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 13, 08:15 UTC
Skip models that cannot currently init on Transformers v5 (#28471)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Nov 12, 2025 at 23:43 UTC
51c599f0ec9c754ddf9f6094f27c1fa2be76b318
3 files changed
+6
-16
.buildkite/test-pipeline.yaml
+4
-4
@@ -872,12 +872,12 @@ steps:
872
optional: true
873
commands:
874
- pip install --upgrade git+https://github.com/huggingface/transformers
875
- - pytest -v -s tests/models/test_initialization.py
875
+ - pytest -v -s tests/models/test_initialization.py -k 'not (Gemma3 or ModernBert or Qwen2_5_VL or Qwen2_5vl or Qwen2VL or TransformersMultiModalEmbeddingModel or TransformersMultiModalForSequenceClassification or Ultravox or Phi4Multimodal or LlavaNextVideo or MiniCPMO or Lfm2Moe or PaliGemma or RobertaForSequenceClassification or Ovis2_5 or Fuyu or DeepseekOCR or KimiVL)'
876
- pytest -v -s tests/models/test_transformers.py
877
- - pytest -v -s tests/models/multimodal/processing/
878
- - pytest -v -s tests/models/multimodal/test_mapping.py
877
+ # - pytest -v -s tests/models/multimodal/processing/
878
+ - pytest -v -s tests/models/multimodal/test_mapping.py -k 'not (Gemma3 or Qwen2VL or Qwen2_5_VL)'
879
- python3 examples/offline_inference/basic/chat.py
880
- - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
880
+ # - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
881
# Whisper needs spawn method to avoid deadlock
882
- VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
883
vllm/model_executor/model_loader/weight_utils.py
+2
-1
@@ -82,7 +82,8 @@ enable_hf_transfer()
82
83
class DisabledTqdm(tqdm):
84
def __init__(self, *args, **kwargs):
85
- super().__init__(*args, **kwargs, disable=True)
85
+ kwargs["disable"] = True
86
+ super().__init__(*args, **kwargs)
87
88
89
def get_lock(model_name_or_path: str | Path, cache_dir: str | None = None):
vllm/model_executor/models/whisper.py
-11
@@ -13,7 +13,6 @@ from transformers import (
13
BatchFeature,
14
WhisperConfig,
15
WhisperFeatureExtractor,
16
- WhisperProcessor,
16
)
17
from transformers.models.whisper.modeling_whisper import sinusoids
18
@@ -660,16 +659,6 @@ class WhisperProcessingInfo(BaseProcessingInfo):
659
def get_hf_config(self) -> WhisperConfig:
660
return self.ctx.get_hf_config(WhisperConfig)
661
663
- def get_hf_processor(self, **kwargs: object) -> WhisperProcessor:
664
- # HACK: Transformers 4.53.2 has issue with whisper tokenizer to
665
- # initialize processor. We use a monkeypatch to fix it here.
666
- # See: https://github.com/vllm-project/vllm/issues/20224
667
- processor_class = WhisperProcessor
668
- tokenizer_class = ("WhisperTokenizer", "WhisperTokenizerFast")
669
- if processor_class.tokenizer_class != tokenizer_class:
670
- processor_class.tokenizer_class = tokenizer_class
671
- return self.ctx.get_hf_processor(processor_class, **kwargs)
672
-
662
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
663
return {"audio": 1}
664