[VLM] Avoid unnecessary dummy multimodal data during processing (#16416)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/aria.py

@@ -21,12 +21,13 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
 from vllm.model_executor.model_loader.weight_utils import (
     default_weight_loader, maybe_remap_kv_scale_name)
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import MultiModalDataItems
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 # yapf: disable
@@ -415,31 +416,31 @@ class AriaProcessingInfo(BaseProcessingInfo):
 
 class AriaDummyInputsBuilder(BaseDummyInputsBuilder[AriaProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token: str = processor.tokenizer.image_token  # type: ignore
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         vision_config = self.info.get_vision_config()
 
         max_image_size = vision_config.image_size
         num_images = mm_counts.get("image", 0)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=max_image_size,
                                    height=max_image_size,
                                    num_images=num_images)
         }
 
-        hf_processor = self.info.get_hf_processor()
-        image_token: str = hf_processor.tokenizer.image_token  # type: ignore
-
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 class AriaMultiModalProcessor(BaseMultiModalProcessor[AriaProcessingInfo]):
 

vllm/model_executor/models/aya_vision.py

@@ -20,7 +20,7 @@ from vllm.jsontree import json_map_leaves
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalKwargs
+from vllm.multimodal.inputs import MultiModalDataDict, MultiModalKwargs
 from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
                                    MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
@@ -28,7 +28,7 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         MultiModalFieldConfig,
                                         PromptReplacement, PromptUpdate,
                                         PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
@@ -146,28 +146,29 @@ class AyaVisionProcessingInfo(BaseProcessingInfo):
 class AyaVisionDummyInputsBuilder(
         BaseDummyInputsBuilder[AyaVisionProcessingInfo]):
 
-    def get_dummy_processor_inputs(
-        self,
-        seq_len: int,
-        mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
         processor = self.info.get_hf_processor()
         image_token = processor.image_token
 
+        return image_token * num_images
+
+    def get_dummy_mm_data(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
         image_size = \
             self.info.get_image_size_with_most_features()
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=image_size.width,
                                    height=image_size.height,
                                    num_images=num_images)
         }
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
 
 
 class AyaVisionMultiModalProcessor(

vllm/model_executor/models/blip2.py

@@ -15,12 +15,13 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import MultiModalDataItems
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptIndexTargets,
                                         PromptInsertion, PromptUpdate)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .blip import BlipVisionModel
@@ -413,29 +414,27 @@ class Blip2ProcessingInfo(BaseProcessingInfo):
 
 class Blip2DummyInputsBuilder(BaseDummyInputsBuilder[Blip2ProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        return ""
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         hf_config = self.info.get_hf_config()
         vision_config = hf_config.vision_config
 
         max_image_size = vision_config.image_size
         num_images = mm_counts.get("image", 0)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=max_image_size,
                                    height=max_image_size,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text="",
-            mm_data=mm_data,
-        )
-
 
 class Blip2MultiModalProcessor(BaseMultiModalProcessor[Blip2ProcessingInfo]):
 

vllm/model_executor/models/chameleon.py

@@ -30,12 +30,13 @@ from vllm.model_executor.model_loader.weight_utils import (
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import MultiModalDataItems
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate, PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import (MultiModalEmbeddings, SupportsMultiModal, SupportsPP,
@@ -72,28 +73,31 @@ class ChameleonProcessingInfo(BaseProcessingInfo):
 class ChameleonDummyInputsBuilder(
         BaseDummyInputsBuilder[ChameleonProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token = processor.image_token
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         config = self.info.get_hf_config()
 
         width = height = config.vq_config.resolution
         num_images = mm_counts.get("image", 0)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=width,
                                    height=height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text="<image>" * num_images,
-            mm_data=mm_data,
-        )
-
 
 class ChameleonMultiModalProcessor(
         BaseMultiModalProcessor[ChameleonProcessingInfo]):

