[Multimodal] Expose `mm_processor_kwargs` for `DummyInputsBuilder` (#34330)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

vllm/model_executor/models/mistral3.py

@@ -237,6 +237,7 @@ class Mistral3DummyInputsBuilder(BaseDummyInputsBuilder[_I]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 

vllm/model_executor/models/mllama4.py

@@ -704,6 +704,7 @@ class Mllama4DummyInputsBuilder(BaseDummyInputsBuilder[Mllama4ProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 

vllm/model_executor/models/molmo.py

@@ -1278,6 +1278,7 @@ class MolmoDummyInputsBuilder(BaseDummyInputsBuilder[MolmoProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         target_width, target_height = self.info.get_image_size_with_most_features()
         num_images = mm_counts.get("image", 0)

vllm/model_executor/models/molmo2.py

@@ -2079,6 +2079,7 @@ class Molmo2DummyInputsBuilder(BaseDummyInputsBuilder[Molmo2ProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
         num_videos = mm_counts.get("video", 0)

vllm/model_executor/models/nano_nemotron_vl.py

@@ -1385,6 +1385,7 @@ class NanoNemotronVLDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
         processor = self.info.get_hf_processor()
@@ -1457,6 +1458,7 @@ class NanoNemotronVLDummyInputsBuilder(
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         dummy_image = super().get_dummy_mm_data(
             seq_len=seq_len, mm_counts=mm_counts, mm_options=mm_options

vllm/model_executor/models/nemotron_parse.py

@@ -642,6 +642,7 @@ class NemotronParseDummyInputsBuilder(
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 

vllm/model_executor/models/nvlm_d.py

@@ -93,6 +93,7 @@ class NVLMDummyInputsBuilder(BaseInternVLDummyInputsBuilder[NVLMProcessingInfo])
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         target_width, target_height = self.info.get_image_size_with_most_features()
         num_images = mm_counts.get("image", 0)

vllm/model_executor/models/ovis.py

@@ -303,6 +303,7 @@ class OvisDummyInputsBuilder(BaseDummyInputsBuilder[OvisProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 

vllm/model_executor/models/ovis2_5.py

@@ -302,6 +302,7 @@ class Ovis2_5DummyInputsBuilder(BaseDummyInputsBuilder[Ovis2_5ProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
         num_videos = mm_counts.get("video", 0)

vllm/model_executor/models/paddleocr_vl.py

@@ -204,6 +204,7 @@ class PaddleOCRVLDummyInputsBuilder(BaseDummyInputsBuilder[PaddleOCRVLProcessing
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 

vllm/model_executor/models/paligemma.py

@@ -128,6 +128,7 @@ class PaliGemmaDummyInputsBuilder(BaseDummyInputsBuilder[PaliGemmaProcessingInfo
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         hf_config = self.info.get_hf_config()
         vision_config = hf_config.vision_config

vllm/model_executor/models/phi3v.py

@@ -380,6 +380,7 @@ class Phi3VDummyInputsBuilder(BaseDummyInputsBuilder[Phi3VProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 

vllm/model_executor/models/phi4mm.py

@@ -826,6 +826,7 @@ class Phi4MMDummyInputsBuilder(BaseDummyInputsBuilder[Phi4MMProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_audios = mm_counts.get("audio", 0)
         num_images = mm_counts.get("image", 0)

vllm/model_executor/models/pixtral.py

@@ -261,6 +261,7 @@ class PixtralDummyInputsBuilder(BaseDummyInputsBuilder[PixtralProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 
@@ -282,6 +283,7 @@ class PixtralDummyInputsBuilder(BaseDummyInputsBuilder[PixtralProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> ProcessorInputs:
         tokenizer = self.info.get_tokenizer()
 

vllm/model_executor/models/qwen2_5_omni_thinker.py

@@ -358,12 +358,14 @@ class Qwen2_5OmniThinkerDummyInputsBuilder(
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_audios = mm_counts.get("audio", 0)
         num_images = mm_counts.get("image", 0)
         num_videos = mm_counts.get("video", 0)
 
