biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Core] [Bugfix] [Multimodal] Fix multimodal profiling and generation for SFT/PTQed models (#20058)

Browse Source 
Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>
This commit is contained in:
Kyle Sayers
2025-06-30 13:26:49 -04:00
committed by GitHub
parent 551ef1631a
commit d8cf819a9a
41 changed files with 207 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/models/qwen2_5_omni_thinker.py
View File

@@ -244,6 +244,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
prompt: str,
mm_data: Mapping[str, object],
mm_kwargs: Mapping[str, object],
tok_kwargs: Mapping[str, object],
) -> BatchFeature:
mm_data = dict(mm_data)
audios = mm_data.pop("audios", [])
@@ -258,6 +259,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
prompt=prompt,
mm_data=mm_data,
mm_kwargs=mm_kwargs,
tok_kwargs=tok_kwargs,
)
input_features = hf_inputs.pop('input_features', None)
@@ -453,6 +455,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
prompt: Union[str, list[int]],
mm_items: MultiModalDataItems,
hf_processor_mm_kwargs: Mapping[str, object],
tokenization_kwargs: Mapping[str, object],
*,
enable_hf_prompt_update: bool,
) -> tuple[list[int], MultiModalKwargs, bool]:
@@ -465,6 +468,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
prompt_text=prompt,
mm_items=mm_items,
hf_processor_mm_kwargs=hf_processor_mm_kwargs,
tokenization_kwargs=tokenization_kwargs,
)
tokenizer = self.info.get_tokenizer()
prompt_ids = encode_tokens(tokenizer, prompt)
@@ -474,6 +478,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
mm_kwargs = self._apply_hf_processor_mm_only(
mm_items=mm_items,
hf_processor_mm_kwargs=hf_processor_mm_kwargs,
tokenization_kwargs=tokenization_kwargs,
)
return prompt_ids, mm_kwargs, False
@@ -482,6 +487,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
self,
mm_items: MultiModalDataItems,
hf_processor_mm_kwargs: Mapping[str, object],
tokenization_kwargs: Mapping[str, object],
) -> MultiModalKwargs:
"""
Qwen2.5-Omni reimplements this function to handle `use_audio_in_video`.
@@ -498,6 +504,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
prompt_text=self.dummy_inputs.get_dummy_text(mm_counts),
mm_items=mm_items,
hf_processor_mm_kwargs=hf_processor_mm_kwargs,
tokenization_kwargs=tokenization_kwargs,
)
return mm_kwargs