[RFC][vllm-API] Support tokenizer registry for customized tokenizer in vLLM (#12518)

Signed-off-by: Keyun Tong <tongkeyun@gmail.com>
This commit is contained in:
Keyun Tong
2025-02-12 12:25:58 +08:00
committed by GitHub
parent 72c2b68dc9
commit 3ee696a63d
11 changed files with 343 additions and 41 deletions
+12 -6
View File
@@ -14,6 +14,8 @@ from transformers import (AutoTokenizer, PreTrainedTokenizer,
from vllm.envs import VLLM_USE_MODELSCOPE
from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
from vllm.transformers_utils.tokenizer_base import (TokenizerBase,
TokenizerRegistry)
from vllm.transformers_utils.tokenizers import MistralTokenizer
from vllm.transformers_utils.utils import check_gguf_file
from vllm.utils import make_async
@@ -21,7 +23,7 @@ from vllm.utils import make_async
logger = init_logger(__name__)
AnyTokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast,
MistralTokenizer]
TokenizerBase]
def decode_tokens(
@@ -47,11 +49,7 @@ def encode_tokens(
Backend-agnostic equivalent of HF's
:code:`tokenizer.encode(text, add_special_tokens=...)`.
"""
if isinstance(tokenizer, MistralTokenizer):
return tokenizer.tokenizer.encode(text,
bos=add_special_tokens,
eos=add_special_tokens)
elif add_special_tokens is not None:
if add_special_tokens is not None:
return tokenizer.encode(text, add_special_tokens=add_special_tokens)
return tokenizer.encode(text)
@@ -183,9 +181,17 @@ def get_tokenizer(
'encoding and decoding.',
FutureWarning,
stacklevel=2)
tokenizer: AnyTokenizer
if tokenizer_mode == "mistral":
tokenizer = MistralTokenizer.from_pretrained(str(tokenizer_name),
revision=revision)
elif tokenizer_mode == "custom":
tokenizer = TokenizerRegistry.get_tokenizer(str(tokenizer_name),
*args,
revision=revision,
download_dir=download_dir,
**kwargs)
else:
try:
tokenizer = AutoTokenizer.from_pretrained(
+146
View File
@@ -0,0 +1,146 @@
# SPDX-License-Identifier: Apache-2.0
import importlib
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
if TYPE_CHECKING:
from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
class TokenizerBase(ABC):
@property
@abstractmethod
def all_special_tokens_extended(self) -> List[str]:
raise NotImplementedError()
@property
@abstractmethod
def all_special_tokens(self) -> List[str]:
raise NotImplementedError()
@property
@abstractmethod
def all_special_ids(self) -> List[int]:
raise NotImplementedError()
@property
@abstractmethod
def bos_token_id(self) -> int:
raise NotImplementedError()
@property
@abstractmethod
def eos_token_id(self) -> int:
raise NotImplementedError()
@property
@abstractmethod
def sep_token(self) -> str:
raise NotImplementedError()
@property
@abstractmethod
def pad_token(self) -> str:
raise NotImplementedError()
@property
@abstractmethod
def is_fast(self) -> bool:
raise NotImplementedError()
@property
@abstractmethod
def vocab_size(self) -> int:
raise NotImplementedError()
@property
@abstractmethod
def max_token_id(self) -> int:
raise NotImplementedError()
def __len__(self) -> int:
return self.vocab_size
@abstractmethod
def __call__(
self,
text: Union[str, List[str], List[int]],
text_pair: Optional[str] = None,
add_special_tokens: bool = False,
truncation: bool = False,
max_length: Optional[int] = None,
):
raise NotImplementedError()
@abstractmethod
def get_vocab(self) -> Dict[str, int]:
raise NotImplementedError()
@abstractmethod
def get_added_vocab(self) -> Dict[str, int]:
raise NotImplementedError()
@abstractmethod
def encode_one(
self,
text: str,
truncation: bool = False,
max_length: Optional[int] = None,
) -> List[int]:
raise NotImplementedError()
@abstractmethod
def encode(self,
text: str,
add_special_tokens: Optional[bool] = None) -> List[int]:
raise NotImplementedError()
@abstractmethod
def apply_chat_template(self,
messages: List["ChatCompletionMessageParam"],
tools: Optional[List[Dict[str, Any]]] = None,
**kwargs) -> List[int]:
raise NotImplementedError()
@abstractmethod
def convert_tokens_to_string(self, tokens: List[str]) -> str:
raise NotImplementedError()
@abstractmethod
def decode(self,
ids: Union[List[int], int],
skip_special_tokens: bool = True) -> str:
raise NotImplementedError()
@abstractmethod
def convert_ids_to_tokens(
self,
ids: List[int],
skip_special_tokens: bool = True,
) -> List[str]:
raise NotImplementedError()
class TokenizerRegistry:
# Tokenizer name -> (tokenizer module, tokenizer class)
REGISTRY: Dict[str, Tuple[str, str]] = {}
@staticmethod
def register(name: str, module: str, class_name: str) -> None:
TokenizerRegistry.REGISTRY[name] = (module, class_name)
@staticmethod
def get_tokenizer(
tokenizer_name: str,
*args,
**kwargs,
) -> TokenizerBase:
tokenizer_cls = TokenizerRegistry.REGISTRY.get(tokenizer_name)
if tokenizer_cls is None:
raise ValueError(f"Tokenizer {tokenizer_name} not found.")
