[VLM] Avoid unnecessary dummy multimodal data during processing

vllm/model_executor/models/aria.py

@@ -21,12 +21,13 @@
 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 @@ def get_num_image_tokens(self) -> int:
 
 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

@@ -20,15 +20,15 @@
 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,
                                         BaseProcessingInfo,
                                         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 @@ def get_num_patches(self, *, image_width: int, image_height: int,
 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.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 @@ def get_num_image_tokens(self) -> int:
 
 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.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 @@ def get_num_image_tokens(self) -> int:
 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.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 @@ def get_image_size_with_most_features(self) -> ImageSize:
 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 @@
                                              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 @@ def get_num_image_tokens(self) -> int:
 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.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 @@ def get_image_size_with_most_features(self) -> ImageSize:
 
 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]):