@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / b0298aa8cc
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 15:45 UTC

[Misc] Remove vLLM patch of `BaichuanTokenizer` (#8921)

Cyrus Leung committed Sep 28, 2024 at 16:11 UTC b0298aa8cc4a54bde659e57271778630785abc9b
3 files changed +3 -273
vllm/transformers_utils/tokenizer.py
+1 -15
@@ -11,8 +11,7 @@ from transformers import (AutoTokenizer, PreTrainedTokenizer,
11 from vllm.envs import VLLM_USE_MODELSCOPE
12 from vllm.logger import init_logger
13 from vllm.lora.request import LoRARequest
14 -from vllm.transformers_utils.tokenizers import (BaichuanTokenizer,
15 - MistralTokenizer)
14 +from vllm.transformers_utils.tokenizers import MistralTokenizer
15 from vllm.transformers_utils.utils import check_gguf_file
16 from vllm.utils import make_async
17
@@ -139,19 +138,6 @@ def get_tokenizer(
138 raise RuntimeError(err_msg) from e
139 else:
140 raise e
142 - except AttributeError as e:
143 - if "BaichuanTokenizer" in str(e):
144 - # This is for the error "'BaichuanTokenizer' object has no
145 - # attribute 'sp_model'".
146 - tokenizer = BaichuanTokenizer.from_pretrained(
147 - tokenizer_name,
148 - *args,
149 - trust_remote_code=trust_remote_code,
150 - revision=revision,
151 - **kwargs,
152 - )
153 - else:
154 - raise e
141
142 # NOTE: We can remove this after https://github.com/THUDM/ChatGLM3/issues/1324
143 if type(tokenizer).__name__ in ("ChatGLMTokenizer",
vllm/transformers_utils/tokenizers/__init__.py
+2 -3
@@ -1,4 +1,3 @@
1 -from vllm.transformers_utils.tokenizers.baichuan import BaichuanTokenizer
2 -from vllm.transformers_utils.tokenizers.mistral import MistralTokenizer
1 +from .mistral import MistralTokenizer
2
4 -__all__ = ["BaichuanTokenizer", "MistralTokenizer"]
3 +__all__ = ["MistralTokenizer"]
vllm/transformers_utils/tokenizers/baichuan.py deleted
-255
@@ -1,255 +0,0 @@
1 -# Adapted from
2 -# https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/blob/8f6e343d545c503b91429582231d1d354dac2740/tokenization_baichuan.py
3 -# This includes a fix suggested in
4 -# https://github.com/vllm-project/vllm/issues/1403#issuecomment-1767503058
5 -# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
6 -
7 -import os
8 -from shutil import copyfile
9 -from typing import Any, Dict, List, Optional, Tuple
10 -
11 -import sentencepiece as spm
12 -from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
13 -from transformers.utils import logging
14 -
15 -logger = logging.get_logger(__name__)
16 -
17 -VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
18 -
19 -PRETRAINED_VOCAB_FILES_MAP = { # type: ignore
20 - "vocab_file": {},
21 - "tokenizer_file": {},
22 -}
23 -PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {} # type: ignore
24 -
25 -
26 -class BaichuanTokenizer(PreTrainedTokenizer):
27 - """
28 - Construct a Baichuan tokenizer. Based on byte-level Byte-Pair-Encoding.
29 -
30 - Args:
31 - vocab_file (`str`):
32 - Path to the vocabulary file.