vllm/model_executor/models/deepseek_vl2.py

@@ -19,14 +19,14 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.model_loader.utils import set_default_torch_dtype
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
-                                    NestedTensors)
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs, NestedTensors)
 from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
                                    ImageSize, MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs.deepseek_vl2 import (DeepseekVLV2Config,
                                                           MlpProjectorConfig,
@@ -172,29 +172,30 @@ class DeepseekVL2ProcessingInfo(BaseProcessingInfo):
 class DeepseekVL2DummyInputsBuilder(
         BaseDummyInputsBuilder[DeepseekVL2ProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token = processor.image_token
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
-        hf_processor = self.info.get_hf_processor()
-        image_token: str = hf_processor.image_token
 
         max_image_size = self.info.get_image_size_with_most_features()
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=max_image_size.width,
                                    height=max_image_size.height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 class DeepseekVL2MultiModalProcessor(
         BaseMultiModalProcessor[DeepseekVL2ProcessingInfo]):

vllm/model_executor/models/florence2.py

@@ -21,13 +21,14 @@ from vllm.model_executor.models.bart import (BartDecoder, BartEncoder,
                                              BartScaledWordEmbedding)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
-from vllm.multimodal.parse import MultiModalDataDict, MultiModalDataItems
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
+from vllm.multimodal.parse import MultiModalDataItems
 from vllm.multimodal.processing import (BaseProcessingInfo,
                                         EncDecMultiModalProcessor,
                                         PromptIndexTargets, PromptInsertion,
                                         PromptUpdate)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import (MultiModalEmbeddings, SupportsMultiModal,
@@ -772,27 +773,25 @@ class Florence2ProcessingInfo(BaseProcessingInfo):
 class Florence2DummyInputsBuilder(
         BaseDummyInputsBuilder[Florence2ProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        return ""
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 
         target_width = target_height = self.info.get_hf_config().projection_dim
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text="",
-            mm_data=mm_data,
-        )
-
 
 class Florence2MultiModalProcessor(
         EncDecMultiModalProcessor[Florence2ProcessingInfo]):

vllm/model_executor/models/fuyu.py

@@ -31,13 +31,14 @@ from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.model_executor.models.persimmon import PersimmonForCausalLM
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
                                    MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate, PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
@@ -125,27 +126,25 @@ class FuyuProcessingInfo(BaseProcessingInfo):
 
 class FuyuDummyInputsBuilder(BaseDummyInputsBuilder[FuyuProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        return ""
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         target_width, target_height = \
             self.info.get_image_size_with_most_features()
         num_images = mm_counts.get("image", 0)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text="",
-            mm_data=mm_data,
-        )
-
 
 class FuyuMultiModalProcessor(BaseMultiModalProcessor[FuyuProcessingInfo]):
 

vllm/model_executor/models/gemma3_mm.py

@@ -15,8 +15,9 @@ from vllm.model_executor.layers.layernorm import GemmaRMSNorm
 from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
-from vllm.multimodal.inputs import MultiModalFieldConfig
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
                                    MultiModalDataItems)
 # yapf: disable
@@ -28,7 +29,7 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         find_mm_placeholders,
                                         replace_token_matches)
 # yapf: enable
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
@@ -224,31 +225,31 @@ class Gemma3ProcessingInfo(BaseProcessingInfo):
 
 class Gemma3DummyInputsBuilder(BaseDummyInputsBuilder[Gemma3ProcessingInfo]):
 
-    def get_dummy_processor_inputs(
-        self,
-        seq_len: int,
-        mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
         processor = self.info.get_hf_processor()
         image_token = processor.boi_token
 
