[Optim] Compute multimodal hash only once per item (#17314)

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

vllm/model_executor/models/deepseek_vl2.py

@@ -22,8 +22,8 @@ from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
 from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
                                    ImageSize, MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
-                                        BaseProcessingInfo, PromptReplacement,
-                                        PromptUpdate)
+                                        BaseProcessingInfo, MultiModalHashes,
+                                        PromptReplacement, PromptUpdate)
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs.deepseek_vl2 import (DeepseekVLV2Config,
@@ -279,24 +279,26 @@ class DeepseekVL2MultiModalProcessor(
         prompt: Union[str, list[int]],
         mm_data_items: MultiModalDataItems,
         hf_processor_mm_kwargs: Mapping[str, object],
-    ) -> tuple[list[int], MultiModalKwargs, bool]:
+        *,
+        return_mm_hashes: bool,
+    ) -> tuple[list[int], MultiModalKwargs, Optional[MultiModalHashes], bool]:
         # The processor logic is different for len(images) <= 2 vs > 2
         # Since the processing cache assumes that the processor output is
         # invariant of how many images are passed per prompt, we only
         # perform caching for the most common case
         if mm_data_items.get_count("image", strict=False) > 2:
-            # This code path corresponds to the cache being disabled
-            return self._apply_hf_processor_main(
+            return self._apply_hf_processor(
                 prompt=prompt,
-                mm_items=mm_data_items,
+                mm_data_items=mm_data_items,
                 hf_processor_mm_kwargs=hf_processor_mm_kwargs,
-                enable_hf_prompt_update=True,
+                return_mm_hashes=return_mm_hashes,
             )
 
         return super()._cached_apply_hf_processor(
             prompt=prompt,
             mm_data_items=mm_data_items,
             hf_processor_mm_kwargs=hf_processor_mm_kwargs,
+            return_mm_hashes=return_mm_hashes,
         )