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

refactor: Add experimental mapping for BPE pre-tokenizers

Browse source
This commit is contained in:
teleprint-me 2024-05-21 22:45:45 -04:00
parent 34e14ae96d
commit 0b43e14030
No known key found for this signature in database
GPG key ID: B0D11345E65C4D48
1 changed files with 129 additions and 46 deletions
Download patch file Download diff file Expand all files Collapse all files

175
gguf-py/gguf/constants.py
View file

@ -959,13 +959,83 @@ class HFModelFileType(IntEnum):
HF_TOKENIZER_BPE_FILES = ("config.json", "tokenizer_config.json", "tokenizer.json",) HF_TOKENIZER_BPE_FILES = ("config.json", "tokenizer_config.json", "tokenizer.json",)
HF_TOKENIZER_SPM_FILES = (HF_TOKENIZER_BPE_FILES + ("tokenizer.model",)) HF_TOKENIZER_SPM_FILES = (HF_TOKENIZER_BPE_FILES + ("tokenizer.model",))
#
# BPE Pre-tokenization Regular Expressions
#
# NOTE: Tokenizers defaults to OpenAI GPT-2 Byte Level Reg-Exp # NOTE: Tokenizers defaults to OpenAI GPT-2 Byte Level Reg-Exp
# The pattern uses perl, is grammatical, and splits are technically arbitrary. # The pattern uses perl, is grammatical, and splits are technically arbitrary.
# https://github.com/openai/gpt-2/blob/master/src/encoder.py#L53 # https://github.com/openai/gpt-2/blob/master/src/encoder.py#L53
# https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/pre_tokenizers/byte_level.rs#L40-L42 # https://github.com/huggingface/tokenizers/blob/main/tokenizers/src/pre_tokenizers/byte_level.rs#L40-L42
UNI_PRE_TOKENIZER_DEFAULT = "\\p{N}"
BPE_PRE_TOKENIZER_DEFAULT = "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)" BPE_PRE_TOKENIZER_DEFAULT = "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)"
GPT_PRE_TOKENIZER_DEFAULT = f"{BPE_PRE_TOKENIZER_DEFAULT}|\\s+" GPT_PRE_TOKENIZER_DEFAULT = f"{BPE_PRE_TOKENIZER_DEFAULT}|\\s+"
# Default pre-processing regex for BPE tokenization
BPE_PRE_PROCESSOR_DEFAULT = (
"[\\p{P}\\$\\+<=>\\^~\\|]+", BPE_PRE_TOKENIZER_DEFAULT, "\\p{N}+", "[0-9][0-9][0-9]",
)
BPE_PRE_TOKENIZERS = {
"default": GPT_PRE_TOKENIZER_DEFAULT,
# NOTE: LLAMA and DBRX (MODEL_ARCH_NAMES[MODEL_ARCH.DBRX]) are the same
MODEL_ARCH_NAMES[MODEL_ARCH.LLAMA]: (
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
),
MODEL_ARCH_NAMES[MODEL_ARCH.FALCON]: (
"[\\p{P}\\$\\+<=>\\^~\\|]+", BPE_PRE_TOKENIZER_DEFAULT, "[0-9][0-9][0-9]",
),
MODEL_ARCH_NAMES[MODEL_ARCH.BAICHUAN]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.GROK]: (),
# NOTE: GPT-2 is the standard default pre-tokenizer for all models
MODEL_ARCH_NAMES[MODEL_ARCH.GPT2]: (
GPT_PRE_TOKENIZER_DEFAULT,
),
MODEL_ARCH_NAMES[MODEL_ARCH.GPTJ]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.GPTNEOX]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.MPT]: (
"\\s?\\p{L}+", "\\s?\\p{P}+", BPE_PRE_TOKENIZER_DEFAULT,
),
MODEL_ARCH_NAMES[MODEL_ARCH.STARCODER]: (
"\\p{N}", BPE_PRE_TOKENIZER_DEFAULT,
),
MODEL_ARCH_NAMES[MODEL_ARCH.REFACT]: (
"\\p{N}", BPE_PRE_TOKENIZER_DEFAULT,
),
MODEL_ARCH_NAMES[MODEL_ARCH.BERT]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.NOMIC_BERT]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.JINA_BERT_V2]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.BLOOM]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.STABLELM]: (
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
),
MODEL_ARCH_NAMES[MODEL_ARCH.QWEN]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.QWEN2]: (
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