+        return image_token * num_images
+
+    def get_dummy_mm_data(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 
         target_width, target_height = \
             self.info.get_image_size_with_most_features()
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 class Gemma3MultiModalProcessor(BaseMultiModalProcessor[Gemma3ProcessingInfo]):
 

vllm/model_executor/models/glm4v.py

@@ -12,7 +12,7 @@ from torch import nn
 from torch.nn import LayerNorm
 from torchvision import transforms
 from torchvision.transforms import InterpolationMode
-from transformers import PreTrainedTokenizer, TensorType
+from transformers import BatchFeature, PreTrainedTokenizer, TensorType
 from transformers.image_utils import ImageInput
 from transformers.tokenization_utils_base import TextInput
 
@@ -28,13 +28,13 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import MultiModalDataItems
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
-                                        BaseProcessingInfo, BatchFeature,
-                                        MultiModalFieldConfig,
-                                        PromptReplacement, PromptUpdate)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+                                        BaseProcessingInfo, PromptReplacement,
+                                        PromptUpdate)
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import ChatGLMConfig
 
@@ -447,31 +447,31 @@ class GLM4VProcessingInfo(BaseProcessingInfo):
 
 class GLM4VDummyInputsBuilder(BaseDummyInputsBuilder[GLM4VProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        base_text = "<|begin_of_image|><|endoftext|><|end_of_image|>"
+
+        return base_text * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         hf_config = self.info.get_hf_config()
         vision_config = hf_config.vision_config
 
         target_width = target_height = vision_config["image_size"]
         num_images = mm_counts.get("image", 0)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        base_text = "<|begin_of_image|><|endoftext|><|end_of_image|>"
-
-        return ProcessorInputs(
-            prompt_text=base_text * num_images,
-            mm_data=mm_data,
-        )
-
 
 class GLM4VMultiModalProcessor(BaseMultiModalProcessor[GLM4VProcessingInfo]):
 

vllm/model_executor/models/idefics3.py

@@ -32,18 +32,18 @@ from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
 from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import ImageProcessorItems, ImageSize
 # yapf conflicts with isort for this block
 # yapf: disable
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo,
-                                        MultiModalDataItems,
-                                        MultiModalFieldConfig,
-                                        PromptReplacement, PromptUpdate,
-                                        PromptUpdateDetails)
+                                        MultiModalDataItems, PromptReplacement,
+                                        PromptUpdate, PromptUpdateDetails)
 # yapf: enable
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 # yapf: disable
@@ -284,29 +284,31 @@ class Idefics3ProcessingInfo(BaseProcessingInfo):
 class Idefics3DummyInputsBuilder(BaseDummyInputsBuilder[Idefics3ProcessingInfo]
                                  ):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token, _, _ = self.info._get_image_token(processor)
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
         hf_processor = self.info.get_hf_processor()
         image_processor: Idefics3ImageProcessor = hf_processor.image_processor
         longest_edge = image_processor.max_image_size['longest_edge']
-        image_token, _, _ = self.info._get_image_token(hf_processor)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=longest_edge,
                                    height=longest_edge,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 class Idefics3MultiModalProcessor(
         BaseMultiModalProcessor[Idefics3ProcessingInfo]):

vllm/model_executor/models/internvl.py

@@ -25,14 +25,14 @@ from vllm.model_executor.models.intern_vit import (InternVisionModel,
                                                    InternVisionPatchModel)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
-                                    NestedTensors)
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs, NestedTensors)
 from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
                                    ImageSize, MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate, PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.tokenizer import AnyTokenizer
 
@@ -504,27 +504,27 @@ _I = TypeVar("_I", bound=BaseInternVLProcessingInfo)
 
 class InternVLDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        return "<image>" * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         target_width, target_height = \
             self.info.get_image_size_with_most_features()
         num_images = mm_counts.get("image", 0)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text="<image>" * num_images,
-            mm_data=mm_data,
-        )
-
 
 class InternVLMultiModalProcessor(BaseMultiModalProcessor[_I]):
 

vllm/model_executor/models/llava.py

@@ -34,7 +34,7 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, ProcessingCache,
                                         PromptReplacement, PromptUpdate,
                                         PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .clip import CLIPVisionModel
@@ -186,30 +186,31 @@ _I = TypeVar("_I", bound=BaseLlavaProcessingInfo)
 
