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

Merge branch 'master' into easier-vocab-conversion

Browse source
This commit is contained in:
20kdc 2024-05-04 15:47:32 +01:00
commit aa814faf61
65 changed files with 1520 additions and 1205 deletions
Download patch file Download diff file Expand all files Collapse all files

3
.flake8
View file

@ -1,3 +1,4 @@
[flake8]
max-line-length = 125
ignore = W503
ignore = E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503
exclude = examples/*,examples/*/**,*/**/__init__.py,scripts/gen-unicode-data.py,tests/test-tokenizer-0.py

3
.github/workflows/python-lint.yml vendored
View file

@ -20,5 +20,4 @@ jobs:
- name: flake8 Lint
uses: py-actions/flake8@v2
with:
ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503"
exclude: "examples/*,examples/*/**,*/**/__init__.py,convert-hf-to-gguf-update.py"
plugins: "flake8-no-print"

5
.pre-commit-config.yaml
View file

@ -3,13 +3,14 @@
exclude: prompts/.*.txt
repos:
- repo: https://github.com/pre-commit/pre-commit-hooks
rev: v3.2.0
rev: v4.6.0
hooks:
- id: trailing-whitespace
- id: end-of-file-fixer
- id: check-yaml
- id: check-added-large-files
- repo: https://github.com/PyCQA/flake8
rev: 6.0.0
rev: 7.0.0
hooks:
- id: flake8
additional_dependencies: [flake8-no-print]

3
Makefile
View file

@ -77,11 +77,10 @@ test: $(TEST_TARGETS)
./$$test_target $(CURDIR)/models/ggml-vocab-llama-bpe.gguf; \
./$$test_target $(CURDIR)/models/ggml-vocab-phi-3.gguf; \
./$$test_target $(CURDIR)/models/ggml-vocab-falcon.gguf; \
./$$test_target $(CURDIR)/models/ggml-vocab-deepseek-coder.gguf; \
./$$test_target $(CURDIR)/models/ggml-vocab-deepseek-llm.gguf; \
./$$test_target $(CURDIR)/models/ggml-vocab-bert-bge.gguf; \
./$$test_target $(CURDIR)/models/ggml-vocab-starcoder.gguf; \
./$$test_target $(CURDIR)/models/ggml-vocab-gpt-2.gguf; \
./$$test_target $(CURDIR)/models/ggml-vocab-refact.gguf; \
elif [ "$$test_target" = "tests/test-tokenizer-1-spm" ]; then \
continue; \
elif [ "$$test_target" = "tests/test-tokenizer-1-bpe" ]; then \

2
common/common.cpp
View file

@ -76,7 +76,7 @@ int32_t get_num_physical_cores() {
// enumerate the set of thread siblings, num entries is num cores
std::unordered_set<std::string> siblings;
for (uint32_t cpu=0; cpu < UINT32_MAX; ++cpu) {
std::ifstream thread_siblings("/sys/devices/system/cpu"
std::ifstream thread_siblings("/sys/devices/system/cpu/cpu"
+ std::to_string(cpu) + "/topology/thread_siblings");
if (!thread_siblings.is_open()) {
break; // no more cpus

4
common/log.h
View file

@ -234,7 +234,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
// INTERNAL, DO NOT USE
// USE LOG() INSTEAD
//
#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER)
#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
#define LOG_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \
@ -257,7 +257,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
// INTERNAL, DO NOT USE
// USE LOG_TEE() INSTEAD
//
#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER)
#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
#define LOG_TEE_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \

157
convert-hf-to-gguf-update.py
View file

@ -21,6 +21,7 @@
# TODO: automate the update of convert-hf-to-gguf.py
#
import logging
import os
import requests
import sys
@ -28,12 +29,18 @@ import json
from hashlib import sha256
from enum import IntEnum, auto
from transformers import AutoTokenizer
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger("convert-hf-to-gguf-update")
class TOKENIZER_TYPE(IntEnum):
SPM = auto()
BPE = auto()
WPM = auto()
# TODO: this string has to exercise as much pre-tokenizer functionality as possible
# will be updated with time - contributions welcome
chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
@ -41,36 +48,39 @@ chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍
if len(sys.argv) == 2:
token = sys.argv[1]
else:
print("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
sys.exit(1)
# TODO: add models here, base models preferred
models = [
{ "name": "llama-spm", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", },
{ "name": "llama-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Meta-Llama-3-8B", },
{ "name": "phi-3", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", },
{ "name": "deepseek-llm", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-llm-7b-base", },
{ "name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
{ "name": "falcon", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
{ "name": "bert-bge", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
{ "name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
{ "name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
{ "name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
]
{"name": "llama-spm", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", },
{"name": "llama-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Meta-Llama-3-8B", },
{"name": "phi-3", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", },
{"name": "deepseek-llm", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-llm-7b-base", },
{"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
{"name": "falcon", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
{"name": "bert-bge", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
{"name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
{"name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
{"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
{"name": "refact", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
]
# make directory "models/tokenizers" if it doesn't exist
if not os.path.exists("models/tokenizers"):
os.makedirs("models/tokenizers")
def download_file_with_auth(url, token, save_path):
headers = {"Authorization": f"Bearer {token}"}
response = requests.get(url, headers=headers)
if response.status_code == 200:
with open(save_path, 'wb') as f:
f.write(response.content)
print(f"File {save_path} downloaded successfully")
logger.info(f"File {save_path} downloaded successfully")
else:
print(f"Failed to download file. Status code: {response.status_code}")
logger.info(f"Failed to download file. Status code: {response.status_code}")
# download the tokenizer models
for model in models:
@ -81,10 +91,10 @@ for model in models:
if not os.path.exists(f"models/tokenizers/{name}"):
os.makedirs(f"models/tokenizers/{name}")
else:
print(f"Directory models/tokenizers/{name} already exists - skipping")
logger.info(f"Directory models/tokenizers/{name} already exists - skipping")
continue
print(f"Downloading {name} to models/tokenizers/{name}")
logger.info(f"Downloading {name} to models/tokenizers/{name}")
url = f"{repo}/raw/main/config.json"
save_path = f"models/tokenizers/{name}/config.json"
@ -115,80 +125,82 @@ for model in models:
continue
# create the tokenizer
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
chktok = tokenizer.encode(chktxt)
chkhsh = sha256(str(chktok).encode()).hexdigest()
print(f"model: {name}")
print(f"tokt: {tokt}")
print(f"repo: {model['repo']}")
print(f"chktok: {chktok}")
print(f"chkhsh: {chkhsh}")
logger.info(f"model: {name}")
logger.info(f"tokt: {tokt}")
logger.info(f"repo: {model['repo']}")
logger.info(f"chktok: {chktok}")
logger.info(f"chkhsh: {chkhsh}")
# print the "pre_tokenizer" content from the tokenizer.json
with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f:
cfg = json.load(f)
pre_tokenizer = cfg["pre_tokenizer"]
print("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
print(f"\n")
logger.info("")
src_ifs += f" if chkhsh == \"{chkhsh}\":\n"
src_ifs += f" # ref: {model['repo']}\n"
src_ifs += f" res = \"{name}\"\n"
src_func = ""
src_func += " def get_vocab_base_pre(self, tokenizer) -> str:\n"
src_func += " # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that\n"
src_func += " # is specific for the BPE pre-tokenizer used by the model\n"
src_func += " # we will use this unique identifier to write a \"tokenizer.ggml.pre\" entry in the GGUF file which we can\n"
src_func += " # use in llama.cpp to implement the same pre-tokenizer\n"
src_func += "\n"
src_func += f" chktxt = {repr(chktxt)}\n"
src_func += "\n"
src_func += " chktok = tokenizer.encode(chktxt)\n"
src_func += " chkhsh = sha256(str(chktok).encode()).hexdigest()\n"
src_func += "\n"
src_func += " print(f\"chktok: {chktok}\")\n"
src_func += " print(f\"chkhsh: {chkhsh}\")\n"
src_func += "\n"
src_func += " res = None\n"
src_func += "\n"
src_func += " # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script\n"
src_func += " # or pull the latest version of the model from Huggingface\n"
src_func += " # don't edit the hashes manually!\n"
src_func += f"{src_ifs}\n"
src_func += " if res is None:\n"
src_func += " print(\"\\n\")\n"
src_func += " print(\"**************************************************************************************\")\n"
src_func += " print(\"** WARNING: The BPE pre-tokenizer was not recognized!\")\n"
src_func += " print(\"** There are 2 possible reasons for this:\")\n"
src_func += " print(\"** - the model has not been added to convert-hf-to-gguf-update.py yet\")\n"
src_func += " print(\"** - the pre-tokenization config has changed upstream\")\n"
src_func += " print(\"** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.\")\n"
src_func += " print(\"** ref: https://github.com/ggerganov/llama.cpp/pull/6920\")\n"
src_func += " print(\"**\")\n"
src_func += " print(f\"** chkhsh: {chkhsh}\")\n"
src_func += " print(\"**************************************************************************************\")\n"
src_func += " print(\"\\n\")\n"
src_func += " raise NotImplementedError(\"BPE pre-tokenizer was not recognized - update get_vocab_base_pre()\")\n"
src_func += "\n"
src_func += " print(f\"tokenizer.ggml.pre: {res}\")\n"
src_func += " print(f\"chkhsh: {chkhsh}\")\n"
src_func += "\n"
src_func += " return res\n"
src_func = f"""
def get_vocab_base_pre(self, tokenizer) -> str:
# encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
# is specific for the BPE pre-tokenizer used by the model
# we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can
# use in llama.cpp to implement the same pre-tokenizer
print(src_func)
chktxt = {repr(chktxt)}
print("\n")
print("!!! Copy-paste the function above into convert-hf-to-gguf.py !!!")
print("\n")
chktok = tokenizer.encode(chktxt)
chkhsh = sha256(str(chktok).encode()).hexdigest()
logger.debug(f"chktok: {{chktok}}")
logger.debug(f"chkhsh: {{chkhsh}}")
res = None
# NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script
# or pull the latest version of the model from Huggingface
# don't edit the hashes manually!
{src_ifs}
if res is None:
logger.warning("\\n")
logger.warning("**************************************************************************************")
logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!")
logger.warning("** There are 2 possible reasons for this:")
logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet")
logger.warning("** - the pre-tokenization config has changed upstream")
logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.")
logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920")
logger.warning("**")
logger.warning(f"** chkhsh: {{chkhsh}}")
logger.warning("**************************************************************************************")
logger.warning("\\n")
raise NotImplementedError("BPE pre-tokenizer was not recognized - update get_vocab_base_pre()")
logger.debug(f"tokenizer.ggml.pre: {{repr(res)}}")
logger.debug(f"chkhsh: {{chkhsh}}")
return res
"""
print(src_func) # noqa: NP100
logger.info("\n")
logger.info("!!! Copy-paste the function above into convert-hf-to-gguf.py !!!")
logger.info("\n")
# generate tests for each tokenizer model
tests = [
"ied 4 ½ months",
"Führer",
"",
" ",
" ",
@ -250,7 +262,6 @@ for model in models:
tokt = model["tokt"]
# create the tokenizer
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
@ -265,15 +276,15 @@ for model in models:
f.write(f" {r}")
f.write("\n")
print(f"Tests for {name} written in ./models/ggml-vocab-{name}.gguf.*")
logger.info(f"Tests for {name} written in ./models/ggml-vocab-{name}.gguf.*")
# generate commands for creating vocab files
print("\nRun the following commands to generate the vocab files for testing:\n")
logger.info("\nRun the following commands to generate the vocab files for testing:\n")
for model in models:
name = model["name"]
print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only")
print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100
print("\n")
logger.info("\n")

276
convert-hf-to-gguf.py
View file

@ -2,6 +2,7 @@
from __future__ import annotations
import logging
import argparse
import contextlib
import json
@ -26,6 +27,8 @@ import gguf
from convert import LlamaHfVocab, permute
logger = logging.getLogger("hf-to-gguf")
###### MODEL DEFINITIONS ######
@ -76,7 +79,7 @@ class Model(ABC):
def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
for part_name in self.part_names:
print(f"gguf: loading model part '{part_name}'")
logger.info(f"gguf: loading model part '{part_name}'")
ctx: ContextManager[Any]
if self.is_safetensors:
from safetensors import safe_open
@ -95,42 +98,42 @@ class Model(ABC):
if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
self.gguf_writer.add_context_length(n_ctx)
print(f"gguf: context length = {n_ctx}")
logger.info(f"gguf: context length = {n_ctx}")
n_embd = self.find_hparam(["hidden_size", "n_embd"])
self.gguf_writer.add_embedding_length(n_embd)
print(f"gguf: embedding length = {n_embd}")
logger.info(f"gguf: embedding length = {n_embd}")
if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
self.gguf_writer.add_feed_forward_length(n_ff)
print(f"gguf: feed forward length = {n_ff}")
logger.info(f"gguf: feed forward length = {n_ff}")
n_head = self.find_hparam(["num_attention_heads", "n_head"])
self.gguf_writer.add_head_count(n_head)
print(f"gguf: head count = {n_head}")
logger.info(f"gguf: head count = {n_head}")
if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
self.gguf_writer.add_head_count_kv(n_head_kv)
print(f"gguf: key-value head count = {n_head_kv}")
logger.info(f"gguf: key-value head count = {n_head_kv}")
if (rope_theta := self.hparams.get("rope_theta")) is not None:
self.gguf_writer.add_rope_freq_base(rope_theta)
print(f"gguf: rope theta = {rope_theta}")
logger.info(f"gguf: rope theta = {rope_theta}")
if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
print(f"gguf: rms norm epsilon = {f_rms_eps}")
logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
self.gguf_writer.add_layer_norm_eps(f_norm_eps)
print(f"gguf: layer norm epsilon = {f_norm_eps}")
logger.info(f"gguf: layer norm epsilon = {f_norm_eps}")
if (n_experts := self.hparams.get("num_local_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
print(f"gguf: expert count = {n_experts}")
logger.info(f"gguf: expert count = {n_experts}")
if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
self.gguf_writer.add_expert_used_count(n_experts_used)
print(f"gguf: experts used count = {n_experts_used}")
logger.info(f"gguf: experts used count = {n_experts_used}")
self.gguf_writer.add_file_type(self.ftype)
print(f"gguf: file type = {self.ftype}")
logger.info(f"gguf: file type = {self.ftype}")
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
@ -151,8 +154,7 @@ class Model(ABC):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -169,7 +171,7 @@ class Model(ABC):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -274,8 +276,8 @@ class Model(ABC):
chktok = tokenizer.encode(chktxt)
chkhsh = sha256(str(chktok).encode()).hexdigest()
print(f"chktok: {chktok}")
print(f"chkhsh: {chkhsh}")
logger.debug(f"chktok: {chktok}")
logger.debug(f"chkhsh: {chkhsh}")
res = None
@ -306,24 +308,27 @@ class Model(ABC):
if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454":
# ref: https://huggingface.co/openai-community/gpt2
res = "gpt-2"
if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff":
# ref: https://huggingface.co/smallcloudai/Refact-1_6-base
res = "refact"
if res is None:
print("\n")
print("**************************************************************************************")
print("** WARNING: The BPE pre-tokenizer was not recognized!")
print("** There are 2 possible reasons for this:")
print("** - the model has not been added to convert-hf-to-gguf-update.py yet")
print("** - the pre-tokenization config has changed upstream")
print("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.")
print("** ref: https://github.com/ggerganov/llama.cpp/pull/6920")
print("**")
print(f"** chkhsh: {chkhsh}")
print("**************************************************************************************")
print("\n")
logger.warning("\n")
logger.warning("**************************************************************************************")
logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!")
logger.warning("** There are 2 possible reasons for this:")
logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet")
logger.warning("** - the pre-tokenization config has changed upstream")
logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.")
logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920")
logger.warning("**")
logger.warning(f"** chkhsh: {chkhsh}")
logger.warning("**************************************************************************************")
logger.warning("\n")
raise NotImplementedError("BPE pre-tokenizer was not recognized - update get_vocab_base_pre()")
print(f"tokenizer.ggml.pre: {res}")
print(f"chkhsh: {chkhsh}")
logger.debug(f"tokenizer.ggml.pre: {repr(res)}")
logger.debug(f"chkhsh: {chkhsh}")
return res
@ -439,9 +444,7 @@ class Model(ABC):
if vocab_size > len(tokens):
pad_count = vocab_size - len(tokens)
print(
f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]"
)
logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
for i in range(1, pad_count + 1):
tokens.append(f"[PAD{i}]")
scores.append(-1000.0)
@ -553,7 +556,7 @@ class BloomModel(Model):
),
axis=0,
)
print("re-format attention.linear_qkv.weight")
logger.info("re-format attention.linear_qkv.weight")
elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
data = np.concatenate(
@ -564,13 +567,12 @@ class BloomModel(Model):
),
axis=0,
)
print("re-format attention.linear_qkv.bias")
logger.info("re-format attention.linear_qkv.bias")
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -587,13 +589,13 @@ class BloomModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
logger.info(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
if not has_lm_head and name == "word_embeddings.weight":
self.gguf_writer.add_tensor("output.weight", data)
print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
logger.info(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
@Model.register("MPTForCausalLM")
@ -653,8 +655,7 @@ class MPTModel(Model):
else:
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -671,7 +672,7 @@ class MPTModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -697,8 +698,7 @@ class OrionModel(Model):
elif "model_max_length" in self.hparams:
ctx_length = self.hparams["model_max_length"]
else:
print("gguf: can not find ctx length parameter.")
sys.exit()
raise ValueError("gguf: can not find ctx length parameter.")
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_name(self.dir_model.name)
@ -736,8 +736,7 @@ class OrionModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -754,7 +753,7 @@ class OrionModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -779,8 +778,7 @@ class BaichuanModel(Model):
elif "model_max_length" in self.hparams:
ctx_length = self.hparams["model_max_length"]
else:
print("gguf: can not find ctx length parameter.")
sys.exit()
raise ValueError("gguf: can not find ctx length parameter.")
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_source_hf_repo(hf_repo)
@ -809,7 +807,7 @@ class BaichuanModel(Model):
for i in range(block_count):
if (w := model_kv.get(f"model.layers.{i}.self_attn.W_pack.weight")) is not None:
print(f"Unpacking and permuting layer {i}")
logger.info(f"Unpacking and permuting layer {i}")
model_kv[f"model.layers.{i}.self_attn.q_proj.weight"] = \
self._reverse_hf_permute_part(w, 0, head_count, head_count)
model_kv[f"model.layers.{i}.self_attn.k_proj.weight"] = \
@ -834,8 +832,7 @@ class BaichuanModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -852,7 +849,7 @@ class BaichuanModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
@ -937,8 +934,7 @@ class XverseModel(Model):
elif "model_max_length" in self.hparams:
ctx_length = self.hparams["model_max_length"]
else:
print("gguf: can not find ctx length parameter.")
sys.exit()
raise ValueError("gguf: can not find ctx length parameter.")
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_source_hf_repo(hf_repo)
@ -987,8 +983,7 @@ class XverseModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1005,7 +1000,7 @@ class XverseModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
@ -1092,8 +1087,7 @@ class FalconModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1110,7 +1104,7 @@ class FalconModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1197,8 +1191,7 @@ class RefactModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1215,7 +1208,7 @@ class RefactModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1264,10 +1257,9 @@ class PersimmonModel(Model):
data = data_torch.to(torch.float32).squeeze().numpy()
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1332,8 +1324,7 @@ class StableLMModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1350,7 +1341,7 @@ class StableLMModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and not new_name.endswith("_norm.weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.debug(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1366,8 +1357,7 @@ class StableLMModel(Model):
merged_name = f"model.layers.{bid}.self_attn.{layer_name}.weight"
new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
data = data.astype(np.float32)
@ -1375,7 +1365,7 @@ class StableLMModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and not new_name.endswith("_norm.weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
logger.debug(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1427,7 +1417,7 @@ class LlamaModel(Model):
experts = dict()
for name, data_torch in self.get_tensors():
# we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
continue
old_dtype = data_torch.dtype
@ -1480,10 +1470,9 @@ class LlamaModel(Model):
new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
continue
@ -1491,8 +1480,7 @@ class LlamaModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1509,7 +1497,7 @@ class LlamaModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1584,10 +1572,9 @@ class GrokModel(Model):
new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
continue
@ -1595,8 +1582,7 @@ class GrokModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1613,7 +1599,7 @@ class GrokModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1646,7 +1632,7 @@ class DbrxModel(Model):
self.gguf_writer.add_layer_norm_eps(1e-5)
self.gguf_writer.add_file_type(self.ftype)
print(f"gguf: file type = {self.ftype}")
logger.info(f"gguf: file type = {self.ftype}")
def write_tensors(self):
block_count = self.hparams.get("n_layers")
@ -1689,8 +1675,7 @@ class DbrxModel(Model):
# https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15
new_name = tensor_map.get_name(name if not experts else name + ".weight", try_suffixes=(".weight",))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1698,8 +1683,7 @@ class DbrxModel(Model):
# Most of the codebase that takes in 1D tensors only handles F32 tensors
# and most of the outputs tensors are F32.
if data_dtype != np.float32 and n_dims == 1:
print(f"Can not map tensor {name!r}: all 1D tensors must be F32")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}: all 1D tensors must be F32")
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
@ -1709,7 +1693,7 @@ class DbrxModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and n_dims > 1:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
logger.debug(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1771,8 +1755,7 @@ class MiniCPMModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1789,7 +1772,7 @@ class MiniCPMModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1855,8 +1838,7 @@ class QwenModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1873,7 +1855,7 @@ class QwenModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -1950,10 +1932,9 @@ class Qwen2MoeModel(Model):
new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
logger.debug(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
continue
@ -1961,8 +1942,7 @@ class Qwen2MoeModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -1979,7 +1959,7 @@ class Qwen2MoeModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
logger.debug(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -2024,8 +2004,7 @@ class GPT2Model(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -2042,13 +2021,13 @@ class GPT2Model(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
# note: GPT2 output is tied to (same as) wte in original model
if new_name == "token_embd.weight":
print(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor("output.weight", data)
@ -2087,8 +2066,7 @@ class Phi3MiniModel(Model):
tokenizer_path = self.dir_model / 'tokenizer.model'
if not tokenizer_path.is_file():
print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
sys.exit(1)
raise ValueError(f'Error: Missing {tokenizer_path}')
tokenizer = SentencePieceProcessor(str(tokenizer_path))
@ -2126,7 +2104,7 @@ class Phi3MiniModel(Model):
for key in added_tokens_json:
token_id = added_tokens_json[key]
if (token_id >= vocab_size):
print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
continue
tokens[token_id] = key.encode("utf-8")
@ -2208,8 +2186,7 @@ class PlamoModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
# shuffle for broadcasting of gqa in ggml_mul_mat
if new_name.endswith("attn_q.weight"):
@ -2240,7 +2217,7 @@ class PlamoModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -2286,8 +2263,7 @@ class CodeShellModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -2304,13 +2280,13 @@ class CodeShellModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
if not has_lm_head and name == "transformer.wte.weight":
self.gguf_writer.add_tensor("output.weight", data)
print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
logger.info(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
@Model.register("InternLM2ForCausalLM")
@ -2332,7 +2308,7 @@ class InternLM2Model(Model):
toktypes: list[int] = []
if not tokenizer_path.is_file():
print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
logger.error(f'Error: Missing {tokenizer_path}')
sys.exit(1)
sentencepiece_model = model.ModelProto()
@ -2349,7 +2325,7 @@ class InternLM2Model(Model):
if text == b"\x00":
# (TODO): fixme
# Hack here and replace the \x00 characters.
print(f"InternLM2 convert token '{text}' to '🐉'!")
logger.debug(f"InternLM2 convert token '{text}' to '🐉'!")
text = "🐉"
toktype = SentencePieceTokenTypes.NORMAL
@ -2390,7 +2366,7 @@ class InternLM2Model(Model):
# TODO: this is a hack, should be fixed
# https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048
special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer)
print(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \
logger.warning(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \
in chat mode so that the conversation can end normally.")
special_vocab.add_to_gguf(self.gguf_writer)
@ -2435,8 +2411,7 @@ in chat mode so that the conversation can end normally.")
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -2453,7 +2428,7 @@ in chat mode so that the conversation can end normally.")
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
def write_tensors(self):
@ -2564,8 +2539,7 @@ class BertModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
@ -2585,7 +2559,7 @@ class BertModel(Model):
# if f32 desired, convert any float16 to float32
new_dtype = np.float32
print(f"{new_name}, n_dims = {n_dims}, {data_torch.dtype} --> {new_dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {data_torch.dtype} --> {new_dtype}")
if data.dtype != new_dtype:
data = data.astype(new_dtype)
@ -2664,7 +2638,7 @@ class GemmaModel(Model):
# lm_head is not used in llama.cpp, while autoawq will include this tensor in model
# To prevent errors, skip loading lm_head.weight.
if name == "lm_head.weight":
print(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
continue
old_dtype = data_torch.dtype
@ -2681,8 +2655,7 @@ class GemmaModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -2693,7 +2666,7 @@ class GemmaModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -2721,7 +2694,7 @@ class MambaModel(Model):
else:
# Use the GPT-NeoX tokenizer when no tokenizer files are present
tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf"
print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
neox_reader = gguf.GGUFReader(tokenizer_path, "r")
field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
@ -2793,17 +2766,16 @@ class MambaModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
if name.endswith(".A_log"):
print("A_log --> A ==> " + new_name)
logger.debug("A_log --> A ==> " + new_name)
data_torch = -torch.exp(data_torch)
# assuming token_embd.weight is seen before output.weight
if tok_embd is not None and new_name == output_name:
if torch.equal(tok_embd, data_torch):
print(f"{output_name} is equivalent to {tok_embd_name}, omitting")
logger.debug(f"{output_name} is equivalent to {tok_embd_name}, omitting")
continue
if new_name == tok_embd_name:
tok_embd = data_torch
@ -2826,7 +2798,7 @@ class MambaModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and new_weight_name.endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -2885,8 +2857,7 @@ class OlmoModel(Model):
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
raise ValueError(f"Can not map tensor {name!r}")
n_dims = len(data.shape)
data_dtype = data.dtype
@ -2903,7 +2874,7 @@ class OlmoModel(Model):
if self.ftype == 1 and data_dtype == np.float32 and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@ -2936,6 +2907,7 @@ def parse_args() -> argparse.Namespace:
)
parser.add_argument("--use-temp-file", action="store_true", help="use the tempfile library while processing (helpful when running out of memory, process killed)")
parser.add_argument("--model-name", type=str, default=None, help="name of the model")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
return parser.parse_args()
@ -2943,6 +2915,8 @@ def parse_args() -> argparse.Namespace:
def main() -> None:
args = parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
dir_model = args.model
if args.awq_path:
@ -2951,15 +2925,15 @@ def main() -> None:
tmp_model_path = args.model / "weighted_model"
dir_model = tmp_model_path
if tmp_model_path.is_dir():
print(f"{tmp_model_path} exists as a weighted model.")
logger.info(f"{tmp_model_path} exists as a weighted model.")
else:
tmp_model_path.mkdir(parents=True, exist_ok=True)
print("Saving new weighted model ...")
logger.info("Saving new weighted model ...")
add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
print(f"Saved weighted model at {tmp_model_path}.")
logger.info(f"Saved weighted model at {tmp_model_path}.")
if not dir_model.is_dir():
print(f'Error: {args.model} is not a directory', file=sys.stderr)
logger.error(f'Error: {args.model} is not a directory')
sys.exit(1)
ftype_map = {
@ -2973,7 +2947,7 @@ def main() -> None:
# output in the same directory as the model by default
fname_out = dir_model / f'ggml-model-{args.outtype}.gguf'
print(f"Loading model: {dir_model.name}")
logger.info(f"Loading model: {dir_model.name}")
hparams = Model.load_hparams(dir_model)
@ -2981,20 +2955,20 @@ def main() -> None:
model_class = Model.from_model_architecture(hparams["architectures"][0])
model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file)
print("Set model parameters")
logger.info("Set model parameters")
model_instance.set_gguf_parameters()
print("Set model tokenizer")
logger.info("Set model tokenizer")
model_instance.set_vocab()
if args.vocab_only:
print(f"Exporting model vocab to '{fname_out}'")
logger.info(f"Exporting model vocab to '{fname_out}'")
model_instance.write_vocab()
else:
print(f"Exporting model to '{fname_out}'")
logger.info(f"Exporting model to '{fname_out}'")
model_instance.write()
print(f"Model successfully exported to '{fname_out}'")
logger.info(f"Model successfully exported to '{fname_out}'")
if __name__ == '__main__':

51
convert-llama-ggml-to-gguf.py
View file

@ -1,6 +1,7 @@
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import os
import struct
@ -14,6 +15,8 @@ if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
import gguf
logger = logging.getLogger("ggml-to-gguf")
class GGMLFormat(IntEnum):
GGML = 0
@ -125,7 +128,6 @@ class Tensor:
self.start_offset = offset
self.len_bytes = n_bytes
offset += n_bytes
# print(n_dims, name_len, dtype, self.dims, self.name, pad)
return offset - orig_offset
@ -175,7 +177,7 @@ class GGMLModel:
offset += self.validate_header(data, offset)
hp = Hyperparameters()
offset += hp.load(data, offset)
print(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}')
logger.info(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}')
self.validate_conversion(hp.ftype)
vocab = Vocab(load_scores = self.file_format > GGMLFormat.GGML)
offset += vocab.load(data, offset, hp.n_vocab)
@ -215,12 +217,12 @@ class GGMLToGGUF:
if float(hp.n_head) / float(x) == gqa:
n_kv_head = x
assert n_kv_head is not None, "Couldn't determine n_kv_head from GQA param"
print(f'- Guessed n_kv_head = {n_kv_head} based on GQA {cfg.gqa}')
logger.info(f'- Guessed n_kv_head = {n_kv_head} based on GQA {cfg.gqa}')
self.n_kv_head = n_kv_head
self.name_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.LLAMA, ggml_model.hyperparameters.n_layer)
def save(self):
print('* Preparing to save GGUF file')
logger.info('* Preparing to save GGUF file')
gguf_writer = gguf.GGUFWriter(
self.cfg.output,
gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
@ -230,11 +232,11 @@ class GGMLToGGUF:
if self.special_vocab is not None:
self.special_vocab.add_to_gguf(gguf_writer)
self.add_tensors(gguf_writer)
print(" gguf: write header")
logger.info(" gguf: write header")
gguf_writer.write_header_to_file()
print(" gguf: write metadata")
logger.info(" gguf: write metadata")
gguf_writer.write_kv_data_to_file()
print(" gguf: write tensors")
logger.info(" gguf: write tensors")
gguf_writer.write_tensors_to_file()
gguf_writer.close()
@ -250,7 +252,7 @@ class GGMLToGGUF:
name = cfg.name if cfg.name is not None else cfg.input.name
except UnicodeDecodeError:
name = None
print('* Adding model parameters and KV items')
logger.info('* Adding model parameters and KV items')
if name is not None:
gguf_writer.add_name(name)
gguf_writer.add_description(desc)
@ -287,7 +289,7 @@ class GGMLToGGUF:
toktypes = []
if self.vocab_override is not None:
vo = self.vocab_override
print('* Adding vocab item(s)')
logger.info('* Adding vocab item(s)')
for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
tokens.append(vbytes)
scores.append(score)
@ -299,7 +301,7 @@ class GGMLToGGUF:
if len(toktypes) > 0:
gguf_writer.add_token_types(toktypes)
return
print(f'* Adding {hp.n_vocab} vocab item(s)')
logger.info(f'* Adding {hp.n_vocab} vocab item(s)')
assert len(self.model.vocab.items) >= 3, 'Cannot handle unexpectedly short model vocab'
for (tokid, (vbytes, vscore)) in enumerate(self.model.vocab.items):
tt = 1 # Normal
@ -334,7 +336,7 @@ class GGMLToGGUF:
def add_tensors(self, gguf_writer):
tensor_map = self.name_map
data = self.data
print(f'* Adding {len(self.model.tensors)} tensor(s)')
logger.info(f'* Adding {len(self.model.tensors)} tensor(s)')
for tensor in self.model.tensors:
name = str(tensor.name, 'UTF-8')
mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
@ -344,7 +346,6 @@ class GGMLToGGUF:
temp = tempdims[1]
tempdims[1] = tempdims[0]
tempdims[0] = temp
# print(f'+ {tensor.name} | {mapped_name} {tensor.dims} :: {tempdims}')
gguf_writer.add_tensor(
mapped_name,
data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
@ -401,33 +402,35 @@ def handle_args():
help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
parser.add_argument("--vocabtype", default="spm,hfft",
help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
return parser.parse_args()
def main():
cfg = handle_args()
print(f'* Using config: {cfg}')
print('\n=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===\n')
logging.basicConfig(level=logging.DEBUG if cfg.verbose else logging.INFO)
logger.info(f'* Using config: {cfg}')
logger.warning('=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===')
if cfg.model_metadata_dir is None and (cfg.gqa == 1 or cfg.eps == '5.0e-06'):
print('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".')
logger.info('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".')
data = np.memmap(cfg.input, mode = 'r')
model = GGMLModel()
print('* Scanning GGML input file')
logger.info('* Scanning GGML input file')
offset = model.load(data, 0) # noqa
print(f'* GGML model hyperparameters: {model.hyperparameters}')
logger.info(f'* GGML model hyperparameters: {model.hyperparameters}')
vocab_override = None
params_override = None
special_vocab = None
if cfg.model_metadata_dir is not None:
(params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters)
print('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
print(f'* Overriding params: {params_override}')
print(f'* Overriding vocab: {vocab_override}')
print(f'* Special vocab: {special_vocab}')
logger.info('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
logger.info(f'* Overriding params: {params_override}')
logger.info(f'* Overriding vocab: {vocab_override}')
logger.info(f'* Special vocab: {special_vocab}')
else:
print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
logger.warning('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
if model.file_format == GGMLFormat.GGML:
print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
logger.info('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
converter = GGMLToGGUF(
model, data, cfg,
params_override = params_override,
@ -435,7 +438,7 @@ def main():
special_vocab = special_vocab
)
converter.save()
print(f'* Successful completion. Output saved to: {cfg.output}')
logger.info(f'* Successful completion. Output saved to: {cfg.output}')
if __name__ == '__main__':

