[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)

vllm/model_executor/models/glm4v.py

@@ -493,6 +493,7 @@ 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:
         hf_config = self.info.get_hf_config()
         vision_config = hf_config.vision_config

vllm/model_executor/models/glmasr.py

@@ -727,8 +727,11 @@ class GlmAsrDummyInputsBuilder(BaseDummyInputsBuilder[GlmAsrProcessingInfo]):
         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
         num_audios = mm_counts.get("audio", 0)
         audio_overrides = mm_options.get("audio") if mm_options else None

vllm/model_executor/models/granite_speech.py

@@ -217,6 +217,7 @@ class GraniteSpeechDummyInputsBuilder(
         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)
         audio_overrides = mm_options.get("audio") if mm_options else None

vllm/model_executor/models/hunyuan_vision.py

@@ -702,6 +702,7 @@ class HunYuanVLDummyInputsBuilder(BaseDummyInputsBuilder[HunYuanVLProcessingInfo
         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", 1)
 

vllm/model_executor/models/hyperclovax_vision.py

@@ -166,6 +166,7 @@ class HCXVisionDummyInputsBuilder(BaseDummyInputsBuilder[HCXVisionProcessingInfo
         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/idefics3.py

@@ -42,7 +42,7 @@ from vllm.multimodal.inputs import (
     MultiModalFieldConfig,
     MultiModalKwargsItems,
 )
-from vllm.multimodal.parse import ImageProcessorItems, ImageSize, MultiModalDataItems
+from vllm.multimodal.parse import ImageProcessorItems, MultiModalDataItems
 from vllm.multimodal.processing import (
     BaseDummyInputsBuilder,
     BaseMultiModalProcessor,
@@ -285,15 +285,6 @@ class Idefics3ProcessingInfo(BaseProcessingInfo):
 
         return num_patches * processor.image_seq_len
 
-    def get_image_size_with_most_features(self) -> ImageSize:
-        processor = self.get_hf_processor()
-        image_processor: Idefics3ImageProcessor = processor.image_processor
-
-        return ImageSize(
-            width=image_processor.size["longest_edge"],
-            height=image_processor.size["longest_edge"],
-        )
-
 
 class Idefics3DummyInputsBuilder(BaseDummyInputsBuilder[Idefics3ProcessingInfo]):
     def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
@@ -309,9 +300,10 @@ class Idefics3DummyInputsBuilder(BaseDummyInputsBuilder[Idefics3ProcessingInfo])
         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_processor = self.info.get_hf_processor()
+        hf_processor = self.info.get_hf_processor(**(mm_processor_kwargs or {}))
         image_processor: Idefics3ImageProcessor = hf_processor.image_processor
         longest_edge = image_processor.max_image_size["longest_edge"]
 

vllm/model_executor/models/interns1.py

@@ -298,6 +298,7 @@ class InternS1DummyInputsBuilder(BaseDummyInputsBuilder[InternS1ProcessingInfo])
         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()
         target_num_frames = self.info.get_num_frames_with_most_features(

vllm/model_executor/models/internvl.py

@@ -766,6 +766,7 @@ class BaseInternVLDummyInputsBuilder(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:
         target_width, target_height = self.info.get_image_size_with_most_features()
         num_images = mm_counts.get("image", 0)
@@ -938,6 +939,7 @@ class InternVLDummyInputsBuilder(
         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/isaac.py

@@ -850,6 +850,7 @@ class IsaacDummyInputsBuilder(BaseDummyInputsBuilder[IsaacProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 

vllm/model_executor/models/kanana_v.py

@@ -445,6 +445,7 @@ class KananaVDummyInputsBuilder(BaseDummyInputsBuilder[KananaVProcessingInfo]):
         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)
         return {

vllm/model_executor/models/keye.py

@@ -1159,6 +1159,7 @@ class KeyeBaseDummyInputsBuilder(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)
         num_videos = mm_counts.get("video", 0)

vllm/model_executor/models/kimi_k25.py

@@ -238,6 +238,7 @@ class KimiK25DummyInputsBuilder(BaseDummyInputsBuilder[KimiK25ProcessingInfo]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> MultiModalDataDict:
         # TODO: Support mm_options for vision_chunk to allow user configuration
         dummy_items = self.get_dummy_mm_items()

vllm/model_executor/models/kimi_vl.py

@@ -216,6 +216,7 @@ class KimiVLDummyInputsBuilder(BaseDummyInputsBuilder[KimiVLProcessingInfo]):
         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/lfm2_vl.py

@@ -319,6 +319,7 @@ class Lfm2VLDummyInputsBuilder(BaseDummyInputsBuilder[Lfm2VLProcessingInfo]):
         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/llava.py

@@ -232,6 +232,7 @@ class LlavaDummyInputsBuilder(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/llava_next_video.py

@@ -166,6 +166,7 @@ class LlavaNextVideoDummyInputsBuilder(
         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_videos = mm_counts.get("video", 0)
 

vllm/model_executor/models/llava_onevision.py

@@ -277,6 +277,7 @@ class LlavaOnevisionDummyInputsBuilder(
         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/midashenglm.py

