0x117FF}, {0x1183C, 0x1189F}, {0x118F3, 0x118FE}, {0x11907, 0x11908}, {0x1190A, 0x1190B}, {0x11914, 0x11914}, {0x11917, 0x11917}, {0x11936, 0x11936},
{0x11939, 0x1193A}, {0x11947, 0x1194F}, {0x1195A, 0x1199F}, {0x119A8, 0x119A9}, {0x119D8, 0x119D9}, {0x119E5, 0x119FF}, {0x11A48, 0x11A4F}, {0x11AA3, 0x11ABF}, {0x11AF9, 0x11BFF}, {0x11C09, 0x11C09},
{0x11C37, 0x11C37}, {0x11C46, 0x11C4F}, {0x11C6D, 0x11C6F}, {0x11C90, 0x11C91}, {0x11CA8, 0x11CA8}, {0x11CB7, 0x11CFF}, {0x11D07, 0x11D07}, {0x11D0A, 0x11D0A}, {0x11D37, 0x11D39}, {0x11D3B, 0x11D3B},
{0x11D3E, 0x11D3E}, {0x11D48, 0x11D4F}, {0x11D5A, 0x11D5F}, {0x11D66, 0x11D66}, {0x11D69, 0x11D69}, {0x11D8F, 0x11D8F}, {0x11D92, 0x11D92}, {0x11D99, 0x11D9F}, {0x11DAA, 0x11EDF}, {0x11EF9, 0x11FAF},
{0x11FB1, 0x11FBF}, {0x11FF2, 0x11FFE}, {0x1239A, 0x123FF}, {0x1246F, 0x1246F}, {0x12475, 0x1247F}, {0x12544, 0x12FFF}, {0x1342F, 0x143FF}, {0x14647, 0x167FF}, {0x16A39, 0x16A3F}, {0x16A5F, 0x16A5F},
{0x16A6A, 0x16A6D}, {0x16A70, 0x16ACF}, {0x16AEE, 0x16AEF}, {0x16AF6, 0x16AFF}, {0x16B46, 0x16B4F}, {0x16B5A, 0x16B5A}, {0x16B62, 0x16B62}, {0x16B78, 0x16B7C}, {0x16B90, 0x16E3F}, {0x16E9B, 0x16EFF},
{0x16F4B, 0x16F4E}, {0x16F88, 0x16F8E}, {0x16FA0, 0x16FDF}, {0x16FE5, 0x16FEF}, {0x16FF2, 0x16FFF}, {0x187F8, 0x187FF}, {0x18CD6, 0x18CFF}, {0x18D09, 0x1AFFF}, {0x1B11F, 0x1B14F}, {0x1B153, 0x1B163},
{0x1B168, 0x1B16F}, {0x1B2FC, 0x1BBFF}, {0x1BC6B, 0x1BC6F}, {0x1BC7D, 0x1BC7F}, {0x1BC89, 0x1BC8F}, {0x1BC9A, 0x1BC9B}, {0x1BCA0, 0x1CFFF}, {0x1D0F6, 0x1D0FF}, {0x1D127, 0x1D128}, {0x1D173, 0x1D17A},
{0x1D1E9, 0x1D1FF}, {0x1D246, 0x1D2DF}, {0x1D2F4, 0x1D2FF}, {0x1D357, 0x1D35F}, {0x1D379, 0x1D3FF}, {0x1D455, 0x1D455}, {0x1D49D, 0x1D49D}, {0x1D4A0, 0x1D4A1}, {0x1D4A3, 0x1D4A4}, {0x1D4A7, 0x1D4A8},
{0x1D4AD, 0x1D4AD}, {0x1D4BA, 0x1D4BA}, {0x1D4BC, 0x1D4BC}, {0x1D4C4, 0x1D4C4}, {0x1D506, 0x1D506}, {0x1D50B, 0x1D50C}, {0x1D515, 0x1D515}, {0x1D51D, 0x1D51D}, {0x1D53A, 0x1D53A}, {0x1D53F, 0x1D53F},
{0x1D545, 0x1D545}, {0x1D547, 0x1D549}, {0x1D551, 0x1D551}, {0x1D6A6, 0x1D6A7}, {0x1D7CC, 0x1D7CD}, {0x1DA8C, 0x1DA9A}, {0x1DAA0, 0x1DAA0}, {0x1DAB0, 0x1DFFF}, {0x1E007, 0x1E007}, {0x1E019, 0x1E01A},
{0x1E022, 0x1E022}, {0x1E025, 0x1E025}, {0x1E02B, 0x1E0FF}, {0x1E12D, 0x1E12F}, {0x1E13E, 0x1E13F}, {0x1E14A, 0x1E14D}, {0x1E150, 0x1E2BF}, {0x1E2FA, 0x1E2FE}, {0x1E300, 0x1E7FF}, {0x1E8C5, 0x1E8C6},
{0x1E8D7, 0x1E8FF}, {0x1E94C, 0x1E94F}, {0x1E95A, 0x1E95D}, {0x1E960, 0x1EC70}, {0x1ECB5, 0x1ED00}, {0x1ED3E, 0x1EDFF}, {0x1EE04, 0x1EE04}, {0x1EE20, 0x1EE20}, {0x1EE23, 0x1EE23}, {0x1EE25, 0x1EE26},
{0x1EE28, 0x1EE28}, {0x1EE33, 0x1EE33}, {0x1EE38, 0x1EE38}, {0x1EE3A, 0x1EE3A}, {0x1EE3C, 0x1EE41}, {0x1EE43, 0x1EE46}, {0x1EE48, 0x1EE48}, {0x1EE4A, 0x1EE4A}, {0x1EE4C, 0x1EE4C}, {0x1EE50, 0x1EE50},