31
convert-lora-to-ggml.py
View file

@ -1,6 +1,7 @@
#!/usr/bin/env python3
from __future__ import annotations
import logging
import json
import os
import struct
@ -15,6 +16,8 @@ if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
import gguf
logger = logging.getLogger("lora-to-gguf")
NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
@ -48,11 +51,9 @@ def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_ty
if __name__ == '__main__':
if len(sys.argv) < 2:
print(f"Usage: python {sys.argv[0]} <path> [arch]")
print(
"Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
)
print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
logger.info(f"Usage: python {sys.argv[0]} <path> [arch]")
logger.info("Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'")
logger.info(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
sys.exit(1)
input_json = os.path.join(sys.argv[1], "adapter_config.json")
@ -70,7 +71,7 @@ if __name__ == '__main__':
arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
if arch_name not in gguf.MODEL_ARCH_NAMES.values():
print(f"Error: unsupported architecture {arch_name}")
logger.error(f"Error: unsupported architecture {arch_name}")
sys.exit(1)
arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
@ -80,21 +81,21 @@ if __name__ == '__main__':
params = json.load(f)
if params["peft_type"] != "LORA":
print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
logger.error(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
sys.exit(1)
if params["fan_in_fan_out"] is True:
print("Error: param fan_in_fan_out is not supported")
logger.error("Error: param fan_in_fan_out is not supported")
sys.exit(1)
if params["bias"] is not None and params["bias"] != "none":
print("Error: param bias is not supported")
logger.error("Error: param bias is not supported")
sys.exit(1)
# TODO: these seem to be layers that have been trained but without lora.
# doesn't seem widely used but eventually should be supported
if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
print("Error: param modules_to_save is not supported")
logger.error("Error: param modules_to_save is not supported")
sys.exit(1)
with open(output_path, "wb") as fout:
@ -125,13 +126,13 @@ if __name__ == '__main__':
suffix = k[-len(lora_suffixes[0]):]
k = k[: -len(lora_suffixes[0])]
else:
print(f"Error: unrecognized tensor name {orig_k}")
logger.error(f"Error: unrecognized tensor name {orig_k}")
sys.exit(1)
tname = name_map.get_name(k)
if tname is None:
print(f"Error: could not map tensor name {orig_k}")
print(" Note: the arch parameter must be specified if the model is not llama")
logger.error(f"Error: could not map tensor name {orig_k}")
logger.error(" Note: the arch parameter must be specified if the model is not llama")
sys.exit(1)
if suffix == ".lora_A.weight":
@ -141,8 +142,8 @@ if __name__ == '__main__':
else:
assert False
print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
logger.info(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
write_tensor_header(fout, tname, t.shape, t.dtype)
t.tofile(fout)
print(f"Converted {input_json} and {input_model} to {output_path}")
logger.info(f"Converted {input_json} and {input_model} to {output_path}")

26
convert-persimmon-to-gguf.py
View file

@ -1,6 +1,7 @@
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import os
import sys
@ -14,6 +15,8 @@ if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
import gguf
logger = logging.getLogger("persimmon-to-gguf")
def _flatten_dict(dct, tensors, prefix=None):
assert isinstance(dct, dict)
@ -30,9 +33,9 @@ def _flatten_dict(dct, tensors, prefix=None):
def _get_sentencepiece_tokenizer_info(dir_model: Path):
tokenizer_path = dir_model / 'adept_vocab.model'
print('gguf: getting sentencepiece tokenizer from', tokenizer_path)
logger.info('getting sentencepiece tokenizer from', tokenizer_path)
tokenizer = SentencePieceProcessor(str(tokenizer_path))
print('gguf: adding tokens')
logger.info('adding tokens')
tokens: list[bytes] = []
scores: list[float] = []
toktypes: list[int] = []
@ -68,7 +71,9 @@ def main():
parser.add_argument("--ckpt-path", type=Path, help="path to persimmon checkpoint .pt file")
parser.add_argument("--model-dir", type=Path, help="directory containing model e.g. 8b_chat_model_release")
parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
sys.path.append(str(args.adept_inference_dir))
persimmon_model = torch.load(args.ckpt_path)
hparams = persimmon_model['args']
@ -107,7 +112,7 @@ def main():
gguf_writer.add_eos_token_id(71013)
tensor_map = gguf.get_tensor_name_map(arch, block_count)
print(tensor_map)
logger.info(tensor_map)
for name in tensors.keys():
data_torch = tensors[name]
if name.endswith(".self_attention.rotary_emb.inv_freq"):
@ -117,22 +122,21 @@ def main():
data = data_torch.to(torch.float32).squeeze().numpy()
new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
if new_name is None:
print("Can not map tensor '" + name + "'")
sys.exit()
raise ValueError(f"Can not map tensor '{name}'")
n_dims = len(data.shape)
print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
logger.debug(f"{new_name}, n_dims = {str(n_dims)}, {str(old_dtype)} --> {str(data.dtype)}")
gguf_writer.add_tensor(new_name, data)
print("gguf: write header")
logger.info("gguf: write header")
gguf_writer.write_header_to_file()
print("gguf: write metadata")
logger.info("gguf: write metadata")
gguf_writer.write_kv_data_to_file()
print("gguf: write tensors")
logger.info("gguf: write tensors")
gguf_writer.write_tensors_to_file()
gguf_writer.close()
print(f"gguf: model successfully exported to '{args.outfile}'")
print("")
logger.info(f"gguf: model successfully exported to '{args.outfile}'")
if __name__ == '__main__':

60
convert.py
View file

@ -1,6 +1,7 @@
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import concurrent.futures
import enum
@ -35,6 +36,8 @@ import gguf
if TYPE_CHECKING:
from typing_extensions import Self, TypeAlias
logger = logging.getLogger("convert")
if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
faulthandler.register(signal.SIGUSR1)
@ -643,7 +646,6 @@ class LlamaHfVocab(Vocab):
def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
# print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
if n_head_kv is not None and n_head != n_head_kv:
n_head = n_head_kv
return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
@ -1033,12 +1035,12 @@ def check_vocab_size(params: Params, vocab: BaseVocab, pad_vocab: bool = False)
# Check for a vocab size mismatch
if params.n_vocab == vocab.vocab_size:
print("Ignoring added_tokens.json since model matches vocab size without it.")
logger.warning("Ignoring added_tokens.json since model matches vocab size without it.")
return
if pad_vocab and params.n_vocab > vocab.vocab_size:
pad_count = params.n_vocab - vocab.vocab_size
print(
logger.debug(
f"Padding vocab with {pad_count} token(s) - <dummy00001> through <dummy{pad_count:05}>"
)
for i in range(1, pad_count + 1):
@ -1166,7 +1168,7 @@ class OutputFile:
elapsed = time.time() - start
size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
padi = len(str(len(model)))
print(
logger.info(
f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
)
self.gguf.write_tensor_data(ndarray)
@ -1281,12 +1283,12 @@ def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) ->
# HF models permut or pack some of the tensors, so we need to undo that
for i in itertools.count():
if f"model.layers.{i}.self_attn.q_proj.weight" in model:
print(f"Permuting layer {i}")
logger.debug(f"Permuting layer {i}")
tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head, params.n_head)
tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_head_kv)
# tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"]
elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
print(f"Unpacking and permuting layer {i}")
logger.debug(f"Unpacking and permuting layer {i}")
tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv)
tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2)
@ -1299,15 +1301,15 @@ def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) ->
tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
if name_new is None:
if skip_unknown:
print(f"Unexpected tensor name: {name} - skipping")
logger.warning(f"Unexpected tensor name: {name} - skipping")
continue
raise ValueError(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
if tensor_type in should_skip:
print(f"skipping tensor {name_new}")
logger.debug(f"skipping tensor {name_new}")
continue
print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
logger.debug(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}")
out[name_new] = lazy_tensor
return out
@ -1372,7 +1374,7 @@ def load_some_model(path: Path) -> ModelPlus:
paths = find_multifile_paths(path)
models_plus: list[ModelPlus] = []
for path in paths:
print(f"Loading model file {path}")
logger.info(f"Loading model file {path}")
models_plus.append(lazy_load_file(path))
model_plus = merge_multifile_models(models_plus)
@ -1413,7 +1415,7 @@ class VocabFactory:
else:
raise FileNotFoundError(f"Could not find a tokenizer matching any of {vocab_types}")
print(f"Loaded vocab file {vocab.fname_tokenizer!r}, type {vocab.name!r}")
logger.info(f"Loaded vocab file {vocab.fname_tokenizer!r}, type {vocab.name!r}")
return vocab
def load_vocab(self, vocab_types: list[str] | None, model_parent_path: Path) -> tuple[BaseVocab, gguf.SpecialVocab]:
@ -1438,19 +1440,19 @@ def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
}[file_type]
ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
if ret in model_paths:
sys.stderr.write(
logger.error(
f"Error: Default output path ({ret}) would overwrite the input. "
"Please explicitly specify a path using --outfile.\n")
"Please explicitly specify a path using --outfile.")
sys.exit(1)
return ret
def do_dump_model(model_plus: ModelPlus) -> None:
print(f"model_plus.paths = {model_plus.paths!r}")
print(f"model_plus.format = {model_plus.format!r}")
print(f"model_plus.vocab = {model_plus.vocab!r}")
print(f"model_plus.paths = {model_plus.paths!r}") # noqa: NP100
print(f"model_plus.format = {model_plus.format!r}") # noqa: NP100
print(f"model_plus.vocab = {model_plus.vocab!r}") # noqa: NP100
for name, lazy_tensor in model_plus.model.items():
print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}")
print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") # noqa: NP100
def main(args_in: list[str] | None = None) -> None:
@ -1473,8 +1475,18 @@ def main(args_in: list[str] | None = None) -> None:
parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine")
parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args(args_in)
if args.verbose:
logging.basicConfig(level=logging.DEBUG)
elif args.dump_single or args.dump:
# Avoid printing anything besides the dump output
logging.basicConfig(level=logging.WARNING)
else:
logging.basicConfig(level=logging.INFO)
if args.no_vocab and args.vocab_only:
raise ValueError("--vocab-only does not make sense with --no-vocab")
@ -1491,6 +1503,7 @@ def main(args_in: list[str] | None = None) -> None:
if args.dump:
do_dump_model(model_plus)
return
endianess = gguf.GGUFEndian.LITTLE
if args.big_endian:
endianess = gguf.GGUFEndian.BIG
@ -1515,7 +1528,7 @@ def main(args_in: list[str] | None = None) -> None:
"q8_0": GGMLFileType.MostlyQ8_0,
}[args.outtype]
print(f"params = {params}")
logger.info(f"params = {params}")
model_parent_path = model_plus.paths[0].parent
vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
@ -1541,15 +1554,14 @@ def main(args_in: list[str] | None = None) -> None:
)
OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
endianess=endianess, pad_vocab=args.pad_vocab)
print(f"Wrote {outfile}")
logger.info(f"Wrote {outfile}")
return
if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
vocab = model_plus.vocab
print(f"Vocab info: {vocab}")
print(f"Special vocab info: {special_vocab}")
logger.info(f"Vocab info: {vocab}")
logger.info(f"Special vocab info: {special_vocab}")
model = model_plus.model
model = convert_model_names(model, params, args.skip_unknown)
ftype = pick_output_type(model, args.outtype)
@ -1557,11 +1569,11 @@ def main(args_in: list[str] | None = None) -> None:
outfile = args.outfile or default_outfile(model_plus.paths, ftype)
params.ftype = ftype
print(f"Writing {outfile}, format {ftype}")
logger.info(f"Writing {outfile}, format {ftype}")
OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab)
print(f"Wrote {outfile}")
logger.info(f"Wrote {outfile}")
if __name__ == '__main__':

2
examples/main/main.cpp
View file

@ -544,7 +544,7 @@ int main(int argc, char ** argv) {
// 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 (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;

3
examples/server/server.cpp
View file

@ -1383,9 +1383,10 @@ struct server_context {
if (!slot.params.stream && slot.stopped_word) {
const std::vector<llama_token> stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false);
size_t safe_offset = std::min(slot.generated_token_probs.size(), stop_word_toks.size());
probs = std::vector<completion_token_output>(
slot.generated_token_probs.begin(),
slot.generated_token_probs.end() - stop_word_toks.size());
slot.generated_token_probs.end() - safe_offset);
} else {
probs = std::vector<completion_token_output>(
slot.generated_token_probs.begin(),

90
examples/server/tests/features/results.feature
View file

@ -7,44 +7,16 @@ Feature: Results
And a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models
And a model file test-model-00001-of-00003.gguf
And 128 as batch size
And 256 KV cache size
And 1024 KV cache size
And 128 max tokens to predict
Scenario Outline: Multi users completion
Given <n_slots> slots
And continuous batching
Scenario Outline: consistent results with same seed
Given <n_slots> slots
Then the server is starting
Then the server is healthy
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 42 as seed
And a prompt:
"""
Write a very long story about AI.
"""
Given 4 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 42
Given concurrent completion requests
Then the server is busy
@ -55,3 +27,55 @@ Feature: Results
| n_slots |
| 1 |
| 2 |
Scenario Outline: different results with different seed
Given <n_slots> slots
Then the server is starting
Then the server is healthy
Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 42
Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 43
Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 44
Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 45
Given concurrent completion requests
Then the server is busy
Then the server is idle
And all slots are idle
Then all predictions are different
Examples:
| n_slots |
| 1 |
| 2 |
Scenario Outline: consistent results with same seed and varying batch size
Given 4 slots
And <temp> temperature
# And 0 as draft
Then the server is starting
Then the server is healthy
Given 1 prompts "Write a very long story about AI." with seed 42
And concurrent completion requests
# Then the server is busy # Not all slots will be utilized.
Then the server is idle
And all slots are idle
Given <n_parallel> prompts "Write a very long story about AI." with seed 42
And concurrent completion requests
# Then the server is busy # Not all slots will be utilized.
Then the server is idle
And all slots are idle
Then all predictions are equal
Examples:
| n_parallel | temp |
| 1 | 0.0 |
| 2 | 0.0 |
| 4 | 0.0 |
| 1 | 1.0 |
# FIXME: These tests fail on master. The problem seems to be the unified KV cache.
# See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
# and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574 .
# | 2 | 1.0 |
# | 4 | 1.0 |

128
examples/server/tests/features/steps/steps.py
View file

@ -65,6 +65,7 @@ def step_server_config(context, server_fqdn, server_port):
context.server_seed = None
context.user_api_key = None
context.response_format = None
context.temperature = None
context.tasks_result = []
context.concurrent_tasks = []
@ -232,15 +233,17 @@ async def step_all_slots_status(context, expected_slot_status_string):
@async_run_until_complete
async def step_request_completion(context, api_error):
expect_api_error = api_error == 'raised'
seeds = await completions_seed(context, num_seeds=1)
completion = await request_completion(context.prompts.pop(),
seeds[0] if seeds is not None else seeds,
context.base_url,
debug=context.debug,
n_predict=context.n_predict,
cache_prompt=context.cache_prompt,
id_slot=context.id_slot,
seed=await completions_seed(context),
expect_api_error=expect_api_error,
user_api_key=context.user_api_key)
user_api_key=context.user_api_key,
temperature=context.temperature)
context.tasks_result.append(completion)
if context.debug:
print(f"Completion response: {completion}")
@ -269,6 +272,15 @@ async def step_predictions_equal(context):
context.tasks_result = []
@step('all predictions are different')
@async_run_until_complete
async def step_predictions_equal(context):
n_completions = await gather_tasks_results(context)
assert n_completions >= 2, "need at least 2 completions"
assert_all_predictions_different(context.tasks_result)
context.tasks_result = []
@step('the completion is truncated')
def step_assert_completion_truncated(context):
step_assert_completion_truncated(context, '')
@ -311,6 +323,11 @@ def step_response_format(context, response_format):
context.response_format = json.loads(response_format)
@step('{temperature:f} temperature')
def step_temperature(context, temperature):
context.temperature = temperature
@step('streaming is {enable_streaming}')
def step_streaming(context, enable_streaming):
context.enable_streaming = enable_streaming == 'enabled'
@ -353,7 +370,10 @@ def step_n_ubatch(context, n_ubatch):
@step('{seed:d} as seed')
def step_seed(context, seed):
context.seed = seed
if context.seed is None:
context.seed = [seed]
else:
context.seed.append(seed)
@step('a prefix prompt')
@ -413,7 +433,9 @@ async def step_oai_chat_completions(context, api_error):
if context.debug:
print(f"Submitting OAI compatible completions request...")
expect_api_error = api_error == 'raised'
seeds = await completions_seed(context, num_seeds=1),
completion = await oai_chat_completions(context.prompts.pop(),
seeds[0] if seeds is not None else seeds,
context.system_prompt,
context.base_url,
'/v1/chat',
@ -429,8 +451,6 @@ async def step_oai_chat_completions(context, api_error):
response_format=context.response_format
if hasattr(context, 'response_format') else None,
seed=await completions_seed(context),
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None,
@ -457,10 +477,21 @@ def step_a_prompt_prompt(context, prompt):
context.n_prompts = len(context.prompts)
@step('{num_prompts:d} prompts {prompt} with seed {seed:d}')
def step_many_prompts(context, num_prompts, prompt, seed):
if context.seed is None:
context.seed = []
for _ in range(num_prompts):
context.seed.append(seed)
context.prompts.append(prompt)
context.n_prompts = len(context.prompts)
@step('concurrent completion requests')
@async_run_until_complete()
async def step_concurrent_completion_requests(context):
await concurrent_requests(context,
await concurrent_requests(
context,
request_completion,
# prompt is inserted automatically
context.base_url,
@ -468,9 +499,9 @@ async def step_concurrent_completion_requests(context):
prompt_prefix=context.prompt_prefix,
prompt_suffix=context.prompt_suffix,
n_predict=context.n_predict if hasattr(context, 'n_predict') else None,
seed=await completions_seed(context),
user_api_key=context.user_api_key if hasattr(context,
'user_api_key') else None)
user_api_key=context.user_api_key if hasattr(context, 'user_api_key') else None,
temperature=context.temperature,
)
@step('concurrent OAI completions requests')
@ -490,7 +521,6 @@ async def step_oai_chat_completions(context):
if hasattr(context, 'enable_streaming') else None,
response_format=context.response_format
if hasattr(context, 'response_format') else None,
seed=await completions_seed(context),
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None)
@ -512,10 +542,6 @@ async def step_oai_chat_completions(context):
if hasattr(context, 'enable_streaming') else None,
response_format=context.response_format
if hasattr(context, 'response_format') else None,
seed=context.seed
if hasattr(context, 'seed') else
context.server_seed
if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None)
@ -544,7 +570,7 @@ async def all_prompts_are_predicted(context, expected_predicted_n=None):
@async_run_until_complete
async def step_compute_embedding(context):
context.n_prompts = 1
context.embeddings = await request_embedding(context_text(context), base_url=context.base_url)
context.embeddings = await request_embedding(context_text(context), None, base_url=context.base_url)
@step('all embeddings are the same')
@ -585,7 +611,7 @@ def step_assert_embeddings(context):
@async_run_until_complete
async def step_oai_compute_embeddings(context):
context.n_prompts = 1
context.embeddings = await request_oai_embeddings(context_text(context),
context.embeddings = await request_oai_embeddings(context_text(context), None,
base_url=context.base_url,
user_api_key=context.user_api_key,
model=context.model)
@ -594,7 +620,7 @@ async def step_oai_compute_embeddings(context):
@step('an OAI compatible embeddings computation request for multiple inputs')
@async_run_until_complete
async def step_oai_compute_embeddings_multiple_inputs(context):
context.embeddings = await request_oai_embeddings(context.prompts,
context.embeddings = await request_oai_embeddings(context.prompts, None,
base_url=context.base_url,
user_api_key=context.user_api_key,
model=context.model)
@ -740,8 +766,9 @@ async def concurrent_requests(context, f_completion, *args, **kwargs):
if context.debug:
print(f"starting {context.n_prompts} concurrent completion requests...")
assert context.n_prompts > 0
seeds = await completions_seed(context)
for prompt_no in range(context.n_prompts):
shifted_args = [context.prompts.pop(), *args]
shifted_args = [context.prompts.pop(), seeds[prompt_no], *args]
context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
await asyncio.sleep(0.1)
@ -781,6 +808,7 @@ def step_server_responds_with_status_code(context, status_code):
async def request_completion(prompt,
seed,
base_url,
debug=False,
prompt_prefix=None,
@ -788,9 +816,9 @@ async def request_completion(prompt,
n_predict=None,
cache_prompt=False,
id_slot=None,
seed=None,
expect_api_error=None,
user_api_key=None):
user_api_key=None,
temperature=None):
if debug:
print(f"Sending completion request: {prompt}")
origin = "my.super.domain"
@ -811,7 +839,8 @@ async def request_completion(prompt,
"n_predict": n_predict if n_predict is not None else -1,
"cache_prompt": cache_prompt,
"id_slot": id_slot,
"seed": seed if seed is not None else 42
"seed": seed if seed is not None else 42,
"temperature": temperature if temperature is not None else "0.8f",
},
headers=headers,
timeout=3600) as response:
@ -824,6 +853,7 @@ async def request_completion(prompt,
async def oai_chat_completions(user_prompt,
seed,
system_prompt,
base_url,
base_path,
@ -833,7 +863,6 @@ async def oai_chat_completions(user_prompt,
n_predict=None,
enable_streaming=None,
response_format=None,
seed=None,
user_api_key=None,
expect_api_error=None):
if debug:
@ -952,7 +981,7 @@ async def oai_chat_completions(user_prompt,
return completion_response
async def request_embedding(content, base_url=None):
async def request_embedding(content, seed, base_url=None):
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/embedding',
json={
@ -963,7 +992,7 @@ async def request_embedding(content, base_url=None):
return [response_json['embedding']]
async def request_oai_embeddings(input,
async def request_oai_embeddings(input, seed,
base_url=None, user_api_key=None,
model=None, async_client=False):
# openai client always expects an api_key
@ -1036,21 +1065,31 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re
f' {n_predicted} <> {expected_predicted_n}')
def assert_all_predictions_equal(completion_responses):
content_0 = completion_responses[0]['content']
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
print(f"content 0: {content_0}")
for i, response_i in enumerate(completion_responses):
content_i = response_i['content']
print(f"content {i}: {content_i}")
for i, response_i in enumerate(completion_responses):
content_i = response_i['content']
for j, response_j in enumerate(completion_responses):
if i == j:
continue
content_j = response_j['content']
assert content_i == content_j, "contents not equal"
i = 1
for response in completion_responses[1:]:
content = response['content']
def assert_all_predictions_different(completion_responses):
if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
print(f"content {i}: {content}")
assert content == content_0, "contents not equal"
i += 1
for i, response_i in enumerate(completion_responses):
content_i = response_i['content']
print(f"content {i}: {content_i}")
for i, response_i in enumerate(completion_responses):
content_i = response_i['content']
for j, response_j in enumerate(completion_responses):
if i == j:
continue
content_j = response_j['content']
assert content_i != content_j, "contents not different"
async def gather_tasks_results(context):
@ -1145,9 +1184,22 @@ def assert_slots_status(slots, expected_slots):
f" = {expected[key]} != {slot[key]}")
async def completions_seed(context):
return context.seed if hasattr(context, 'seed') and context.seed is not None \
else context.server_seed if hasattr(context, 'server_seed') else None
async def completions_seed(context, num_seeds=None):
if hasattr(context, "seed") and context.seed is not None:
assert len(context.seed) == context.n_prompts
if num_seeds is None:
num_seeds = context.n_prompts
assert num_seeds <= context.n_prompts
seeds = context.seed[:num_seeds]
context.seed = context.seed[num_seeds:] if num_seeds < context.n_prompts else None
return seeds
if hasattr(context, "server_seed") and context.server_seed is not None:
if num_seeds is None:
return [context.server_seed] * context.n_prompts
else:
return [context.server_seed] * num_seeds
return None
def context_text(context):

12
ggml_vk_generate_shaders.py
View file

@ -1,11 +1,14 @@
#!/usr/bin/env python
import logging
import argparse
import asyncio
import os
import sys
from tempfile import gettempdir, NamedTemporaryFile
logger = logging.getLogger("ggml-vk-generate-shaders")
shader_f32 = """
#define FLOAT_TYPE float
"""
@ -2498,7 +2501,7 @@ async def string_to_spv(name, code, defines, fp16=True):
stdout, stderr = await proc.communicate()
print(" ".join(cmd))
logger.info(" ".join(cmd))
if proc.returncode:
raise RuntimeError(f"{name=} {f.name=} {stdout=} {stderr=}")
@ -2507,7 +2510,7 @@ async def string_to_spv(name, code, defines, fp16=True):
cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
code_with_lines = "\n".join([f"{i + 1}: {line}" for i, line in enumerate(preprocessed_code.splitlines())])
print(f"ERROR compiling {name}\n\n{code_with_lines}\n\n{error}")
logger.error(f"cannot compile {name}\n\n{code_with_lines}\n\n{error}")
f.close()
os.remove(f.name)
sys.exit(proc.returncode)
@ -2520,7 +2523,7 @@ async def string_to_spv(name, code, defines, fp16=True):
async def main():
print("ggml_vulkan: Generating and compiling shaders to SPIR-V")
logger.info("ggml_vulkan: Generating and compiling shaders to SPIR-V")
tasks = []
@ -2768,9 +2771,12 @@ if __name__ == "__main__":
parser = argparse.ArgumentParser(description="GGML Vulkan Shader Generator")
parser.add_argument("--glslc", help="Path to glslc")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
if args.glslc:
GLSLC = args.glslc

18
gguf-py/examples/reader.py
View file

@ -1,8 +1,10 @@
#!/usr/bin/env python3
import logging
import sys
from pathlib import Path
from gguf.gguf_reader import GGUFReader
logger = logging.getLogger("reader")
sys.path.insert(0, str(Path(__file__).parent.parent))
@ -18,28 +20,28 @@ def read_gguf_file(gguf_file_path):
reader = GGUFReader(gguf_file_path)
# List all key-value pairs in a columnized format
print("Key-Value Pairs:")
print("Key-Value Pairs:") # noqa: NP100
max_key_length = max(len(key) for key in reader.fields.keys())
for key, field in reader.fields.items():
value = field.parts[field.data[0]]
print(f"{key:{max_key_length}} : {value}")
print("----")
print(f"{key:{max_key_length}} : {value}") # noqa: NP100
print("----") # noqa: NP100
# List all tensors
print("Tensors:")
print("Tensors:") # noqa: NP100
tensor_info_format = "{:<30} | Shape: {:<15} | Size: {:<12} | Quantization: {}"
print(tensor_info_format.format("Tensor Name", "Shape", "Size", "Quantization"))
print("-" * 80)
print(tensor_info_format.format("Tensor Name", "Shape", "Size", "Quantization")) # noqa: NP100
print("-" * 80) # noqa: NP100
for tensor in reader.tensors:
shape_str = "x".join(map(str, tensor.shape))
size_str = str(tensor.n_elements)
quantization_str = tensor.tensor_type.name
print(tensor_info_format.format(tensor.name, shape_str, size_str, quantization_str))
print(tensor_info_format.format(tensor.name, shape_str, size_str, quantization_str)) # noqa: NP100
if __name__ == '__main__':
if len(sys.argv) < 2:
print("Usage: reader.py <path_to_gguf_file>")
logger.info("Usage: reader.py <path_to_gguf_file>")
sys.exit(1)
gguf_file_path = sys.argv[1]
read_gguf_file(gguf_file_path)

4
gguf-py/gguf/constants.py
View file

@ -1,6 +1,5 @@
from __future__ import annotations
import sys
from enum import Enum, IntEnum, auto
from typing import Any
@ -854,8 +853,7 @@ class GGUFValueType(IntEnum):
return GGUFValueType.INT32
# TODO: need help with 64-bit types in Python
else:
print("Unknown type:", type(val))
sys.exit()
raise ValueError(f"Unknown type: {type(val)}")
# Note: Does not support GGML_QKK_64

4
gguf-py/gguf/gguf_reader.py
View file

@ -4,6 +4,7 @@
#
from __future__ import annotations
import logging
import os
from collections import OrderedDict
from typing import Any, Literal, NamedTuple, TypeVar, Union
@ -27,6 +28,7 @@ from gguf.constants import (
GGUFValueType,
)
logger = logging.getLogger(__name__)
READER_SUPPORTED_VERSIONS = [2, GGUF_VERSION]
@ -142,7 +144,7 @@ class GGUFReader:
# TODO: add option to generate error on duplicate keys
# raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
print(f'Warning: Duplicate key {field.name} at offset {field.offset}')
logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
self.fields[field.name + '_{}'.format(field.offset)] = field
else:
self.fields[field.name] = field

5
gguf-py/gguf/gguf_writer.py
View file

@ -1,5 +1,6 @@
from __future__ import annotations
import logging
import os
import shutil
import struct
@ -24,6 +25,8 @@ from .constants import (
TokenType,
)
logger = logging.getLogger(__name__)
class WriterState(Enum):
EMPTY = auto()
@ -67,7 +70,7 @@ class GGUFWriter:
self.use_temp_file = use_temp_file
self.temp_file = None
self.tensors = []
print("gguf: This GGUF file is for {0} Endian only".format(
logger.info("gguf: This GGUF file is for {0} Endian only".format(
"Big" if self.endianess == GGUFEndian.BIG else "Little",
))
self.state = WriterState.EMPTY

42
gguf-py/gguf/vocab.py
View file

@ -1,13 +1,15 @@
from __future__ import annotations
import logging
import json
import os
import sys
from pathlib import Path
from typing import Any, Callable
from .gguf_writer import GGUFWriter
logger = logging.getLogger(__name__)
class SpecialVocab:
merges: list[str]
@ -40,38 +42,29 @@ class SpecialVocab:
def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None:
if self.merges:
if not quiet:
print(f'gguf: Adding {len(self.merges)} merge(s).')
logger.info(f'Adding {len(self.merges)} merge(s).')
gw.add_token_merges(self.merges)
elif self.load_merges:
print(
'gguf: WARNING: Adding merges requested but no merges found, output may be non-functional.',
file = sys.stderr,
)
logger.warning('Adding merges requested but no merges found, output may be non-functional.')
for typ, tokid in self.special_token_ids.items():
id_handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None)
if id_handler is None:
print(
f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping',
file = sys.stderr,
)
logger.warning(f'No handler for special token type {typ} with id {tokid} - skipping')
continue
if not quiet:
print(f'gguf: Setting special token type {typ} to {tokid}')
logger.info(f'Setting special token type {typ} to {tokid}')
id_handler(tokid)
for typ, value in self.add_special_token.items():
add_handler: Callable[[bool], None] | None = getattr(gw, f'add_add_{typ}_token', None)
if add_handler is None:
print(
f'gguf: WARNING: No handler for add_{typ}_token with value {value} - skipping',
file = sys.stderr,
)
logger.warning(f'No handler for add_{typ}_token with value {value} - skipping')
continue
if not quiet:
print(f'gguf: Setting add_{typ}_token to {value}')
logger.info(f'Setting add_{typ}_token to {value}')
add_handler(value)
if self.chat_template is not None:
if not quiet:
print(f'gguf: Setting chat_template to {self.chat_template}')
logger.info(f'Setting chat_template to {self.chat_template}')
gw.add_chat_template(self.chat_template)
def _load(self, path: Path) -> None:
@ -99,10 +92,7 @@ class SpecialVocab:
continue
parts = line.split(None, 3)
if len(parts) != 2:
print(
f'gguf: WARNING: {merges_file.name}: Line {line_num}: Entry malformed, ignoring',
file = sys.stderr,
)
logger.warning(f'{merges_file.name}: Line {line_num}: Entry malformed, ignoring')
continue
merges.append(f'{parts[0]} {parts[1]}')
self.merges = merges
@ -118,10 +108,7 @@ class SpecialVocab:
return
self.special_token_ids[typ] = tid
return
print(
f'gguf: WARNING: Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping',
file = sys.stderr,
)
logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping')
def _try_load_from_tokenizer_json(self, path: Path) -> bool:
tokenizer_file = path / 'tokenizer.json'
@ -144,10 +131,7 @@ class SpecialVocab:
if chat_template is None or isinstance(chat_template, (str, list)):
self.chat_template = chat_template
else:
print(
f'gguf: WARNING: Bad type for chat_template field in {tokenizer_config_file!r} - ignoring',
file = sys.stderr
)
logger.warning(f'Bad type for chat_template field in {tokenizer_config_file!r} - ignoring')
for typ in self.special_token_types:
add_entry = tokenizer_config.get(f'add_{typ}_token')
if isinstance(add_entry, bool):