@@ -566,6 +566,7 @@ class MiDashengLMDummyInputsBuilder(BaseDummyInputsBuilder[MiDashengLMProcessing
         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)
 

vllm/model_executor/models/minicpmo.py

@@ -302,6 +302,7 @@ class MiniCPMODummyInputsBuilder(MiniCPMVDummyInputsBuilder[MiniCPMOProcessingIn
         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)
         audio_len = (

vllm/model_executor/models/minicpmv.py

@@ -708,6 +708,7 @@ class MiniCPMVDummyInputsBuilder(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)
         num_videos = mm_counts.get("video", 0)

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.

vllm/model_executor/models/qwen_vl.py

@@ -618,6 +618,7 @@ class QwenVLDummyInputsBuilder(BaseDummyInputsBuilder[QwenVLProcessingInfo]):
         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.visual

vllm/model_executor/models/rvl.py

@@ -41,6 +41,7 @@ class RVLDummyInputsBuilder(LlavaDummyInputsBuilder[RVLProcessingInfo]):
         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/siglip.py

@@ -155,6 +155,7 @@ class SiglipDummyInputsBuilder(BaseDummyInputsBuilder[SiglipProcessingInfo]):
         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/skyworkr1v.py

@@ -533,6 +533,7 @@ class SkyworkR1VDummyInputsBuilder(BaseDummyInputsBuilder[SkyworkR1VProcessingIn
         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/step3_vl.py

@@ -565,6 +565,7 @@ class Step3VLDummyInputsBuilder(BaseDummyInputsBuilder[Step3VLProcessingInfo]):
         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/terratorch.py

@@ -154,6 +154,7 @@ class TerratorchInputBuilder(BaseDummyInputsBuilder[TerratorchProcessingInfo]):
         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 data is generated based on the 'input' section
         # defined in the HF configuration file

vllm/model_executor/models/transformers/multimodal.py

@@ -98,6 +98,7 @@ class MultiModalDummyInputsBuilder(BaseDummyInputsBuilder[MultiModalProcessingIn
         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/ultravox.py

@@ -161,8 +161,11 @@ class UltravoxDummyInputsBuilder(BaseDummyInputsBuilder[UltravoxProcessingInfo])
         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 = (

vllm/model_executor/models/voxtral.py

@@ -220,6 +220,7 @@ class VoxtralDummyInputsBuilder(BaseDummyInputsBuilder[VoxtralProcessingInfo]):
         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)
 
@@ -238,6 +239,7 @@ class VoxtralDummyInputsBuilder(BaseDummyInputsBuilder[VoxtralProcessingInfo]):
         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/whisper.py

@@ -685,8 +685,11 @@ class WhisperDummyInputsBuilder(BaseDummyInputsBuilder[WhisperProcessingInfo]):
         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/multimodal/processing/dummy_inputs.py

@@ -63,6 +63,7 @@ class BaseDummyInputsBuilder(ABC, Generic[_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:
         """
         Build the multimodal input which, after processing, results in
@@ -83,6 +84,7 @@ class BaseDummyInputsBuilder(ABC, Generic[_I]):
         seq_len: int,
         mm_counts: Mapping[str, int],
         mm_options: Mapping[str, BaseDummyOptions] | None = None,
+        mm_processor_kwargs: Mapping[str, object] | None = None,
     ) -> ProcessorInputs:
         """
         Build the input which, after processing, results in
@@ -92,9 +94,16 @@ class BaseDummyInputsBuilder(ABC, Generic[_I]):
             seq_len: Sequence length
             mm_counts: Count of items per modality
             mm_options: Configurable options per modality (optional)
+            mm_processor_kwargs: Additional keyword arguments
+                                for hf_processor (optional)
         """
         dummy_text = self.get_dummy_text(mm_counts)
-        dummy_mm_data = self.get_dummy_mm_data(seq_len, mm_counts, mm_options)
+        dummy_mm_data = self.get_dummy_mm_data(
+            seq_len,
+            mm_counts,
+            mm_options,
+            mm_processor_kwargs=mm_processor_kwargs,
+        )
         dummy_mm_items = self.info.parse_mm_data(dummy_mm_data, validate=False)
 
         tokenization_kwargs = {"truncation": False}
@@ -102,6 +111,7 @@ class BaseDummyInputsBuilder(ABC, Generic[_I]):
         return ProcessorInputs(
             prompt=dummy_text,
             mm_items=dummy_mm_items,
+            hf_processor_mm_kwargs=mm_processor_kwargs or {},
             tokenization_kwargs=tokenization_kwargs,
         )
 

vllm/multimodal/registry.py

@@ -257,10 +257,12 @@ class MultiModalRegistry:
         if processor is None:
             processor = self.create_processor(model_config, cache=cache)
 
+        mm_config = model_config.get_multimodal_config()
         processor_inputs = processor.dummy_inputs.get_dummy_processor_inputs(
             seq_len=seq_len,
             mm_counts=mm_counts,
             mm_options=self._extract_mm_options(model_config),
+            mm_processor_kwargs=mm_config.mm_processor_kwargs,
         )
         mm_inputs = processor.apply(
             prompt=processor_inputs.prompt,