{0x1EE53, 0x1EE53}, {0x1EE55, 0x1EE56}, {0x1EE58, 0x1EE58}, {0x1EE5A, 0x1EE5A}, {0x1EE5C, 0x1EE5C}, {0x1EE5E, 0x1EE5E}, {0x1EE60, 0x1EE60}, {0x1EE63, 0x1EE63}, {0x1EE65, 0x1EE66}, {0x1EE6B, 0x1EE6B},
{0x1EE73, 0x1EE73}, {0x1EE78, 0x1EE78}, {0x1EE7D, 0x1EE7D}, {0x1EE7F, 0x1EE7F}, {0x1EE8A, 0x1EE8A}, {0x1EE9C, 0x1EEA0}, {0x1EEA4, 0x1EEA4}, {0x1EEAA, 0x1EEAA}, {0x1EEBC, 0x1EEEF}, {0x1EEF2, 0x1EFFF},
{0x1F02C, 0x1F02F}, {0x1F094, 0x1F09F}, {0x1F0AF, 0x1F0B0}, {0x1F0C0, 0x1F0C0}, {0x1F0D0, 0x1F0D0}, {0x1F0F6, 0x1F0FF}, {0x1F1AE, 0x1F1E5}, {0x1F203, 0x1F20F}, {0x1F23C, 0x1F23F}, {0x1F249, 0x1F24F},
{0x1F252, 0x1F25F}, {0x1F266, 0x1F2FF}, {0x1F6D8, 0x1F6DF}, {0x1F6ED, 0x1F6EF}, {0x1F6FD, 0x1F6FF}, {0x1F774, 0x1F77F}, {0x1F7D9, 0x1F7DF}, {0x1F7EC, 0x1F7FF}, {0x1F80C, 0x1F80F}, {0x1F848, 0x1F84F},
{0x1F85A, 0x1F85F}, {0x1F888, 0x1F88F}, {0x1F8AE, 0x1F8AF}, {0x1F8B2, 0x1F8FF}, {0x1F979, 0x1F979}, {0x1F9CC, 0x1F9CC}, {0x1FA54, 0x1FA5F}, {0x1FA6E, 0x1FA6F}, {0x1FA75, 0x1FA77}, {0x1FA7B, 0x1FA7F},
{0x1FA87, 0x1FA8F}, {0x1FAA9, 0x1FAAF}, {0x1FAB7, 0x1FABF}, {0x1FAC3, 0x1FACF}, {0x1FAD7, 0x1FAFF}, {0x1FB93, 0x1FB93}, {0x1FBCB, 0x1FBEF}, {0x1FBFA, 0x1FFFF}, {0x2A6DE, 0x2A6FF}, {0x2B735, 0x2B73F},
{0x2B81E, 0x2B81F}, {0x2CEA2, 0x2CEAF}, {0x2EBE1, 0x2F7FF}, {0x2FA1E, 0x2FFFF}, {0x3134B, 0xE00FF}, {0xE01F0, 0x10FFFF},
};
static std::string codepoint_to_utf8(uint32_t cp) {
std::string result;
if (/* 0x00 <= cp && */ cp <= 0x7f) {
result.push_back(cp);
}
else if (0x80 <= cp && cp <= 0x7ff) {
result.push_back(0xc0 | ((cp >> 6) & 0x1f));
result.push_back(0x80 | (cp & 0x3f));
}
else if (0x800 <= cp && cp <= 0xffff) {
result.push_back(0xe0 | ((cp >> 12) & 0x0f));
result.push_back(0x80 | ((cp >> 6) & 0x3f));
result.push_back(0x80 | (cp & 0x3f));
}
else if (0x10000 <= cp && cp <= 0x10ffff) {
result.push_back(0xf0 | ((cp >> 18) & 0x07));
result.push_back(0x80 | ((cp >> 12) & 0x3f));
result.push_back(0x80 | ((cp >> 6) & 0x3f));
result.push_back(0x80 | (cp & 0x3f));
}
else {
throw std::invalid_argument("invalid codepoint");
}
return result;
}
static std::string codepoints_to_utf8(const std::vector<uint32_t> & cps) {
std::string result;
for (size_t i = 0; i < cps.size(); ++i) {
result.append(codepoint_to_utf8(cps[i]));
}
return result;
}
static uint32_t codepoint_from_utf8(const std::string & utf8, size_t & offset) {
assert(offset < utf8.size());
if (!(utf8[offset + 0] & 0x80)) {
auto result = utf8[offset + 0];
offset += 1;
return result;
}
else if (!(utf8[offset + 0] & 0x40)) {
throw std::invalid_argument("invalid character");
}
else if (!(utf8[offset + 0] & 0x20)) {
if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80))