86
gguf-py/scripts/gguf-convert-endian.py
View file

@ -1,9 +1,11 @@
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import os
import sys
from tqdm import tqdm
from pathlib import Path
import numpy as np
@ -14,6 +16,8 @@ if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent /
import gguf
logger = logging.getLogger("gguf-convert-endian")
def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None:
if np.uint32(1) == np.uint32(1).newbyteorder("<"):
@ -29,11 +33,11 @@ def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None
else:
file_endian = host_endian
order = host_endian if args.order == "native" else args.order
print(f"* Host is {host_endian.upper()} endian, GGUF file seems to be {file_endian.upper()} endian")
logger.info(f"* Host is {host_endian.upper()} endian, GGUF file seems to be {file_endian.upper()} endian")
if file_endian == order:
print(f"* File is already {order.upper()} endian. Nothing to do.")
logger.info(f"* File is already {order.upper()} endian. Nothing to do.")
sys.exit(0)
print("* Checking tensors for conversion compatibility")
logger.info("* Checking tensors for conversion compatibility")
for tensor in reader.tensors:
if tensor.tensor_type not in (
gguf.GGMLQuantizationType.F32,
@ -41,51 +45,64 @@ def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None
gguf.GGMLQuantizationType.Q8_0,
):
raise ValueError(f"Cannot handle type {tensor.tensor_type.name} for tensor {repr(tensor.name)}")
print(f"* Preparing to convert from {file_endian.upper()} to {order.upper()}")
logger.info(f"* Preparing to convert from {file_endian.upper()} to {order.upper()}")
if args.dry_run:
return
print("\n*** Warning *** Warning *** Warning **")
print("* This conversion process may damage the file. Ensure you have a backup.")
logger.warning("*** Warning *** Warning *** Warning **")
logger.warning("* This conversion process may damage the file. Ensure you have a backup.")
if order != host_endian:
print("* Requested endian differs from host, you will not be able to load the model on this machine.")
print("* The file will be modified immediately, so if conversion fails or is interrupted")
print("* the file will be corrupted. Enter exactly YES if you are positive you want to proceed:")
logger.warning("* Requested endian differs from host, you will not be able to load the model on this machine.")
logger.warning("* The file will be modified immediately, so if conversion fails or is interrupted")
logger.warning("* the file will be corrupted. Enter exactly YES if you are positive you want to proceed:")
response = input("YES, I am sure> ")
if response != "YES":
print("You didn't enter YES. Okay then, see ya!")
logger.warning("You didn't enter YES. Okay then, see ya!")
sys.exit(0)
print(f"\n* Converting fields ({len(reader.fields)})")
logger.info(f"* Converting fields ({len(reader.fields)})")
for idx, field in enumerate(reader.fields.values()):
print(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}")
logger.info(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}")
for part in field.parts:
part.byteswap(inplace=True)
print(f"\n* Converting tensors ({len(reader.tensors)})")
for idx, tensor in enumerate(reader.tensors):
print(
f" - {idx:4}: Converting tensor {repr(tensor.name)}, type={tensor.tensor_type.name}, "
f"elements={tensor.n_elements}... ",
end="",
logger.info(f"* Converting tensors ({len(reader.tensors)})")
for idx, tensor in enumerate(pbar := tqdm(reader.tensors, desc="Converting tensor")):
log_message = (
f"Converting tensor {repr(tensor.name)}, "
f"type={tensor.tensor_type.name}, "
f"elements={tensor.n_elements} "
)
tensor_type = tensor.tensor_type
# Byte-swap each part of the tensor's field
for part in tensor.field.parts:
part.byteswap(inplace=True)
if tensor_type != gguf.GGMLQuantizationType.Q8_0:
tensor.data.byteswap(inplace=True)
print()
continue
# A Q8_0 block consists of a f16 delta followed by 32 int8 quants, so 34 bytes
block_size = 34
# Byte-swap tensor data if necessary
if tensor.tensor_type == gguf.GGMLQuantizationType.Q8_0:
# Handle Q8_0 tensor blocks (block_q8_0)
# Specific handling of block_q8_0 is required.
# Each block_q8_0 consists of an f16 delta (scaling factor) followed by 32 int8 quantizations.
block_size = 34 # 34 bytes = <f16 delta scaling factor> + 32 * <int8 quant>
n_blocks = len(tensor.data) // block_size
for block_num in range(n_blocks):
for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)):
block_offs = block_num * block_size
# I know I said f16, but it doesn't matter here - any simple 16 bit type works.
# Byte-Swap f16 sized delta field
delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16)
delta.byteswap(inplace=True)
# Byte-Swap Q8 weights
if block_num % 100000 == 0:
print(f"[{(n_blocks - block_num) // 1000}K]", end="")
sys.stdout.flush()
print()
print("* Completion")
inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]")
else:
# Handle other tensor types
tensor.data.byteswap(inplace=True)
pbar.set_description(log_message)
logger.info("* Completion")
def main() -> None:
@ -102,8 +119,13 @@ def main() -> None:
"--dry-run", action="store_true",
help="Don't actually change anything",
)
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
print(f'* Loading: {args.model}')
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
logger.info(f'* Loading: {args.model}')
reader = gguf.GGUFReader(args.model, 'r' if args.dry_run else 'r+')
convert_byteorder(reader, args)

29
gguf-py/scripts/gguf-dump.py
View file

@ -1,6 +1,7 @@
#!/usr/bin/env python3
from __future__ import annotations
import logging
import argparse
import os
import sys
@ -15,6 +16,8 @@ if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent /
from gguf import GGUFReader, GGUFValueType # noqa: E402
logger = logging.getLogger("gguf-dump")
def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]:
host_endian = 'LITTLE' if np.uint32(1) == np.uint32(1).newbyteorder("<") else 'BIG'
@ -29,8 +32,8 @@ def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]:
# please see the comments in the modify_gguf.py example.
def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
host_endian, file_endian = get_file_host_endian(reader)
print(f'* File is {file_endian} endian, script is running on a {host_endian} endian host.')
print(f'\n* Dumping {len(reader.fields)} key/value pair(s)')
print(f'* File is {file_endian} endian, script is running on a {host_endian} endian host.') # noqa: NP100
print(f'* Dumping {len(reader.fields)} key/value pair(s)') # noqa: NP100
for n, field in enumerate(reader.fields.values(), 1):
if not field.types:
pretty_type = 'N/A'
@ -39,20 +42,21 @@ def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
pretty_type = '[' * nest_count + str(field.types[-1].name) + ']' * nest_count
else:
pretty_type = str(field.types[-1].name)
print(f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}', end = '')
log_message = f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}'
if len(field.types) == 1:
curr_type = field.types[0]
if curr_type == GGUFValueType.STRING:
print(' = {0}'.format(repr(str(bytes(field.parts[-1]), encoding='utf8')[:60])), end = '')
log_message += ' = {0}'.format(repr(str(bytes(field.parts[-1]), encoding='utf8')[:60]))
elif field.types[0] in reader.gguf_scalar_to_np:
print(' = {0}'.format(field.parts[-1][0]), end = '')
print()
log_message += ' = {0}'.format(field.parts[-1][0])
print(log_message) # noqa: NP100
if args.no_tensors:
return
print(f'\n* Dumping {len(reader.tensors)} tensor(s)')
print(f'* Dumping {len(reader.tensors)} tensor(s)') # noqa: NP100
for n, tensor in enumerate(reader.tensors, 1):
prettydims = ', '.join('{0:5}'.format(d) for d in list(tensor.shape) + [1] * (4 - len(tensor.shape)))
print(f' {n:5}: {tensor.n_elements:10} | {prettydims} | {tensor.tensor_type.name:7} | {tensor.name}')
print(f' {n:5}: {tensor.n_elements:10} | {prettydims} | {tensor.tensor_type.name:7} | {tensor.name}') # noqa: NP100
def dump_metadata_json(reader: GGUFReader, args: argparse.Namespace) -> None:
@ -103,10 +107,17 @@ def main() -> None:
parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata")
parser.add_argument("--json", action="store_true", help="Produce JSON output")
parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
if not args.json:
print(f'* Loading: {args.model}')
logger.info(f'* Loading: {args.model}')
reader = GGUFReader(args.model, 'r')
if args.json:
dump_metadata_json(reader, args)
else:

31
gguf-py/scripts/gguf-set-metadata.py
View file

@ -1,4 +1,5 @@
#!/usr/bin/env python3
import logging
import argparse
import os
import sys
@ -10,6 +11,8 @@ if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent /
from gguf import GGUFReader # noqa: E402
logger = logging.getLogger("gguf-set-metadata")
def minimal_example(filename: str) -> None:
reader = GGUFReader(filename, 'r+')
@ -41,36 +44,33 @@ def minimal_example(filename: str) -> None:
def set_metadata(reader: GGUFReader, args: argparse.Namespace) -> None:
field = reader.get_field(args.key)
if field is None:
print(f'! Field {repr(args.key)} not found', file = sys.stderr)
logger.error(f'! Field {repr(args.key)} not found')
sys.exit(1)
# Note that field.types is a list of types. This is because the GGUF
# format supports arrays. For example, an array of UINT32 would
# look like [GGUFValueType.ARRAY, GGUFValueType.UINT32]
handler = reader.gguf_scalar_to_np.get(field.types[0]) if field.types else None
if handler is None:
print(
f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}',
file = sys.stderr,
)
logger.error(f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}')
sys.exit(1)
current_value = field.parts[field.data[0]][0]
new_value = handler(args.value)
print(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}')
logger.info(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}')
if current_value == new_value:
print(f'- Key {repr(args.key)} already set to requested value {current_value}')
logger.info(f'- Key {repr(args.key)} already set to requested value {current_value}')
sys.exit(0)
if args.dry_run:
sys.exit(0)
if not args.force:
print('*** Warning *** Warning *** Warning **')
print('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.')
print('* Enter exactly YES if you are positive you want to proceed:')
logger.warning('*** Warning *** Warning *** Warning **')
logger.warning('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.')
logger.warning('* Enter exactly YES if you are positive you want to proceed:')
response = input('YES, I am sure> ')
if response != 'YES':
print("You didn't enter YES. Okay then, see ya!")
logger.info("You didn't enter YES. Okay then, see ya!")
sys.exit(0)
field.parts[field.data[0]][0] = new_value
print('* Field changed. Successful completion.')
logger.info('* Field changed. Successful completion.')
def main() -> None:
@ -80,8 +80,13 @@ def main() -> None:
parser.add_argument("value", type=str, help="Metadata value to set")
parser.add_argument("--dry-run", action="store_true", help="Don't actually change anything")
parser.add_argument("--force", action="store_true", help="Change the field without confirmation")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"])
print(f'* Loading: {args.model}')
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
logger.info(f'* Loading: {args.model}')
reader = GGUFReader(args.model, 'r' if args.dry_run else 'r+')
set_metadata(reader, args)

23
llama.cpp
View file

@ -2359,7 +2359,7 @@ static bool llama_kv_cache_init(
cache.recurrent = model.arch == LLM_ARCH_MAMBA;
cache.v_trans = !cparams.flash_attn;
// TODO: support mixed reccurent Transformer architectues
// TODO: support mixed recurrent Transformer architectures
// NOTE: (!a || b) is a logical implication (a -> b)
GGML_ASSERT(!cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_s());
GGML_ASSERT(!cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_s());
@ -4383,6 +4383,9 @@ static void llm_load_vocab(
} else if (
tokenizer_pre == "gpt-2") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2;
} else if (
tokenizer_pre == "refact") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_REFACT;
} else {
throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
}
@ -11952,7 +11955,7 @@ static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id
static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE);
GGML_ASSERT(llama_is_byte_token(vocab, id));
const auto& token_data = vocab.id_to_token.at(id);
const auto & token_data = vocab.id_to_token.at(id);
switch (llama_vocab_get_type(vocab)) {
case LLAMA_VOCAB_TYPE_SPM: {
auto buf = token_data.text.substr(3, 2);
@ -12212,14 +12215,13 @@ struct llm_tokenizer_bpe {
"\\s?\\p{L}+",
"\\s?\\p{P}+",
"[一-龥ࠀ-一가-퟿]+",
"\\p{N}+",
"\\p{N}",
});
break;
case LLAMA_VOCAB_PRE_TYPE_FALCON:
word_collection = unicode_regex_split(text, {
"[\\p{P}\\$\\+<=>\\^~\\|]+",
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
"\\p{N}+",
"[0-9][0-9][0-9]",
});
break;
@ -12235,6 +12237,12 @@ struct llm_tokenizer_bpe {
});
break;
case LLAMA_VOCAB_PRE_TYPE_STARCODER:
case LLAMA_VOCAB_PRE_TYPE_REFACT:
word_collection = unicode_regex_split(text, {
"\\p{N}",
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
});
break;
case LLAMA_VOCAB_PRE_TYPE_GPT2:
word_collection = unicode_regex_split(text, {
"'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
@ -17466,9 +17474,10 @@ int32_t llama_tokenize(
static std::string llama_decode_text(const std::string & text) {
std::string decoded_text;
auto unicode_sequences = unicode_cpts_from_utf8(text);
for (auto & unicode_sequence : unicode_sequences) {
decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(unicode_sequence));
const auto cpts = unicode_cpts_from_utf8(text);
for (const auto cpt : cpts) {
decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(cpt));
}
return decoded_text;

3
llama.h
View file

@ -79,6 +79,7 @@ extern "C" {
LLAMA_VOCAB_PRE_TYPE_MPT = 5,
LLAMA_VOCAB_PRE_TYPE_STARCODER = 6,
LLAMA_VOCAB_PRE_TYPE_GPT2 = 7,
LLAMA_VOCAB_PRE_TYPE_REFACT = 8,
};
// note: these values should be synchronized with ggml_rope
@ -171,7 +172,7 @@ extern "C" {
bool sorted;
} llama_token_data_array;
typedef bool (*llama_progress_callback)(float progress, void *ctx);
typedef bool (*llama_progress_callback)(float progress, void * user_data);
// Input data for llama_decode
// A llama_batch object can contain input about one or many sequences

4
models/ggml-vocab-bert-bge.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-bert-bge.gguf.out
View file

@ -1,3 +1,5 @@
29464 2094 1018 1092 2706
11865 17875

4
models/ggml-vocab-deepseek-coder.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-deepseek-coder.gguf.out
View file

@ -1,3 +1,5 @@
1050 207 19 207 19192 4217
37 32009 71 6247
207
243

4
models/ggml-vocab-deepseek-llm.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-deepseek-llm.gguf.out
View file

@ -1,3 +1,5 @@
1052 207 19 207 19109 4223
37 100014 71 6245
207
243

4
models/ggml-vocab-falcon.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-falcon.gguf.out
View file

@ -1,3 +1,5 @@
878 204 31 3068 133 2137
28611 132 30042
204
258

4
models/ggml-vocab-gpt-2.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-gpt-2.gguf.out
View file

@ -1,3 +1,5 @@
798 604 25208 1933
37 9116 71 11751
220
220 220

4
models/ggml-vocab-llama-bpe.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-llama-bpe.gguf.out
View file

@ -1,3 +1,5 @@
1142 220 19 220 27154 4038
37 51853 261
220
256

4
models/ggml-vocab-llama-spm.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-llama-spm.gguf.out
View file

@ -1,3 +1,5 @@
474 287 29871 29946 29871 30226 7378
383 4000 261
259
1678

4
models/ggml-vocab-mpt.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-mpt.gguf.out
View file

@ -1,3 +1,5 @@
728 577 24142 2607
39 26288 6554
209
50276

BIN
models/ggml-vocab-phi-3.gguf
View file

Binary file not shown.

4
models/ggml-vocab-phi-3.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-phi-3.gguf.out
View file

@ -1,3 +1,5 @@
474 287 29871 29946 29871 30226 7378
383 4000 261
259
1678

BIN
models/ggml-vocab-refact.gguf
View file

Binary file not shown.

106
models/ggml-vocab-refact.gguf.inp Normal file
View file

@ -0,0 +1,106 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__
__ggml_vocab_test__
__ggml_vocab_test__
__ggml_vocab_test__
__ggml_vocab_test__
__ggml_vocab_test__
__ggml_vocab_test__
__ggml_vocab_test__
__ggml_vocab_test__
Hello world
__ggml_vocab_test__
Hello world
__ggml_vocab_test__
Hello World
__ggml_vocab_test__
Hello World
__ggml_vocab_test__
Hello World!
__ggml_vocab_test__
Hello, world!
__ggml_vocab_test__
Hello, world!
__ggml_vocab_test__
this is 🦙.cpp
__ggml_vocab_test__
w048 7tuijk dsdfhu
__ggml_vocab_test__
нещо на Български
__ggml_vocab_test__
កាន់តែពិសេសអាចខលចេញ
__ggml_vocab_test__
🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)
__ggml_vocab_test__
Hello
__ggml_vocab_test__
Hello
__ggml_vocab_test__
Hello
__ggml_vocab_test__
Hello
__ggml_vocab_test__
Hello
__ggml_vocab_test__
Hello
Hello
__ggml_vocab_test__
(
__ggml_vocab_test__
=
__ggml_vocab_test__
' era
__ggml_vocab_test__
Hello, y'all! How are you 😁 ?我想在apple工作1314151天～
__ggml_vocab_test__
3
__ggml_vocab_test__
33
__ggml_vocab_test__
333
__ggml_vocab_test__
3333
__ggml_vocab_test__
33333
__ggml_vocab_test__
333333
__ggml_vocab_test__
3333333
__ggml_vocab_test__
33333333
__ggml_vocab_test__
333333333
__ggml_vocab_test__
🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL
__ggml_vocab_test__

43
models/ggml-vocab-refact.gguf.out Normal file
View file

@ -0,0 +1,43 @@
4833 225 38 225 143 140 17723
56 2006 3935 265
225
261
264
202
203
478
2831
15773
8279 5788
12000 5788
8279 10896
12000 10896
12000 10896 19
8279 30 5788 19
12000 30 5788 19
458 438 5945 118 252 32 3766
105 34 38 42 225 41 102 1707 12530 10180 1479 8278
39862 8372 1039 9446 40242 13852 2053 8949 12531 1520 10700
14574 227 14574 133 14574 246 30457 238 14574 242 30457 229 14574 249 14574 134 14574 258 30457 228 14574 258 14574 114 14574 133 14574 232 14574 228 14574 254 14574 232 30457 228 14574 236
3807 253 227 308 4382 27 18458 133 46113 44967 123 13868 308 12565 19775 33071 40824 733 27 41889 308 2585 22680 688 1401 2819 4369 2404 27
8279
12000
225 12000
261 12000
264 12000
264 12000 284 12000
308
203 280
25 34666
8279 30 533 25 464 19 4971 884 844 18458 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838
37
37 37
37 37 37
37 37 37 37
37 37 37 37 37
37 37 37 37 37 37
37 37 37 37 37 37 37
37 37 37 37 37 37 37 37
37 37 37 37 37 37 37 37 37
334 719 8878 202 10885 4222 16104 28570 203 3807 253 227 308 4382 27 18458 133 46113 44967 123 13868 308 12565 19775 33071 40824 733 27 41889 5945 118 252 3807 118 252 225 37 225 37 37 225 37 37 37 225 37 37 37 37 225 37 37 37 37 37 225 37 37 37 37 37 37 225 37 37 37 37 37 37 37 225 37 37 37 37 37 37 37 37 225 37 32 37 225 37 497 37 225 37 1179 37 225 14574 227 14574 133 14574 246 30457 238 14574 242 30457 229 14574 249 14574 134 14574 258 30457 228 14574 258 14574 114 14574 133 14574 232 36628 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838 20921 16623 13028 8372 1039 9446 40242 13852 2053 8949 12531 1520 10700 5881 9592 13299 914 31753 31359 9163 3202 35472 10397 439 4763 2583 330 102 1455 938 1182 2017 30 330 613 844 3654 49 330 63 646 3654 439 4621 1930 561 30 330 54 844 2124 1629 35993 49 2688 25 7709 312 25 94 62

4
models/ggml-vocab-starcoder.gguf.inp
View file

@ -1,3 +1,7 @@
ied 4 ½ months
__ggml_vocab_test__
Führer
__ggml_vocab_test__
__ggml_vocab_test__

2
models/ggml-vocab-starcoder.gguf.out
View file

@ -1,3 +1,5 @@
4850 244 57 244 162 159 17722
75 2022 3943 284
244
280

40
scripts/compare-llama-bench.py
View file

@ -1,5 +1,6 @@
#!/usr/bin/env python3
import logging
import argparse
import heapq
import sys
@ -11,9 +12,11 @@ try:
import git
from tabulate import tabulate
except ImportError as e:
print("ERROR: the following Python libraries are required: GitPython, tabulate.")
print("the following Python libraries are required: GitPython, tabulate.") # noqa: NP100
raise e
logger = logging.getLogger("compare-llama-bench")
# Properties by which to differentiate results per commit:
KEY_PROPERTIES = [
"cpu_info", "gpu_info", "n_gpu_layers", "main_gpu", "cuda", "opencl", "metal", "gpu_blas",
@ -94,8 +97,7 @@ parser.add_argument("-s", "--show", help=help_s)
known_args, unknown_args = parser.parse_known_args()
if unknown_args:
print(f"ERROR: Received unknown args: {unknown_args}.")
print()
logger.error(f"Received unknown args: {unknown_args}.")
parser.print_help()
sys.exit(1)
@ -108,8 +110,7 @@ if input_file is None:
input_file = sqlite_files[0]
if input_file is None:
print("ERROR: Cannot find a suitable input file, please provide one.")
print()
logger.error("Cannot find a suitable input file, please provide one.")
parser.print_help()
sys.exit(1)
@ -194,23 +195,19 @@ if known_args.baseline is not None:
hexsha8_baseline = get_commit_hexsha8(known_args.baseline)
name_baseline = known_args.baseline
if hexsha8_baseline is None:
print(f"ERROR: cannot find data for baseline={known_args.baseline}.")
logger.error(f"cannot find data for baseline={known_args.baseline}.")
sys.exit(1)
# Otherwise, search for the most recent parent of master for which there is data:
elif repo is not None:
hexsha8_baseline = find_parent_in_data(repo.heads.master.commit)
if hexsha8_baseline is None:
print("ERROR: No baseline was provided and did not find data for any master branch commits.")
print()
logger.error("No baseline was provided and did not find data for any master branch commits.")
parser.print_help()
sys.exit(1)
else:
print(
"ERROR: No baseline was provided and the current working directory "
"is not part of a git repository from which a baseline could be inferred."
)
print()
logger.error("No baseline was provided and the current working directory "
"is not part of a git repository from which a baseline could be inferred.")
parser.print_help()
sys.exit(1)
@ -227,7 +224,7 @@ if known_args.compare is not None:
hexsha8_compare = get_commit_hexsha8(known_args.compare)
name_compare = known_args.compare
if hexsha8_compare is None:
print(f"ERROR: cannot find data for compare={known_args.compare}.")
logger.error(f"cannot find data for compare={known_args.compare}.")
sys.exit(1)
# Otherwise, search for the commit for llama-bench was most recently run
# and that is not a parent of master:
@ -241,16 +238,12 @@ elif repo is not None:
break
if hexsha8_compare is None:
print("ERROR: No compare target was provided and did not find data for any non-master commits.")
print()
logger.error("No compare target was provided and did not find data for any non-master commits.")
parser.print_help()
sys.exit(1)
else:
print(
"ERROR: No compare target was provided and the current working directory "
"is not part of a git repository from which a compare target could be inferred."
)
print()
logger.error("No compare target was provided and the current working directory "
"is not part of a git repository from which a compare target could be inferred.\n")
parser.print_help()
sys.exit(1)
@ -284,8 +277,7 @@ if known_args.show is not None:
if prop not in KEY_PROPERTIES[:-2]: # Last two values are n_prompt, n_gen.
unknown_cols.append(prop)
if unknown_cols:
print(f"ERROR: Unknown values for --show: {', '.join(unknown_cols)}")
print()
logger.error(f"Unknown values for --show: {', '.join(unknown_cols)}")
parser.print_usage()
sys.exit(1)
rows_show = get_rows(show)
@ -369,7 +361,7 @@ if "gpu_info" in show:
headers = [PRETTY_NAMES[p] for p in show]
headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
print(tabulate(
logger.info(tabulate(
table,
headers=headers,
floatfmt=".2f",

66
scripts/gen-unicode-data.py Normal file
View file

@ -0,0 +1,66 @@
import regex
def cpt_to_utf8_str(cpt):
if cpt <= 0xFF:
return bytes([cpt, 0, 0, 0])
elif cpt <= 0xFFFF:
return bytes([cpt & 0xFF, cpt >> 8, 0, 0])
elif cpt <= 0xFFFFFF:
return bytes([cpt & 0xFF, (cpt >> 8) & 0xFF, (cpt >> 16) & 0xFF, 0])
else:
return bytes([cpt & 0xFF, (cpt >> 8) & 0xFF, (cpt >> 16) & 0xFF, cpt >> 24])
def is_match(codepoint, regex_expr):
try:
res = regex.match(regex_expr, cpt_to_utf8_str(codepoint).decode('utf-32'))
return res is not None
except Exception:
return False
def get_matches(regex_expr):
unicode_ranges = []
current_range = None
for codepoint in range(0x110000):
if is_match(codepoint, regex_expr):
if current_range is None:
current_range = [codepoint, codepoint]
else:
current_range[1] = codepoint
elif current_range is not None:
unicode_ranges.append(tuple(current_range))
current_range = None
if current_range is not None:
unicode_ranges.append(tuple(current_range))
return unicode_ranges
def print_cat(cat, ranges):
print("const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_{} = {{".format(cat))
cnt = 0
for start, end in ranges:
if cnt % 4 != 0:
print(" ", end="")
print("{{0x{:08X}, 0x{:08X}}},".format(start, end), end="")
if cnt % 4 == 3:
print("")
cnt += 1
if cnt % 4 != 0:
print("")
print("};")
print("")
print_cat("number", get_matches(r'\p{N}'))
print_cat("letter", get_matches(r'\p{L}'))
print_cat("whitespace", get_matches(r'\p{Z}'))
print_cat("accent_mark", get_matches(r'\p{M}'))
print_cat("punctuation", get_matches(r'\p{P}'))
print_cat("symbol", get_matches(r'\p{S}'))
print_cat("control", get_matches(r'\p{C}'))

14
scripts/run-with-preset.py
View file

@ -1,5 +1,6 @@
#!/usr/bin/env python3
import logging
import argparse
import os
import subprocess
@ -7,6 +8,8 @@ import sys
import yaml
logger = logging.getLogger("run-with-preset")
CLI_ARGS_MAIN_PERPLEXITY = [
"batch-size", "cfg-negative-prompt", "cfg-scale", "chunks", "color", "ctx-size", "escape",
"export", "file", "frequency-penalty", "grammar", "grammar-file", "hellaswag",
@ -56,6 +59,7 @@ parser.add_argument("-bin", "--binary", help="The binary to run.")
parser.add_argument("yaml_files", nargs="*",
help="Arbitrary number of YAML files from which to read preset values. "
"If two files specify the same values the later one will be used.")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
known_args, unknown_args = parser.parse_known_args()
@ -63,6 +67,8 @@ if not known_args.yaml_files and not unknown_args:
parser.print_help()
sys.exit(0)
logging.basicConfig(level=logging.DEBUG if known_args.verbose else logging.INFO)
props = dict()
for yaml_file in known_args.yaml_files:
@ -85,7 +91,7 @@ elif binary.lower().endswith("llama-bench"):
elif binary.lower().endswith("server"):
cli_args = CLI_ARGS_SERVER
else:
print(f"Unknown binary: {binary}")
logger.error(f"Unknown binary: {binary}")
sys.exit(1)
command_list = [binary]
@ -121,11 +127,11 @@ for cli_arg in cli_args:
num_unused = len(props)
if num_unused > 10:
print(f"The preset file contained a total of {num_unused} unused properties.")
logger.info(f"The preset file contained a total of {num_unused} unused properties.")
elif num_unused > 0:
print("The preset file contained the following unused properties:")
logger.info("The preset file contained the following unused properties:")
for prop, value in props.items():
print(f" {prop}: {value}")
logger.info(f" {prop}: {value}")
command_list += unknown_args

13
scripts/verify-checksum-models.py
View file

@ -1,8 +1,11 @@
#!/usr/bin/env python3
import logging
import os
import hashlib
logger = logging.getLogger("verify-checksum-models")
def sha256sum(file):
block_size = 16 * 1024 * 1024 # 16 MB block size
@ -27,7 +30,7 @@ hash_list_file = os.path.join(llama_path, "SHA256SUMS")
# Check if the hash list file exists
if not os.path.exists(hash_list_file):
print(f"Hash list file not found: {hash_list_file}")
logger.error(f"Hash list file not found: {hash_list_file}")
exit(1)
# Read the hash file content and split it into an array of lines
@ -46,7 +49,7 @@ for line in hash_list:
file_path = os.path.join(llama_path, filename)
# Informing user of the progress of the integrity check
print(f"Verifying the checksum of {file_path}")
logger.info(f"Verifying the checksum of {file_path}")
# Check if the file exists
if os.path.exists(file_path):
@ -73,9 +76,9 @@ for line in hash_list:
# Print column headers for results table
print("\n" + "filename".ljust(40) + "valid checksum".center(20) + "file missing".center(20))
print("-" * 80)
print("filename".ljust(40) + "valid checksum".center(20) + "file missing".center(20)) # noqa: NP100
print("-" * 80) # noqa: NP100
# Output the results as a table
for r in results:
print(f"{r['filename']:40} {r['valid checksum']:^20} {r['file missing']:^20}")
print(f"{r['filename']:40} {r['valid checksum']:^20} {r['file missing']:^20}") # noqa: NP100

6
tests/CMakeLists.txt
View file

@ -74,13 +74,15 @@ llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-spm ARGS ${CMAKE
llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-bpe ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
llama_test(test-tokenizer-0 NAME test-tokenizer-0-phi-3 ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-phi-3.gguf)
llama_test(test-tokenizer-0 NAME test-tokenizer-0-falcon ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-llm.gguf)
llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-coder.gguf)
llama_test(test-tokenizer-0 NAME test-tokenizer-0-bert-bge ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bert-bge.gguf)
# TODO: enable when fixed
# https://github.com/ggerganov/llama.cpp/pull/7036
#llama_test(test-tokenizer-0 NAME test-tokenizer-0-mpt ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
#llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-llm.gguf)
#llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-coder.gguf)
llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
llama_test(test-tokenizer-0 NAME test-tokenizer-0-gpt-2 ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf)
llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
# build test-tokenizer-1-bpe target once and add many tests
add_executable(test-tokenizer-1-bpe test-tokenizer-1-bpe.cpp)

117
tests/test-tokenizer-0-bpe.py
View file

@ -1,117 +0,0 @@
# tests with BPE tokenizer
#
# sample usage:
#
# python3 tests/test-tokenizer-0-bpe.py ~/Data/huggingface/Meta-Llama-3-8B-Instruct/
# python3 tests/test-tokenizer-0-bpe.py ~/Data/huggingface/falcon-7b/
# python3 tests/test-tokenizer-0-bpe.py ~/Data/huggingface/deepseek-coder-6.7b-instruct/
#
import argparse
from transformers import AutoTokenizer
parser = argparse.ArgumentParser()
parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
parser.add_argument("--fname-tok", help="path to a text file to tokenize")
args = parser.parse_args()
dir_tokenizer = args.dir_tokenizer
tokenizer = AutoTokenizer.from_pretrained(dir_tokenizer)
tests = [
"",
" ",
" ",
" ",
"\t",
"\n",
"\n\n",
"\n\n\n",
"\t\n",
"Hello world",
" Hello world",
"Hello World",
" Hello World",
" Hello World!",
"Hello, world!",
" Hello, world!",
" this is 🦙.cpp",
"w048 7tuijk dsdfhu",
"нещо на Български",
"កាន់តែពិសេសអាចខលចេញ",
"🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
"Hello",
" Hello",
" Hello",
" Hello",
" Hello",
" Hello\n Hello",
" (",
"\n =",
"' era",
"Hello, y'all! How are you 😁 ?我想在apple工作1314151天～",
"3",
"33",
"333",
"3333",
"33333",
"333333",
"3333333",
"33333333",
"333333333",
]
for text in tests:
print('text: ', text)
print(tokenizer.encode(text))
print(tokenizer.decode(tokenizer.encode(text)))
print("\n\ntests for C++:\n")
for text in tests:
res = tokenizer.encode(text)
k = text.replace('\n', '\\n')
k = k.replace('\t', '\\t')
k = '"' + k + '"'
print("{ %-24s, { " % k, end='')
for x in res:
print("%7d," % x, end='')
print(" }, },")
print(tokenizer.encode('hello'))
print(tokenizer.encode('world'))
print(tokenizer.encode(' world'))
print(tokenizer.encode('hello world'))
fname_tok = args.fname_tok
if fname_tok:
print('tokenizing file: ', fname_tok)
fname_out = fname_tok + '.tok'
with open(fname_tok, 'r', encoding='utf-8') as f:
lines = f.readlines()
s = ''.join(lines)
res = tokenizer.encode(s)
# write to file
with open(fname_out, 'w', encoding='utf-8') as f:
for x in res:
# LLaMA v3 for some reason strips the space for these tokens (and others)
# if x == 662:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 1174:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 2564:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 758:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 949:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 5354:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# else:
# f.write(str(x) + ' \'' + tokenizer.decode(x) + '\'\n')
f.write(str(x) + ' \'' + tokenizer.decode(x).strip() + '\'\n')
print('len(res): ', len(res))
print('len(lines): ', len(lines))
print('results written to: ', fname_out)

114
tests/test-tokenizer-0-spm.py
View file

@ -1,114 +0,0 @@
# tests with SPM tokenizer
#
# sample usage:
#
# python3 tests/test-tokenizer-0-spm.py ~/Data/huggingface/Llama-2-7b-hf/
# python3 tests/test-tokenizer-0-spm.py ~/Data/huggingface/CodeLlama-34b-Instruct-hf/
#
import argparse
from sentencepiece import SentencePieceProcessor
parser = argparse.ArgumentParser()
parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
parser.add_argument("--fname-tok", help="path to a text file to tokenize")
args = parser.parse_args()
dir_tokenizer = args.dir_tokenizer
tokenizer = SentencePieceProcessor(dir_tokenizer + '/tokenizer.model')
tests = [
"",
" ",
" ",
" ",
"\t",
"\n",
"\n\n",
"\n\n\n",
"\t\n",
"Hello world",
" Hello world",
"Hello World",
" Hello World",
" Hello World!",
"Hello, world!",
" Hello, world!",
" this is 🦙.cpp",
"w048 7tuijk dsdfhu",
"нещо на Български",
"កាន់តែពិសេសអាចខលចេញ",
"🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
"Hello",
" Hello",
" Hello",
" Hello",
" Hello",
" Hello\n Hello",
" (",
"\n =",
"' era",
"Hello, y'all! How are you 😁 ?我想在apple工作1314151天～",
"3",
"33",
"333",
"3333",
"33333",
"333333",
"3333333",
"33333333",
"333333333",
]
for text in tests:
print('text: ', text)
print('\nwith bos:')
print(tokenizer.encode(text, add_bos=True))
print(tokenizer.decode(tokenizer.encode(text, add_bos=True)))
print('\nwithout bos:')
print(tokenizer.encode(text, add_bos=False))
print(tokenizer.decode(tokenizer.encode(text, add_bos=False)))
print("'" + tokenizer.id_to_piece(15043) + "'") # '_Hello'
print("'" + tokenizer.id_to_piece(29871) + "'") # '_'
print("'" + tokenizer.decode([15043]) + "'") # 'Hello'
print("'" + tokenizer.decode([15043, 15043]) + "'") # 'Hello Hello'
print("'" + tokenizer.decode([29871, 15043]) + "'") # ' Hello'
print("'" + tokenizer.decode([29871, 15043, 29871, 15043]) + "'") # ' Hello Hello'
print("\n\ntests for C++:\n")
for text in tests:
res = tokenizer.encode(text, add_bos=False)
k = text.replace('\n', '\\n')
k = k.replace('\t', '\\t')
k = '"' + k + '"'
print("{ %-24s, { " % k, end='')
for x in res:
print("%7d," % x, end='')
print(" }, },")
print(tokenizer.encode('hello'))
print(tokenizer.encode('world'))
print(tokenizer.encode(' world'))
print(tokenizer.encode('hello world'))
fname_tok = args.fname_tok
if fname_tok:
print('tokenizing file: ', fname_tok)
fname_out = fname_tok + '.tok'
with open(fname_tok, 'r', encoding='utf-8') as f:
lines = f.readlines()
s = ''.join(lines)
res = tokenizer.encode(s, add_bos=True)
# write to file
with open(fname_out, 'w', encoding='utf-8') as f:
for x in res:
f.write(str(x) + ' \'' + tokenizer.decode(x) + '\'\n')
print('len(res): ', len(res))
print('len(lines): ', len(lines))
print('results written to: ', fname_out)

