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
[Bugfix] Fix processor initialization for model from modelscope instead of HF (#27461)
Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
rongfu.leng committed
Oct 26, 2025 at 15:44 UTC
87c41c26ad5bae18da8e393cd7e82f708816fd1e
2 files changed
+16
-1
vllm/transformers_utils/processor.py
+5
-1
@@ -16,6 +16,7 @@ from transformers.processing_utils import ProcessorMixin
16
from transformers.video_processing_utils import BaseVideoProcessor
17
from typing_extensions import TypeVar
18
19
+from vllm.transformers_utils.utils import convert_model_repo_to_path
20
from vllm.utils.func_utils import get_allowed_kwarg_only_overrides
21
22
if TYPE_CHECKING:
@@ -94,8 +95,8 @@ def get_processor(
95
"""Load a processor for the given model name via HuggingFace."""
96
if revision is None:
97
revision = "main"
97
-
98
try:
99
+ processor_name = convert_model_repo_to_path(processor_name)
100
if isinstance(processor_cls, tuple) or processor_cls == ProcessorMixin:
101
processor = AutoProcessor.from_pretrained(
102
processor_name,
@@ -168,6 +169,7 @@ def get_feature_extractor(
169
"""Load an audio feature extractor for the given model name
170
via HuggingFace."""
171
try:
172
+ processor_name = convert_model_repo_to_path(processor_name)
173
feature_extractor = AutoFeatureExtractor.from_pretrained(
174
processor_name,
175
*args,
@@ -217,6 +219,7 @@ def get_image_processor(
219
):
220
"""Load an image processor for the given model name via HuggingFace."""
221
try:
222
+ processor_name = convert_model_repo_to_path(processor_name)
223
processor = AutoImageProcessor.from_pretrained(
224
processor_name,
225
*args,
@@ -268,6 +271,7 @@ def get_video_processor(
271
):
272
"""Load a video processor for the given model name via HuggingFace."""
273
try:
274
+ processor_name = convert_model_repo_to_path(processor_name)
275
processor_cls = processor_cls_overrides or AutoVideoProcessor
276
processor = processor_cls.from_pretrained(
277
processor_name,
vllm/transformers_utils/utils.py
+11
@@ -2,6 +2,7 @@
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
import json
5
+import os
6
import struct
7
from functools import cache
8
from os import PathLike
@@ -109,3 +110,13 @@ def parse_safetensors_file_metadata(path: str | PathLike) -> dict[str, Any]:
110
length_of_metadata = struct.unpack("<Q", f.read(8))[0]
111
metadata = json.loads(f.read(length_of_metadata).decode("utf-8"))
112
return metadata
113
+
114
+
115
+def convert_model_repo_to_path(model_repo: str) -> str:
116
+ """When VLLM_USE_MODELSCOPE is True convert a model
117
+ repository string to a Path str."""
118
+ if not envs.VLLM_USE_MODELSCOPE or Path(model_repo).exists():
119
+ return model_repo
120
+ from modelscope.utils.file_utils import get_model_cache_root
121
+
122
+ return os.path.join(get_model_cache_root(), model_repo)