),
MODEL_ARCH_NAMES[MODEL_ARCH.QWEN2MOE]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.PHI2]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.PHI3]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.PLAMO]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.CODESHELL]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.ORION]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.INTERNLM2]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.MINICPM]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.GEMMA]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.STARCODER2]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.MAMBA]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.XVERSE]: (),
MODEL_ARCH_NAMES[MODEL_ARCH.COMMAND_R]: (
"\\p{N}", BPE_PRE_TOKENIZER_DEFAULT,
),
MODEL_ARCH_NAMES[MODEL_ARCH.DBRX]: (
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
),
MODEL_ARCH_NAMES[MODEL_ARCH.OLMO]: (
GPT_PRE_TOKENIZER_DEFAULT,
),
}
# #
# HuggingFace Model Map # HuggingFace Model Map
# #
@ -979,22 +1049,78 @@ GPT_PRE_TOKENIZER_DEFAULT = f"{BPE_PRE_TOKENIZER_DEFAULT}|\\s+"
# - `model_parts` (int): Number of parts required to join the model during conversion # - `model_parts` (int): Number of parts required to join the model during conversion
# - `model_type` (FileFormatType): File format for the Hugging Face model files # - `model_type` (FileFormatType): File format for the Hugging Face model files
# - `vocab_type` (VocabType): Vocabulary type used by the tokenizer # - `vocab_type` (VocabType): Vocabulary type used by the tokenizer
# - `vocab_pre` (Optional[Tuple[str]]): List of pre-tokenizer pattern strings for this model # - `vocab_pre` (Optional[Tuple[str]]): Tuple of pre-tokenizer pattern strings for this model
# - `vocab_files` (Tuple[str]): List of file names required to extract vocabulary and other metadata # - `vocab_files` (Tuple[str]): Tuple of file names required to extract vocabulary and other metadata
# #
# NOTES # NOTES
# - Possible algorithms are WordLevel, BPE, WordPiece, or Unigram # - Possible algorithms are WordLevel, BPE, WordPiece, or Unigram
# - Possible LLaMa tokenizer model types are: None, SPM, BPE, or WPM # - Possible LLaMa tokenizer model types are: None, SPM, BPE, or WPM
HF_MODEL_MAP = ( HF_MODEL_MAP = (
# Sentence Piece Models
{ {
"model_repo": "meta-llama/Llama-2-7b-hf", "model_repo": "meta-llama/Llama-2-7b-hf",
"model_arch": MODEL_ARCH.LLAMA, "model_arch": MODEL_ARCH_NAMES[MODEL_ARCH.LLAMA],
"model_parts": 2, "model_parts": 2,
"model_type": HFModelFileType.SFT, "model_type": HFModelFileType.SFT,
"vocab_type": LLaMaVocabType.SPM, "vocab_type": LLaMaVocabType.SPM,
"vocab_pre": (), "vocab_pre": (),
"vocab_files": HF_TOKENIZER_SPM_FILES, "vocab_files": HF_TOKENIZER_SPM_FILES,
}, },
{
"model_repo": "mistralai/Mistral-7B-Instruct-v0.2",
"model_arch": MODEL_ARCH_NAMES[MODEL_ARCH.LLAMA],
"model_parts": 3,
"model_type": HFModelFileType.SFT,
"vocab_type": LLaMaVocabType.SPM,
"vocab_pre": (),
"vocab_files": HF_TOKENIZER_SPM_FILES,
},
{
"model_repo": "mistralai/Mixtral-8x7B-Instruct-v0.1",
"model_arch": MODEL_ARCH_NAMES[MODEL_ARCH.LLAMA],
"model_parts": 8,
"model_type": HFModelFileType.SFT,
"vocab_type": LLaMaVocabType.SPM,
"vocab_pre": (),
"vocab_files": HF_TOKENIZER_SPM_FILES,
},
{
"model_repo": "microsoft/Phi-3-mini-4k-instruct",
"model_arch": MODEL_ARCH_NAMES[MODEL_ARCH.PHI3],
"model_parts": 2,
"model_type": HFModelFileType.SFT,