 class LlavaDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token = processor.image_token
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 
-        processor = self.info.get_hf_processor()
-        image_token = processor.image_token
         target_width, target_height = \
             self.info.get_image_size_with_most_features()
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 class LlavaProcessingInfo(BaseLlavaProcessingInfo):
 

vllm/model_executor/models/llava_next_video.py

@@ -16,13 +16,14 @@ from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.models.clip import CLIPVisionModel
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import (ImageSize, MultiModalDataItems,
                                    VideoEmbeddingItems, VideoProcessorItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 from vllm.utils import is_list_of
 
@@ -130,22 +131,27 @@ class LlavaNextVideoProcessingInfo(BaseProcessingInfo):
 class LlavaNextVideoDummyInputsBuilder(
         BaseDummyInputsBuilder[LlavaNextVideoProcessingInfo]):
 
-    def get_dummy_processor_inputs(
-        self,
-        seq_len: int,
-        mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
         num_videos = mm_counts.get("video", 0)
 
         processor = self.info.get_hf_processor()
         video_token = processor.video_token
 
+        return video_token * num_videos
+
+    def get_dummy_mm_data(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> MultiModalDataDict:
+        num_videos = mm_counts.get("video", 0)
+
         target_width, target_height = \
             self.info.get_image_size_with_most_features()
         target_num_frames = \
             self.info.get_num_frames_with_most_features(seq_len, mm_counts)
 
-        mm_data = {
+        return {
             "video":
             self._get_dummy_videos(
                 width=target_width,
@@ -155,11 +161,6 @@ class LlavaNextVideoDummyInputsBuilder(
             )
         }
 
-        return ProcessorInputs(
-            prompt_text=video_token * num_videos,
-            mm_data=mm_data,
-        )
-
 
 class LlavaNextVideoMultiModalProcessor(
         BaseMultiModalProcessor[LlavaNextVideoProcessingInfo]):

vllm/model_executor/models/llava_onevision.py

@@ -19,11 +19,11 @@ from vllm.model_executor.layers.activation import get_act_fn
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import (ImageSize, MultiModalDataItems,
                                    VideoEmbeddingItems, VideoProcessorItems)
 from vllm.multimodal.processing import PromptReplacement, PromptUpdate
-from vllm.multimodal.profiling import ProcessorInputs
 from vllm.sequence import IntermediateTensors
 
 from .clip import CLIPVisionModel
@@ -226,11 +226,7 @@ class LlavaOnevisionProcessingInfo(LlavaNextProcessingInfo):
 class LlavaOnevisionDummyInputsBuilder(
         LlavaDummyInputsBuilder[LlavaOnevisionProcessingInfo]):
 
-    def get_dummy_processor_inputs(
-        self,
-        seq_len: int,
-        mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
         num_images = mm_counts.get("image", 0)
         num_videos = mm_counts.get("video", 0)
 
@@ -238,13 +234,23 @@ class LlavaOnevisionDummyInputsBuilder(
         image_token = processor.image_token
         video_token = processor.video_token
 
+        return image_token * num_images + video_token * num_videos
+
+    def get_dummy_mm_data(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> 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()
         target_num_frames = \
             self.info.get_num_frames_with_most_features(seq_len,
                                                         mm_counts)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
@@ -258,11 +264,6 @@ class LlavaOnevisionDummyInputsBuilder(
             )
         }
 
