[Refactor] Define MultiModalKwargsItems separate from MultiModalKwargs (#23053)

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

vllm/model_executor/models/deepseek_vl2.py

@@ -21,7 +21,7 @@ from vllm.model_executor.model_loader.utils import set_default_torch_dtype
 from vllm.model_executor.models.transformers import replace_linear_class
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
-                                    MultiModalKwargs, NestedTensors)
+                                    MultiModalKwargsItems, NestedTensors)
 from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
                                    ImageSize, MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
@@ -252,7 +252,7 @@ class DeepseekVL2MultiModalProcessor(
         self,
         mm_items: MultiModalDataItems,
         hf_processor_mm_kwargs: Mapping[str, object],
-        out_mm_kwargs: MultiModalKwargs,
+        out_mm_kwargs: MultiModalKwargsItems,
     ) -> Sequence[PromptUpdate]:
         hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
 
@@ -291,7 +291,8 @@ class DeepseekVL2MultiModalProcessor(
         tokenization_kwargs: Mapping[str, object],
         *,
         return_mm_hashes: bool,
-    ) -> tuple[list[int], MultiModalKwargs, Optional[MultiModalHashes], bool]:
+    ) -> tuple[list[int], MultiModalKwargsItems, 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