tokenizer_module = importlib.import_module(tokenizer_cls[0])
class_ = getattr(tokenizer_module, tokenizer_cls[1])
return class_.from_pretrained(*args, **kwargs)
+28 -11
View File
@@ -10,6 +10,7 @@ import huggingface_hub
from huggingface_hub import HfApi, hf_hub_download
from vllm.logger import init_logger
from vllm.transformers_utils.tokenizer_base import TokenizerBase
from vllm.utils import is_list_of
if TYPE_CHECKING:
@@ -140,7 +141,7 @@ def make_mistral_chat_completion_request(
tools=tools) # type: ignore[type-var]
class MistralTokenizer:
class MistralTokenizer(TokenizerBase):
def __init__(self, tokenizer: "PublicMistralTokenizer") -> None:
self.mistral = tokenizer
@@ -251,6 +252,14 @@ class MistralTokenizer:
def eos_token_id(self) -> int:
return self.tokenizer.eos_id
@property
def sep_token(self) -> str:
raise NotImplementedError()
@property
def pad_token(self) -> str:
raise NotImplementedError()
@property
def is_fast(self) -> bool:
return True
@@ -268,25 +277,26 @@ class MistralTokenizer:
def __call__(
self,
prompt: Union[str, List[str], List[int]],
text: Union[str, List[str], List[int]],
text_pair: Optional[str] = None,
add_special_tokens: bool = False,
truncation: bool = False,
max_length: Optional[int] = None,
):
input_ids: Union[List[int], List[List[int]]]
# For List[str], original prompt text
if is_list_of(prompt, str):
if is_list_of(text, str):
input_ids_: List[List[int]] = []
for p in prompt:
for p in text:
each_input_ids = self.encode_one(p, truncation, max_length)
input_ids_.append(each_input_ids)
input_ids = input_ids_
# For List[int], apply chat template output, already tokens.
elif is_list_of(prompt, int):
input_ids = prompt
elif is_list_of(text, int):
input_ids = text
# For str, single prompt text
else:
input_ids = self.encode_one(prompt, truncation, max_length)
input_ids = self.encode_one(text, truncation, max_length)
return Encoding(input_ids=input_ids)
def get_vocab(self) -> Dict[str, int]:
@@ -300,22 +310,29 @@ class MistralTokenizer:
def encode_one(
self,
prompt: str,
text: str,
truncation: bool = False,
max_length: Optional[int] = None,
) -> List[int]:
# Mistral Tokenizers should not add special tokens
input_ids = self.encode(prompt)
input_ids = self.encode(text)
if truncation:
input_ids = input_ids[:max_length]
return input_ids
def encode(self, prompt: str) -> List[int]:
def encode(self,
text: str,
add_special_tokens: Optional[bool] = None) -> List[int]:
# `encode` should only be used for prompt completion
# it should never be used for chat_completion.
# For chat completion use `apply_chat_template`
return self.tokenizer.encode(prompt, bos=True, eos=False)
if add_special_tokens is not None:
return self.tokenizer.encode(text,
bos=add_special_tokens,
eos=add_special_tokens)
else:
return self.tokenizer.encode(text, bos=True, eos=False)
def apply_chat_template(self,
messages: List["ChatCompletionMessageParam"],