-        feature_extractor = self.info.get_feature_extractor()
+        mm_processor_kwargs = mm_processor_kwargs or {}
+        feature_extractor = self.info.get_feature_extractor(**mm_processor_kwargs)
 
         target_audio_length = (
             min(
@@ -372,7 +374,10 @@ class Qwen2_5OmniThinkerDummyInputsBuilder(
             )
             * feature_extractor.sampling_rate
         )
-        target_width, target_height = self.info.get_image_size_with_most_features()
+
+        target_width, target_height = self.info.get_image_size_with_most_features(
+            max_pixels=mm_processor_kwargs.get("max_pixels", None),
+        )
         target_num_frames = self.info.get_num_frames_with_most_features(
             seq_len, mm_counts
         )

vllm/model_executor/models/qwen2_audio.py

@@ -195,8 +195,11 @@ class Qwen2AudioDummyInputsBuilder(BaseDummyInputsBuilder[Qwen2AudioProcessingIn
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
-        feature_extractor = self.info.get_feature_extractor()
+        feature_extractor = self.info.get_feature_extractor(
+            **(mm_processor_kwargs or {})
+        )
 
         sampling_rate = feature_extractor.sampling_rate
         audio_len = feature_extractor.chunk_length * sampling_rate

vllm/model_executor/models/qwen2_vl.py

@@ -1016,11 +1016,15 @@ class Qwen2VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen2VLProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
         num_videos = mm_counts.get("video", 0)
 
-        target_width, target_height = self.info.get_image_size_with_most_features()
+        mm_processor_kwargs = mm_processor_kwargs or {}
+        target_width, target_height = self.info.get_image_size_with_most_features(
+            max_pixels=mm_processor_kwargs.get("max_pixels", None)
+        )
         target_num_frames = self.info.get_num_frames_with_most_features(
             seq_len, mm_counts
         )

vllm/model_executor/models/qwen3_asr.py

@@ -147,10 +147,13 @@ class Qwen3ASRDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3ASRProcessingInfo])
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_audios = mm_counts.get("audio", 0)
 
-        feature_extractor = self.info.get_feature_extractor()
+        feature_extractor = self.info.get_feature_extractor(
+            **(mm_processor_kwargs or {})
+        )
 
         target_audio_length = (
             min(

vllm/model_executor/models/qwen3_omni_moe_thinker.py

@@ -1169,7 +1169,7 @@ class Qwen3OmniMoeThinkerMultiModalProcessor(
             return x
 
         # NOTE: WhisperFeatureExtractor cannot handle empty list of audios
-        feature_extractor = self.info.get_feature_extractor()
+        feature_extractor = self.info.get_feature_extractor(**mm_kwargs)
         hop_length = feature_extractor.hop_length
         if audios:
             # NOTE: Qwen3-Omni processor accept "audio"

vllm/model_executor/models/qwen3_vl.py

@@ -796,14 +796,18 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
         num_videos = mm_counts.get("video", 0)
         image_overrides = mm_options.get("image") if mm_options else None
         video_overrides = mm_options.get("video") if mm_options else None
 
+        mm_processor_kwargs = mm_processor_kwargs or {}
         target_image_width, target_image_height = (
-            self.info.get_image_size_with_most_features()
+            self.info.get_image_size_with_most_features(
+                max_pixels=mm_processor_kwargs.get("max_pixels", None),
+            )
         )
 
         # treat videos as special images
@@ -828,7 +832,7 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
                 target_num_frames = min(target_num_frames, num_frames_override)
         target_num_frames = max(target_num_frames, 2)
 
-        video_processor = self.info.get_video_processor()
+        video_processor = self.info.get_video_processor(**(mm_processor_kwargs or {}))
         video_max_pixels = video_processor.size["longest_edge"]
         # video_max_pixels contains the temporal compression factor,
         # so we divide by 2 to get the maximum number of image pixels.