39
tests/test-tokenizer-0.cpp
View file

@ -55,8 +55,10 @@
// return _k_tests;
//}
static std::map<std::string, std::vector<llama_token>> read_tests(const std::string & fname_inp, const std::string & fname_out) {
std::map<std::string, std::vector<llama_token>> tests;
using llama_tests = std::map<std::string, std::vector<llama_token>>;
static llama_tests read_tests(const std::string & fname_inp, const std::string & fname_out) {
llama_tests tests;
std::ifstream ifs_inp(fname_inp);
if (!ifs_inp) {
@ -175,13 +177,21 @@ int main(int argc, char **argv) {
bool success = true;
const auto k_tests = read_tests(fname_inp, fname_out);
if (k_tests.empty()) {
fprintf(stderr, "%s : error: no tests found\n", __func__);
return 1;
const auto k_tests = [&]() -> llama_tests {
if (!fname_text.empty()) {
return {};
}
const auto res = read_tests(fname_inp, fname_out);
if (res.empty()) {
fprintf(stderr, "%s : error: no tests found\n", __func__);
exit(1);
}
return res;
}();
const bool add_special = false;
for (const auto & test_kv : k_tests) {
@ -238,7 +248,17 @@ int main(int argc, char **argv) {
fprintf(stderr, "%s : text size: %zu\n", __func__, text.size());
const std::vector<llama_token> res = llama_tokenize(ctx, text, add_special);
std::vector<llama_token> res;
{
const auto t_start = ggml_time_us();
res = llama_tokenize(ctx, text, add_special);
const auto t_end = ggml_time_us();
fprintf(stderr, "%s : tokenized in %.3f ms (cpp)\n", __func__, (t_end - t_start) / 1000.0);
}
fprintf(stderr, "%s : tokens: %zu\n", __func__, res.size());
@ -252,7 +272,8 @@ int main(int argc, char **argv) {
}
for (const auto & tok : res) {
ofs << tok << " '" << string_strip(llama_detokenize_bpe(ctx, std::vector<int>{tok})) << "'" << std::endl;
//ofs << tok << " '" << string_strip(llama_detokenize_bpe(ctx, std::vector<int>{tok})) << "'" << std::endl;
ofs << tok << "\n";
}
}

46
tests/test-tokenizer-0.py Normal file
View file

@ -0,0 +1,46 @@
import time
import argparse
from transformers import AutoTokenizer
parser = argparse.ArgumentParser()
parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
parser.add_argument("--fname-tok", help="path to a text file to tokenize", required=True)
args = parser.parse_args()
dir_tokenizer = args.dir_tokenizer
fname_tok = args.fname_tok
tokenizer = AutoTokenizer.from_pretrained(dir_tokenizer)
print('tokenizing file: ', fname_tok)
fname_out = fname_tok + '.tok'
with open(fname_tok, 'r', encoding='utf-8') as f:
lines = f.readlines()
s = ''.join(lines)
t_start = time.time()
res = tokenizer.encode(s, add_special_tokens=False)
t_end = time.time()
print('\nmain : tokenized in', "{:.3f}".format(1000.0 * (t_end - t_start)), 'ms (py)')
with open(fname_out, 'w', encoding='utf-8') as f:
for x in res:
# LLaMA v3 for some reason strips the space for these tokens (and others)
# if x == 662:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 1174:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 2564:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 758:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 949:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# elif x == 5354:
# f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n')
# else:
# f.write(str(x) + ' \'' + tokenizer.decode(x) + '\'\n')
# f.write(str(x) + ' \'' + tokenizer.decode(x).strip() + '\'\n')
f.write(str(x) + '\n')
print('len(res): ', len(res))
print('len(lines): ', len(lines))
print('results written to: ', fname_out)

34
tests/test-tokenizer-0.sh Executable file
View file

@ -0,0 +1,34 @@
#!/bin/bash
#
# Usage:
#
# test-tokenizer-0.sh <name> <input>
#
if [ $# -ne 2 ]; then
printf "Usage: $0 <name> <input>\n"
exit 1
fi
name=$1
input=$2
make -j tests/test-tokenizer-0
printf "Testing %s on %s ...\n" $name $input
python3 ./tests/test-tokenizer-0.py ./models/tokenizers/$name --fname-tok $input > /tmp/test-tokenizer-0-$name-py.log 2>&1
cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in"
./tests/test-tokenizer-0 ./models/ggml-vocab-$name.gguf $input > /tmp/test-tokenizer-0-$name-cpp.log 2>&1
cat /tmp/test-tokenizer-0-$name-cpp.log | grep "tokenized in"
diff $input.tok $input.tokcpp > /dev/null 2>&1
if [ $? -eq 0 ]; then
printf "Tokenization is correct!\n"
else
diff $input.tok $input.tokcpp | head -n 32
printf "Tokenization differs!\n"
fi

