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

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

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]):