33 - """
34 -
35 - vocab_files_names = VOCAB_FILES_NAMES
36 - pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
37 - max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
38 - model_input_names = ["input_ids", "attention_mask"]
39 -
40 - def __init__(
41 - self,
42 - vocab_file,
43 - unk_token="<unk>",
44 - bos_token="<s>",
45 - eos_token="</s>",
46 - pad_token=None,
47 - sp_model_kwargs: Optional[Dict[str, Any]] = None,
48 - add_bos_token=True,
49 - add_eos_token=False,
50 - clean_up_tokenization_spaces=False,
51 - **kwargs,
52 - ):
53 - self.sp_model_kwargs = ({} if sp_model_kwargs is None else
54 - sp_model_kwargs)
55 - bos_token = (AddedToken(bos_token, lstrip=False, rstrip=False)
56 - if isinstance(bos_token, str) else bos_token)
57 - eos_token = (AddedToken(eos_token, lstrip=False, rstrip=False)
58 - if isinstance(eos_token, str) else eos_token)
59 - unk_token = (AddedToken(unk_token, lstrip=False, rstrip=False)
60 - if isinstance(unk_token, str) else unk_token)
61 - pad_token = (AddedToken(pad_token, lstrip=False, rstrip=False)
62 - if isinstance(pad_token, str) else pad_token)
63 - self.vocab_file = vocab_file
64 - self.add_bos_token = add_bos_token
65 - self.add_eos_token = add_eos_token
66 - self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
67 - self.sp_model.Load(vocab_file)
68 - super().__init__(
69 - bos_token=bos_token,
70 - eos_token=eos_token,
71 - unk_token=unk_token,
72 - pad_token=pad_token,
73 - add_bos_token=add_bos_token,
74 - add_eos_token=add_eos_token,
75 - sp_model_kwargs=self.sp_model_kwargs,
76 - clean_up_tokenization_spaces=clean_up_tokenization_spaces,
77 - **kwargs,
78 - )
79 -
80 - def __getstate__(self):
81 - state = self.__dict__.copy()
82 - state["sp_model"] = None
83 - return state
84 -
85 - def __setstate__(self, d):
86 - self.__dict__ = d
87 - self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
88 - self.sp_model.Load(self.vocab_file)
89 -
90 - @property
91 - def vocab_size(self):
92 - """Returns vocab size"""
93 - return self.sp_model.get_piece_size()
94 -
95 - def get_vocab(self):
96 - """Returns vocab as a dict"""
97 - vocab = {
98 - self.convert_ids_to_tokens(i): i
99 - for i in range(self.vocab_size)
100 - }
101 - vocab.update(self.added_tokens_encoder)
102 - return vocab
103 -
104 - def _tokenize(self, text):
105 - """Returns a tokenized string."""
106 - return self.sp_model.encode(text, out_type=str)
107 -
108 - def _convert_token_to_id(self, token):
109 - """Converts a token (str) in an id using the vocab."""
110 - return self.sp_model.piece_to_id(token)
111 -
112 - def _convert_id_to_token(self, index):
113 - """Converts an index (integer) in a token (str) using the vocab."""
114 - token = self.sp_model.IdToPiece(index)
115 - return token
116 -
117 - def convert_tokens_to_string(self, tokens: List[str]):
118 - """Converts a sequence of tokens (string) in a single string."""
119 - current_sub_tokens: List[str] = []
120 - out_string = ""
121 - prev_is_special = False
122 - for i, token in enumerate(tokens):
123 - # make sure that special tokens are not decoded using
124 - # sentencepiece model
125 - if token in self.all_special_tokens:
126 - if not prev_is_special and i != 0:
127 - out_string += " "
128 - out_string += self.sp_model.decode(current_sub_tokens) + token
129 - prev_is_special = True
130 - current_sub_tokens = []
131 - else:
132 - current_sub_tokens.append(token)
133 - prev_is_special = False
134 - out_string += self.sp_model.decode(current_sub_tokens)
135 - return out_string
136 -
137 - def save_vocabulary(self,
138 - save_directory,
139 - filename_prefix: Optional[str] = None) -> Tuple[str]:
140 - """
141 - Save the vocabulary and special tokens file to a directory.
142 -
143 - Args:
144 - save_directory (`str`):
145 - The directory in which to save the vocabulary.
146 -
147 - Returns:
148 - `Tuple(str)`: Paths to the files saved.
