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

6
vllm/model_executor/models/deepseek_vl2.py
View File

@@ -204,12 +204,13 @@ class DeepseekVL2MultiModalProcessor(
prompt: str,
mm_data: Mapping[str, object],
mm_kwargs: Mapping[str, object],
tok_kwargs: Mapping[str, object],
) -> BatchFeature:
if mm_data:
processed_outputs = self.info.ctx.call_hf_processor(
self.info.get_hf_processor(**mm_kwargs),
dict(prompt=prompt, **mm_data),
mm_kwargs,
dict(**mm_kwargs, **tok_kwargs),
)
pixel_values = processed_outputs["pixel_values"]
# split pixel values into patches corresponding to each image
@@ -278,6 +279,7 @@ class DeepseekVL2MultiModalProcessor(
prompt: Union[str, list[int]],
mm_data_items: MultiModalDataItems,
hf_processor_mm_kwargs: Mapping[str, object],
tokenization_kwargs: Mapping[str, object],
*,
return_mm_hashes: bool,
) -> tuple[list[int], MultiModalKwargs, Optional[MultiModalHashes], bool]:
@@ -290,6 +292,7 @@ class DeepseekVL2MultiModalProcessor(
prompt=prompt,
mm_data_items=mm_data_items,
hf_processor_mm_kwargs=hf_processor_mm_kwargs,
tokenization_kwargs=tokenization_kwargs,
return_mm_hashes=return_mm_hashes,
)
@@ -297,6 +300,7 @@ class DeepseekVL2MultiModalProcessor(
prompt=prompt,
mm_data_items=mm_data_items,
hf_processor_mm_kwargs=hf_processor_mm_kwargs,
tokenization_kwargs=tokenization_kwargs,
return_mm_hashes=return_mm_hashes,
)