874
unicode-data.cpp
View file

@ -5,27 +5,47 @@
#include <utility>
#include <vector>
const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_digit = {
{0x00000030, 0x00000039}, {0x000000B2, 0x000000B3}, {0x000000B9, 0x000000B9}, {0x00000660, 0x00000669},
{0x000006F0, 0x000006F9}, {0x000007C0, 0x000007C9}, {0x00000966, 0x0000096F}, {0x000009E6, 0x000009EF},
{0x00000A66, 0x00000A6F}, {0x00000AE6, 0x00000AEF}, {0x00000B66, 0x00000B6F}, {0x00000BE6, 0x00000BEF},
{0x00000C66, 0x00000C6F}, {0x00000CE6, 0x00000CEF}, {0x00000D66, 0x00000D6F}, {0x00000DE6, 0x00000DEF},
{0x00000E50, 0x00000E59}, {0x00000ED0, 0x00000ED9}, {0x00000F20, 0x00000F29}, {0x00001040, 0x00001049},
{0x00001090, 0x00001099}, {0x00001369, 0x00001371}, {0x000017E0, 0x000017E9}, {0x00001810, 0x00001819},
{0x00001946, 0x0000194F}, {0x000019D0, 0x000019DA}, {0x00001A80, 0x00001A89}, {0x00001A90, 0x00001A99},
{0x00001B50, 0x00001B59}, {0x00001BB0, 0x00001BB9}, {0x00001C40, 0x00001C49}, {0x00001C50, 0x00001C59},
{0x00002070, 0x00002070}, {0x00002074, 0x00002079}, {0x00002080, 0x00002089}, {0x00002460, 0x00002468},
{0x00002474, 0x0000247C}, {0x00002488, 0x00002490}, {0x000024EA, 0x000024EA}, {0x000024F5, 0x000024FD},
{0x000024FF, 0x000024FF}, {0x00002776, 0x0000277E}, {0x00002780, 0x00002788}, {0x0000278A, 0x00002792},
{0x0000A620, 0x0000A629}, {0x0000A8D0, 0x0000A8D9}, {0x0000A900, 0x0000A909}, {0x0000A9D0, 0x0000A9D9},
{0x0000A9F0, 0x0000A9F9}, {0x0000AA50, 0x0000AA59}, {0x0000ABF0, 0x0000ABF9}, {0x0000FF10, 0x0000FF19},
{0x000104A0, 0x000104A9}, {0x00010A40, 0x00010A43}, {0x00010D30, 0x00010D39}, {0x00010E60, 0x00010E68},
{0x00011052, 0x0001105A}, {0x00011066, 0x0001106F}, {0x000110F0, 0x000110F9}, {0x00011136, 0x0001113F},
{0x000111D0, 0x000111D9}, {0x000112F0, 0x000112F9}, {0x00011450, 0x00011459}, {0x000114D0, 0x000114D9},
{0x00011650, 0x00011659}, {0x000116C0, 0x000116C9}, {0x00011730, 0x00011739}, {0x000118E0, 0x000118E9},
{0x00011950, 0x00011959}, {0x00011C50, 0x00011C59}, {0x00011D50, 0x00011D59}, {0x00011DA0, 0x00011DA9},
{0x00016A60, 0x00016A69}, {0x00016B50, 0x00016B59}, {0x0001D7CE, 0x0001D7FF}, {0x0001E140, 0x0001E149},
{0x0001E2F0, 0x0001E2F9}, {0x0001E950, 0x0001E959}, {0x0001F100, 0x0001F10A}, {0x0001FBF0, 0x0001FBF9},
// generated with scripts/gen-unicode-data.py
//
// TODO: generate unicode_map_lowercase
// TODO: generate unicode_map_nfd
const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_number = {
{0x00000030, 0x00000039}, {0x000000B2, 0x000000B3}, {0x000000B9, 0x000000B9}, {0x000000BC, 0x000000BE},
{0x00000660, 0x00000669}, {0x000006F0, 0x000006F9}, {0x000007C0, 0x000007C9}, {0x00000966, 0x0000096F},
{0x000009E6, 0x000009EF}, {0x000009F4, 0x000009F9}, {0x00000A66, 0x00000A6F}, {0x00000AE6, 0x00000AEF},
{0x00000B66, 0x00000B6F}, {0x00000B72, 0x00000B77}, {0x00000BE6, 0x00000BF2}, {0x00000C66, 0x00000C6F},
{0x00000C78, 0x00000C7E}, {0x00000CE6, 0x00000CEF}, {0x00000D58, 0x00000D5E}, {0x00000D66, 0x00000D78},
{0x00000DE6, 0x00000DEF}, {0x00000E50, 0x00000E59}, {0x00000ED0, 0x00000ED9}, {0x00000F20, 0x00000F33},
{0x00001040, 0x00001049}, {0x00001090, 0x00001099}, {0x00001369, 0x0000137C}, {0x000016EE, 0x000016F0},
{0x000017E0, 0x000017E9}, {0x000017F0, 0x000017F9}, {0x00001810, 0x00001819}, {0x00001946, 0x0000194F},
{0x000019D0, 0x000019DA}, {0x00001A80, 0x00001A89}, {0x00001A90, 0x00001A99}, {0x00001B50, 0x00001B59},
{0x00001BB0, 0x00001BB9}, {0x00001C40, 0x00001C49}, {0x00001C50, 0x00001C59}, {0x00002070, 0x00002070},
{0x00002074, 0x00002079}, {0x00002080, 0x00002089}, {0x00002150, 0x00002182}, {0x00002185, 0x00002189},
{0x00002460, 0x0000249B}, {0x000024EA, 0x000024FF}, {0x00002776, 0x00002793}, {0x00002CFD, 0x00002CFD},
{0x00003007, 0x00003007}, {0x00003021, 0x00003029}, {0x00003038, 0x0000303A}, {0x00003192, 0x00003195},
{0x00003220, 0x00003229}, {0x00003248, 0x0000324F}, {0x00003251, 0x0000325F}, {0x00003280, 0x00003289},
{0x000032B1, 0x000032BF}, {0x0000A620, 0x0000A629}, {0x0000A6E6, 0x0000A6EF}, {0x0000A830, 0x0000A835},
{0x0000A8D0, 0x0000A8D9}, {0x0000A900, 0x0000A909}, {0x0000A9D0, 0x0000A9D9}, {0x0000A9F0, 0x0000A9F9},
{0x0000AA50, 0x0000AA59}, {0x0000ABF0, 0x0000ABF9}, {0x0000FF10, 0x0000FF19}, {0x00010107, 0x00010133},
{0x00010140, 0x00010178}, {0x0001018A, 0x0001018B}, {0x000102E1, 0x000102FB}, {0x00010320, 0x00010323},
{0x00010341, 0x00010341}, {0x0001034A, 0x0001034A}, {0x000103D1, 0x000103D5}, {0x000104A0, 0x000104A9},
{0x00010858, 0x0001085F}, {0x00010879, 0x0001087F}, {0x000108A7, 0x000108AF}, {0x000108FB, 0x000108FF},
{0x00010916, 0x0001091B}, {0x000109BC, 0x000109BD}, {0x000109C0, 0x000109CF}, {0x000109D2, 0x000109FF},
{0x00010A40, 0x00010A48}, {0x00010A7D, 0x00010A7E}, {0x00010A9D, 0x00010A9F}, {0x00010AEB, 0x00010AEF},
{0x00010B58, 0x00010B5F}, {0x00010B78, 0x00010B7F}, {0x00010BA9, 0x00010BAF}, {0x00010CFA, 0x00010CFF},
{0x00010D30, 0x00010D39}, {0x00010E60, 0x00010E7E}, {0x00010F1D, 0x00010F26}, {0x00010F51, 0x00010F54},
{0x00010FC5, 0x00010FCB}, {0x00011052, 0x0001106F}, {0x000110F0, 0x000110F9}, {0x00011136, 0x0001113F},
{0x000111D0, 0x000111D9}, {0x000111E1, 0x000111F4}, {0x000112F0, 0x000112F9}, {0x00011450, 0x00011459},
{0x000114D0, 0x000114D9}, {0x00011650, 0x00011659}, {0x000116C0, 0x000116C9}, {0x00011730, 0x0001173B},
{0x000118E0, 0x000118F2}, {0x00011950, 0x00011959}, {0x00011C50, 0x00011C6C}, {0x00011D50, 0x00011D59},
{0x00011DA0, 0x00011DA9}, {0x00011F50, 0x00011F59}, {0x00011FC0, 0x00011FD4}, {0x00012400, 0x0001246E},
{0x00016A60, 0x00016A69}, {0x00016AC0, 0x00016AC9}, {0x00016B50, 0x00016B59}, {0x00016B5B, 0x00016B61},
{0x00016E80, 0x00016E96}, {0x0001D2C0, 0x0001D2D3}, {0x0001D2E0, 0x0001D2F3}, {0x0001D360, 0x0001D378},
{0x0001D7CE, 0x0001D7FF}, {0x0001E140, 0x0001E149}, {0x0001E2F0, 0x0001E2F9}, {0x0001E4F0, 0x0001E4F9},
{0x0001E8C7, 0x0001E8CF}, {0x0001E950, 0x0001E959}, {0x0001EC71, 0x0001ECAB}, {0x0001ECAD, 0x0001ECAF},
{0x0001ECB1, 0x0001ECB4}, {0x0001ED01, 0x0001ED2D}, {0x0001ED2F, 0x0001ED3D}, {0x0001F100, 0x0001F10C},
{0x0001FBF0, 0x0001FBF9},
};
const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter = {
@ -41,73 +61,73 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter = {
{0x00000710, 0x00000710}, {0x00000712, 0x0000072F}, {0x0000074D, 0x000007A5}, {0x000007B1, 0x000007B1},
{0x000007CA, 0x000007EA}, {0x000007F4, 0x000007F5}, {0x000007FA, 0x000007FA}, {0x00000800, 0x00000815},
{0x0000081A, 0x0000081A}, {0x00000824, 0x00000824}, {0x00000828, 0x00000828}, {0x00000840, 0x00000858},
{0x00000860, 0x0000086A}, {0x000008A0, 0x000008B4}, {0x000008B6, 0x000008C7}, {0x00000904, 0x00000939},
{0x0000093D, 0x0000093D}, {0x00000950, 0x00000950}, {0x00000958, 0x00000961}, {0x00000971, 0x00000980},
{0x00000985, 0x0000098C}, {0x0000098F, 0x00000990}, {0x00000993, 0x000009A8}, {0x000009AA, 0x000009B0},
{0x000009B2, 0x000009B2}, {0x000009B6, 0x000009B9}, {0x000009BD, 0x000009BD}, {0x000009CE, 0x000009CE},
{0x000009DC, 0x000009DD}, {0x000009DF, 0x000009E1}, {0x000009F0, 0x000009F1}, {0x000009FC, 0x000009FC},
{0x00000A05, 0x00000A0A}, {0x00000A0F, 0x00000A10}, {0x00000A13, 0x00000A28}, {0x00000A2A, 0x00000A30},
{0x00000A32, 0x00000A33}, {0x00000A35, 0x00000A36}, {0x00000A38, 0x00000A39}, {0x00000A59, 0x00000A5C},
{0x00000A5E, 0x00000A5E}, {0x00000A72, 0x00000A74}, {0x00000A85, 0x00000A8D}, {0x00000A8F, 0x00000A91},
{0x00000A93, 0x00000AA8}, {0x00000AAA, 0x00000AB0}, {0x00000AB2, 0x00000AB3}, {0x00000AB5, 0x00000AB9},
{0x00000ABD, 0x00000ABD}, {0x00000AD0, 0x00000AD0}, {0x00000AE0, 0x00000AE1}, {0x00000AF9, 0x00000AF9},
{0x00000B05, 0x00000B0C}, {0x00000B0F, 0x00000B10}, {0x00000B13, 0x00000B28}, {0x00000B2A, 0x00000B30},
{0x00000B32, 0x00000B33}, {0x00000B35, 0x00000B39}, {0x00000B3D, 0x00000B3D}, {0x00000B5C, 0x00000B5D},
{0x00000B5F, 0x00000B61}, {0x00000B71, 0x00000B71}, {0x00000B83, 0x00000B83}, {0x00000B85, 0x00000B8A},
{0x00000B8E, 0x00000B90}, {0x00000B92, 0x00000B95}, {0x00000B99, 0x00000B9A}, {0x00000B9C, 0x00000B9C},
{0x00000B9E, 0x00000B9F}, {0x00000BA3, 0x00000BA4}, {0x00000BA8, 0x00000BAA}, {0x00000BAE, 0x00000BB9},
{0x00000BD0, 0x00000BD0}, {0x00000C05, 0x00000C0C}, {0x00000C0E, 0x00000C10}, {0x00000C12, 0x00000C28},
{0x00000C2A, 0x00000C39}, {0x00000C3D, 0x00000C3D}, {0x00000C58, 0x00000C5A}, {0x00000C60, 0x00000C61},
{0x00000C80, 0x00000C80}, {0x00000C85, 0x00000C8C}, {0x00000C8E, 0x00000C90}, {0x00000C92, 0x00000CA8},
{0x00000CAA, 0x00000CB3}, {0x00000CB5, 0x00000CB9}, {0x00000CBD, 0x00000CBD}, {0x00000CDE, 0x00000CDE},
{0x00000CE0, 0x00000CE1}, {0x00000CF1, 0x00000CF2}, {0x00000D04, 0x00000D0C}, {0x00000D0E, 0x00000D10},
{0x00000D12, 0x00000D3A}, {0x00000D3D, 0x00000D3D}, {0x00000D4E, 0x00000D4E}, {0x00000D54, 0x00000D56},
{0x00000D5F, 0x00000D61}, {0x00000D7A, 0x00000D7F}, {0x00000D85, 0x00000D96}, {0x00000D9A, 0x00000DB1},
{0x00000DB3, 0x00000DBB}, {0x00000DBD, 0x00000DBD}, {0x00000DC0, 0x00000DC6}, {0x00000E01, 0x00000E30},
{0x00000E32, 0x00000E33}, {0x00000E40, 0x00000E46}, {0x00000E81, 0x00000E82}, {0x00000E84, 0x00000E84},
{0x00000E86, 0x00000E8A}, {0x00000E8C, 0x00000EA3}, {0x00000EA5, 0x00000EA5}, {0x00000EA7, 0x00000EB0},
{0x00000EB2, 0x00000EB3}, {0x00000EBD, 0x00000EBD}, {0x00000EC0, 0x00000EC4}, {0x00000EC6, 0x00000EC6},
{0x00000EDC, 0x00000EDF}, {0x00000F00, 0x00000F00}, {0x00000F40, 0x00000F47}, {0x00000F49, 0x00000F6C},
{0x00000F88, 0x00000F8C}, {0x00001000, 0x0000102A}, {0x0000103F, 0x0000103F}, {0x00001050, 0x00001055},
{0x0000105A, 0x0000105D}, {0x00001061, 0x00001061}, {0x00001065, 0x00001066}, {0x0000106E, 0x00001070},
{0x00001075, 0x00001081}, {0x0000108E, 0x0000108E}, {0x000010A0, 0x000010C5}, {0x000010C7, 0x000010C7},
{0x000010CD, 0x000010CD}, {0x000010D0, 0x000010FA}, {0x000010FC, 0x00001248}, {0x0000124A, 0x0000124D},
{0x00001250, 0x00001256}, {0x00001258, 0x00001258}, {0x0000125A, 0x0000125D}, {0x00001260, 0x00001288},
{0x0000128A, 0x0000128D}, {0x00001290, 0x000012B0}, {0x000012B2, 0x000012B5}, {0x000012B8, 0x000012BE},
{0x000012C0, 0x000012C0}, {0x000012C2, 0x000012C5}, {0x000012C8, 0x000012D6}, {0x000012D8, 0x00001310},
{0x00001312, 0x00001315}, {0x00001318, 0x0000135A}, {0x00001380, 0x0000138F}, {0x000013A0, 0x000013F5},
{0x000013F8, 0x000013FD}, {0x00001401, 0x0000166C}, {0x0000166F, 0x0000167F}, {0x00001681, 0x0000169A},
{0x000016A0, 0x000016EA}, {0x000016F1, 0x000016F8}, {0x00001700, 0x0000170C}, {0x0000170E, 0x00001711},
{0x00001720, 0x00001731}, {0x00001740, 0x00001751}, {0x00001760, 0x0000176C}, {0x0000176E, 0x00001770},
{0x00001780, 0x000017B3}, {0x000017D7, 0x000017D7}, {0x000017DC, 0x000017DC}, {0x00001820, 0x00001878},
{0x00001880, 0x00001884}, {0x00001887, 0x000018A8}, {0x000018AA, 0x000018AA}, {0x000018B0, 0x000018F5},
{0x00001900, 0x0000191E}, {0x00001950, 0x0000196D}, {0x00001970, 0x00001974}, {0x00001980, 0x000019AB},
{0x000019B0, 0x000019C9}, {0x00001A00, 0x00001A16}, {0x00001A20, 0x00001A54}, {0x00001AA7, 0x00001AA7},
{0x00001B05, 0x00001B33}, {0x00001B45, 0x00001B4B}, {0x00001B83, 0x00001BA0}, {0x00001BAE, 0x00001BAF},
{0x00001BBA, 0x00001BE5}, {0x00001C00, 0x00001C23}, {0x00001C4D, 0x00001C4F}, {0x00001C5A, 0x00001C7D},
{0x00001C80, 0x00001C88}, {0x00001C90, 0x00001CBA}, {0x00001CBD, 0x00001CBF}, {0x00001CE9, 0x00001CEC},
{0x00001CEE, 0x00001CF3}, {0x00001CF5, 0x00001CF6}, {0x00001CFA, 0x00001CFA}, {0x00001D00, 0x00001DBF},
{0x00001E00, 0x00001F15}, {0x00001F18, 0x00001F1D}, {0x00001F20, 0x00001F45}, {0x00001F48, 0x00001F4D},
{0x00001F50, 0x00001F57}, {0x00001F59, 0x00001F59}, {0x00001F5B, 0x00001F5B}, {0x00001F5D, 0x00001F5D},
{0x00001F5F, 0x00001F7D}, {0x00001F80, 0x00001FB4}, {0x00001FB6, 0x00001FBC}, {0x00001FBE, 0x00001FBE},
{0x00001FC2, 0x00001FC4}, {0x00001FC6, 0x00001FCC}, {0x00001FD0, 0x00001FD3}, {0x00001FD6, 0x00001FDB},
{0x00001FE0, 0x00001FEC}, {0x00001FF2, 0x00001FF4}, {0x00001FF6, 0x00001FFC}, {0x00002071, 0x00002071},
{0x0000207F, 0x0000207F}, {0x00002090, 0x0000209C}, {0x00002102, 0x00002102}, {0x00002107, 0x00002107},
{0x0000210A, 0x00002113}, {0x00002115, 0x00002115}, {0x00002119, 0x0000211D}, {0x00002124, 0x00002124},
{0x00002126, 0x00002126}, {0x00002128, 0x00002128}, {0x0000212A, 0x0000212D}, {0x0000212F, 0x00002139},
{0x0000213C, 0x0000213F}, {0x00002145, 0x00002149}, {0x0000214E, 0x0000214E}, {0x00002183, 0x00002184},
{0x00002C00, 0x00002C2E}, {0x00002C30, 0x00002C5E}, {0x00002C60, 0x00002CE4}, {0x00002CEB, 0x00002CEE},
{0x00002CF2, 0x00002CF3}, {0x00002D00, 0x00002D25}, {0x00002D27, 0x00002D27}, {0x00002D2D, 0x00002D2D},
{0x00002D30, 0x00002D67}, {0x00002D6F, 0x00002D6F}, {0x00002D80, 0x00002D96}, {0x00002DA0, 0x00002DA6},
{0x00002DA8, 0x00002DAE}, {0x00002DB0, 0x00002DB6}, {0x00002DB8, 0x00002DBE}, {0x00002DC0, 0x00002DC6},
{0x00002DC8, 0x00002DCE}, {0x00002DD0, 0x00002DD6}, {0x00002DD8, 0x00002DDE}, {0x00002E2F, 0x00002E2F},
{0x00003005, 0x00003006}, {0x00003031, 0x00003035}, {0x0000303B, 0x0000303C}, {0x00003041, 0x00003096},
{0x0000309D, 0x0000309F}, {0x000030A1, 0x000030FA}, {0x000030FC, 0x000030FF}, {0x00003105, 0x0000312F},
{0x00003131, 0x0000318E}, {0x000031A0, 0x000031BF}, {0x000031F0, 0x000031FF}, {0x00003400, 0x00004DBF},
{0x00004E00, 0x00009FFC}, {0x0000A000, 0x0000A48C}, {0x0000A4D0, 0x0000A4FD}, {0x0000A500, 0x0000A60C},
{0x0000A610, 0x0000A61F}, {0x0000A62A, 0x0000A62B}, {0x0000A640, 0x0000A66E}, {0x0000A67F, 0x0000A69D},
{0x0000A6A0, 0x0000A6E5}, {0x0000A717, 0x0000A71F}, {0x0000A722, 0x0000A788}, {0x0000A78B, 0x0000A7BF},
{0x0000A7C2, 0x0000A7CA}, {0x0000A7F5, 0x0000A801}, {0x0000A803, 0x0000A805}, {0x0000A807, 0x0000A80A},
{0x00000860, 0x0000086A}, {0x00000870, 0x00000887}, {0x00000889, 0x0000088E}, {0x000008A0, 0x000008C9},
{0x00000904, 0x00000939}, {0x0000093D, 0x0000093D}, {0x00000950, 0x00000950}, {0x00000958, 0x00000961},
{0x00000971, 0x00000980}, {0x00000985, 0x0000098C}, {0x0000098F, 0x00000990}, {0x00000993, 0x000009A8},
{0x000009AA, 0x000009B0}, {0x000009B2, 0x000009B2}, {0x000009B6, 0x000009B9}, {0x000009BD, 0x000009BD},
{0x000009CE, 0x000009CE}, {0x000009DC, 0x000009DD}, {0x000009DF, 0x000009E1}, {0x000009F0, 0x000009F1},
{0x000009FC, 0x000009FC}, {0x00000A05, 0x00000A0A}, {0x00000A0F, 0x00000A10}, {0x00000A13, 0x00000A28},
{0x00000A2A, 0x00000A30}, {0x00000A32, 0x00000A33}, {0x00000A35, 0x00000A36}, {0x00000A38, 0x00000A39},
{0x00000A59, 0x00000A5C}, {0x00000A5E, 0x00000A5E}, {0x00000A72, 0x00000A74}, {0x00000A85, 0x00000A8D},
{0x00000A8F, 0x00000A91}, {0x00000A93, 0x00000AA8}, {0x00000AAA, 0x00000AB0}, {0x00000AB2, 0x00000AB3},
{0x00000AB5, 0x00000AB9}, {0x00000ABD, 0x00000ABD}, {0x00000AD0, 0x00000AD0}, {0x00000AE0, 0x00000AE1},
{0x00000AF9, 0x00000AF9}, {0x00000B05, 0x00000B0C}, {0x00000B0F, 0x00000B10}, {0x00000B13, 0x00000B28},
{0x00000B2A, 0x00000B30}, {0x00000B32, 0x00000B33}, {0x00000B35, 0x00000B39}, {0x00000B3D, 0x00000B3D},
{0x00000B5C, 0x00000B5D}, {0x00000B5F, 0x00000B61}, {0x00000B71, 0x00000B71}, {0x00000B83, 0x00000B83},
{0x00000B85, 0x00000B8A}, {0x00000B8E, 0x00000B90}, {0x00000B92, 0x00000B95}, {0x00000B99, 0x00000B9A},
{0x00000B9C, 0x00000B9C}, {0x00000B9E, 0x00000B9F}, {0x00000BA3, 0x00000BA4}, {0x00000BA8, 0x00000BAA},
{0x00000BAE, 0x00000BB9}, {0x00000BD0, 0x00000BD0}, {0x00000C05, 0x00000C0C}, {0x00000C0E, 0x00000C10},
{0x00000C12, 0x00000C28}, {0x00000C2A, 0x00000C39}, {0x00000C3D, 0x00000C3D}, {0x00000C58, 0x00000C5A},
{0x00000C5D, 0x00000C5D}, {0x00000C60, 0x00000C61}, {0x00000C80, 0x00000C80}, {0x00000C85, 0x00000C8C},
{0x00000C8E, 0x00000C90}, {0x00000C92, 0x00000CA8}, {0x00000CAA, 0x00000CB3}, {0x00000CB5, 0x00000CB9},
{0x00000CBD, 0x00000CBD}, {0x00000CDD, 0x00000CDE}, {0x00000CE0, 0x00000CE1}, {0x00000CF1, 0x00000CF2},
{0x00000D04, 0x00000D0C}, {0x00000D0E, 0x00000D10}, {0x00000D12, 0x00000D3A}, {0x00000D3D, 0x00000D3D},
{0x00000D4E, 0x00000D4E}, {0x00000D54, 0x00000D56}, {0x00000D5F, 0x00000D61}, {0x00000D7A, 0x00000D7F},
{0x00000D85, 0x00000D96}, {0x00000D9A, 0x00000DB1}, {0x00000DB3, 0x00000DBB}, {0x00000DBD, 0x00000DBD},
{0x00000DC0, 0x00000DC6}, {0x00000E01, 0x00000E30}, {0x00000E32, 0x00000E33}, {0x00000E40, 0x00000E46},
{0x00000E81, 0x00000E82}, {0x00000E84, 0x00000E84}, {0x00000E86, 0x00000E8A}, {0x00000E8C, 0x00000EA3},
{0x00000EA5, 0x00000EA5}, {0x00000EA7, 0x00000EB0}, {0x00000EB2, 0x00000EB3}, {0x00000EBD, 0x00000EBD},
{0x00000EC0, 0x00000EC4}, {0x00000EC6, 0x00000EC6}, {0x00000EDC, 0x00000EDF}, {0x00000F00, 0x00000F00},
{0x00000F40, 0x00000F47}, {0x00000F49, 0x00000F6C}, {0x00000F88, 0x00000F8C}, {0x00001000, 0x0000102A},
{0x0000103F, 0x0000103F}, {0x00001050, 0x00001055}, {0x0000105A, 0x0000105D}, {0x00001061, 0x00001061},
{0x00001065, 0x00001066}, {0x0000106E, 0x00001070}, {0x00001075, 0x00001081}, {0x0000108E, 0x0000108E},
{0x000010A0, 0x000010C5}, {0x000010C7, 0x000010C7}, {0x000010CD, 0x000010CD}, {0x000010D0, 0x000010FA},
{0x000010FC, 0x00001248}, {0x0000124A, 0x0000124D}, {0x00001250, 0x00001256}, {0x00001258, 0x00001258},
{0x0000125A, 0x0000125D}, {0x00001260, 0x00001288}, {0x0000128A, 0x0000128D}, {0x00001290, 0x000012B0},
{0x000012B2, 0x000012B5}, {0x000012B8, 0x000012BE}, {0x000012C0, 0x000012C0}, {0x000012C2, 0x000012C5},
{0x000012C8, 0x000012D6}, {0x000012D8, 0x00001310}, {0x00001312, 0x00001315}, {0x00001318, 0x0000135A},
{0x00001380, 0x0000138F}, {0x000013A0, 0x000013F5}, {0x000013F8, 0x000013FD}, {0x00001401, 0x0000166C},
{0x0000166F, 0x0000167F}, {0x00001681, 0x0000169A}, {0x000016A0, 0x000016EA}, {0x000016F1, 0x000016F8},
{0x00001700, 0x00001711}, {0x0000171F, 0x00001731}, {0x00001740, 0x00001751}, {0x00001760, 0x0000176C},
{0x0000176E, 0x00001770}, {0x00001780, 0x000017B3}, {0x000017D7, 0x000017D7}, {0x000017DC, 0x000017DC},
{0x00001820, 0x00001878}, {0x00001880, 0x00001884}, {0x00001887, 0x000018A8}, {0x000018AA, 0x000018AA},
{0x000018B0, 0x000018F5}, {0x00001900, 0x0000191E}, {0x00001950, 0x0000196D}, {0x00001970, 0x00001974},
{0x00001980, 0x000019AB}, {0x000019B0, 0x000019C9}, {0x00001A00, 0x00001A16}, {0x00001A20, 0x00001A54},
{0x00001AA7, 0x00001AA7}, {0x00001B05, 0x00001B33}, {0x00001B45, 0x00001B4C}, {0x00001B83, 0x00001BA0},
{0x00001BAE, 0x00001BAF}, {0x00001BBA, 0x00001BE5}, {0x00001C00, 0x00001C23}, {0x00001C4D, 0x00001C4F},
{0x00001C5A, 0x00001C7D}, {0x00001C80, 0x00001C88}, {0x00001C90, 0x00001CBA}, {0x00001CBD, 0x00001CBF},
{0x00001CE9, 0x00001CEC}, {0x00001CEE, 0x00001CF3}, {0x00001CF5, 0x00001CF6}, {0x00001CFA, 0x00001CFA},
{0x00001D00, 0x00001DBF}, {0x00001E00, 0x00001F15}, {0x00001F18, 0x00001F1D}, {0x00001F20, 0x00001F45},
{0x00001F48, 0x00001F4D}, {0x00001F50, 0x00001F57}, {0x00001F59, 0x00001F59}, {0x00001F5B, 0x00001F5B},
{0x00001F5D, 0x00001F5D}, {0x00001F5F, 0x00001F7D}, {0x00001F80, 0x00001FB4}, {0x00001FB6, 0x00001FBC},
{0x00001FBE, 0x00001FBE}, {0x00001FC2, 0x00001FC4}, {0x00001FC6, 0x00001FCC}, {0x00001FD0, 0x00001FD3},
{0x00001FD6, 0x00001FDB}, {0x00001FE0, 0x00001FEC}, {0x00001FF2, 0x00001FF4}, {0x00001FF6, 0x00001FFC},
{0x00002071, 0x00002071}, {0x0000207F, 0x0000207F}, {0x00002090, 0x0000209C}, {0x00002102, 0x00002102},
{0x00002107, 0x00002107}, {0x0000210A, 0x00002113}, {0x00002115, 0x00002115}, {0x00002119, 0x0000211D},
{0x00002124, 0x00002124}, {0x00002126, 0x00002126}, {0x00002128, 0x00002128}, {0x0000212A, 0x0000212D},
{0x0000212F, 0x00002139}, {0x0000213C, 0x0000213F}, {0x00002145, 0x00002149}, {0x0000214E, 0x0000214E},
{0x00002183, 0x00002184}, {0x00002C00, 0x00002CE4}, {0x00002CEB, 0x00002CEE}, {0x00002CF2, 0x00002CF3},
{0x00002D00, 0x00002D25}, {0x00002D27, 0x00002D27}, {0x00002D2D, 0x00002D2D}, {0x00002D30, 0x00002D67},
{0x00002D6F, 0x00002D6F}, {0x00002D80, 0x00002D96}, {0x00002DA0, 0x00002DA6}, {0x00002DA8, 0x00002DAE},
{0x00002DB0, 0x00002DB6}, {0x00002DB8, 0x00002DBE}, {0x00002DC0, 0x00002DC6}, {0x00002DC8, 0x00002DCE},
{0x00002DD0, 0x00002DD6}, {0x00002DD8, 0x00002DDE}, {0x00002E2F, 0x00002E2F}, {0x00003005, 0x00003006},
{0x00003031, 0x00003035}, {0x0000303B, 0x0000303C}, {0x00003041, 0x00003096}, {0x0000309D, 0x0000309F},
{0x000030A1, 0x000030FA}, {0x000030FC, 0x000030FF}, {0x00003105, 0x0000312F}, {0x00003131, 0x0000318E},
{0x000031A0, 0x000031BF}, {0x000031F0, 0x000031FF}, {0x00003400, 0x00004DBF}, {0x00004E00, 0x0000A48C},
{0x0000A4D0, 0x0000A4FD}, {0x0000A500, 0x0000A60C}, {0x0000A610, 0x0000A61F}, {0x0000A62A, 0x0000A62B},
{0x0000A640, 0x0000A66E}, {0x0000A67F, 0x0000A69D}, {0x0000A6A0, 0x0000A6E5}, {0x0000A717, 0x0000A71F},
{0x0000A722, 0x0000A788}, {0x0000A78B, 0x0000A7CA}, {0x0000A7D0, 0x0000A7D1}, {0x0000A7D3, 0x0000A7D3},
{0x0000A7D5, 0x0000A7D9}, {0x0000A7F2, 0x0000A801}, {0x0000A803, 0x0000A805}, {0x0000A807, 0x0000A80A},
{0x0000A80C, 0x0000A822}, {0x0000A840, 0x0000A873}, {0x0000A882, 0x0000A8B3}, {0x0000A8F2, 0x0000A8F7},
{0x0000A8FB, 0x0000A8FB}, {0x0000A8FD, 0x0000A8FE}, {0x0000A90A, 0x0000A925}, {0x0000A930, 0x0000A946},
{0x0000A960, 0x0000A97C}, {0x0000A984, 0x0000A9B2}, {0x0000A9CF, 0x0000A9CF}, {0x0000A9E0, 0x0000A9E4},
@ -129,51 +149,60 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter = {
{0x000102A0, 0x000102D0}, {0x00010300, 0x0001031F}, {0x0001032D, 0x00010340}, {0x00010342, 0x00010349},
{0x00010350, 0x00010375}, {0x00010380, 0x0001039D}, {0x000103A0, 0x000103C3}, {0x000103C8, 0x000103CF},
{0x00010400, 0x0001049D}, {0x000104B0, 0x000104D3}, {0x000104D8, 0x000104FB}, {0x00010500, 0x00010527},
{0x00010530, 0x00010563}, {0x00010600, 0x00010736}, {0x00010740, 0x00010755}, {0x00010760, 0x00010767},
{0x00010800, 0x00010805}, {0x00010808, 0x00010808}, {0x0001080A, 0x00010835}, {0x00010837, 0x00010838},
{0x0001083C, 0x0001083C}, {0x0001083F, 0x00010855}, {0x00010860, 0x00010876}, {0x00010880, 0x0001089E},
{0x000108E0, 0x000108F2}, {0x000108F4, 0x000108F5}, {0x00010900, 0x00010915}, {0x00010920, 0x00010939},
{0x00010980, 0x000109B7}, {0x000109BE, 0x000109BF}, {0x00010A00, 0x00010A00}, {0x00010A10, 0x00010A13},
{0x00010A15, 0x00010A17}, {0x00010A19, 0x00010A35}, {0x00010A60, 0x00010A7C}, {0x00010A80, 0x00010A9C},
{0x00010AC0, 0x00010AC7}, {0x00010AC9, 0x00010AE4}, {0x00010B00, 0x00010B35}, {0x00010B40, 0x00010B55},
{0x00010B60, 0x00010B72}, {0x00010B80, 0x00010B91}, {0x00010C00, 0x00010C48}, {0x00010C80, 0x00010CB2},
{0x00010CC0, 0x00010CF2}, {0x00010D00, 0x00010D23}, {0x00010E80, 0x00010EA9}, {0x00010EB0, 0x00010EB1},
{0x00010F00, 0x00010F1C}, {0x00010F27, 0x00010F27}, {0x00010F30, 0x00010F45}, {0x00010FB0, 0x00010FC4},
{0x00010FE0, 0x00010FF6}, {0x00011003, 0x00011037}, {0x00011083, 0x000110AF}, {0x000110D0, 0x000110E8},
{0x00011103, 0x00011126}, {0x00011144, 0x00011144}, {0x00011147, 0x00011147}, {0x00011150, 0x00011172},
{0x00011176, 0x00011176}, {0x00011183, 0x000111B2}, {0x000111C1, 0x000111C4}, {0x000111DA, 0x000111DA},
{0x000111DC, 0x000111DC}, {0x00011200, 0x00011211}, {0x00011213, 0x0001122B}, {0x00011280, 0x00011286},
{0x00011288, 0x00011288}, {0x0001128A, 0x0001128D}, {0x0001128F, 0x0001129D}, {0x0001129F, 0x000112A8},
{0x000112B0, 0x000112DE}, {0x00011305, 0x0001130C}, {0x0001130F, 0x00011310}, {0x00011313, 0x00011328},
{0x0001132A, 0x00011330}, {0x00011332, 0x00011333}, {0x00011335, 0x00011339}, {0x0001133D, 0x0001133D},
{0x00011350, 0x00011350}, {0x0001135D, 0x00011361}, {0x00011400, 0x00011434}, {0x00011447, 0x0001144A},
{0x0001145F, 0x00011461}, {0x00011480, 0x000114AF}, {0x000114C4, 0x000114C5}, {0x000114C7, 0x000114C7},
{0x00011580, 0x000115AE}, {0x000115D8, 0x000115DB}, {0x00011600, 0x0001162F}, {0x00011644, 0x00011644},
{0x00011680, 0x000116AA}, {0x000116B8, 0x000116B8}, {0x00011700, 0x0001171A}, {0x00011800, 0x0001182B},
{0x00010530, 0x00010563}, {0x00010570, 0x0001057A}, {0x0001057C, 0x0001058A}, {0x0001058C, 0x00010592},
{0x00010594, 0x00010595}, {0x00010597, 0x000105A1}, {0x000105A3, 0x000105B1}, {0x000105B3, 0x000105B9},
{0x000105BB, 0x000105BC}, {0x00010600, 0x00010736}, {0x00010740, 0x00010755}, {0x00010760, 0x00010767},
{0x00010780, 0x00010785}, {0x00010787, 0x000107B0}, {0x000107B2, 0x000107BA}, {0x00010800, 0x00010805},
{0x00010808, 0x00010808}, {0x0001080A, 0x00010835}, {0x00010837, 0x00010838}, {0x0001083C, 0x0001083C},
{0x0001083F, 0x00010855}, {0x00010860, 0x00010876}, {0x00010880, 0x0001089E}, {0x000108E0, 0x000108F2},
{0x000108F4, 0x000108F5}, {0x00010900, 0x00010915}, {0x00010920, 0x00010939}, {0x00010980, 0x000109B7},
{0x000109BE, 0x000109BF}, {0x00010A00, 0x00010A00}, {0x00010A10, 0x00010A13}, {0x00010A15, 0x00010A17},
{0x00010A19, 0x00010A35}, {0x00010A60, 0x00010A7C}, {0x00010A80, 0x00010A9C}, {0x00010AC0, 0x00010AC7},
{0x00010AC9, 0x00010AE4}, {0x00010B00, 0x00010B35}, {0x00010B40, 0x00010B55}, {0x00010B60, 0x00010B72},
{0x00010B80, 0x00010B91}, {0x00010C00, 0x00010C48}, {0x00010C80, 0x00010CB2}, {0x00010CC0, 0x00010CF2},
{0x00010D00, 0x00010D23}, {0x00010E80, 0x00010EA9}, {0x00010EB0, 0x00010EB1}, {0x00010F00, 0x00010F1C},
{0x00010F27, 0x00010F27}, {0x00010F30, 0x00010F45}, {0x00010F70, 0x00010F81}, {0x00010FB0, 0x00010FC4},
{0x00010FE0, 0x00010FF6}, {0x00011003, 0x00011037}, {0x00011071, 0x00011072}, {0x00011075, 0x00011075},
{0x00011083, 0x000110AF}, {0x000110D0, 0x000110E8}, {0x00011103, 0x00011126}, {0x00011144, 0x00011144},
{0x00011147, 0x00011147}, {0x00011150, 0x00011172}, {0x00011176, 0x00011176}, {0x00011183, 0x000111B2},
{0x000111C1, 0x000111C4}, {0x000111DA, 0x000111DA}, {0x000111DC, 0x000111DC}, {0x00011200, 0x00011211},
{0x00011213, 0x0001122B}, {0x0001123F, 0x00011240}, {0x00011280, 0x00011286}, {0x00011288, 0x00011288},
{0x0001128A, 0x0001128D}, {0x0001128F, 0x0001129D}, {0x0001129F, 0x000112A8}, {0x000112B0, 0x000112DE},
{0x00011305, 0x0001130C}, {0x0001130F, 0x00011310}, {0x00011313, 0x00011328}, {0x0001132A, 0x00011330},
{0x00011332, 0x00011333}, {0x00011335, 0x00011339}, {0x0001133D, 0x0001133D}, {0x00011350, 0x00011350},
{0x0001135D, 0x00011361}, {0x00011400, 0x00011434}, {0x00011447, 0x0001144A}, {0x0001145F, 0x00011461},
{0x00011480, 0x000114AF}, {0x000114C4, 0x000114C5}, {0x000114C7, 0x000114C7}, {0x00011580, 0x000115AE},
{0x000115D8, 0x000115DB}, {0x00011600, 0x0001162F}, {0x00011644, 0x00011644}, {0x00011680, 0x000116AA},
{0x000116B8, 0x000116B8}, {0x00011700, 0x0001171A}, {0x00011740, 0x00011746}, {0x00011800, 0x0001182B},
{0x000118A0, 0x000118DF}, {0x000118FF, 0x00011906}, {0x00011909, 0x00011909}, {0x0001190C, 0x00011913},
{0x00011915, 0x00011916}, {0x00011918, 0x0001192F}, {0x0001193F, 0x0001193F}, {0x00011941, 0x00011941},
{0x000119A0, 0x000119A7}, {0x000119AA, 0x000119D0}, {0x000119E1, 0x000119E1}, {0x000119E3, 0x000119E3},
{0x00011A00, 0x00011A00}, {0x00011A0B, 0x00011A32}, {0x00011A3A, 0x00011A3A}, {0x00011A50, 0x00011A50},
{0x00011A5C, 0x00011A89}, {0x00011A9D, 0x00011A9D}, {0x00011AC0, 0x00011AF8}, {0x00011C00, 0x00011C08},
{0x00011A5C, 0x00011A89}, {0x00011A9D, 0x00011A9D}, {0x00011AB0, 0x00011AF8}, {0x00011C00, 0x00011C08},
{0x00011C0A, 0x00011C2E}, {0x00011C40, 0x00011C40}, {0x00011C72, 0x00011C8F}, {0x00011D00, 0x00011D06},
{0x00011D08, 0x00011D09}, {0x00011D0B, 0x00011D30}, {0x00011D46, 0x00011D46}, {0x00011D60, 0x00011D65},
{0x00011D67, 0x00011D68}, {0x00011D6A, 0x00011D89}, {0x00011D98, 0x00011D98}, {0x00011EE0, 0x00011EF2},
{0x00011FB0, 0x00011FB0}, {0x00012000, 0x00012399}, {0x00012480, 0x00012543}, {0x00013000, 0x0001342E},
{0x00014400, 0x00014646}, {0x00016800, 0x00016A38}, {0x00016A40, 0x00016A5E}, {0x00016AD0, 0x00016AED},
{0x00016B00, 0x00016B2F}, {0x00016B40, 0x00016B43}, {0x00016B63, 0x00016B77}, {0x00016B7D, 0x00016B8F},
{0x00016E40, 0x00016E7F}, {0x00016F00, 0x00016F4A}, {0x00016F50, 0x00016F50}, {0x00016F93, 0x00016F9F},
{0x00016FE0, 0x00016FE1}, {0x00016FE3, 0x00016FE3}, {0x00017000, 0x000187F7}, {0x00018800, 0x00018CD5},
{0x00018D00, 0x00018D08}, {0x0001B000, 0x0001B11E}, {0x0001B150, 0x0001B152}, {0x0001B164, 0x0001B167},
{0x0001B170, 0x0001B2FB}, {0x0001BC00, 0x0001BC6A}, {0x0001BC70, 0x0001BC7C}, {0x0001BC80, 0x0001BC88},
{0x0001BC90, 0x0001BC99}, {0x0001D400, 0x0001D454}, {0x0001D456, 0x0001D49C}, {0x0001D49E, 0x0001D49F},
{0x0001D4A2, 0x0001D4A2}, {0x0001D4A5, 0x0001D4A6}, {0x0001D4A9, 0x0001D4AC}, {0x0001D4AE, 0x0001D4B9},
{0x0001D4BB, 0x0001D4BB}, {0x0001D4BD, 0x0001D4C3}, {0x0001D4C5, 0x0001D505}, {0x0001D507, 0x0001D50A},
{0x0001D50D, 0x0001D514}, {0x0001D516, 0x0001D51C}, {0x0001D51E, 0x0001D539}, {0x0001D53B, 0x0001D53E},
{0x0001D540, 0x0001D544}, {0x0001D546, 0x0001D546}, {0x0001D54A, 0x0001D550}, {0x0001D552, 0x0001D6A5},
{0x0001D6A8, 0x0001D6C0}, {0x0001D6C2, 0x0001D6DA}, {0x0001D6DC, 0x0001D6FA}, {0x0001D6FC, 0x0001D714},
{0x0001D716, 0x0001D734}, {0x0001D736, 0x0001D74E}, {0x0001D750, 0x0001D76E}, {0x0001D770, 0x0001D788},
{0x0001D78A, 0x0001D7A8}, {0x0001D7AA, 0x0001D7C2}, {0x0001D7C4, 0x0001D7CB}, {0x0001E100, 0x0001E12C},
{0x0001E137, 0x0001E13D}, {0x0001E14E, 0x0001E14E}, {0x0001E2C0, 0x0001E2EB}, {0x0001E800, 0x0001E8C4},
{0x00011F02, 0x00011F02}, {0x00011F04, 0x00011F10}, {0x00011F12, 0x00011F33}, {0x00011FB0, 0x00011FB0},
{0x00012000, 0x00012399}, {0x00012480, 0x00012543}, {0x00012F90, 0x00012FF0}, {0x00013000, 0x0001342F},
{0x00013441, 0x00013446}, {0x00014400, 0x00014646}, {0x00016800, 0x00016A38}, {0x00016A40, 0x00016A5E},
{0x00016A70, 0x00016ABE}, {0x00016AD0, 0x00016AED}, {0x00016B00, 0x00016B2F}, {0x00016B40, 0x00016B43},
{0x00016B63, 0x00016B77}, {0x00016B7D, 0x00016B8F}, {0x00016E40, 0x00016E7F}, {0x00016F00, 0x00016F4A},
{0x00016F50, 0x00016F50}, {0x00016F93, 0x00016F9F}, {0x00016FE0, 0x00016FE1}, {0x00016FE3, 0x00016FE3},
{0x00017000, 0x000187F7}, {0x00018800, 0x00018CD5}, {0x00018D00, 0x00018D08}, {0x0001AFF0, 0x0001AFF3},
{0x0001AFF5, 0x0001AFFB}, {0x0001AFFD, 0x0001AFFE}, {0x0001B000, 0x0001B122}, {0x0001B132, 0x0001B132},
{0x0001B150, 0x0001B152}, {0x0001B155, 0x0001B155}, {0x0001B164, 0x0001B167}, {0x0001B170, 0x0001B2FB},
{0x0001BC00, 0x0001BC6A}, {0x0001BC70, 0x0001BC7C}, {0x0001BC80, 0x0001BC88}, {0x0001BC90, 0x0001BC99},
{0x0001D400, 0x0001D454}, {0x0001D456, 0x0001D49C}, {0x0001D49E, 0x0001D49F}, {0x0001D4A2, 0x0001D4A2},
{0x0001D4A5, 0x0001D4A6}, {0x0001D4A9, 0x0001D4AC}, {0x0001D4AE, 0x0001D4B9}, {0x0001D4BB, 0x0001D4BB},
{0x0001D4BD, 0x0001D4C3}, {0x0001D4C5, 0x0001D505}, {0x0001D507, 0x0001D50A}, {0x0001D50D, 0x0001D514},
{0x0001D516, 0x0001D51C}, {0x0001D51E, 0x0001D539}, {0x0001D53B, 0x0001D53E}, {0x0001D540, 0x0001D544},
{0x0001D546, 0x0001D546}, {0x0001D54A, 0x0001D550}, {0x0001D552, 0x0001D6A5}, {0x0001D6A8, 0x0001D6C0},
{0x0001D6C2, 0x0001D6DA}, {0x0001D6DC, 0x0001D6FA}, {0x0001D6FC, 0x0001D714}, {0x0001D716, 0x0001D734},
{0x0001D736, 0x0001D74E}, {0x0001D750, 0x0001D76E}, {0x0001D770, 0x0001D788}, {0x0001D78A, 0x0001D7A8},
{0x0001D7AA, 0x0001D7C2}, {0x0001D7C4, 0x0001D7CB}, {0x0001DF00, 0x0001DF1E}, {0x0001DF25, 0x0001DF2A},
{0x0001E030, 0x0001E06D}, {0x0001E100, 0x0001E12C}, {0x0001E137, 0x0001E13D}, {0x0001E14E, 0x0001E14E},
{0x0001E290, 0x0001E2AD}, {0x0001E2C0, 0x0001E2EB}, {0x0001E4D0, 0x0001E4EB}, {0x0001E7E0, 0x0001E7E6},
{0x0001E7E8, 0x0001E7EB}, {0x0001E7ED, 0x0001E7EE}, {0x0001E7F0, 0x0001E7FE}, {0x0001E800, 0x0001E8C4},
{0x0001E900, 0x0001E943}, {0x0001E94B, 0x0001E94B}, {0x0001EE00, 0x0001EE03}, {0x0001EE05, 0x0001EE1F},
{0x0001EE21, 0x0001EE22}, {0x0001EE24, 0x0001EE24}, {0x0001EE27, 0x0001EE27}, {0x0001EE29, 0x0001EE32},
{0x0001EE34, 0x0001EE37}, {0x0001EE39, 0x0001EE39}, {0x0001EE3B, 0x0001EE3B}, {0x0001EE42, 0x0001EE42},
@ -182,15 +211,14 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter = {
{0x0001EE5B, 0x0001EE5B}, {0x0001EE5D, 0x0001EE5D}, {0x0001EE5F, 0x0001EE5F}, {0x0001EE61, 0x0001EE62},
{0x0001EE64, 0x0001EE64}, {0x0001EE67, 0x0001EE6A}, {0x0001EE6C, 0x0001EE72}, {0x0001EE74, 0x0001EE77},
{0x0001EE79, 0x0001EE7C}, {0x0001EE7E, 0x0001EE7E}, {0x0001EE80, 0x0001EE89}, {0x0001EE8B, 0x0001EE9B},
{0x0001EEA1, 0x0001EEA3}, {0x0001EEA5, 0x0001EEA9}, {0x0001EEAB, 0x0001EEBB}, {0x00020000, 0x0002A6DD},
{0x0002A700, 0x0002B734}, {0x0002B740, 0x0002B81D}, {0x0002B820, 0x0002CEA1}, {0x0002CEB0, 0x0002EBE0},
{0x0002F800, 0x0002FA1D}, {0x00030000, 0x0003134A},
{0x0001EEA1, 0x0001EEA3}, {0x0001EEA5, 0x0001EEA9}, {0x0001EEAB, 0x0001EEBB}, {0x00020000, 0x0002A6DF},
{0x0002A700, 0x0002B739}, {0x0002B740, 0x0002B81D}, {0x0002B820, 0x0002CEA1}, {0x0002CEB0, 0x0002EBE0},
{0x0002F800, 0x0002FA1D}, {0x00030000, 0x0003134A}, {0x00031350, 0x000323AF},
};
const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace = {
{0x00000009, 0x0000000D}, {0x0000001C, 0x00000020}, {0x00000085, 0x00000085}, {0x000000A0, 0x000000A0},
{0x00001680, 0x00001680}, {0x00002000, 0x0000200A}, {0x00002028, 0x00002029}, {0x0000202F, 0x0000202F},
{0x0000205F, 0x0000205F}, {0x00003000, 0x00003000},
{0x00000020, 0x00000020}, {0x000000A0, 0x000000A0}, {0x00001680, 0x00001680}, {0x00002000, 0x0000200A},
{0x00002028, 0x00002029}, {0x0000202F, 0x0000202F}, {0x0000205F, 0x0000205F}, {0x00003000, 0x00003000},
};
const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark = {
@ -200,72 +228,77 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark = {
{0x000006E7, 0x000006E8}, {0x000006EA, 0x000006ED}, {0x00000711, 0x00000711}, {0x00000730, 0x0000074A},