throw std::invalid_argument("invalid character");
auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
offset += 2;
return result;
}
else if (!(utf8[offset + 0] & 0x10)) {
if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80))
throw std::invalid_argument("invalid character");
auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
offset += 3;
return result;
}
else if (!(utf8[offset + 0] & 0x08)) {
if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80))
throw std::invalid_argument("invalid character");
auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
offset += 4;
return result;
}
throw std::invalid_argument("invalid string");
}
static std::vector<uint32_t> codepoints_from_utf8(const std::string & utf8) {
std::vector<uint32_t> result;
size_t offset = 0;
while (offset < utf8.size()) {
result.push_back(codepoint_from_utf8(utf8, offset));
}
return result;
}
static std::vector<uint16_t> codepoint_to_utf16(uint32_t cp) {
std::vector<uint16_t> result;
if (/* 0x0000 <= cp && */ cp <= 0xffff) {
result.emplace_back(cp);
}
else if (0x10000 <= cp && cp <= 0x10ffff) {
result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
}
else {
throw std::invalid_argument("invalid codepoint");
}
return result;
}
static std::vector<uint16_t> codepoints_to_utf16(const std::vector<uint32_t> & cps) {
std::vector<uint16_t> result;
for (size_t i = 0; i < cps.size(); ++i) {
auto temp = codepoint_to_utf16(cps[i]);
result.insert(result.end(), temp.begin(), temp.end());
}
return result;
}
static uint32_t codepoint_from_utf16(const std::vector<uint16_t> & utf16, size_t & offset) {
assert(offset < utf16.size());
if (((utf16[0] >> 10) << 10) != 0xd800) {
auto result = utf16[offset + 0];
offset += 1;
return result;
}
else {
if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00))
throw std::invalid_argument("invalid character");
auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
offset += 2;
return result;
}
throw std::invalid_argument("invalid string");
}
static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> & utf16) {
std::vector<uint32_t> result;
size_t offset = 0;
while (offset < utf16.size())
result.push_back(codepoint_from_utf16(utf16, offset));
return result;
}
#define CODEPOINT_TYPE_UNIDENTIFIED 0
#define CODEPOINT_TYPE_DIGIT 1
#define CODEPOINT_TYPE_LETTER 2
#define CODEPOINT_TYPE_WHITESPACE 3
#define CODEPOINT_TYPE_ACCENT_MARK 4
#define CODEPOINT_TYPE_PUNCTUATION 5
#define CODEPOINT_TYPE_SYMBOL 6
#define CODEPOINT_TYPE_CONTROL 7
static std::unordered_map<uint32_t, int> codepoint_type_map() {
std::unordered_map<uint32_t, int> codepoint_types;
for (auto p : digit_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