"vocab_type": LLaMaVocabType.SPM,
"vocab_pre": (),
"vocab_files": HF_TOKENIZER_SPM_FILES,
},
# Word Piece Models
{
"model_repo": "BAAI/bge-small-en-v1.5",
"model_arch": MODEL_ARCH_NAMES[MODEL_ARCH.BERT],
"model_parts": 1,
"model_type": HFModelFileType.BIN,
"vocab_type": LLaMaVocabType.WPM,
"vocab_pre": (),
"vocab_files": HF_TOKENIZER_BPE_FILES,
},
{
"model_repo": "jinaai/jina-embeddings-v2-base-en",
"model_arch": MODEL_ARCH.JINA_BERT_V2,
"vocab_type": LLaMaVocabType.WPM,
},
# Byte Pair Encoding Models
{
"model_repo": "meta-llama/Meta-Llama-3-8B",
"model_arch": MODEL_ARCH.LLAMA,
"model_parts": 4,
"model_type": HFModelFileType.SFT,
"vocab_type": LLaMaVocabType.BPE,
# PR#6920: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989
"vocab_pre": (
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
),
"vocab_files": HF_TOKENIZER_BPE_FILES,
},
{ {
"model_repo": "tiiuae/falcon-7b", "model_repo": "tiiuae/falcon-7b",
"model_arch": MODEL_ARCH.FALCON, "model_arch": MODEL_ARCH.FALCON,
@ -1004,25 +1130,6 @@ HF_MODEL_MAP = (
"vocab_pre": BPE_PRE_TOKENIZERS[MODEL_ARCH_NAMES[MODEL_ARCH.FALCON]], "vocab_pre": BPE_PRE_TOKENIZERS[MODEL_ARCH_NAMES[MODEL_ARCH.FALCON]],
"vocab_files": HF_TOKENIZER_BPE_FILES, "vocab_files": HF_TOKENIZER_BPE_FILES,
}, },
{
"model_repo": "meta-llama/Meta-Llama-3-8B",
"model_arch": MODEL_ARCH.LLAMA,
"model_parts": 4,
"model_type": HFModelFileType.SFT,
"vocab_type": LLaMaVocabType.BPE,
# PR#6920: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989
"vocab_pre": BPE_PRE_TOKENIZERS[MODEL_ARCH_NAMES[MODEL_ARCH.LLAMA]],
"vocab_files": HF_TOKENIZER_BPE_FILES,
},
{
"model_repo": "microsoft/Phi-3-mini-4k-instruct",
"model_arch": MODEL_ARCH.PHI3,
"model_parts": 2,
"model_type": HFModelFileType.SFT,
"vocab_type": LLaMaVocabType.SPM,
"vocab_pre": (),
"vocab_files": HF_TOKENIZER_SPM_FILES,
},
{ {
"model_repo": "deepseek-ai/deepseek-llm-7b-base", "model_repo": "deepseek-ai/deepseek-llm-7b-base",
"model_arch": MODEL_ARCH.LLAMA, "model_arch": MODEL_ARCH.LLAMA,
@ -1054,15 +1161,6 @@ HF_MODEL_MAP = (
), ),
"vocab_files": HF_TOKENIZER_BPE_FILES, "vocab_files": HF_TOKENIZER_BPE_FILES,
}, },
{
"model_repo": "BAAI/bge-small-en-v1.5",
"model_arch": MODEL_ARCH.BERT,
"model_parts": 1,
"model_type": HFModelFileType.BIN,
"vocab_type": LLaMaVocabType.WPM,
"vocab_pre": (),
"vocab_files": HF_TOKENIZER_BPE_FILES,
},
{ {
"model_repo": "mosaicml/mpt-7b", "model_repo": "mosaicml/mpt-7b",
"model_arch": MODEL_ARCH.MPT, "model_arch": MODEL_ARCH.MPT,
@ -1118,11 +1216,6 @@ HF_MODEL_MAP = (
"model_arch": MODEL_ARCH.DBRX, "model_arch": MODEL_ARCH.DBRX,
"vocab_type": LLaMaVocabType.BPE, "vocab_type": LLaMaVocabType.BPE,
}, },
{
"model_repo": "jinaai/jina-embeddings-v2-base-en",
"model_arch": MODEL_ARCH.JINA_BERT_V2,
"vocab_type": LLaMaVocabType.WPM,
},
{ {
"model_repo": "jinaai/jina-embeddings-v2-base-es", "model_repo": "jinaai/jina-embeddings-v2-base-es",
"model_arch": MODEL_ARCH.JINA_BERT_V2, "model_arch": MODEL_ARCH.JINA_BERT_V2,
@ -1143,16 +1236,6 @@ HF_MODEL_MAP = (
"model_arch": MODEL_ARCH.STABLELM, "model_arch": MODEL_ARCH.STABLELM,
"vocab_type": LLaMaVocabType.BPE, "vocab_type": LLaMaVocabType.BPE,
}, },
{
"model_repo": "mistralai/Mistral-7B-Instruct-v0.2",
"model_arch": MODEL_ARCH.LLAMA,
"vocab_type": LLaMaVocabType.SPM,
},
{
"model_repo": "mistralai/Mixtral-8x7B-Instruct-v0.1",
"model_arch": MODEL_ARCH.LLAMA,
"vocab_type": LLaMaVocabType.SPM,
},
) )
# Aliases for backward compatibility. # Aliases for backward compatibility.