-        return ProcessorInputs(
-            prompt_text=image_token * num_images + video_token * num_videos,
-            mm_data=mm_data,
-        )
-
 
 class LlavaOnevisionMultiModalProcessor(
         BaseLlavaNextMultiModalProcessor[LlavaOnevisionProcessingInfo]):

vllm/model_executor/models/minicpmo.py

@@ -35,14 +35,14 @@ from transformers.models.whisper.modeling_whisper import (
 
 from vllm.config import VllmConfig
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
-from vllm.multimodal.inputs import MultiModalFieldConfig, NestedTensors
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    NestedTensors)
 from vllm.multimodal.parse import (AudioItem, AudioProcessorItems,
                                    DictEmbeddingItems, ModalityData,
                                    ModalityDataItems, MultiModalDataItems,
                                    MultiModalDataParser)
 from vllm.multimodal.processing import (PromptReplacement, PromptUpdate,
                                         PromptUpdateDetails)
-from vllm.multimodal.profiling import ProcessorInputs
 
 from .minicpmv import (_MAX_FRAMES_PER_VIDEO, MiniCPMV2_6,
                        MiniCPMVDummyInputsBuilder,
@@ -206,29 +206,31 @@ class MiniCPMOProcessingInfo(MiniCPMVProcessingInfo):
 class MiniCPMODummyInputsBuilder(
         MiniCPMVDummyInputsBuilder[MiniCPMOProcessingInfo]):
 
-    def get_dummy_processor_inputs(
-            self, seq_len: int, mm_counts: Mapping[str,
-                                                   int]) -> ProcessorInputs:
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_audios = mm_counts.get("audio", 0)
+
+        audio_prompt_texts = self.info.audio_pattern * num_audios
+
+        return super().get_dummy_text(mm_counts) + audio_prompt_texts
+
+    def get_dummy_mm_data(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> MultiModalDataDict:
         num_audios = mm_counts.get("audio", 0)
         audio_len = self.info.get_max_audio_chunks_with_most_features() * \
             self.info.get_default_audio_sampling_rate()
 
-        processor_inputs = super().get_dummy_processor_inputs(
-            seq_len, mm_counts)
-
-        audio_prompt_texts = self.info.audio_pattern * num_audios
         audio_mm_data = {
             "audio":
             self._get_dummy_audios(length=audio_len, num_audios=num_audios)
         }
 
-        return ProcessorInputs(
-            prompt_text=processor_inputs.prompt_text + audio_prompt_texts,
-            mm_data={
-                **processor_inputs.mm_data,
-                **audio_mm_data,
-            },
-        )
+        return {
+            **super().get_dummy_mm_data(seq_len, mm_counts),
+            **audio_mm_data,
+        }
 
 
 class MiniCPMOMultiModalProcessor(

vllm/model_executor/models/minicpmv.py

@@ -48,7 +48,8 @@ from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.model_executor.models.qwen2 import Qwen2ForCausalLM
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
-from vllm.multimodal.inputs import MultiModalFieldConfig, NestedTensors
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    NestedTensors)
 from vllm.multimodal.parse import (DictEmbeddingItems, ImageItem,
                                    ImageProcessorItems, ImageSize,
                                    ModalityData, ModalityDataItems,
@@ -57,7 +58,7 @@ from vllm.multimodal.parse import (DictEmbeddingItems, ImageItem,
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate, PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
 from vllm.utils import flatten_2d_lists
@@ -471,11 +472,20 @@ _I = TypeVar("_I",
 
 class MiniCPMVDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+        num_videos = mm_counts.get("video", 0)
+
+        image_prompt_texts = self.info.image_pattern * num_images
+        video_prompt_texts = self.info.video_pattern * num_videos
+
+        return image_prompt_texts + video_prompt_texts
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
         num_videos = mm_counts.get("video", 0)
 
@@ -486,7 +496,7 @@ class MiniCPMVDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
         num_video_frames = \
             self.info.get_num_frames_with_most_features(seq_len, mm_counts)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=image_width,
                                    height=image_height,
@@ -498,13 +508,6 @@ class MiniCPMVDummyInputsBuilder(BaseDummyInputsBuilder[_I]):
             ] * num_videos,
         }
 