{0x000007A6, 0x000007B0}, {0x000007EB, 0x000007F3}, {0x000007FD, 0x000007FD}, {0x00000816, 0x00000819},
{0x0000081B, 0x00000823}, {0x00000825, 0x00000827}, {0x00000829, 0x0000082D}, {0x00000859, 0x0000085B},
{0x000008D3, 0x000008E1}, {0x000008E3, 0x00000903}, {0x0000093A, 0x0000093C}, {0x0000093E, 0x0000094F},
{0x00000951, 0x00000957}, {0x00000962, 0x00000963}, {0x00000981, 0x00000983}, {0x000009BC, 0x000009BC},
{0x000009BE, 0x000009C4}, {0x000009C7, 0x000009C8}, {0x000009CB, 0x000009CD}, {0x000009D7, 0x000009D7},
{0x000009E2, 0x000009E3}, {0x000009FE, 0x000009FE}, {0x00000A01, 0x00000A03}, {0x00000A3C, 0x00000A3C},
{0x00000A3E, 0x00000A42}, {0x00000A47, 0x00000A48}, {0x00000A4B, 0x00000A4D}, {0x00000A51, 0x00000A51},
{0x00000A70, 0x00000A71}, {0x00000A75, 0x00000A75}, {0x00000A81, 0x00000A83}, {0x00000ABC, 0x00000ABC},
{0x00000ABE, 0x00000AC5}, {0x00000AC7, 0x00000AC9}, {0x00000ACB, 0x00000ACD}, {0x00000AE2, 0x00000AE3},
{0x00000AFA, 0x00000AFF}, {0x00000B01, 0x00000B03}, {0x00000B3C, 0x00000B3C}, {0x00000B3E, 0x00000B44},
{0x00000B47, 0x00000B48}, {0x00000B4B, 0x00000B4D}, {0x00000B55, 0x00000B57}, {0x00000B62, 0x00000B63},
{0x00000B82, 0x00000B82}, {0x00000BBE, 0x00000BC2}, {0x00000BC6, 0x00000BC8}, {0x00000BCA, 0x00000BCD},
{0x00000BD7, 0x00000BD7}, {0x00000C00, 0x00000C04}, {0x00000C3E, 0x00000C44}, {0x00000C46, 0x00000C48},
{0x00000C4A, 0x00000C4D}, {0x00000C55, 0x00000C56}, {0x00000C62, 0x00000C63}, {0x00000C81, 0x00000C83},
{0x00000CBC, 0x00000CBC}, {0x00000CBE, 0x00000CC4}, {0x00000CC6, 0x00000CC8}, {0x00000CCA, 0x00000CCD},
{0x00000CD5, 0x00000CD6}, {0x00000CE2, 0x00000CE3}, {0x00000D00, 0x00000D03}, {0x00000D3B, 0x00000D3C},
{0x00000D3E, 0x00000D44}, {0x00000D46, 0x00000D48}, {0x00000D4A, 0x00000D4D}, {0x00000D57, 0x00000D57},
{0x00000D62, 0x00000D63}, {0x00000D81, 0x00000D83}, {0x00000DCA, 0x00000DCA}, {0x00000DCF, 0x00000DD4},
{0x00000DD6, 0x00000DD6}, {0x00000DD8, 0x00000DDF}, {0x00000DF2, 0x00000DF3}, {0x00000E31, 0x00000E31},
{0x00000E34, 0x00000E3A}, {0x00000E47, 0x00000E4E}, {0x00000EB1, 0x00000EB1}, {0x00000EB4, 0x00000EBC},
{0x00000EC8, 0x00000ECD}, {0x00000F18, 0x00000F19}, {0x00000F35, 0x00000F35}, {0x00000F37, 0x00000F37},
{0x00000F39, 0x00000F39}, {0x00000F3E, 0x00000F3F}, {0x00000F71, 0x00000F84}, {0x00000F86, 0x00000F87},
{0x00000F8D, 0x00000F97}, {0x00000F99, 0x00000FBC}, {0x00000FC6, 0x00000FC6}, {0x0000102B, 0x0000103E},
{0x00001056, 0x00001059}, {0x0000105E, 0x00001060}, {0x00001062, 0x00001064}, {0x00001067, 0x0000106D},
{0x00001071, 0x00001074}, {0x00001082, 0x0000108D}, {0x0000108F, 0x0000108F}, {0x0000109A, 0x0000109D},
{0x0000135D, 0x0000135F}, {0x00001712, 0x00001714}, {0x00001732, 0x00001734}, {0x00001752, 0x00001753},
{0x00001772, 0x00001773}, {0x000017B4, 0x000017D3}, {0x000017DD, 0x000017DD}, {0x0000180B, 0x0000180D},
{0x00000898, 0x0000089F}, {0x000008CA, 0x000008E1}, {0x000008E3, 0x00000903}, {0x0000093A, 0x0000093C},
{0x0000093E, 0x0000094F}, {0x00000951, 0x00000957}, {0x00000962, 0x00000963}, {0x00000981, 0x00000983},
{0x000009BC, 0x000009BC}, {0x000009BE, 0x000009C4}, {0x000009C7, 0x000009C8}, {0x000009CB, 0x000009CD},
{0x000009D7, 0x000009D7}, {0x000009E2, 0x000009E3}, {0x000009FE, 0x000009FE}, {0x00000A01, 0x00000A03},
{0x00000A3C, 0x00000A3C}, {0x00000A3E, 0x00000A42}, {0x00000A47, 0x00000A48}, {0x00000A4B, 0x00000A4D},
{0x00000A51, 0x00000A51}, {0x00000A70, 0x00000A71}, {0x00000A75, 0x00000A75}, {0x00000A81, 0x00000A83},
{0x00000ABC, 0x00000ABC}, {0x00000ABE, 0x00000AC5}, {0x00000AC7, 0x00000AC9}, {0x00000ACB, 0x00000ACD},
{0x00000AE2, 0x00000AE3}, {0x00000AFA, 0x00000AFF}, {0x00000B01, 0x00000B03}, {0x00000B3C, 0x00000B3C},
{0x00000B3E, 0x00000B44}, {0x00000B47, 0x00000B48}, {0x00000B4B, 0x00000B4D}, {0x00000B55, 0x00000B57},
{0x00000B62, 0x00000B63}, {0x00000B82, 0x00000B82}, {0x00000BBE, 0x00000BC2}, {0x00000BC6, 0x00000BC8},
{0x00000BCA, 0x00000BCD}, {0x00000BD7, 0x00000BD7}, {0x00000C00, 0x00000C04}, {0x00000C3C, 0x00000C3C},
{0x00000C3E, 0x00000C44}, {0x00000C46, 0x00000C48}, {0x00000C4A, 0x00000C4D}, {0x00000C55, 0x00000C56},
{0x00000C62, 0x00000C63}, {0x00000C81, 0x00000C83}, {0x00000CBC, 0x00000CBC}, {0x00000CBE, 0x00000CC4},
{0x00000CC6, 0x00000CC8}, {0x00000CCA, 0x00000CCD}, {0x00000CD5, 0x00000CD6}, {0x00000CE2, 0x00000CE3},
{0x00000CF3, 0x00000CF3}, {0x00000D00, 0x00000D03}, {0x00000D3B, 0x00000D3C}, {0x00000D3E, 0x00000D44},
{0x00000D46, 0x00000D48}, {0x00000D4A, 0x00000D4D}, {0x00000D57, 0x00000D57}, {0x00000D62, 0x00000D63},
{0x00000D81, 0x00000D83}, {0x00000DCA, 0x00000DCA}, {0x00000DCF, 0x00000DD4}, {0x00000DD6, 0x00000DD6},
{0x00000DD8, 0x00000DDF}, {0x00000DF2, 0x00000DF3}, {0x00000E31, 0x00000E31}, {0x00000E34, 0x00000E3A},
{0x00000E47, 0x00000E4E}, {0x00000EB1, 0x00000EB1}, {0x00000EB4, 0x00000EBC}, {0x00000EC8, 0x00000ECE},
{0x00000F18, 0x00000F19}, {0x00000F35, 0x00000F35}, {0x00000F37, 0x00000F37}, {0x00000F39, 0x00000F39},
{0x00000F3E, 0x00000F3F}, {0x00000F71, 0x00000F84}, {0x00000F86, 0x00000F87}, {0x00000F8D, 0x00000F97},
{0x00000F99, 0x00000FBC}, {0x00000FC6, 0x00000FC6}, {0x0000102B, 0x0000103E}, {0x00001056, 0x00001059},
{0x0000105E, 0x00001060}, {0x00001062, 0x00001064}, {0x00001067, 0x0000106D}, {0x00001071, 0x00001074},
{0x00001082, 0x0000108D}, {0x0000108F, 0x0000108F}, {0x0000109A, 0x0000109D}, {0x0000135D, 0x0000135F},
{0x00001712, 0x00001715}, {0x00001732, 0x00001734}, {0x00001752, 0x00001753}, {0x00001772, 0x00001773},
{0x000017B4, 0x000017D3}, {0x000017DD, 0x000017DD}, {0x0000180B, 0x0000180D}, {0x0000180F, 0x0000180F},
{0x00001885, 0x00001886}, {0x000018A9, 0x000018A9}, {0x00001920, 0x0000192B}, {0x00001930, 0x0000193B},
{0x00001A17, 0x00001A1B}, {0x00001A55, 0x00001A5E}, {0x00001A60, 0x00001A7C}, {0x00001A7F, 0x00001A7F},
{0x00001AB0, 0x00001AC0}, {0x00001B00, 0x00001B04}, {0x00001B34, 0x00001B44}, {0x00001B6B, 0x00001B73},
{0x00001AB0, 0x00001ACE}, {0x00001B00, 0x00001B04}, {0x00001B34, 0x00001B44}, {0x00001B6B, 0x00001B73},
{0x00001B80, 0x00001B82}, {0x00001BA1, 0x00001BAD}, {0x00001BE6, 0x00001BF3}, {0x00001C24, 0x00001C37},
{0x00001CD0, 0x00001CD2}, {0x00001CD4, 0x00001CE8}, {0x00001CED, 0x00001CED}, {0x00001CF4, 0x00001CF4},
{0x00001CF7, 0x00001CF9}, {0x00001DC0, 0x00001DF9}, {0x00001DFB, 0x00001DFF}, {0x000020D0, 0x000020F0},
{0x00002CEF, 0x00002CF1}, {0x00002D7F, 0x00002D7F}, {0x00002DE0, 0x00002DFF}, {0x0000302A, 0x0000302F},
{0x00003099, 0x0000309A}, {0x0000A66F, 0x0000A672}, {0x0000A674, 0x0000A67D}, {0x0000A69E, 0x0000A69F},
{0x0000A6F0, 0x0000A6F1}, {0x0000A802, 0x0000A802}, {0x0000A806, 0x0000A806}, {0x0000A80B, 0x0000A80B},
{0x0000A823, 0x0000A827}, {0x0000A82C, 0x0000A82C}, {0x0000A880, 0x0000A881}, {0x0000A8B4, 0x0000A8C5},
{0x0000A8E0, 0x0000A8F1}, {0x0000A8FF, 0x0000A8FF}, {0x0000A926, 0x0000A92D}, {0x0000A947, 0x0000A953},
{0x0000A980, 0x0000A983}, {0x0000A9B3, 0x0000A9C0}, {0x0000A9E5, 0x0000A9E5}, {0x0000AA29, 0x0000AA36},
{0x0000AA43, 0x0000AA43}, {0x0000AA4C, 0x0000AA4D}, {0x0000AA7B, 0x0000AA7D}, {0x0000AAB0, 0x0000AAB0},
{0x0000AAB2, 0x0000AAB4}, {0x0000AAB7, 0x0000AAB8}, {0x0000AABE, 0x0000AABF}, {0x0000AAC1, 0x0000AAC1},
{0x0000AAEB, 0x0000AAEF}, {0x0000AAF5, 0x0000AAF6}, {0x0000ABE3, 0x0000ABEA}, {0x0000ABEC, 0x0000ABED},
{0x0000FB1E, 0x0000FB1E}, {0x0000FE00, 0x0000FE0F}, {0x0000FE20, 0x0000FE2F}, {0x000101FD, 0x000101FD},
{0x000102E0, 0x000102E0}, {0x00010376, 0x0001037A}, {0x00010A01, 0x00010A03}, {0x00010A05, 0x00010A06},
{0x00010A0C, 0x00010A0F}, {0x00010A38, 0x00010A3A}, {0x00010A3F, 0x00010A3F}, {0x00010AE5, 0x00010AE6},
{0x00010D24, 0x00010D27}, {0x00010EAB, 0x00010EAC}, {0x00010F46, 0x00010F50}, {0x00011000, 0x00011002},
{0x00011038, 0x00011046}, {0x0001107F, 0x00011082}, {0x000110B0, 0x000110BA}, {0x00011100, 0x00011102},
{0x00001CF7, 0x00001CF9}, {0x00001DC0, 0x00001DFF}, {0x000020D0, 0x000020F0}, {0x00002CEF, 0x00002CF1},
{0x00002D7F, 0x00002D7F}, {0x00002DE0, 0x00002DFF}, {0x0000302A, 0x0000302F}, {0x00003099, 0x0000309A},
{0x0000A66F, 0x0000A672}, {0x0000A674, 0x0000A67D}, {0x0000A69E, 0x0000A69F}, {0x0000A6F0, 0x0000A6F1},
{0x0000A802, 0x0000A802}, {0x0000A806, 0x0000A806}, {0x0000A80B, 0x0000A80B}, {0x0000A823, 0x0000A827},
{0x0000A82C, 0x0000A82C}, {0x0000A880, 0x0000A881}, {0x0000A8B4, 0x0000A8C5}, {0x0000A8E0, 0x0000A8F1},
{0x0000A8FF, 0x0000A8FF}, {0x0000A926, 0x0000A92D}, {0x0000A947, 0x0000A953}, {0x0000A980, 0x0000A983},
{0x0000A9B3, 0x0000A9C0}, {0x0000A9E5, 0x0000A9E5}, {0x0000AA29, 0x0000AA36}, {0x0000AA43, 0x0000AA43},
{0x0000AA4C, 0x0000AA4D}, {0x0000AA7B, 0x0000AA7D}, {0x0000AAB0, 0x0000AAB0}, {0x0000AAB2, 0x0000AAB4},
{0x0000AAB7, 0x0000AAB8}, {0x0000AABE, 0x0000AABF}, {0x0000AAC1, 0x0000AAC1}, {0x0000AAEB, 0x0000AAEF},
{0x0000AAF5, 0x0000AAF6}, {0x0000ABE3, 0x0000ABEA}, {0x0000ABEC, 0x0000ABED}, {0x0000FB1E, 0x0000FB1E},
{0x0000FE00, 0x0000FE0F}, {0x0000FE20, 0x0000FE2F}, {0x000101FD, 0x000101FD}, {0x000102E0, 0x000102E0},
{0x00010376, 0x0001037A}, {0x00010A01, 0x00010A03}, {0x00010A05, 0x00010A06}, {0x00010A0C, 0x00010A0F},
{0x00010A38, 0x00010A3A}, {0x00010A3F, 0x00010A3F}, {0x00010AE5, 0x00010AE6}, {0x00010D24, 0x00010D27},
{0x00010EAB, 0x00010EAC}, {0x00010EFD, 0x00010EFF}, {0x00010F46, 0x00010F50}, {0x00010F82, 0x00010F85},
{0x00011000, 0x00011002}, {0x00011038, 0x00011046}, {0x00011070, 0x00011070}, {0x00011073, 0x00011074},
{0x0001107F, 0x00011082}, {0x000110B0, 0x000110BA}, {0x000110C2, 0x000110C2}, {0x00011100, 0x00011102},
{0x00011127, 0x00011134}, {0x00011145, 0x00011146}, {0x00011173, 0x00011173}, {0x00011180, 0x00011182},
{0x000111B3, 0x000111C0}, {0x000111C9, 0x000111CC}, {0x000111CE, 0x000111CF}, {0x0001122C, 0x00011237},
{0x0001123E, 0x0001123E}, {0x000112DF, 0x000112EA}, {0x00011300, 0x00011303}, {0x0001133B, 0x0001133C},
{0x0001133E, 0x00011344}, {0x00011347, 0x00011348}, {0x0001134B, 0x0001134D}, {0x00011357, 0x00011357},
{0x00011362, 0x00011363}, {0x00011366, 0x0001136C}, {0x00011370, 0x00011374}, {0x00011435, 0x00011446},
{0x0001145E, 0x0001145E}, {0x000114B0, 0x000114C3}, {0x000115AF, 0x000115B5}, {0x000115B8, 0x000115C0},
{0x000115DC, 0x000115DD}, {0x00011630, 0x00011640}, {0x000116AB, 0x000116B7}, {0x0001171D, 0x0001172B},
{0x0001182C, 0x0001183A}, {0x00011930, 0x00011935}, {0x00011937, 0x00011938}, {0x0001193B, 0x0001193E},
{0x00011940, 0x00011940}, {0x00011942, 0x00011943}, {0x000119D1, 0x000119D7}, {0x000119DA, 0x000119E0},
{0x000119E4, 0x000119E4}, {0x00011A01, 0x00011A0A}, {0x00011A33, 0x00011A39}, {0x00011A3B, 0x00011A3E},
{0x00011A47, 0x00011A47}, {0x00011A51, 0x00011A5B}, {0x00011A8A, 0x00011A99}, {0x00011C2F, 0x00011C36},
{0x00011C38, 0x00011C3F}, {0x00011C92, 0x00011CA7}, {0x00011CA9, 0x00011CB6}, {0x00011D31, 0x00011D36},
{0x00011D3A, 0x00011D3A}, {0x00011D3C, 0x00011D3D}, {0x00011D3F, 0x00011D45}, {0x00011D47, 0x00011D47},
{0x00011D8A, 0x00011D8E}, {0x00011D90, 0x00011D91}, {0x00011D93, 0x00011D97}, {0x00011EF3, 0x00011EF6},
{0x00016AF0, 0x00016AF4}, {0x00016B30, 0x00016B36}, {0x00016F4F, 0x00016F4F}, {0x00016F51, 0x00016F87},
{0x00016F8F, 0x00016F92}, {0x00016FE4, 0x00016FE4}, {0x00016FF0, 0x00016FF1}, {0x0001BC9D, 0x0001BC9E},
{0x0001D165, 0x0001D169}, {0x0001D16D, 0x0001D172}, {0x0001D17B, 0x0001D182}, {0x0001D185, 0x0001D18B},
{0x0001D1AA, 0x0001D1AD}, {0x0001D242, 0x0001D244}, {0x0001DA00, 0x0001DA36}, {0x0001DA3B, 0x0001DA6C},
{0x0001DA75, 0x0001DA75}, {0x0001DA84, 0x0001DA84}, {0x0001DA9B, 0x0001DA9F}, {0x0001DAA1, 0x0001DAAF},
{0x0001E000, 0x0001E006}, {0x0001E008, 0x0001E018}, {0x0001E01B, 0x0001E021}, {0x0001E023, 0x0001E024},
{0x0001E026, 0x0001E02A}, {0x0001E130, 0x0001E136}, {0x0001E2EC, 0x0001E2EF}, {0x0001E8D0, 0x0001E8D6},
{0x0001123E, 0x0001123E}, {0x00011241, 0x00011241}, {0x000112DF, 0x000112EA}, {0x00011300, 0x00011303},
{0x0001133B, 0x0001133C}, {0x0001133E, 0x00011344}, {0x00011347, 0x00011348}, {0x0001134B, 0x0001134D},
{0x00011357, 0x00011357}, {0x00011362, 0x00011363}, {0x00011366, 0x0001136C}, {0x00011370, 0x00011374},
{0x00011435, 0x00011446}, {0x0001145E, 0x0001145E}, {0x000114B0, 0x000114C3}, {0x000115AF, 0x000115B5},
{0x000115B8, 0x000115C0}, {0x000115DC, 0x000115DD}, {0x00011630, 0x00011640}, {0x000116AB, 0x000116B7},
{0x0001171D, 0x0001172B}, {0x0001182C, 0x0001183A}, {0x00011930, 0x00011935}, {0x00011937, 0x00011938},
{0x0001193B, 0x0001193E}, {0x00011940, 0x00011940}, {0x00011942, 0x00011943}, {0x000119D1, 0x000119D7},
{0x000119DA, 0x000119E0}, {0x000119E4, 0x000119E4}, {0x00011A01, 0x00011A0A}, {0x00011A33, 0x00011A39},
{0x00011A3B, 0x00011A3E}, {0x00011A47, 0x00011A47}, {0x00011A51, 0x00011A5B}, {0x00011A8A, 0x00011A99},
{0x00011C2F, 0x00011C36}, {0x00011C38, 0x00011C3F}, {0x00011C92, 0x00011CA7}, {0x00011CA9, 0x00011CB6},
{0x00011D31, 0x00011D36}, {0x00011D3A, 0x00011D3A}, {0x00011D3C, 0x00011D3D}, {0x00011D3F, 0x00011D45},
{0x00011D47, 0x00011D47}, {0x00011D8A, 0x00011D8E}, {0x00011D90, 0x00011D91}, {0x00011D93, 0x00011D97},
{0x00011EF3, 0x00011EF6}, {0x00011F00, 0x00011F01}, {0x00011F03, 0x00011F03}, {0x00011F34, 0x00011F3A},
{0x00011F3E, 0x00011F42}, {0x00013440, 0x00013440}, {0x00013447, 0x00013455}, {0x00016AF0, 0x00016AF4},
{0x00016B30, 0x00016B36}, {0x00016F4F, 0x00016F4F}, {0x00016F51, 0x00016F87}, {0x00016F8F, 0x00016F92},
{0x00016FE4, 0x00016FE4}, {0x00016FF0, 0x00016FF1}, {0x0001BC9D, 0x0001BC9E}, {0x0001CF00, 0x0001CF2D},
{0x0001CF30, 0x0001CF46}, {0x0001D165, 0x0001D169}, {0x0001D16D, 0x0001D172}, {0x0001D17B, 0x0001D182},
{0x0001D185, 0x0001D18B}, {0x0001D1AA, 0x0001D1AD}, {0x0001D242, 0x0001D244}, {0x0001DA00, 0x0001DA36},
{0x0001DA3B, 0x0001DA6C}, {0x0001DA75, 0x0001DA75}, {0x0001DA84, 0x0001DA84}, {0x0001DA9B, 0x0001DA9F},
{0x0001DAA1, 0x0001DAAF}, {0x0001E000, 0x0001E006}, {0x0001E008, 0x0001E018}, {0x0001E01B, 0x0001E021},
{0x0001E023, 0x0001E024}, {0x0001E026, 0x0001E02A}, {0x0001E08F, 0x0001E08F}, {0x0001E130, 0x0001E136},
{0x0001E2AE, 0x0001E2AE}, {0x0001E2EC, 0x0001E2EF}, {0x0001E4EC, 0x0001E4EF}, {0x0001E8D0, 0x0001E8D6},
{0x0001E944, 0x0001E94A}, {0x000E0100, 0x000E01EF},
};
@ -276,7 +309,7 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation = {
{0x000000B6, 0x000000B7}, {0x000000BB, 0x000000BB}, {0x000000BF, 0x000000BF}, {0x0000037E, 0x0000037E},
{0x00000387, 0x00000387}, {0x0000055A, 0x0000055F}, {0x00000589, 0x0000058A}, {0x000005BE, 0x000005BE},
{0x000005C0, 0x000005C0}, {0x000005C3, 0x000005C3}, {0x000005C6, 0x000005C6}, {0x000005F3, 0x000005F4},
{0x00000609, 0x0000060A}, {0x0000060C, 0x0000060D}, {0x0000061B, 0x0000061B}, {0x0000061E, 0x0000061F},
{0x00000609, 0x0000060A}, {0x0000060C, 0x0000060D}, {0x0000061B, 0x0000061B}, {0x0000061D, 0x0000061F},
{0x0000066A, 0x0000066D}, {0x000006D4, 0x000006D4}, {0x00000700, 0x0000070D}, {0x000007F7, 0x000007F9},
{0x00000830, 0x0000083E}, {0x0000085E, 0x0000085E}, {0x00000964, 0x00000965}, {0x00000970, 0x00000970},
{0x000009FD, 0x000009FD}, {0x00000A76, 0x00000A76}, {0x00000AF0, 0x00000AF0}, {0x00000C77, 0x00000C77},
@ -286,37 +319,38 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation = {
{0x00001360, 0x00001368}, {0x00001400, 0x00001400}, {0x0000166E, 0x0000166E}, {0x0000169B, 0x0000169C},
{0x000016EB, 0x000016ED}, {0x00001735, 0x00001736}, {0x000017D4, 0x000017D6}, {0x000017D8, 0x000017DA},
{0x00001800, 0x0000180A}, {0x00001944, 0x00001945}, {0x00001A1E, 0x00001A1F}, {0x00001AA0, 0x00001AA6},
{0x00001AA8, 0x00001AAD}, {0x00001B5A, 0x00001B60}, {0x00001BFC, 0x00001BFF}, {0x00001C3B, 0x00001C3F},
{0x00001C7E, 0x00001C7F}, {0x00001CC0, 0x00001CC7}, {0x00001CD3, 0x00001CD3}, {0x00002010, 0x00002027},
{0x00002030, 0x00002043}, {0x00002045, 0x00002051}, {0x00002053, 0x0000205E}, {0x0000207D, 0x0000207E},
{0x0000208D, 0x0000208E}, {0x00002308, 0x0000230B}, {0x00002329, 0x0000232A}, {0x00002768, 0x00002775},
{0x000027C5, 0x000027C6}, {0x000027E6, 0x000027EF}, {0x00002983, 0x00002998}, {0x000029D8, 0x000029DB},
{0x000029FC, 0x000029FD}, {0x00002CF9, 0x00002CFC}, {0x00002CFE, 0x00002CFF}, {0x00002D70, 0x00002D70},
{0x00002E00, 0x00002E2E}, {0x00002E30, 0x00002E4F}, {0x00002E52, 0x00002E52}, {0x00003001, 0x00003003},
{0x00003008, 0x00003011}, {0x00003014, 0x0000301F}, {0x00003030, 0x00003030}, {0x0000303D, 0x0000303D},
{0x000030A0, 0x000030A0}, {0x000030FB, 0x000030FB}, {0x0000A4FE, 0x0000A4FF}, {0x0000A60D, 0x0000A60F},
{0x0000A673, 0x0000A673}, {0x0000A67E, 0x0000A67E}, {0x0000A6F2, 0x0000A6F7}, {0x0000A874, 0x0000A877},
{0x0000A8CE, 0x0000A8CF}, {0x0000A8F8, 0x0000A8FA}, {0x0000A8FC, 0x0000A8FC}, {0x0000A92E, 0x0000A92F},
{0x0000A95F, 0x0000A95F}, {0x0000A9C1, 0x0000A9CD}, {0x0000A9DE, 0x0000A9DF}, {0x0000AA5C, 0x0000AA5F},
{0x0000AADE, 0x0000AADF}, {0x0000AAF0, 0x0000AAF1}, {0x0000ABEB, 0x0000ABEB}, {0x0000FD3E, 0x0000FD3F},
{0x0000FE10, 0x0000FE19}, {0x0000FE30, 0x0000FE52}, {0x0000FE54, 0x0000FE61}, {0x0000FE63, 0x0000FE63},
{0x0000FE68, 0x0000FE68}, {0x0000FE6A, 0x0000FE6B}, {0x0000FF01, 0x0000FF03}, {0x0000FF05, 0x0000FF0A},
{0x0000FF0C, 0x0000FF0F}, {0x0000FF1A, 0x0000FF1B}, {0x0000FF1F, 0x0000FF20}, {0x0000FF3B, 0x0000FF3D},
{0x0000FF3F, 0x0000FF3F}, {0x0000FF5B, 0x0000FF5B}, {0x0000FF5D, 0x0000FF5D}, {0x0000FF5F, 0x0000FF65},
{0x00010100, 0x00010102}, {0x0001039F, 0x0001039F}, {0x000103D0, 0x000103D0}, {0x0001056F, 0x0001056F},
{0x00010857, 0x00010857}, {0x0001091F, 0x0001091F}, {0x0001093F, 0x0001093F}, {0x00010A50, 0x00010A58},
{0x00010A7F, 0x00010A7F}, {0x00010AF0, 0x00010AF6}, {0x00010B39, 0x00010B3F}, {0x00010B99, 0x00010B9C},
{0x00010EAD, 0x00010EAD}, {0x00010F55, 0x00010F59}, {0x00011047, 0x0001104D}, {0x000110BB, 0x000110BC},
{0x000110BE, 0x000110C1}, {0x00011140, 0x00011143}, {0x00011174, 0x00011175}, {0x000111C5, 0x000111C8},
{0x000111CD, 0x000111CD}, {0x000111DB, 0x000111DB}, {0x000111DD, 0x000111DF}, {0x00011238, 0x0001123D},
{0x000112A9, 0x000112A9}, {0x0001144B, 0x0001144F}, {0x0001145A, 0x0001145B}, {0x0001145D, 0x0001145D},
{0x000114C6, 0x000114C6}, {0x000115C1, 0x000115D7}, {0x00011641, 0x00011643}, {0x00011660, 0x0001166C},
{0x0001173C, 0x0001173E}, {0x0001183B, 0x0001183B}, {0x00011944, 0x00011946}, {0x000119E2, 0x000119E2},
{0x00011A3F, 0x00011A46}, {0x00011A9A, 0x00011A9C}, {0x00011A9E, 0x00011AA2}, {0x00011C41, 0x00011C45},
{0x00011C70, 0x00011C71}, {0x00011EF7, 0x00011EF8}, {0x00011FFF, 0x00011FFF}, {0x00012470, 0x00012474},
{0x00016A6E, 0x00016A6F}, {0x00016AF5, 0x00016AF5}, {0x00016B37, 0x00016B3B}, {0x00016B44, 0x00016B44},
{0x00016E97, 0x00016E9A}, {0x00016FE2, 0x00016FE2}, {0x0001BC9F, 0x0001BC9F}, {0x0001DA87, 0x0001DA8B},
{0x0001E95E, 0x0001E95F},
{0x00001AA8, 0x00001AAD}, {0x00001B5A, 0x00001B60}, {0x00001B7D, 0x00001B7E}, {0x00001BFC, 0x00001BFF},
{0x00001C3B, 0x00001C3F}, {0x00001C7E, 0x00001C7F}, {0x00001CC0, 0x00001CC7}, {0x00001CD3, 0x00001CD3},
{0x00002010, 0x00002027}, {0x00002030, 0x00002043}, {0x00002045, 0x00002051}, {0x00002053, 0x0000205E},
{0x0000207D, 0x0000207E}, {0x0000208D, 0x0000208E}, {0x00002308, 0x0000230B}, {0x00002329, 0x0000232A},
{0x00002768, 0x00002775}, {0x000027C5, 0x000027C6}, {0x000027E6, 0x000027EF}, {0x00002983, 0x00002998},
{0x000029D8, 0x000029DB}, {0x000029FC, 0x000029FD}, {0x00002CF9, 0x00002CFC}, {0x00002CFE, 0x00002CFF},
{0x00002D70, 0x00002D70}, {0x00002E00, 0x00002E2E}, {0x00002E30, 0x00002E4F}, {0x00002E52, 0x00002E5D},
{0x00003001, 0x00003003}, {0x00003008, 0x00003011}, {0x00003014, 0x0000301F}, {0x00003030, 0x00003030},
{0x0000303D, 0x0000303D}, {0x000030A0, 0x000030A0}, {0x000030FB, 0x000030FB}, {0x0000A4FE, 0x0000A4FF},
{0x0000A60D, 0x0000A60F}, {0x0000A673, 0x0000A673}, {0x0000A67E, 0x0000A67E}, {0x0000A6F2, 0x0000A6F7},
{0x0000A874, 0x0000A877}, {0x0000A8CE, 0x0000A8CF}, {0x0000A8F8, 0x0000A8FA}, {0x0000A8FC, 0x0000A8FC},
{0x0000A92E, 0x0000A92F}, {0x0000A95F, 0x0000A95F}, {0x0000A9C1, 0x0000A9CD}, {0x0000A9DE, 0x0000A9DF},
{0x0000AA5C, 0x0000AA5F}, {0x0000AADE, 0x0000AADF}, {0x0000AAF0, 0x0000AAF1}, {0x0000ABEB, 0x0000ABEB},
{0x0000FD3E, 0x0000FD3F}, {0x0000FE10, 0x0000FE19}, {0x0000FE30, 0x0000FE52}, {0x0000FE54, 0x0000FE61},
{0x0000FE63, 0x0000FE63}, {0x0000FE68, 0x0000FE68}, {0x0000FE6A, 0x0000FE6B}, {0x0000FF01, 0x0000FF03},
{0x0000FF05, 0x0000FF0A}, {0x0000FF0C, 0x0000FF0F}, {0x0000FF1A, 0x0000FF1B}, {0x0000FF1F, 0x0000FF20},
{0x0000FF3B, 0x0000FF3D}, {0x0000FF3F, 0x0000FF3F}, {0x0000FF5B, 0x0000FF5B}, {0x0000FF5D, 0x0000FF5D},
{0x0000FF5F, 0x0000FF65}, {0x00010100, 0x00010102}, {0x0001039F, 0x0001039F}, {0x000103D0, 0x000103D0},
{0x0001056F, 0x0001056F}, {0x00010857, 0x00010857}, {0x0001091F, 0x0001091F}, {0x0001093F, 0x0001093F},
{0x00010A50, 0x00010A58}, {0x00010A7F, 0x00010A7F}, {0x00010AF0, 0x00010AF6}, {0x00010B39, 0x00010B3F},
{0x00010B99, 0x00010B9C}, {0x00010EAD, 0x00010EAD}, {0x00010F55, 0x00010F59}, {0x00010F86, 0x00010F89},
{0x00011047, 0x0001104D}, {0x000110BB, 0x000110BC}, {0x000110BE, 0x000110C1}, {0x00011140, 0x00011143},
{0x00011174, 0x00011175}, {0x000111C5, 0x000111C8}, {0x000111CD, 0x000111CD}, {0x000111DB, 0x000111DB},
{0x000111DD, 0x000111DF}, {0x00011238, 0x0001123D}, {0x000112A9, 0x000112A9}, {0x0001144B, 0x0001144F},
{0x0001145A, 0x0001145B}, {0x0001145D, 0x0001145D}, {0x000114C6, 0x000114C6}, {0x000115C1, 0x000115D7},
{0x00011641, 0x00011643}, {0x00011660, 0x0001166C}, {0x000116B9, 0x000116B9}, {0x0001173C, 0x0001173E},
{0x0001183B, 0x0001183B}, {0x00011944, 0x00011946}, {0x000119E2, 0x000119E2}, {0x00011A3F, 0x00011A46},
{0x00011A9A, 0x00011A9C}, {0x00011A9E, 0x00011AA2}, {0x00011B00, 0x00011B09}, {0x00011C41, 0x00011C45},
{0x00011C70, 0x00011C71}, {0x00011EF7, 0x00011EF8}, {0x00011F43, 0x00011F4F}, {0x00011FFF, 0x00011FFF},
{0x00012470, 0x00012474}, {0x00012FF1, 0x00012FF2}, {0x00016A6E, 0x00016A6F}, {0x00016AF5, 0x00016AF5},
{0x00016B37, 0x00016B3B}, {0x00016B44, 0x00016B44}, {0x00016E97, 0x00016E9A}, {0x00016FE2, 0x00016FE2},
{0x0001BC9F, 0x0001BC9F}, {0x0001DA87, 0x0001DA8B}, {0x0001E95E, 0x0001E95F},
};
const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol = {
@ -328,40 +362,41 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol = {
{0x00000375, 0x00000375}, {0x00000384, 0x00000385}, {0x000003F6, 0x000003F6}, {0x00000482, 0x00000482},
{0x0000058D, 0x0000058F}, {0x00000606, 0x00000608}, {0x0000060B, 0x0000060B}, {0x0000060E, 0x0000060F},
{0x000006DE, 0x000006DE}, {0x000006E9, 0x000006E9}, {0x000006FD, 0x000006FE}, {0x000007F6, 0x000007F6},
{0x000007FE, 0x000007FF}, {0x000009F2, 0x000009F3}, {0x000009FA, 0x000009FB}, {0x00000AF1, 0x00000AF1},
{0x00000B70, 0x00000B70}, {0x00000BF3, 0x00000BFA}, {0x00000C7F, 0x00000C7F}, {0x00000D4F, 0x00000D4F},
{0x00000D79, 0x00000D79}, {0x00000E3F, 0x00000E3F}, {0x00000F01, 0x00000F03}, {0x00000F13, 0x00000F13},
{0x00000F15, 0x00000F17}, {0x00000F1A, 0x00000F1F}, {0x00000F34, 0x00000F34}, {0x00000F36, 0x00000F36},
{0x00000F38, 0x00000F38}, {0x00000FBE, 0x00000FC5}, {0x00000FC7, 0x00000FCC}, {0x00000FCE, 0x00000FCF},
{0x00000FD5, 0x00000FD8}, {0x0000109E, 0x0000109F}, {0x00001390, 0x00001399}, {0x0000166D, 0x0000166D},
{0x000017DB, 0x000017DB}, {0x00001940, 0x00001940}, {0x000019DE, 0x000019FF}, {0x00001B61, 0x00001B6A},
{0x00001B74, 0x00001B7C}, {0x00001FBD, 0x00001FBD}, {0x00001FBF, 0x00001FC1}, {0x00001FCD, 0x00001FCF},
{0x00001FDD, 0x00001FDF}, {0x00001FED, 0x00001FEF}, {0x00001FFD, 0x00001FFE}, {0x00002044, 0x00002044},
{0x00002052, 0x00002052}, {0x0000207A, 0x0000207C}, {0x0000208A, 0x0000208C}, {0x000020A0, 0x000020BF},
{0x00002100, 0x00002101}, {0x00002103, 0x00002106}, {0x00002108, 0x00002109}, {0x00002114, 0x00002114},
{0x00002116, 0x00002118}, {0x0000211E, 0x00002123}, {0x00002125, 0x00002125}, {0x00002127, 0x00002127},
{0x00002129, 0x00002129}, {0x0000212E, 0x0000212E}, {0x0000213A, 0x0000213B}, {0x00002140, 0x00002144},
{0x0000214A, 0x0000214D}, {0x0000214F, 0x0000214F}, {0x0000218A, 0x0000218B}, {0x00002190, 0x00002307},
{0x0000230C, 0x00002328}, {0x0000232B, 0x00002426}, {0x00002440, 0x0000244A}, {0x0000249C, 0x000024E9},
{0x00002500, 0x00002767}, {0x00002794, 0x000027C4}, {0x000027C7, 0x000027E5}, {0x000027F0, 0x00002982},
{0x00002999, 0x000029D7}, {0x000029DC, 0x000029FB}, {0x000029FE, 0x00002B73}, {0x00002B76, 0x00002B95},
{0x00002B97, 0x00002BFF}, {0x00002CE5, 0x00002CEA}, {0x00002E50, 0x00002E51}, {0x00002E80, 0x00002E99},
{0x00002E9B, 0x00002EF3}, {0x00002F00, 0x00002FD5}, {0x00002FF0, 0x00002FFB}, {0x00003004, 0x00003004},
{0x00003012, 0x00003013}, {0x00003020, 0x00003020}, {0x00003036, 0x00003037}, {0x0000303E, 0x0000303F},
{0x0000309B, 0x0000309C}, {0x00003190, 0x00003191}, {0x00003196, 0x0000319F}, {0x000031C0, 0x000031E3},
{0x00003200, 0x0000321E}, {0x0000322A, 0x00003247}, {0x00003250, 0x00003250}, {0x00003260, 0x0000327F},
{0x0000328A, 0x000032B0}, {0x000032C0, 0x000033FF}, {0x00004DC0, 0x00004DFF}, {0x0000A490, 0x0000A4C6},
{0x0000A700, 0x0000A716}, {0x0000A720, 0x0000A721}, {0x0000A789, 0x0000A78A}, {0x0000A828, 0x0000A82B},
{0x0000A836, 0x0000A839}, {0x0000AA77, 0x0000AA79}, {0x0000AB5B, 0x0000AB5B}, {0x0000AB6A, 0x0000AB6B},
{0x0000FB29, 0x0000FB29}, {0x0000FBB2, 0x0000FBC1}, {0x0000FDFC, 0x0000FDFD}, {0x0000FE62, 0x0000FE62},
{0x0000FE64, 0x0000FE66}, {0x0000FE69, 0x0000FE69}, {0x0000FF04, 0x0000FF04}, {0x0000FF0B, 0x0000FF0B},
{0x0000FF1C, 0x0000FF1E}, {0x0000FF3E, 0x0000FF3E}, {0x0000FF40, 0x0000FF40}, {0x0000FF5C, 0x0000FF5C},
{0x0000FF5E, 0x0000FF5E}, {0x0000FFE0, 0x0000FFE6}, {0x0000FFE8, 0x0000FFEE}, {0x0000FFFC, 0x0000FFFD},
{0x00010137, 0x0001013F}, {0x00010179, 0x00010189}, {0x0001018C, 0x0001018E}, {0x00010190, 0x0001019C},
{0x000101A0, 0x000101A0}, {0x000101D0, 0x000101FC}, {0x00010877, 0x00010878}, {0x00010AC8, 0x00010AC8},
{0x0001173F, 0x0001173F}, {0x00011FD5, 0x00011FF1}, {0x00016B3C, 0x00016B3F}, {0x00016B45, 0x00016B45},
{0x0001BC9C, 0x0001BC9C}, {0x0001D000, 0x0001D0F5}, {0x0001D100, 0x0001D126}, {0x0001D129, 0x0001D164},
{0x0001D16A, 0x0001D16C}, {0x0001D183, 0x0001D184}, {0x0001D18C, 0x0001D1A9}, {0x0001D1AE, 0x0001D1E8},
{0x000007FE, 0x000007FF}, {0x00000888, 0x00000888}, {0x000009F2, 0x000009F3}, {0x000009FA, 0x000009FB},
{0x00000AF1, 0x00000AF1}, {0x00000B70, 0x00000B70}, {0x00000BF3, 0x00000BFA}, {0x00000C7F, 0x00000C7F},
{0x00000D4F, 0x00000D4F}, {0x00000D79, 0x00000D79}, {0x00000E3F, 0x00000E3F}, {0x00000F01, 0x00000F03},
{0x00000F13, 0x00000F13}, {0x00000F15, 0x00000F17}, {0x00000F1A, 0x00000F1F}, {0x00000F34, 0x00000F34},
{0x00000F36, 0x00000F36}, {0x00000F38, 0x00000F38}, {0x00000FBE, 0x00000FC5}, {0x00000FC7, 0x00000FCC},
{0x00000FCE, 0x00000FCF}, {0x00000FD5, 0x00000FD8}, {0x0000109E, 0x0000109F}, {0x00001390, 0x00001399},
{0x0000166D, 0x0000166D}, {0x000017DB, 0x000017DB}, {0x00001940, 0x00001940}, {0x000019DE, 0x000019FF},
{0x00001B61, 0x00001B6A}, {0x00001B74, 0x00001B7C}, {0x00001FBD, 0x00001FBD}, {0x00001FBF, 0x00001FC1},
{0x00001FCD, 0x00001FCF}, {0x00001FDD, 0x00001FDF}, {0x00001FED, 0x00001FEF}, {0x00001FFD, 0x00001FFE},
{0x00002044, 0x00002044}, {0x00002052, 0x00002052}, {0x0000207A, 0x0000207C}, {0x0000208A, 0x0000208C},
{0x000020A0, 0x000020C0}, {0x00002100, 0x00002101}, {0x00002103, 0x00002106}, {0x00002108, 0x00002109},
{0x00002114, 0x00002114}, {0x00002116, 0x00002118}, {0x0000211E, 0x00002123}, {0x00002125, 0x00002125},
{0x00002127, 0x00002127}, {0x00002129, 0x00002129}, {0x0000212E, 0x0000212E}, {0x0000213A, 0x0000213B},
{0x00002140, 0x00002144}, {0x0000214A, 0x0000214D}, {0x0000214F, 0x0000214F}, {0x0000218A, 0x0000218B},
{0x00002190, 0x00002307}, {0x0000230C, 0x00002328}, {0x0000232B, 0x00002426}, {0x00002440, 0x0000244A},
{0x0000249C, 0x000024E9}, {0x00002500, 0x00002767}, {0x00002794, 0x000027C4}, {0x000027C7, 0x000027E5},
{0x000027F0, 0x00002982}, {0x00002999, 0x000029D7}, {0x000029DC, 0x000029FB}, {0x000029FE, 0x00002B73},
{0x00002B76, 0x00002B95}, {0x00002B97, 0x00002BFF}, {0x00002CE5, 0x00002CEA}, {0x00002E50, 0x00002E51},
{0x00002E80, 0x00002E99}, {0x00002E9B, 0x00002EF3}, {0x00002F00, 0x00002FD5}, {0x00002FF0, 0x00002FFB},
{0x00003004, 0x00003004}, {0x00003012, 0x00003013}, {0x00003020, 0x00003020}, {0x00003036, 0x00003037},
{0x0000303E, 0x0000303F}, {0x0000309B, 0x0000309C}, {0x00003190, 0x00003191}, {0x00003196, 0x0000319F},
{0x000031C0, 0x000031E3}, {0x00003200, 0x0000321E}, {0x0000322A, 0x00003247}, {0x00003250, 0x00003250},
{0x00003260, 0x0000327F}, {0x0000328A, 0x000032B0}, {0x000032C0, 0x000033FF}, {0x00004DC0, 0x00004DFF},
{0x0000A490, 0x0000A4C6}, {0x0000A700, 0x0000A716}, {0x0000A720, 0x0000A721}, {0x0000A789, 0x0000A78A},
{0x0000A828, 0x0000A82B}, {0x0000A836, 0x0000A839}, {0x0000AA77, 0x0000AA79}, {0x0000AB5B, 0x0000AB5B},
{0x0000AB6A, 0x0000AB6B}, {0x0000FB29, 0x0000FB29}, {0x0000FBB2, 0x0000FBC2}, {0x0000FD40, 0x0000FD4F},
{0x0000FDCF, 0x0000FDCF}, {0x0000FDFC, 0x0000FDFF}, {0x0000FE62, 0x0000FE62}, {0x0000FE64, 0x0000FE66},
{0x0000FE69, 0x0000FE69}, {0x0000FF04, 0x0000FF04}, {0x0000FF0B, 0x0000FF0B}, {0x0000FF1C, 0x0000FF1E},
{0x0000FF3E, 0x0000FF3E}, {0x0000FF40, 0x0000FF40}, {0x0000FF5C, 0x0000FF5C}, {0x0000FF5E, 0x0000FF5E},
{0x0000FFE0, 0x0000FFE6}, {0x0000FFE8, 0x0000FFEE}, {0x0000FFFC, 0x0000FFFD}, {0x00010137, 0x0001013F},
{0x00010179, 0x00010189}, {0x0001018C, 0x0001018E}, {0x00010190, 0x0001019C}, {0x000101A0, 0x000101A0},
{0x000101D0, 0x000101FC}, {0x00010877, 0x00010878}, {0x00010AC8, 0x00010AC8}, {0x0001173F, 0x0001173F},
{0x00011FD5, 0x00011FF1}, {0x00016B3C, 0x00016B3F}, {0x00016B45, 0x00016B45}, {0x0001BC9C, 0x0001BC9C},
{0x0001CF50, 0x0001CFC3}, {0x0001D000, 0x0001D0F5}, {0x0001D100, 0x0001D126}, {0x0001D129, 0x0001D164},
{0x0001D16A, 0x0001D16C}, {0x0001D183, 0x0001D184}, {0x0001D18C, 0x0001D1A9}, {0x0001D1AE, 0x0001D1EA},
{0x0001D200, 0x0001D241}, {0x0001D245, 0x0001D245}, {0x0001D300, 0x0001D356}, {0x0001D6C1, 0x0001D6C1},
{0x0001D6DB, 0x0001D6DB}, {0x0001D6FB, 0x0001D6FB}, {0x0001D715, 0x0001D715}, {0x0001D735, 0x0001D735},
{0x0001D74F, 0x0001D74F}, {0x0001D76F, 0x0001D76F}, {0x0001D789, 0x0001D789}, {0x0001D7A9, 0x0001D7A9},
@ -371,102 +406,100 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_symbol = {
{0x0001F000, 0x0001F02B}, {0x0001F030, 0x0001F093}, {0x0001F0A0, 0x0001F0AE}, {0x0001F0B1, 0x0001F0BF},
{0x0001F0C1, 0x0001F0CF}, {0x0001F0D1, 0x0001F0F5}, {0x0001F10D, 0x0001F1AD}, {0x0001F1E6, 0x0001F202},
{0x0001F210, 0x0001F23B}, {0x0001F240, 0x0001F248}, {0x0001F250, 0x0001F251}, {0x0001F260, 0x0001F265},
{0x0001F300, 0x0001F6D7}, {0x0001F6E0, 0x0001F6EC}, {0x0001F6F0, 0x0001F6FC}, {0x0001F700, 0x0001F773},
{0x0001F780, 0x0001F7D8}, {0x0001F7E0, 0x0001F7EB}, {0x0001F800, 0x0001F80B}, {0x0001F810, 0x0001F847},
{0x0001F850, 0x0001F859}, {0x0001F860, 0x0001F887}, {0x0001F890, 0x0001F8AD}, {0x0001F8B0, 0x0001F8B1},
{0x0001F900, 0x0001F978}, {0x0001F97A, 0x0001F9CB}, {0x0001F9CD, 0x0001FA53}, {0x0001FA60, 0x0001FA6D},
{0x0001FA70, 0x0001FA74}, {0x0001FA78, 0x0001FA7A}, {0x0001FA80, 0x0001FA86}, {0x0001FA90, 0x0001FAA8},
{0x0001FAB0, 0x0001FAB6}, {0x0001FAC0, 0x0001FAC2}, {0x0001FAD0, 0x0001FAD6}, {0x0001FB00, 0x0001FB92},
{0x0001FB94, 0x0001FBCA},
{0x0001F300, 0x0001F6D7}, {0x0001F6DC, 0x0001F6EC}, {0x0001F6F0, 0x0001F6FC}, {0x0001F700, 0x0001F776},
{0x0001F77B, 0x0001F7D9}, {0x0001F7E0, 0x0001F7EB}, {0x0001F7F0, 0x0001F7F0}, {0x0001F800, 0x0001F80B},
{0x0001F810, 0x0001F847}, {0x0001F850, 0x0001F859}, {0x0001F860, 0x0001F887}, {0x0001F890, 0x0001F8AD},
{0x0001F8B0, 0x0001F8B1}, {0x0001F900, 0x0001FA53}, {0x0001FA60, 0x0001FA6D}, {0x0001FA70, 0x0001FA7C},
{0x0001FA80, 0x0001FA88}, {0x0001FA90, 0x0001FABD}, {0x0001FABF, 0x0001FAC5}, {0x0001FACE, 0x0001FADB},
{0x0001FAE0, 0x0001FAE8}, {0x0001FAF0, 0x0001FAF8}, {0x0001FB00, 0x0001FB92}, {0x0001FB94, 0x0001FBCA},
};
const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control = {
{0x00000000, 0x00000008}, {0x0000000E, 0x0000001B}, {0x0000007F, 0x00000084}, {0x00000086, 0x0000009F},
{0x000000AD, 0x000000AD}, {0x00000378, 0x00000379}, {0x00000380, 0x00000383}, {0x0000038B, 0x0000038B},
{0x0000038D, 0x0000038D}, {0x000003A2, 0x000003A2}, {0x00000530, 0x00000530}, {0x00000557, 0x00000558},
{0x0000058B, 0x0000058C}, {0x00000590, 0x00000590}, {0x000005C8, 0x000005CF}, {0x000005EB, 0x000005EE},
{0x000005F5, 0x00000605}, {0x0000061C, 0x0000061D}, {0x000006DD, 0x000006DD}, {0x0000070E, 0x0000070F},
{0x0000074B, 0x0000074C}, {0x000007B2, 0x000007BF}, {0x000007FB, 0x000007FC}, {0x0000082E, 0x0000082F},
{0x0000083F, 0x0000083F}, {0x0000085C, 0x0000085D}, {0x0000085F, 0x0000085F}, {0x0000086B, 0x0000089F},
{0x000008B5, 0x000008B5}, {0x000008C8, 0x000008D2}, {0x000008E2, 0x000008E2}, {0x00000984, 0x00000984},
{0x0000098D, 0x0000098E}, {0x00000991, 0x00000992}, {0x000009A9, 0x000009A9}, {0x000009B1, 0x000009B1},
{0x000009B3, 0x000009B5}, {0x000009BA, 0x000009BB}, {0x000009C5, 0x000009C6}, {0x000009C9, 0x000009CA},
{0x000009CF, 0x000009D6}, {0x000009D8, 0x000009DB}, {0x000009DE, 0x000009DE}, {0x000009E4, 0x000009E5},
{0x000009FF, 0x00000A00}, {0x00000A04, 0x00000A04}, {0x00000A0B, 0x00000A0E}, {0x00000A11, 0x00000A12},
{0x00000A29, 0x00000A29}, {0x00000A31, 0x00000A31}, {0x00000A34, 0x00000A34}, {0x00000A37, 0x00000A37},