codepoint_types[i] = CODEPOINT_TYPE_DIGIT;
}
for(auto p : letter_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
codepoint_types[i] = CODEPOINT_TYPE_LETTER;
}
for(auto p : whitespace_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
codepoint_types[i] = CODEPOINT_TYPE_WHITESPACE;
}
for(auto p : accent_mark_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
codepoint_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
}
for(auto p : punctuation_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
codepoint_types[i] = CODEPOINT_TYPE_PUNCTUATION;
}
for (auto p : symbol_ranges) {
for (auto i = p.first; i <= p.second; ++i)
codepoint_types[i] = CODEPOINT_TYPE_SYMBOL;
}
for(auto p : control_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
codepoint_types[i] = CODEPOINT_TYPE_CONTROL;
}
return codepoint_types;
}
static int codepoint_type(uint32_t cp) {
static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
return codepoint_types[cp];
}
static int codepoint_type(const std::string & utf8) {
if (utf8.length() == 0)
return CODEPOINT_TYPE_UNIDENTIFIED;
size_t offset = 0;
return codepoint_type(codepoint_from_utf8(utf8, offset));
}
static std::unordered_map<uint8_t, std::string> bytes_to_unicode_map_bpe() {
std::unordered_map<uint8_t, std::string> map;
for (int ch = u'!'; ch <= u'~'; ++ch) {
assert(0 <= ch && ch < 256);
map[ch] = codepoint_to_utf8(ch);
}
for (int ch = u'¡'; ch <= u'¬'; ++ch) {
assert(0 <= ch && ch < 256);
map[ch] = codepoint_to_utf8(ch);
}
for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
assert(0 <= ch && ch < 256);
map[ch] = codepoint_to_utf8(ch);
}
auto n = 0;
for (int ch = 0; ch < 256; ++ch) {
if (map.find(ch) == map.end()) {
map[ch] = codepoint_to_utf8(256 + n);
++n;
}
}
return map;
}
static std::string bytes_to_unicode_bpe(uint8_t byte) {
static std::unordered_map<uint8_t, std::string> map = bytes_to_unicode_map_bpe();
return map.at(byte);
}
static std::unordered_map<std::string, uint8_t> unicode_to_bytes_map_bpe() {
std::unordered_map<std::string, uint8_t> map;
for (int ch = u'!'; ch <= u'~'; ++ch) {
assert(0 <= ch && ch < 256);
map[codepoint_to_utf8(ch)] = ch;
}
for (int ch = u'¡'; ch <= u'¬'; ++ch) {
assert(0 <= ch && ch < 256);
map[codepoint_to_utf8(ch)] = ch;
}
for (int ch = u'®'; ch <= u'ÿ'; ++ch) {
assert(0 <= ch && ch < 256);
map[codepoint_to_utf8(ch)] = ch;
}
auto n = 0;
for (int ch = 0; ch < 256; ++ch) {
if (map.find(codepoint_to_utf8(ch)) == map.end()) {
map[codepoint_to_utf8(256 + n)] = ch;
++n;
}
}
return map;
}
static uint8_t unicode_to_bytes_bpe(const std::string & utf8) {
static std::unordered_map<std::string, uint8_t> map = unicode_to_bytes_map_bpe();
return map.at(utf8);
}