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

[Refactor] Define MM data parser in processing info instead of processor itself (#33260)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-29 13:55:17 +08:00
committed by GitHub
parent 07ea184f00
commit 51550179fc
34 changed files with 399 additions and 347 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/model_executor/models/qwen3_vl.py
View File

@@ -81,7 +81,7 @@ from vllm.multimodal.inputs import (
PlaceholderRange,
VideoItem,
)
from vllm.multimodal.parse import ImageSize, MultiModalDataItems, MultiModalDataParser
from vllm.multimodal.parse import ImageSize, MultiModalDataItems
from vllm.multimodal.processing import (
BaseDummyInputsBuilder,
BaseMultiModalProcessor,
@@ -624,6 +624,13 @@ class Qwen3VLProcessingInfo(Qwen2VLProcessingInfo):
def get_video_processor(self, **kwargs: object) -> Qwen3VLVideoProcessor:
return self.get_hf_processor(**kwargs).video_processor
def get_data_parser(self):
return Qwen2VLMultiModalDataParser(
self.get_hf_config().vision_config.spatial_merge_size,
video_needs_metadata=True,
expected_hidden_size=self._get_expected_hidden_size(),
)
def _get_vision_info(
self,
*,
@@ -901,12 +908,6 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo]):
def _get_data_parser(self) -> MultiModalDataParser:
return Qwen2VLMultiModalDataParser(
self.info.get_hf_config().vision_config.spatial_merge_size,
video_needs_metadata=True,
)
def _call_hf_processor(
self,
prompt: str,