-        image_prompt_texts = self.info.image_pattern * num_images
-        video_prompt_texts = self.info.video_pattern * num_videos
-
-        return ProcessorInputs(prompt_text=image_prompt_texts +
-                               video_prompt_texts,
-                               mm_data=mm_data)
-
 
 class MiniCPMVMultiModalProcessor(BaseMultiModalProcessor[_I]):
 

vllm/model_executor/models/mistral3.py

@@ -22,14 +22,15 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
                                    MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, ProcessingCache,
                                         PromptReplacement, PromptUpdate,
                                         PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
@@ -185,30 +186,31 @@ _I = TypeVar("_I", bound=BaseLlavaProcessingInfo)
 
 class Mistral3DummyInputsBuilder(BaseDummyInputsBuilder[_I]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token = processor.image_token
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 
-        processor = self.info.get_hf_processor()
-        image_token = processor.image_token
         target_width, target_height = \
             self.info.get_image_size_with_most_features()
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 class Mistral3ProcessingInfo(BaseLlavaProcessingInfo):
 

vllm/model_executor/models/mllama.py

@@ -54,14 +54,14 @@ from vllm.model_executor.model_loader.weight_utils import (
     default_weight_loader, maybe_remap_kv_scale_name)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import (MultiModalEncDecInputs,
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalEncDecInputs,
                                     MultiModalFieldConfig, MultiModalKwargs)
 from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
-                                   MultiModalDataDict, MultiModalDataItems)
+                                   MultiModalDataItems)
 from vllm.multimodal.processing import (BaseProcessingInfo,
                                         EncDecMultiModalProcessor,
                                         PromptReplacement, PromptUpdate)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 
 from .clip import CLIPMLP
 from .interfaces import SupportsMultiModal, SupportsV0Only
@@ -131,31 +131,31 @@ class MllamaProcessingInfo(BaseProcessingInfo):
 
 class MllamaDummyInputsBuilder(BaseDummyInputsBuilder[MllamaProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token = processor.image_token
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 
         target_width, target_height = \
             self.info.get_image_size_with_most_features()
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        hf_processor = self.info.get_hf_processor()
-        image_token: str = hf_processor.image_token
-
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 class MllamaMultiModalProcessor(EncDecMultiModalProcessor[MllamaProcessingInfo]
                                 ):

vllm/model_executor/models/mllama4.py

@@ -43,14 +43,14 @@ from vllm.model_executor.model_loader.loader import _initialize_model
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
-                                    NestedTensors)
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs, NestedTensors)
 from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
                                    MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate, PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
@@ -619,29 +619,31 @@ class Mllama4MultiModalProcessor(BaseMultiModalProcessor[Mllama4ProcessingInfo]
 
 class Mllama4DummyInputsBuilder(BaseDummyInputsBuilder[Mllama4ProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token = processor.fake_image_token
+
+        return image_token * num_images
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 
         (target_width,
          target_height) = self.info.get_image_size_with_most_features()
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        image_token = self.info.get_hf_processor().fake_image_token
-        return ProcessorInputs(
-            prompt_text=image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 @MULTIMODAL_REGISTRY.register_processor(
     Mllama4MultiModalProcessor,

vllm/model_executor/models/molmo.py

@@ -41,14 +41,15 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs)
 from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
                                    MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptIndexTargets,
                                         PromptInsertion, PromptUpdate,
                                         PromptUpdateDetails)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
@@ -1216,27 +1217,25 @@ class MolmoProcessingInfo(BaseProcessingInfo):
 
 class MolmoDummyInputsBuilder(BaseDummyInputsBuilder[MolmoProcessingInfo]):
 
-    def get_dummy_processor_inputs(
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        return ""
+
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         target_width, target_height = \
             self.info.get_image_size_with_most_features()
         num_images = mm_counts.get("image", 0)
 
-        mm_data = {
+        return {
             "image":
             self._get_dummy_images(width=target_width,
                                    height=target_height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text="",
-            mm_data=mm_data,
-        )
-
 
 class MolmoMultiModalProcessor(BaseMultiModalProcessor[MolmoProcessingInfo]):