{0x00000A3A, 0x00000A3B}, {0x00000A3D, 0x00000A3D}, {0x00000A43, 0x00000A46}, {0x00000A49, 0x00000A4A},
{0x00000A4E, 0x00000A50}, {0x00000A52, 0x00000A58}, {0x00000A5D, 0x00000A5D}, {0x00000A5F, 0x00000A65},
{0x00000A77, 0x00000A80}, {0x00000A84, 0x00000A84}, {0x00000A8E, 0x00000A8E}, {0x00000A92, 0x00000A92},
{0x00000AA9, 0x00000AA9}, {0x00000AB1, 0x00000AB1}, {0x00000AB4, 0x00000AB4}, {0x00000ABA, 0x00000ABB},
{0x00000AC6, 0x00000AC6}, {0x00000ACA, 0x00000ACA}, {0x00000ACE, 0x00000ACF}, {0x00000AD1, 0x00000ADF},
{0x00000AE4, 0x00000AE5}, {0x00000AF2, 0x00000AF8}, {0x00000B00, 0x00000B00}, {0x00000B04, 0x00000B04},
{0x00000B0D, 0x00000B0E}, {0x00000B11, 0x00000B12}, {0x00000B29, 0x00000B29}, {0x00000B31, 0x00000B31},
{0x00000B34, 0x00000B34}, {0x00000B3A, 0x00000B3B}, {0x00000B45, 0x00000B46}, {0x00000B49, 0x00000B4A},
{0x00000B4E, 0x00000B54}, {0x00000B58, 0x00000B5B}, {0x00000B5E, 0x00000B5E}, {0x00000B64, 0x00000B65},
{0x00000B78, 0x00000B81}, {0x00000B84, 0x00000B84}, {0x00000B8B, 0x00000B8D}, {0x00000B91, 0x00000B91},
{0x00000B96, 0x00000B98}, {0x00000B9B, 0x00000B9B}, {0x00000B9D, 0x00000B9D}, {0x00000BA0, 0x00000BA2},
{0x00000BA5, 0x00000BA7}, {0x00000BAB, 0x00000BAD}, {0x00000BBA, 0x00000BBD}, {0x00000BC3, 0x00000BC5},
{0x00000BC9, 0x00000BC9}, {0x00000BCE, 0x00000BCF}, {0x00000BD1, 0x00000BD6}, {0x00000BD8, 0x00000BE5},
{0x00000BFB, 0x00000BFF}, {0x00000C0D, 0x00000C0D}, {0x00000C11, 0x00000C11}, {0x00000C29, 0x00000C29},
{0x00000C3A, 0x00000C3C}, {0x00000C45, 0x00000C45}, {0x00000C49, 0x00000C49}, {0x00000C4E, 0x00000C54},
{0x00000C57, 0x00000C57}, {0x00000C5B, 0x00000C5F}, {0x00000C64, 0x00000C65}, {0x00000C70, 0x00000C76},
{0x00000C8D, 0x00000C8D}, {0x00000C91, 0x00000C91}, {0x00000CA9, 0x00000CA9}, {0x00000CB4, 0x00000CB4},
{0x00000CBA, 0x00000CBB}, {0x00000CC5, 0x00000CC5}, {0x00000CC9, 0x00000CC9}, {0x00000CCE, 0x00000CD4},
{0x00000CD7, 0x00000CDD}, {0x00000CDF, 0x00000CDF}, {0x00000CE4, 0x00000CE5}, {0x00000CF0, 0x00000CF0},
{0x00000CF3, 0x00000CFF}, {0x00000D0D, 0x00000D0D}, {0x00000D11, 0x00000D11}, {0x00000D45, 0x00000D45},
{0x00000D49, 0x00000D49}, {0x00000D50, 0x00000D53}, {0x00000D64, 0x00000D65}, {0x00000D80, 0x00000D80},
{0x00000D84, 0x00000D84}, {0x00000D97, 0x00000D99}, {0x00000DB2, 0x00000DB2}, {0x00000DBC, 0x00000DBC},
{0x00000DBE, 0x00000DBF}, {0x00000DC7, 0x00000DC9}, {0x00000DCB, 0x00000DCE}, {0x00000DD5, 0x00000DD5},
{0x00000DD7, 0x00000DD7}, {0x00000DE0, 0x00000DE5}, {0x00000DF0, 0x00000DF1}, {0x00000DF5, 0x00000E00},
{0x00000E3B, 0x00000E3E}, {0x00000E5C, 0x00000E80}, {0x00000E83, 0x00000E83}, {0x00000E85, 0x00000E85},
{0x00000E8B, 0x00000E8B}, {0x00000EA4, 0x00000EA4}, {0x00000EA6, 0x00000EA6}, {0x00000EBE, 0x00000EBF},
{0x00000EC5, 0x00000EC5}, {0x00000EC7, 0x00000EC7}, {0x00000ECE, 0x00000ECF}, {0x00000EDA, 0x00000EDB},
{0x00000EE0, 0x00000EFF}, {0x00000F48, 0x00000F48}, {0x00000F6D, 0x00000F70}, {0x00000F98, 0x00000F98},
{0x00000FBD, 0x00000FBD}, {0x00000FCD, 0x00000FCD}, {0x00000FDB, 0x00000FFF}, {0x000010C6, 0x000010C6},
{0x000010C8, 0x000010CC}, {0x000010CE, 0x000010CF}, {0x00001249, 0x00001249}, {0x0000124E, 0x0000124F},
{0x00001257, 0x00001257}, {0x00001259, 0x00001259}, {0x0000125E, 0x0000125F}, {0x00001289, 0x00001289},
{0x0000128E, 0x0000128F}, {0x000012B1, 0x000012B1}, {0x000012B6, 0x000012B7}, {0x000012BF, 0x000012BF},
{0x000012C1, 0x000012C1}, {0x000012C6, 0x000012C7}, {0x000012D7, 0x000012D7}, {0x00001311, 0x00001311},
{0x00001316, 0x00001317}, {0x0000135B, 0x0000135C}, {0x0000137D, 0x0000137F}, {0x0000139A, 0x0000139F},
{0x000013F6, 0x000013F7}, {0x000013FE, 0x000013FF}, {0x0000169D, 0x0000169F}, {0x000016F9, 0x000016FF},
{0x0000170D, 0x0000170D}, {0x00001715, 0x0000171F}, {0x00001737, 0x0000173F}, {0x00001754, 0x0000175F},
{0x0000176D, 0x0000176D}, {0x00001771, 0x00001771}, {0x00001774, 0x0000177F}, {0x000017DE, 0x000017DF},
{0x000017EA, 0x000017EF}, {0x000017FA, 0x000017FF}, {0x0000180E, 0x0000180F}, {0x0000181A, 0x0000181F},
{0x00001879, 0x0000187F}, {0x000018AB, 0x000018AF}, {0x000018F6, 0x000018FF}, {0x0000191F, 0x0000191F},
{0x0000192C, 0x0000192F}, {0x0000193C, 0x0000193F}, {0x00001941, 0x00001943}, {0x0000196E, 0x0000196F},
{0x00001975, 0x0000197F}, {0x000019AC, 0x000019AF}, {0x000019CA, 0x000019CF}, {0x000019DB, 0x000019DD},
{0x00001A1C, 0x00001A1D}, {0x00001A5F, 0x00001A5F}, {0x00001A7D, 0x00001A7E}, {0x00001A8A, 0x00001A8F},
{0x00001A9A, 0x00001A9F}, {0x00001AAE, 0x00001AAF}, {0x00001AC1, 0x00001AFF}, {0x00001B4C, 0x00001B4F},
{0x00001B7D, 0x00001B7F}, {0x00001BF4, 0x00001BFB}, {0x00001C38, 0x00001C3A}, {0x00001C4A, 0x00001C4C},
{0x00001C89, 0x00001C8F}, {0x00001CBB, 0x00001CBC}, {0x00001CC8, 0x00001CCF}, {0x00001CFB, 0x00001CFF},
{0x00001DFA, 0x00001DFA}, {0x00001F16, 0x00001F17}, {0x00001F1E, 0x00001F1F}, {0x00001F46, 0x00001F47},
{0x00000000, 0x0000001F}, {0x0000007F, 0x0000009F}, {0x000000AD, 0x000000AD}, {0x00000378, 0x00000379},
{0x00000380, 0x00000383}, {0x0000038B, 0x0000038B}, {0x0000038D, 0x0000038D}, {0x000003A2, 0x000003A2},
{0x00000530, 0x00000530}, {0x00000557, 0x00000558}, {0x0000058B, 0x0000058C}, {0x00000590, 0x00000590},
{0x000005C8, 0x000005CF}, {0x000005EB, 0x000005EE}, {0x000005F5, 0x00000605}, {0x0000061C, 0x0000061C},
{0x000006DD, 0x000006DD}, {0x0000070E, 0x0000070F}, {0x0000074B, 0x0000074C}, {0x000007B2, 0x000007BF},
{0x000007FB, 0x000007FC}, {0x0000082E, 0x0000082F}, {0x0000083F, 0x0000083F}, {0x0000085C, 0x0000085D},
{0x0000085F, 0x0000085F}, {0x0000086B, 0x0000086F}, {0x0000088F, 0x00000897}, {0x000008E2, 0x000008E2},
{0x00000984, 0x00000984}, {0x0000098D, 0x0000098E}, {0x00000991, 0x00000992}, {0x000009A9, 0x000009A9},
{0x000009B1, 0x000009B1}, {0x000009B3, 0x000009B5}, {0x000009BA, 0x000009BB}, {0x000009C5, 0x000009C6},
{0x000009C9, 0x000009CA}, {0x000009CF, 0x000009D6}, {0x000009D8, 0x000009DB}, {0x000009DE, 0x000009DE},
{0x000009E4, 0x000009E5}, {0x000009FF, 0x00000A00}, {0x00000A04, 0x00000A04}, {0x00000A0B, 0x00000A0E},
{0x00000A11, 0x00000A12}, {0x00000A29, 0x00000A29}, {0x00000A31, 0x00000A31}, {0x00000A34, 0x00000A34},
{0x00000A37, 0x00000A37}, {0x00000A3A, 0x00000A3B}, {0x00000A3D, 0x00000A3D}, {0x00000A43, 0x00000A46},
{0x00000A49, 0x00000A4A}, {0x00000A4E, 0x00000A50}, {0x00000A52, 0x00000A58}, {0x00000A5D, 0x00000A5D},
{0x00000A5F, 0x00000A65}, {0x00000A77, 0x00000A80}, {0x00000A84, 0x00000A84}, {0x00000A8E, 0x00000A8E},
{0x00000A92, 0x00000A92}, {0x00000AA9, 0x00000AA9}, {0x00000AB1, 0x00000AB1}, {0x00000AB4, 0x00000AB4},
{0x00000ABA, 0x00000ABB}, {0x00000AC6, 0x00000AC6}, {0x00000ACA, 0x00000ACA}, {0x00000ACE, 0x00000ACF},
{0x00000AD1, 0x00000ADF}, {0x00000AE4, 0x00000AE5}, {0x00000AF2, 0x00000AF8}, {0x00000B00, 0x00000B00},
{0x00000B04, 0x00000B04}, {0x00000B0D, 0x00000B0E}, {0x00000B11, 0x00000B12}, {0x00000B29, 0x00000B29},
{0x00000B31, 0x00000B31}, {0x00000B34, 0x00000B34}, {0x00000B3A, 0x00000B3B}, {0x00000B45, 0x00000B46},
{0x00000B49, 0x00000B4A}, {0x00000B4E, 0x00000B54}, {0x00000B58, 0x00000B5B}, {0x00000B5E, 0x00000B5E},
{0x00000B64, 0x00000B65}, {0x00000B78, 0x00000B81}, {0x00000B84, 0x00000B84}, {0x00000B8B, 0x00000B8D},
{0x00000B91, 0x00000B91}, {0x00000B96, 0x00000B98}, {0x00000B9B, 0x00000B9B}, {0x00000B9D, 0x00000B9D},
{0x00000BA0, 0x00000BA2}, {0x00000BA5, 0x00000BA7}, {0x00000BAB, 0x00000BAD}, {0x00000BBA, 0x00000BBD},
{0x00000BC3, 0x00000BC5}, {0x00000BC9, 0x00000BC9}, {0x00000BCE, 0x00000BCF}, {0x00000BD1, 0x00000BD6},
{0x00000BD8, 0x00000BE5}, {0x00000BFB, 0x00000BFF}, {0x00000C0D, 0x00000C0D}, {0x00000C11, 0x00000C11},
{0x00000C29, 0x00000C29}, {0x00000C3A, 0x00000C3B}, {0x00000C45, 0x00000C45}, {0x00000C49, 0x00000C49},
{0x00000C4E, 0x00000C54}, {0x00000C57, 0x00000C57}, {0x00000C5B, 0x00000C5C}, {0x00000C5E, 0x00000C5F},
{0x00000C64, 0x00000C65}, {0x00000C70, 0x00000C76}, {0x00000C8D, 0x00000C8D}, {0x00000C91, 0x00000C91},
{0x00000CA9, 0x00000CA9}, {0x00000CB4, 0x00000CB4}, {0x00000CBA, 0x00000CBB}, {0x00000CC5, 0x00000CC5},
{0x00000CC9, 0x00000CC9}, {0x00000CCE, 0x00000CD4}, {0x00000CD7, 0x00000CDC}, {0x00000CDF, 0x00000CDF},
{0x00000CE4, 0x00000CE5}, {0x00000CF0, 0x00000CF0}, {0x00000CF4, 0x00000CFF}, {0x00000D0D, 0x00000D0D},
{0x00000D11, 0x00000D11}, {0x00000D45, 0x00000D45}, {0x00000D49, 0x00000D49}, {0x00000D50, 0x00000D53},
{0x00000D64, 0x00000D65}, {0x00000D80, 0x00000D80}, {0x00000D84, 0x00000D84}, {0x00000D97, 0x00000D99},
{0x00000DB2, 0x00000DB2}, {0x00000DBC, 0x00000DBC}, {0x00000DBE, 0x00000DBF}, {0x00000DC7, 0x00000DC9},
{0x00000DCB, 0x00000DCE}, {0x00000DD5, 0x00000DD5}, {0x00000DD7, 0x00000DD7}, {0x00000DE0, 0x00000DE5},
{0x00000DF0, 0x00000DF1}, {0x00000DF5, 0x00000E00}, {0x00000E3B, 0x00000E3E}, {0x00000E5C, 0x00000E80},
{0x00000E83, 0x00000E83}, {0x00000E85, 0x00000E85}, {0x00000E8B, 0x00000E8B}, {0x00000EA4, 0x00000EA4},
{0x00000EA6, 0x00000EA6}, {0x00000EBE, 0x00000EBF}, {0x00000EC5, 0x00000EC5}, {0x00000EC7, 0x00000EC7},
{0x00000ECF, 0x00000ECF}, {0x00000EDA, 0x00000EDB}, {0x00000EE0, 0x00000EFF}, {0x00000F48, 0x00000F48},
{0x00000F6D, 0x00000F70}, {0x00000F98, 0x00000F98}, {0x00000FBD, 0x00000FBD}, {0x00000FCD, 0x00000FCD},
{0x00000FDB, 0x00000FFF}, {0x000010C6, 0x000010C6}, {0x000010C8, 0x000010CC}, {0x000010CE, 0x000010CF},
{0x00001249, 0x00001249}, {0x0000124E, 0x0000124F}, {0x00001257, 0x00001257}, {0x00001259, 0x00001259},
{0x0000125E, 0x0000125F}, {0x00001289, 0x00001289}, {0x0000128E, 0x0000128F}, {0x000012B1, 0x000012B1},
{0x000012B6, 0x000012B7}, {0x000012BF, 0x000012BF}, {0x000012C1, 0x000012C1}, {0x000012C6, 0x000012C7},
{0x000012D7, 0x000012D7}, {0x00001311, 0x00001311}, {0x00001316, 0x00001317}, {0x0000135B, 0x0000135C},
{0x0000137D, 0x0000137F}, {0x0000139A, 0x0000139F}, {0x000013F6, 0x000013F7}, {0x000013FE, 0x000013FF},
{0x0000169D, 0x0000169F}, {0x000016F9, 0x000016FF}, {0x00001716, 0x0000171E}, {0x00001737, 0x0000173F},
{0x00001754, 0x0000175F}, {0x0000176D, 0x0000176D}, {0x00001771, 0x00001771}, {0x00001774, 0x0000177F},
{0x000017DE, 0x000017DF}, {0x000017EA, 0x000017EF}, {0x000017FA, 0x000017FF}, {0x0000180E, 0x0000180E},
{0x0000181A, 0x0000181F}, {0x00001879, 0x0000187F}, {0x000018AB, 0x000018AF}, {0x000018F6, 0x000018FF},
{0x0000191F, 0x0000191F}, {0x0000192C, 0x0000192F}, {0x0000193C, 0x0000193F}, {0x00001941, 0x00001943},
{0x0000196E, 0x0000196F}, {0x00001975, 0x0000197F}, {0x000019AC, 0x000019AF}, {0x000019CA, 0x000019CF},
{0x000019DB, 0x000019DD}, {0x00001A1C, 0x00001A1D}, {0x00001A5F, 0x00001A5F}, {0x00001A7D, 0x00001A7E},
{0x00001A8A, 0x00001A8F}, {0x00001A9A, 0x00001A9F}, {0x00001AAE, 0x00001AAF}, {0x00001ACF, 0x00001AFF},
{0x00001B4D, 0x00001B4F}, {0x00001B7F, 0x00001B7F}, {0x00001BF4, 0x00001BFB}, {0x00001C38, 0x00001C3A},
{0x00001C4A, 0x00001C4C}, {0x00001C89, 0x00001C8F}, {0x00001CBB, 0x00001CBC}, {0x00001CC8, 0x00001CCF},
{0x00001CFB, 0x00001CFF}, {0x00001F16, 0x00001F17}, {0x00001F1E, 0x00001F1F}, {0x00001F46, 0x00001F47},
{0x00001F4E, 0x00001F4F}, {0x00001F58, 0x00001F58}, {0x00001F5A, 0x00001F5A}, {0x00001F5C, 0x00001F5C},
{0x00001F5E, 0x00001F5E}, {0x00001F7E, 0x00001F7F}, {0x00001FB5, 0x00001FB5}, {0x00001FC5, 0x00001FC5},
{0x00001FD4, 0x00001FD5}, {0x00001FDC, 0x00001FDC}, {0x00001FF0, 0x00001FF1}, {0x00001FF5, 0x00001FF5},
{0x00001FFF, 0x00001FFF}, {0x0000200B, 0x0000200F}, {0x0000202A, 0x0000202E}, {0x00002060, 0x0000206F},
{0x00002072, 0x00002073}, {0x0000208F, 0x0000208F}, {0x0000209D, 0x0000209F}, {0x000020C0, 0x000020CF},
{0x00002072, 0x00002073}, {0x0000208F, 0x0000208F}, {0x0000209D, 0x0000209F}, {0x000020C1, 0x000020CF},
{0x000020F1, 0x000020FF}, {0x0000218C, 0x0000218F}, {0x00002427, 0x0000243F}, {0x0000244B, 0x0000245F},
{0x00002B74, 0x00002B75}, {0x00002B96, 0x00002B96}, {0x00002C2F, 0x00002C2F}, {0x00002C5F, 0x00002C5F},
{0x00002CF4, 0x00002CF8}, {0x00002D26, 0x00002D26}, {0x00002D28, 0x00002D2C}, {0x00002D2E, 0x00002D2F},
{0x00002D68, 0x00002D6E}, {0x00002D71, 0x00002D7E}, {0x00002D97, 0x00002D9F}, {0x00002DA7, 0x00002DA7},
{0x00002DAF, 0x00002DAF}, {0x00002DB7, 0x00002DB7}, {0x00002DBF, 0x00002DBF}, {0x00002DC7, 0x00002DC7},
{0x00002DCF, 0x00002DCF}, {0x00002DD7, 0x00002DD7}, {0x00002DDF, 0x00002DDF}, {0x00002E53, 0x00002E7F},
{0x00002E9A, 0x00002E9A}, {0x00002EF4, 0x00002EFF}, {0x00002FD6, 0x00002FEF}, {0x00002FFC, 0x00002FFF},
{0x00003040, 0x00003040}, {0x00003097, 0x00003098}, {0x00003100, 0x00003104}, {0x00003130, 0x00003130},
{0x0000318F, 0x0000318F}, {0x000031E4, 0x000031EF}, {0x0000321F, 0x0000321F}, {0x00009FFD, 0x00009FFF},
{0x0000A48D, 0x0000A48F}, {0x0000A4C7, 0x0000A4CF}, {0x0000A62C, 0x0000A63F}, {0x0000A6F8, 0x0000A6FF},
{0x0000A7C0, 0x0000A7C1}, {0x0000A7CB, 0x0000A7F4}, {0x0000A82D, 0x0000A82F}, {0x0000A83A, 0x0000A83F},
{0x0000A878, 0x0000A87F}, {0x0000A8C6, 0x0000A8CD}, {0x0000A8DA, 0x0000A8DF}, {0x0000A954, 0x0000A95E},
{0x0000A97D, 0x0000A97F}, {0x0000A9CE, 0x0000A9CE}, {0x0000A9DA, 0x0000A9DD}, {0x0000A9FF, 0x0000A9FF},
{0x0000AA37, 0x0000AA3F}, {0x0000AA4E, 0x0000AA4F}, {0x0000AA5A, 0x0000AA5B}, {0x0000AAC3, 0x0000AADA},
{0x0000AAF7, 0x0000AB00}, {0x0000AB07, 0x0000AB08}, {0x0000AB0F, 0x0000AB10}, {0x0000AB17, 0x0000AB1F},
{0x0000AB27, 0x0000AB27}, {0x0000AB2F, 0x0000AB2F}, {0x0000AB6C, 0x0000AB6F}, {0x0000ABEE, 0x0000ABEF},
{0x0000ABFA, 0x0000ABFF}, {0x0000D7A4, 0x0000D7AF}, {0x0000D7C7, 0x0000D7CA}, {0x0000D7FC, 0x0000F8FF},
{0x00002B74, 0x00002B75}, {0x00002B96, 0x00002B96}, {0x00002CF4, 0x00002CF8}, {0x00002D26, 0x00002D26},
{0x00002D28, 0x00002D2C}, {0x00002D2E, 0x00002D2F}, {0x00002D68, 0x00002D6E}, {0x00002D71, 0x00002D7E},
{0x00002D97, 0x00002D9F}, {0x00002DA7, 0x00002DA7}, {0x00002DAF, 0x00002DAF}, {0x00002DB7, 0x00002DB7},
{0x00002DBF, 0x00002DBF}, {0x00002DC7, 0x00002DC7}, {0x00002DCF, 0x00002DCF}, {0x00002DD7, 0x00002DD7},
{0x00002DDF, 0x00002DDF}, {0x00002E5E, 0x00002E7F}, {0x00002E9A, 0x00002E9A}, {0x00002EF4, 0x00002EFF},
{0x00002FD6, 0x00002FEF}, {0x00002FFC, 0x00002FFF}, {0x00003040, 0x00003040}, {0x00003097, 0x00003098},
{0x00003100, 0x00003104}, {0x00003130, 0x00003130}, {0x0000318F, 0x0000318F}, {0x000031E4, 0x000031EF},
{0x0000321F, 0x0000321F}, {0x0000A48D, 0x0000A48F}, {0x0000A4C7, 0x0000A4CF}, {0x0000A62C, 0x0000A63F},
{0x0000A6F8, 0x0000A6FF}, {0x0000A7CB, 0x0000A7CF}, {0x0000A7D2, 0x0000A7D2}, {0x0000A7D4, 0x0000A7D4},
{0x0000A7DA, 0x0000A7F1}, {0x0000A82D, 0x0000A82F}, {0x0000A83A, 0x0000A83F}, {0x0000A878, 0x0000A87F},
{0x0000A8C6, 0x0000A8CD}, {0x0000A8DA, 0x0000A8DF}, {0x0000A954, 0x0000A95E}, {0x0000A97D, 0x0000A97F},
{0x0000A9CE, 0x0000A9CE}, {0x0000A9DA, 0x0000A9DD}, {0x0000A9FF, 0x0000A9FF}, {0x0000AA37, 0x0000AA3F},
{0x0000AA4E, 0x0000AA4F}, {0x0000AA5A, 0x0000AA5B}, {0x0000AAC3, 0x0000AADA}, {0x0000AAF7, 0x0000AB00},
{0x0000AB07, 0x0000AB08}, {0x0000AB0F, 0x0000AB10}, {0x0000AB17, 0x0000AB1F}, {0x0000AB27, 0x0000AB27},
{0x0000AB2F, 0x0000AB2F}, {0x0000AB6C, 0x0000AB6F}, {0x0000ABEE, 0x0000ABEF}, {0x0000ABFA, 0x0000ABFF},
{0x0000D7A4, 0x0000D7AF}, {0x0000D7C7, 0x0000D7CA}, {0x0000D7FC, 0x0000D7FF}, {0x0000E000, 0x0000F8FF},
{0x0000FA6E, 0x0000FA6F}, {0x0000FADA, 0x0000FAFF}, {0x0000FB07, 0x0000FB12}, {0x0000FB18, 0x0000FB1C},
{0x0000FB37, 0x0000FB37}, {0x0000FB3D, 0x0000FB3D}, {0x0000FB3F, 0x0000FB3F}, {0x0000FB42, 0x0000FB42},
{0x0000FB45, 0x0000FB45}, {0x0000FBC2, 0x0000FBD2}, {0x0000FD40, 0x0000FD4F}, {0x0000FD90, 0x0000FD91},
{0x0000FDC8, 0x0000FDEF}, {0x0000FDFE, 0x0000FDFF}, {0x0000FE1A, 0x0000FE1F}, {0x0000FE53, 0x0000FE53},
{0x0000FE67, 0x0000FE67}, {0x0000FE6C, 0x0000FE6F}, {0x0000FE75, 0x0000FE75}, {0x0000FEFD, 0x0000FF00},
{0x0000FB45, 0x0000FB45}, {0x0000FBC3, 0x0000FBD2}, {0x0000FD90, 0x0000FD91}, {0x0000FDC8, 0x0000FDCE},
{0x0000FDD0, 0x0000FDEF}, {0x0000FE1A, 0x0000FE1F}, {0x0000FE53, 0x0000FE53}, {0x0000FE67, 0x0000FE67},
{0x0000FE6C, 0x0000FE6F}, {0x0000FE75, 0x0000FE75}, {0x0000FEFD, 0x0000FEFE}, {0x0000FF00, 0x0000FF00},
{0x0000FFBF, 0x0000FFC1}, {0x0000FFC8, 0x0000FFC9}, {0x0000FFD0, 0x0000FFD1}, {0x0000FFD8, 0x0000FFD9},
{0x0000FFDD, 0x0000FFDF}, {0x0000FFE7, 0x0000FFE7}, {0x0000FFEF, 0x0000FFFB}, {0x0000FFFE, 0x0000FFFF},
{0x0001000C, 0x0001000C}, {0x00010027, 0x00010027}, {0x0001003B, 0x0001003B}, {0x0001003E, 0x0001003E},
@ -476,82 +509,91 @@ const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_control = {
{0x00010324, 0x0001032C}, {0x0001034B, 0x0001034F}, {0x0001037B, 0x0001037F}, {0x0001039E, 0x0001039E},
{0x000103C4, 0x000103C7}, {0x000103D6, 0x000103FF}, {0x0001049E, 0x0001049F}, {0x000104AA, 0x000104AF},
{0x000104D4, 0x000104D7}, {0x000104FC, 0x000104FF}, {0x00010528, 0x0001052F}, {0x00010564, 0x0001056E},
{0x00010570, 0x000105FF}, {0x00010737, 0x0001073F}, {0x00010756, 0x0001075F}, {0x00010768, 0x000107FF},
{0x00010806, 0x00010807}, {0x00010809, 0x00010809}, {0x00010836, 0x00010836}, {0x00010839, 0x0001083B},
{0x0001083D, 0x0001083E}, {0x00010856, 0x00010856}, {0x0001089F, 0x000108A6}, {0x000108B0, 0x000108DF},
{0x000108F3, 0x000108F3}, {0x000108F6, 0x000108FA}, {0x0001091C, 0x0001091E}, {0x0001093A, 0x0001093E},
{0x00010940, 0x0001097F}, {0x000109B8, 0x000109BB}, {0x000109D0, 0x000109D1}, {0x00010A04, 0x00010A04},
{0x00010A07, 0x00010A0B}, {0x00010A14, 0x00010A14}, {0x00010A18, 0x00010A18}, {0x00010A36, 0x00010A37},
{0x00010A3B, 0x00010A3E}, {0x00010A49, 0x00010A4F}, {0x00010A59, 0x00010A5F}, {0x00010AA0, 0x00010ABF},
{0x00010AE7, 0x00010AEA}, {0x00010AF7, 0x00010AFF}, {0x00010B36, 0x00010B38}, {0x00010B56, 0x00010B57},
{0x00010B73, 0x00010B77}, {0x00010B92, 0x00010B98}, {0x00010B9D, 0x00010BA8}, {0x00010BB0, 0x00010BFF},
{0x00010C49, 0x00010C7F}, {0x00010CB3, 0x00010CBF}, {0x00010CF3, 0x00010CF9}, {0x00010D28, 0x00010D2F},
{0x00010D3A, 0x00010E5F}, {0x00010E7F, 0x00010E7F}, {0x00010EAA, 0x00010EAA}, {0x00010EAE, 0x00010EAF},
{0x00010EB2, 0x00010EFF}, {0x00010F28, 0x00010F2F}, {0x00010F5A, 0x00010FAF}, {0x00010FCC, 0x00010FDF},
{0x00010FF7, 0x00010FFF}, {0x0001104E, 0x00011051}, {0x00011070, 0x0001107E}, {0x000110BD, 0x000110BD},
{0x000110C2, 0x000110CF}, {0x000110E9, 0x000110EF}, {0x000110FA, 0x000110FF}, {0x00011135, 0x00011135},
{0x00011148, 0x0001114F}, {0x00011177, 0x0001117F}, {0x000111E0, 0x000111E0}, {0x000111F5, 0x000111FF},
{0x00011212, 0x00011212}, {0x0001123F, 0x0001127F}, {0x00011287, 0x00011287}, {0x00011289, 0x00011289},
{0x0001128E, 0x0001128E}, {0x0001129E, 0x0001129E}, {0x000112AA, 0x000112AF}, {0x000112EB, 0x000112EF},
{0x000112FA, 0x000112FF}, {0x00011304, 0x00011304}, {0x0001130D, 0x0001130E}, {0x00011311, 0x00011312},
{0x00011329, 0x00011329}, {0x00011331, 0x00011331}, {0x00011334, 0x00011334}, {0x0001133A, 0x0001133A},
{0x00011345, 0x00011346}, {0x00011349, 0x0001134A}, {0x0001134E, 0x0001134F}, {0x00011351, 0x00011356},
{0x00011358, 0x0001135C}, {0x00011364, 0x00011365}, {0x0001136D, 0x0001136F}, {0x00011375, 0x000113FF},
{0x0001145C, 0x0001145C}, {0x00011462, 0x0001147F}, {0x000114C8, 0x000114CF}, {0x000114DA, 0x0001157F},
{0x000115B6, 0x000115B7}, {0x000115DE, 0x000115FF}, {0x00011645, 0x0001164F}, {0x0001165A, 0x0001165F},
{0x0001166D, 0x0001167F}, {0x000116B9, 0x000116BF}, {0x000116CA, 0x000116FF}, {0x0001171B, 0x0001171C},
{0x0001172C, 0x0001172F}, {0x00011740, 0x000117FF}, {0x0001183C, 0x0001189F}, {0x000118F3, 0x000118FE},
{0x00011907, 0x00011908}, {0x0001190A, 0x0001190B}, {0x00011914, 0x00011914}, {0x00011917, 0x00011917},
{0x00011936, 0x00011936}, {0x00011939, 0x0001193A}, {0x00011947, 0x0001194F}, {0x0001195A, 0x0001199F},
{0x000119A8, 0x000119A9}, {0x000119D8, 0x000119D9}, {0x000119E5, 0x000119FF}, {0x00011A48, 0x00011A4F},
{0x00011AA3, 0x00011ABF}, {0x00011AF9, 0x00011BFF}, {0x00011C09, 0x00011C09}, {0x00011C37, 0x00011C37},
{0x0001057B, 0x0001057B}, {0x0001058B, 0x0001058B}, {0x00010593, 0x00010593}, {0x00010596, 0x00010596},
{0x000105A2, 0x000105A2}, {0x000105B2, 0x000105B2}, {0x000105BA, 0x000105BA}, {0x000105BD, 0x000105FF},
{0x00010737, 0x0001073F}, {0x00010756, 0x0001075F}, {0x00010768, 0x0001077F}, {0x00010786, 0x00010786},
{0x000107B1, 0x000107B1}, {0x000107BB, 0x000107FF}, {0x00010806, 0x00010807}, {0x00010809, 0x00010809},
{0x00010836, 0x00010836}, {0x00010839, 0x0001083B}, {0x0001083D, 0x0001083E}, {0x00010856, 0x00010856},
{0x0001089F, 0x000108A6}, {0x000108B0, 0x000108DF}, {0x000108F3, 0x000108F3}, {0x000108F6, 0x000108FA},
{0x0001091C, 0x0001091E}, {0x0001093A, 0x0001093E}, {0x00010940, 0x0001097F}, {0x000109B8, 0x000109BB},
{0x000109D0, 0x000109D1}, {0x00010A04, 0x00010A04}, {0x00010A07, 0x00010A0B}, {0x00010A14, 0x00010A14},
{0x00010A18, 0x00010A18}, {0x00010A36, 0x00010A37}, {0x00010A3B, 0x00010A3E}, {0x00010A49, 0x00010A4F},
{0x00010A59, 0x00010A5F}, {0x00010AA0, 0x00010ABF}, {0x00010AE7, 0x00010AEA}, {0x00010AF7, 0x00010AFF},
{0x00010B36, 0x00010B38}, {0x00010B56, 0x00010B57}, {0x00010B73, 0x00010B77}, {0x00010B92, 0x00010B98},
{0x00010B9D, 0x00010BA8}, {0x00010BB0, 0x00010BFF}, {0x00010C49, 0x00010C7F}, {0x00010CB3, 0x00010CBF},
{0x00010CF3, 0x00010CF9}, {0x00010D28, 0x00010D2F}, {0x00010D3A, 0x00010E5F}, {0x00010E7F, 0x00010E7F},
{0x00010EAA, 0x00010EAA}, {0x00010EAE, 0x00010EAF}, {0x00010EB2, 0x00010EFC}, {0x00010F28, 0x00010F2F},
{0x00010F5A, 0x00010F6F}, {0x00010F8A, 0x00010FAF}, {0x00010FCC, 0x00010FDF}, {0x00010FF7, 0x00010FFF},
{0x0001104E, 0x00011051}, {0x00011076, 0x0001107E}, {0x000110BD, 0x000110BD}, {0x000110C3, 0x000110CF},
{0x000110E9, 0x000110EF}, {0x000110FA, 0x000110FF}, {0x00011135, 0x00011135}, {0x00011148, 0x0001114F},
{0x00011177, 0x0001117F}, {0x000111E0, 0x000111E0}, {0x000111F5, 0x000111FF}, {0x00011212, 0x00011212},
{0x00011242, 0x0001127F}, {0x00011287, 0x00011287}, {0x00011289, 0x00011289}, {0x0001128E, 0x0001128E},
{0x0001129E, 0x0001129E}, {0x000112AA, 0x000112AF}, {0x000112EB, 0x000112EF}, {0x000112FA, 0x000112FF},
{0x00011304, 0x00011304}, {0x0001130D, 0x0001130E}, {0x00011311, 0x00011312}, {0x00011329, 0x00011329},
{0x00011331, 0x00011331}, {0x00011334, 0x00011334}, {0x0001133A, 0x0001133A}, {0x00011345, 0x00011346},
{0x00011349, 0x0001134A}, {0x0001134E, 0x0001134F}, {0x00011351, 0x00011356}, {0x00011358, 0x0001135C},
{0x00011364, 0x00011365}, {0x0001136D, 0x0001136F}, {0x00011375, 0x000113FF}, {0x0001145C, 0x0001145C},
{0x00011462, 0x0001147F}, {0x000114C8, 0x000114CF}, {0x000114DA, 0x0001157F}, {0x000115B6, 0x000115B7},
{0x000115DE, 0x000115FF}, {0x00011645, 0x0001164F}, {0x0001165A, 0x0001165F}, {0x0001166D, 0x0001167F},
{0x000116BA, 0x000116BF}, {0x000116CA, 0x000116FF}, {0x0001171B, 0x0001171C}, {0x0001172C, 0x0001172F},
{0x00011747, 0x000117FF}, {0x0001183C, 0x0001189F}, {0x000118F3, 0x000118FE}, {0x00011907, 0x00011908},
{0x0001190A, 0x0001190B}, {0x00011914, 0x00011914}, {0x00011917, 0x00011917}, {0x00011936, 0x00011936},
{0x00011939, 0x0001193A}, {0x00011947, 0x0001194F}, {0x0001195A, 0x0001199F}, {0x000119A8, 0x000119A9},
{0x000119D8, 0x000119D9}, {0x000119E5, 0x000119FF}, {0x00011A48, 0x00011A4F}, {0x00011AA3, 0x00011AAF},
{0x00011AF9, 0x00011AFF}, {0x00011B0A, 0x00011BFF}, {0x00011C09, 0x00011C09}, {0x00011C37, 0x00011C37},
{0x00011C46, 0x00011C4F}, {0x00011C6D, 0x00011C6F}, {0x00011C90, 0x00011C91}, {0x00011CA8, 0x00011CA8},
{0x00011CB7, 0x00011CFF}, {0x00011D07, 0x00011D07}, {0x00011D0A, 0x00011D0A}, {0x00011D37, 0x00011D39},
{0x00011D3B, 0x00011D3B}, {0x00011D3E, 0x00011D3E}, {0x00011D48, 0x00011D4F}, {0x00011D5A, 0x00011D5F},
{0x00011D66, 0x00011D66}, {0x00011D69, 0x00011D69}, {0x00011D8F, 0x00011D8F}, {0x00011D92, 0x00011D92},
{0x00011D99, 0x00011D9F}, {0x00011DAA, 0x00011EDF}, {0x00011EF9, 0x00011FAF}, {0x00011FB1, 0x00011FBF},
{0x00011FF2, 0x00011FFE}, {0x0001239A, 0x000123FF}, {0x0001246F, 0x0001246F}, {0x00012475, 0x0001247F},
{0x00012544, 0x00012FFF}, {0x0001342F, 0x000143FF}, {0x00014647, 0x000167FF}, {0x00016A39, 0x00016A3F},
{0x00016A5F, 0x00016A5F}, {0x00016A6A, 0x00016A6D}, {0x00016A70, 0x00016ACF}, {0x00016AEE, 0x00016AEF},
{0x00016AF6, 0x00016AFF}, {0x00016B46, 0x00016B4F}, {0x00016B5A, 0x00016B5A}, {0x00016B62, 0x00016B62},
{0x00016B78, 0x00016B7C}, {0x00016B90, 0x00016E3F}, {0x00016E9B, 0x00016EFF}, {0x00016F4B, 0x00016F4E},
{0x00016F88, 0x00016F8E}, {0x00016FA0, 0x00016FDF}, {0x00016FE5, 0x00016FEF}, {0x00016FF2, 0x00016FFF},
{0x000187F8, 0x000187FF}, {0x00018CD6, 0x00018CFF}, {0x00018D09, 0x0001AFFF}, {0x0001B11F, 0x0001B14F},
{0x0001B153, 0x0001B163}, {0x0001B168, 0x0001B16F}, {0x0001B2FC, 0x0001BBFF}, {0x0001BC6B, 0x0001BC6F},
{0x0001BC7D, 0x0001BC7F}, {0x0001BC89, 0x0001BC8F}, {0x0001BC9A, 0x0001BC9B}, {0x0001BCA0, 0x0001CFFF},
{0x0001D0F6, 0x0001D0FF}, {0x0001D127, 0x0001D128}, {0x0001D173, 0x0001D17A}, {0x0001D1E9, 0x0001D1FF},
{0x0001D246, 0x0001D2DF}, {0x0001D2F4, 0x0001D2FF}, {0x0001D357, 0x0001D35F}, {0x0001D379, 0x0001D3FF},
{0x0001D455, 0x0001D455}, {0x0001D49D, 0x0001D49D}, {0x0001D4A0, 0x0001D4A1}, {0x0001D4A3, 0x0001D4A4},
{0x0001D4A7, 0x0001D4A8}, {0x0001D4AD, 0x0001D4AD}, {0x0001D4BA, 0x0001D4BA}, {0x0001D4BC, 0x0001D4BC},
{0x0001D4C4, 0x0001D4C4}, {0x0001D506, 0x0001D506}, {0x0001D50B, 0x0001D50C}, {0x0001D515, 0x0001D515},
{0x0001D51D, 0x0001D51D}, {0x0001D53A, 0x0001D53A}, {0x0001D53F, 0x0001D53F}, {0x0001D545, 0x0001D545},
{0x0001D547, 0x0001D549}, {0x0001D551, 0x0001D551}, {0x0001D6A6, 0x0001D6A7}, {0x0001D7CC, 0x0001D7CD},
{0x0001DA8C, 0x0001DA9A}, {0x0001DAA0, 0x0001DAA0}, {0x0001DAB0, 0x0001DFFF}, {0x0001E007, 0x0001E007},
{0x0001E019, 0x0001E01A}, {0x0001E022, 0x0001E022}, {0x0001E025, 0x0001E025}, {0x0001E02B, 0x0001E0FF},
{0x0001E12D, 0x0001E12F}, {0x0001E13E, 0x0001E13F}, {0x0001E14A, 0x0001E14D}, {0x0001E150, 0x0001E2BF},
{0x0001E2FA, 0x0001E2FE}, {0x0001E300, 0x0001E7FF}, {0x0001E8C5, 0x0001E8C6}, {0x0001E8D7, 0x0001E8FF},
{0x0001E94C, 0x0001E94F}, {0x0001E95A, 0x0001E95D}, {0x0001E960, 0x0001EC70}, {0x0001ECB5, 0x0001ED00},
{0x0001ED3E, 0x0001EDFF}, {0x0001EE04, 0x0001EE04}, {0x0001EE20, 0x0001EE20}, {0x0001EE23, 0x0001EE23},
{0x0001EE25, 0x0001EE26}, {0x0001EE28, 0x0001EE28}, {0x0001EE33, 0x0001EE33}, {0x0001EE38, 0x0001EE38},
{0x0001EE3A, 0x0001EE3A}, {0x0001EE3C, 0x0001EE41}, {0x0001EE43, 0x0001EE46}, {0x0001EE48, 0x0001EE48},
{0x0001EE4A, 0x0001EE4A}, {0x0001EE4C, 0x0001EE4C}, {0x0001EE50, 0x0001EE50}, {0x0001EE53, 0x0001EE53},
{0x0001EE55, 0x0001EE56}, {0x0001EE58, 0x0001EE58}, {0x0001EE5A, 0x0001EE5A}, {0x0001EE5C, 0x0001EE5C},
{0x0001EE5E, 0x0001EE5E}, {0x0001EE60, 0x0001EE60}, {0x0001EE63, 0x0001EE63}, {0x0001EE65, 0x0001EE66},
{0x0001EE6B, 0x0001EE6B}, {0x0001EE73, 0x0001EE73}, {0x0001EE78, 0x0001EE78}, {0x0001EE7D, 0x0001EE7D},
{0x0001EE7F, 0x0001EE7F}, {0x0001EE8A, 0x0001EE8A}, {0x0001EE9C, 0x0001EEA0}, {0x0001EEA4, 0x0001EEA4},
{0x0001EEAA, 0x0001EEAA}, {0x0001EEBC, 0x0001EEEF}, {0x0001EEF2, 0x0001EFFF}, {0x0001F02C, 0x0001F02F},
{0x0001F094, 0x0001F09F}, {0x0001F0AF, 0x0001F0B0}, {0x0001F0C0, 0x0001F0C0}, {0x0001F0D0, 0x0001F0D0},
{0x0001F0F6, 0x0001F0FF}, {0x0001F1AE, 0x0001F1E5}, {0x0001F203, 0x0001F20F}, {0x0001F23C, 0x0001F23F},
{0x0001F249, 0x0001F24F}, {0x0001F252, 0x0001F25F}, {0x0001F266, 0x0001F2FF}, {0x0001F6D8, 0x0001F6DF},
{0x0001F6ED, 0x0001F6EF}, {0x0001F6FD, 0x0001F6FF}, {0x0001F774, 0x0001F77F}, {0x0001F7D9, 0x0001F7DF},
{0x0001F7EC, 0x0001F7FF}, {0x0001F80C, 0x0001F80F}, {0x0001F848, 0x0001F84F}, {0x0001F85A, 0x0001F85F},
{0x0001F888, 0x0001F88F}, {0x0001F8AE, 0x0001F8AF}, {0x0001F8B2, 0x0001F8FF}, {0x0001F979, 0x0001F979},
{0x0001F9CC, 0x0001F9CC}, {0x0001FA54, 0x0001FA5F}, {0x0001FA6E, 0x0001FA6F}, {0x0001FA75, 0x0001FA77},
{0x0001FA7B, 0x0001FA7F}, {0x0001FA87, 0x0001FA8F}, {0x0001FAA9, 0x0001FAAF}, {0x0001FAB7, 0x0001FABF},
{0x0001FAC3, 0x0001FACF}, {0x0001FAD7, 0x0001FAFF}, {0x0001FB93, 0x0001FB93}, {0x0001FBCB, 0x0001FBEF},
{0x0001FBFA, 0x0001FFFF}, {0x0002A6DE, 0x0002A6FF}, {0x0002B735, 0x0002B73F}, {0x0002B81E, 0x0002B81F},
{0x0002CEA2, 0x0002CEAF}, {0x0002EBE1, 0x0002F7FF}, {0x0002FA1E, 0x0002FFFF}, {0x0003134B, 0x000E00FF},
{0x000E01F0, 0x0010FFFF},
{0x00011D99, 0x00011D9F}, {0x00011DAA, 0x00011EDF}, {0x00011EF9, 0x00011EFF}, {0x00011F11, 0x00011F11},
{0x00011F3B, 0x00011F3D}, {0x00011F5A, 0x00011FAF}, {0x00011FB1, 0x00011FBF}, {0x00011FF2, 0x00011FFE},
{0x0001239A, 0x000123FF}, {0x0001246F, 0x0001246F}, {0x00012475, 0x0001247F}, {0x00012544, 0x00012F8F},
{0x00012FF3, 0x00012FFF}, {0x00013430, 0x0001343F}, {0x00013456, 0x000143FF}, {0x00014647, 0x000167FF},
{0x00016A39, 0x00016A3F}, {0x00016A5F, 0x00016A5F}, {0x00016A6A, 0x00016A6D}, {0x00016ABF, 0x00016ABF},
{0x00016ACA, 0x00016ACF}, {0x00016AEE, 0x00016AEF}, {0x00016AF6, 0x00016AFF}, {0x00016B46, 0x00016B4F},
{0x00016B5A, 0x00016B5A}, {0x00016B62, 0x00016B62}, {0x00016B78, 0x00016B7C}, {0x00016B90, 0x00016E3F},
{0x00016E9B, 0x00016EFF}, {0x00016F4B, 0x00016F4E}, {0x00016F88, 0x00016F8E}, {0x00016FA0, 0x00016FDF},
{0x00016FE5, 0x00016FEF}, {0x00016FF2, 0x00016FFF}, {0x000187F8, 0x000187FF}, {0x00018CD6, 0x00018CFF},
{0x00018D09, 0x0001AFEF}, {0x0001AFF4, 0x0001AFF4}, {0x0001AFFC, 0x0001AFFC}, {0x0001AFFF, 0x0001AFFF},
{0x0001B123, 0x0001B131}, {0x0001B133, 0x0001B14F}, {0x0001B153, 0x0001B154}, {0x0001B156, 0x0001B163},
{0x0001B168, 0x0001B16F}, {0x0001B2FC, 0x0001BBFF}, {0x0001BC6B, 0x0001BC6F}, {0x0001BC7D, 0x0001BC7F},
{0x0001BC89, 0x0001BC8F}, {0x0001BC9A, 0x0001BC9B}, {0x0001BCA0, 0x0001CEFF}, {0x0001CF2E, 0x0001CF2F},
{0x0001CF47, 0x0001CF4F}, {0x0001CFC4, 0x0001CFFF}, {0x0001D0F6, 0x0001D0FF}, {0x0001D127, 0x0001D128},
{0x0001D173, 0x0001D17A}, {0x0001D1EB, 0x0001D1FF}, {0x0001D246, 0x0001D2BF}, {0x0001D2D4, 0x0001D2DF},
{0x0001D2F4, 0x0001D2FF}, {0x0001D357, 0x0001D35F}, {0x0001D379, 0x0001D3FF}, {0x0001D455, 0x0001D455},
{0x0001D49D, 0x0001D49D}, {0x0001D4A0, 0x0001D4A1}, {0x0001D4A3, 0x0001D4A4}, {0x0001D4A7, 0x0001D4A8},
{0x0001D4AD, 0x0001D4AD}, {0x0001D4BA, 0x0001D4BA}, {0x0001D4BC, 0x0001D4BC}, {0x0001D4C4, 0x0001D4C4},
{0x0001D506, 0x0001D506}, {0x0001D50B, 0x0001D50C}, {0x0001D515, 0x0001D515}, {0x0001D51D, 0x0001D51D},
{0x0001D53A, 0x0001D53A}, {0x0001D53F, 0x0001D53F}, {0x0001D545, 0x0001D545}, {0x0001D547, 0x0001D549},
{0x0001D551, 0x0001D551}, {0x0001D6A6, 0x0001D6A7}, {0x0001D7CC, 0x0001D7CD}, {0x0001DA8C, 0x0001DA9A},
{0x0001DAA0, 0x0001DAA0}, {0x0001DAB0, 0x0001DEFF}, {0x0001DF1F, 0x0001DF24}, {0x0001DF2B, 0x0001DFFF},
{0x0001E007, 0x0001E007}, {0x0001E019, 0x0001E01A}, {0x0001E022, 0x0001E022}, {0x0001E025, 0x0001E025},
{0x0001E02B, 0x0001E02F}, {0x0001E06E, 0x0001E08E}, {0x0001E090, 0x0001E0FF}, {0x0001E12D, 0x0001E12F},
{0x0001E13E, 0x0001E13F}, {0x0001E14A, 0x0001E14D}, {0x0001E150, 0x0001E28F}, {0x0001E2AF, 0x0001E2BF},
{0x0001E2FA, 0x0001E2FE}, {0x0001E300, 0x0001E4CF}, {0x0001E4FA, 0x0001E7DF}, {0x0001E7E7, 0x0001E7E7},
{0x0001E7EC, 0x0001E7EC}, {0x0001E7EF, 0x0001E7EF}, {0x0001E7FF, 0x0001E7FF}, {0x0001E8C5, 0x0001E8C6},
{0x0001E8D7, 0x0001E8FF}, {0x0001E94C, 0x0001E94F}, {0x0001E95A, 0x0001E95D}, {0x0001E960, 0x0001EC70},
{0x0001ECB5, 0x0001ED00}, {0x0001ED3E, 0x0001EDFF}, {0x0001EE04, 0x0001EE04}, {0x0001EE20, 0x0001EE20},
{0x0001EE23, 0x0001EE23}, {0x0001EE25, 0x0001EE26}, {0x0001EE28, 0x0001EE28}, {0x0001EE33, 0x0001EE33},
{0x0001EE38, 0x0001EE38}, {0x0001EE3A, 0x0001EE3A}, {0x0001EE3C, 0x0001EE41}, {0x0001EE43, 0x0001EE46},
{0x0001EE48, 0x0001EE48}, {0x0001EE4A, 0x0001EE4A}, {0x0001EE4C, 0x0001EE4C}, {0x0001EE50, 0x0001EE50},
{0x0001EE53, 0x0001EE53}, {0x0001EE55, 0x0001EE56}, {0x0001EE58, 0x0001EE58}, {0x0001EE5A, 0x0001EE5A},
{0x0001EE5C, 0x0001EE5C}, {0x0001EE5E, 0x0001EE5E}, {0x0001EE60, 0x0001EE60}, {0x0001EE63, 0x0001EE63},
{0x0001EE65, 0x0001EE66}, {0x0001EE6B, 0x0001EE6B}, {0x0001EE73, 0x0001EE73}, {0x0001EE78, 0x0001EE78},
{0x0001EE7D, 0x0001EE7D}, {0x0001EE7F, 0x0001EE7F}, {0x0001EE8A, 0x0001EE8A}, {0x0001EE9C, 0x0001EEA0},
{0x0001EEA4, 0x0001EEA4}, {0x0001EEAA, 0x0001EEAA}, {0x0001EEBC, 0x0001EEEF}, {0x0001EEF2, 0x0001EFFF},
{0x0001F02C, 0x0001F02F}, {0x0001F094, 0x0001F09F}, {0x0001F0AF, 0x0001F0B0}, {0x0001F0C0, 0x0001F0C0},
{0x0001F0D0, 0x0001F0D0}, {0x0001F0F6, 0x0001F0FF}, {0x0001F1AE, 0x0001F1E5}, {0x0001F203, 0x0001F20F},
{0x0001F23C, 0x0001F23F}, {0x0001F249, 0x0001F24F}, {0x0001F252, 0x0001F25F}, {0x0001F266, 0x0001F2FF},
{0x0001F6D8, 0x0001F6DB}, {0x0001F6ED, 0x0001F6EF}, {0x0001F6FD, 0x0001F6FF}, {0x0001F777, 0x0001F77A},
{0x0001F7DA, 0x0001F7DF}, {0x0001F7EC, 0x0001F7EF}, {0x0001F7F1, 0x0001F7FF}, {0x0001F80C, 0x0001F80F},
{0x0001F848, 0x0001F84F}, {0x0001F85A, 0x0001F85F}, {0x0001F888, 0x0001F88F}, {0x0001F8AE, 0x0001F8AF},
{0x0001F8B2, 0x0001F8FF}, {0x0001FA54, 0x0001FA5F}, {0x0001FA6E, 0x0001FA6F}, {0x0001FA7D, 0x0001FA7F},
{0x0001FA89, 0x0001FA8F}, {0x0001FABE, 0x0001FABE}, {0x0001FAC6, 0x0001FACD}, {0x0001FADC, 0x0001FADF},
{0x0001FAE9, 0x0001FAEF}, {0x0001FAF9, 0x0001FAFF}, {0x0001FB93, 0x0001FB93}, {0x0001FBCB, 0x0001FBEF},
{0x0001FBFA, 0x0001FFFF}, {0x0002A6E0, 0x0002A6FF}, {0x0002B73A, 0x0002B73F}, {0x0002B81E, 0x0002B81F},
{0x0002CEA2, 0x0002CEAF}, {0x0002EBE1, 0x0002F7FF}, {0x0002FA1E, 0x0002FFFF}, {0x0003134B, 0x0003134F},
{0x000323B0, 0x000E00FF}, {0x000E01F0, 0x0010FFFF},
};
const std::multimap<uint32_t, uint32_t> unicode_map_nfd = {