149 - """
150 - if not os.path.isdir(save_directory):
151 - raise ValueError(f"Vocabulary path ({save_directory}) "
152 - "should be a directory")
153 -
154 - out_vocab_file = os.path.join(
155 - save_directory,
156 - (filename_prefix + "-" if filename_prefix else "") +
157 - VOCAB_FILES_NAMES["vocab_file"],
158 - )
159 -
160 - if os.path.abspath(self.vocab_file) != os.path.abspath(
161 - out_vocab_file) and os.path.isfile(self.vocab_file):
162 - copyfile(self.vocab_file, out_vocab_file)
163 - elif not os.path.isfile(self.vocab_file):
164 - with open(out_vocab_file, "wb") as fi:
165 - content_spiece_model = self.sp_model.serialized_model_proto()
166 - fi.write(content_spiece_model)
167 -
168 - return (out_vocab_file, )
169 -
170 - def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
171 - bos_token_id = [self.bos_token_id] if self.add_bos_token else []
172 - eos_token_id = [self.eos_token_id] if self.add_eos_token else []
173 -
174 - output = bos_token_id + token_ids_0 + eos_token_id
175 -
176 - if token_ids_1 is not None:
177 - output = output + bos_token_id + token_ids_1 + eos_token_id
178 -
179 - return output
180 -
181 - def get_special_tokens_mask(
182 - self,
183 - token_ids_0: List[int],
184 - token_ids_1: Optional[List[int]] = None,
185 - already_has_special_tokens: bool = False,
186 - ) -> List[int]:
187 - """
188 - Retrieve sequence ids from a token list that has no special tokens
189 - added. This method is called when adding
190 - special tokens using the tokenizer `prepare_for_model` method.
191 -
192 - Args:
193 - token_ids_0 (`List[int]`):
194 - List of IDs.
195 - token_ids_1 (`List[int]`, *optional*):
196 - Optional second list of IDs for sequence pairs.
197 - already_has_special_tokens (`bool`, *optional*, defaults to
198 - `False`):
199 - Whether or not the token list is already formatted with
200 - special tokens for the model.
201 -
202 - Returns:
203 - `List[int]`: A list of integers in the range [0, 1]:
204 - 1 for a special token, 0 for a sequence token.
205 - """
206 - if already_has_special_tokens:
207 - return super().get_special_tokens_mask(
208 - token_ids_0=token_ids_0,
209 - token_ids_1=token_ids_1,
210 - already_has_special_tokens=True,
211 - )
212 -
213 - bos_token_id = [1] if self.add_bos_token else []
214 - eos_token_id = [1] if self.add_eos_token else []
215 -
216 - if token_ids_1 is None:
217 - return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
218 - return (bos_token_id + ([0] * len(token_ids_0)) + eos_token_id +
219 - bos_token_id + ([0] * len(token_ids_1)) + eos_token_id)
220 -
221 - def create_token_type_ids_from_sequences(
222 - self,
223 - token_ids_0: List[int],
224 - token_ids_1: Optional[List[int]] = None) -> List[int]:
225 - """
226 - Creates a mask from the two sequences passed to be used in a
227 - sequence-pair classification task. An ALBERT
228 - sequence pair mask has the following format:
229 -
230 - ```
231 - 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
232 - | first sequence | second sequence |
233 - ```
234 -
235 - if token_ids_1 is None, only returns the first portion of the mask (0s).
236 -
237 - Args:
238 - token_ids_0 (`List[int]`):
239 - List of ids.
240 - token_ids_1 (`List[int]`, *optional*):
241 - Optional second list of IDs for sequence pairs.
242 -
243 - Returns:
244 - `List[int]`: List of [token type IDs](../glossary#token-type-ids)
245 - according to the given sequence(s).
246 - """
247 - bos_token_id = [self.bos_token_id] if self.add_bos_token else []
248 - eos_token_id = [self.eos_token_id] if self.add_eos_token else []
249 -
250 - output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
251 -
252 - if token_ids_1 is not None:
253 - output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
254 -
255 - return output