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 14, 05:30 UTC
chore: remove unnecessary patch_padding_side for the chatglm model (#23090)
Signed-off-by: carlory <baofa.fan@daocloud.io> Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
杨朱 · Kiki committed
Aug 18, 2025 at 20:32 UTC
569aefd13437ea0c49c0302f1100c3997fe067eb
2 files changed
-29
tests/models/multimodal/generation/vlm_utils/model_utils.py
-2
@@ -19,7 +19,6 @@ from transformers import (AutoConfig, AutoTokenizer, BatchFeature,
19
from transformers.video_utils import VideoMetadata
20
21
from vllm.sequence import SampleLogprobs
22
-from vllm.transformers_utils.tokenizer import patch_padding_side
22
from vllm.utils import is_list_of
23
24
from .....conftest import HfRunner, ImageAsset, ImageTestAssets
@@ -343,7 +342,6 @@ def gemma3_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
342
def glm4v_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
343
"""Patches and returns an instance of the HfRunner to use for GLM4V."""
344
hf_processor = hf_model.processor
346
- patch_padding_side(hf_processor)
345
346
def processor(*args, text="", images=None, **kwargs):
347
if images is None:
vllm/transformers_utils/tokenizer.py
-27
@@ -7,7 +7,6 @@ import os
7
import warnings
8
from functools import lru_cache
9
from pathlib import Path
10
-from types import MethodType
10
from typing import TYPE_CHECKING, Any, Optional, Union
11
12
import huggingface_hub
@@ -144,26 +143,6 @@ def get_cached_tokenizer(tokenizer: AnyTokenizer) -> AnyTokenizer:
143
return cached_tokenizer
144
145
147
-def patch_padding_side(tokenizer: PreTrainedTokenizer) -> None:
148
- """Patch _pad method to accept `padding_side` for older tokenizers."""
149
- orig_pad = tokenizer._pad
150
-
151
- def _pad(
152
- self: PreTrainedTokenizer,
153
- *args,
154
- padding_side: Optional[str] = None,
155
- **kwargs,
156
- ):
157
- if padding_side is not None and padding_side != self.padding_side:
158
- msg = ("`padding_side` argument is not supported by "
159
- f"{type(tokenizer).__name__} and will be ignored.")
160
- warnings.warn(msg, stacklevel=2)
161
-
162
- return orig_pad(*args, **kwargs)
163
-
164
- tokenizer._pad = MethodType(_pad, tokenizer)
165
-
166
-
146
def get_tokenizer(
147
tokenizer_name: Union[str, Path],
148
*args,
@@ -271,12 +250,6 @@ def get_tokenizer(
250
}
251
tokenizer.add_special_tokens(special_tokens_map)
252
274
- # NOTE: We can remove this after https://github.com/zai-org/ChatGLM3/issues/1324
275
- if type(tokenizer).__name__ in ("ChatGLMTokenizer",
276
- "ChatGLM4Tokenizer"):
277
- assert isinstance(tokenizer, PreTrainedTokenizer)
278
- patch_padding_side(tokenizer)
279
-
253
if not isinstance(tokenizer, PreTrainedTokenizerFast):
254
logger.warning(
255
"Using a slow tokenizer. This might cause a significant "