2
unicode-data.h
View file

@ -5,7 +5,7 @@
#include <utility>
#include <vector>
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_digit;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_number;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace;
extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark;

22
unicode.cpp
View file

@ -110,9 +110,9 @@ static uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset)
static std::unordered_map<uint32_t, int> unicode_cpt_type_map() {
std::unordered_map<uint32_t, int> cpt_types;
for (auto p : unicode_ranges_digit) {
for (auto p : unicode_ranges_number) {
for (auto i = p.first; i <= p.second; ++ i) {
cpt_types[i] = CODEPOINT_TYPE_DIGIT;
cpt_types[i] = CODEPOINT_TYPE_NUMBER;
}
}
for (auto p : unicode_ranges_letter) {
@ -300,13 +300,13 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
collecting_letter = true;
collecting = true;
}
else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_DIGIT || (token.empty() && utf_char == " " && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_NUMBER || (token.empty() && utf_char == " " && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_NUMBER)) {
collecting_numeric = true;
collecting = true;
}
else if (
((unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_DIGIT) && (unicode_cpt_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) ||
(token.empty() && utf_char == " " && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_DIGIT && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE)
((unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_NUMBER) && (unicode_cpt_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) ||
(token.empty() && utf_char == " " && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_NUMBER && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE)
) {
collecting_special = true;
collecting = true;
@ -323,13 +323,13 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
if (collecting_letter && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER) {
split_condition = true;
}
else if (collecting_numeric && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_DIGIT) {
else if (collecting_numeric && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_NUMBER) {
split_condition = true;
}
else if (collecting_special && (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_DIGIT || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) {
else if (collecting_special && (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_NUMBER || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) {
split_condition = true;
}
else if (collecting_whitespace_lookahead && (unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_DIGIT)) {
else if (collecting_whitespace_lookahead && (unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_NUMBER)) {
split_condition = true;
}
}
@ -524,19 +524,19 @@ char32_t unicode_tolower(char32_t cp) {
std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
// unicode categories
static const std::map<std::string, int> k_ucat_enum = {
{ "\\p{N}", CODEPOINT_TYPE_DIGIT },
{ "\\p{N}", CODEPOINT_TYPE_NUMBER },
{ "\\p{L}", CODEPOINT_TYPE_LETTER },
{ "\\p{P}", CODEPOINT_TYPE_PUNCTUATION },
};
static const std::map<int, int> k_ucat_cpt = {
{ CODEPOINT_TYPE_DIGIT, 0xD1 },
{ CODEPOINT_TYPE_NUMBER, 0xD1 },
{ CODEPOINT_TYPE_LETTER, 0xD2 },
{ CODEPOINT_TYPE_PUNCTUATION, 0xD3 },
};
static const std::map<int, std::string> k_ucat_map = {
{ CODEPOINT_TYPE_DIGIT, "\x30-\x39" }, // 0-9
{ CODEPOINT_TYPE_NUMBER, "\x30-\x39" }, // 0-9
{ CODEPOINT_TYPE_LETTER, "\x41-\x5A\x61-\x7A" }, // A-Za-z
{ CODEPOINT_TYPE_PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
};

2
unicode.h
View file

@ -5,7 +5,7 @@
#include <vector>
#define CODEPOINT_TYPE_UNIDENTIFIED 0
#define CODEPOINT_TYPE_DIGIT 1
#define CODEPOINT_TYPE_NUMBER 1
#define CODEPOINT_TYPE_LETTER 2
#define CODEPOINT_TYPE_WHITESPACE 3
#define CODEPOINT_TYPE_ACCENT_MARK 4