[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]):