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

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

vllm/model_executor/models/aria.py

@@ -445,6 +445,7 @@ class AriaDummyInputsBuilder(BaseDummyInputsBuilder[AriaProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         vision_config = self.info.get_vision_config()
 

vllm/model_executor/models/audioflamingo3.py

@@ -253,8 +253,11 @@ class AudioFlamingo3DummyInputsBuilder(
         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 = MAX_AUDIO_LEN * sampling_rate
         num_audios = mm_counts.get("audio", 0)

vllm/model_executor/models/aya_vision.py

@@ -192,6 +192,7 @@ class AyaVisionDummyInputsBuilder(BaseDummyInputsBuilder[AyaVisionProcessingInfo
         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)
         image_size = self.info.get_image_size_with_most_features()

vllm/model_executor/models/bagel.py

@@ -250,6 +250,7 @@ class BagelDummyInputsBuilder(BaseDummyInputsBuilder[BagelProcessingInfo]):
         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)
         hf_config = self.info.get_hf_config()

vllm/model_executor/models/bee.py

@@ -91,6 +91,7 @@ class BeeDummyInputsBuilder(LlavaDummyInputsBuilder[BeeProcessingInfo]):
         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/blip2.py

@@ -446,6 +446,7 @@ class Blip2DummyInputsBuilder(BaseDummyInputsBuilder[Blip2ProcessingInfo]):
         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/chameleon.py

@@ -117,6 +117,7 @@ class ChameleonDummyInputsBuilder(BaseDummyInputsBuilder[ChameleonProcessingInfo
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         config = self.info.get_hf_config()
 

vllm/model_executor/models/clip.py

@@ -171,6 +171,7 @@ class CLIPDummyInputsBuilder(BaseDummyInputsBuilder[CLIPProcessingInfo]):
         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/cohere2_vision.py

@@ -221,6 +221,7 @@ class Cohere2VisionDummyInputsBuilder(
         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)
         image_size = self.info.get_image_size_with_most_features()

vllm/model_executor/models/deepseek_ocr.py

@@ -256,6 +256,7 @@ class DeepseekOCRDummyInputsBuilder(BaseDummyInputsBuilder[DeepseekOCRProcessing
         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/deepseek_ocr2.py

@@ -138,6 +138,7 @@ class DeepseekOCR2DummyInputsBuilder(
         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/deepseek_vl2.py

@@ -215,6 +215,7 @@ class DeepseekVL2DummyInputsBuilder(BaseDummyInputsBuilder[DeepseekVL2Processing
         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/dots_ocr.py

@@ -107,10 +107,13 @@ class DotsOCRDummyInputsBuilder(Qwen2VLDummyInputsBuilder):
         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)
 
+        mm_processor_kwargs = mm_processor_kwargs or {}
         target_width, target_height = self.info.get_image_size_with_most_features(  # noqa: E501
+            mm_processor_kwargs.get("max_pixels", None)
         )
 
         image_overrides = mm_options.get("image") if mm_options else None

vllm/model_executor/models/ernie45_vl.py

@@ -1153,6 +1153,7 @@ class Ernie4_5_VLDummyInputsBuilder(BaseDummyInputsBuilder[Ernie4_5_VLProcessing
         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/funasr.py

@@ -745,8 +745,11 @@ class FunASRDummyInputsBuilder(BaseDummyInputsBuilder[FunASRProcessingInfo]):
         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/funaudiochat.py

@@ -611,8 +611,11 @@ class FunAudioChatDummyInputsBuilder(
         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 = int(feature_extractor.sampling_rate)
 
         # Dummy inputs are used for profiling; construct the worst-case audio
@@ -656,7 +659,7 @@ class FunAudioChatMultiModalProcessor(
         if not audios:
             return BatchFeature({"input_ids": input_ids})
 
-        feature_extractor = self.info.get_feature_extractor()
+        feature_extractor = self.info.get_feature_extractor(**mm_kwargs)
         sr = int(feature_extractor.sampling_rate)
         min_samples = int(getattr(feature_extractor, "n_fft", 400) or 400)
 

vllm/model_executor/models/fuyu.py

@@ -143,6 +143,7 @@ class FuyuDummyInputsBuilder(BaseDummyInputsBuilder[FuyuProcessingInfo]):
         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/gemma3_mm.py

@@ -256,6 +256,7 @@ class Gemma3DummyInputsBuilder(BaseDummyInputsBuilder[Gemma3ProcessingInfo]):
         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/gemma3n_mm.py

@@ -182,6 +182,7 @@ class Gemma3nDummyInputsBuilder(BaseDummyInputsBuilder[Gemma3nProcessingInfo]):
         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_audios = mm_counts.get("audio", 0)

vllm/model_executor/models/glm4_1v.py

@@ -1143,6 +1143,7 @@ class Glm4vDummyInputsBuilder(BaseDummyInputsBuilder[Glm4vProcessingInfo]):
         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)