[Misc] Clean up type annotation for SupportsMultiModal (#14794)

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

vllm/model_executor/models/phi3v.py

@@ -31,8 +31,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
-                                    NestedTensors)
+from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargs
 from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
                                    ImageSize, MultiModalDataItems)
 # yapf conflicts with isort for this block
@@ -48,7 +47,8 @@ from vllm.sequence import IntermediateTensors
 from vllm.utils import is_list_of
 
 from .clip import CLIPVisionModel
-from .interfaces import SupportsMultiModal, SupportsPP, SupportsQuant
+from .interfaces import (MultiModalEmbeddings, SupportsMultiModal, SupportsPP,
+                         SupportsQuant)
 from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
                     init_vllm_registered_model, maybe_prefix,
                     merge_multimodal_embeddings)
@@ -649,8 +649,7 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP,
         return image_embeds
 
     def get_multimodal_embeddings(
-        self, **kwargs
-    ) -> Union[list[torch.Tensor], torch.Tensor, tuple[torch.Tensor, ...]]:
+            self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
         image_input = self._parse_and_validate_image_input(**kwargs)
         if image_input is None:
             return None
@@ -660,7 +659,7 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP,
     def get_input_embeddings(
         self,
         input_ids: torch.Tensor,
-        multimodal_embeddings: Optional[NestedTensors] = None,
+        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
     ) -> torch.Tensor:
         inputs_embeds = self.embed_tokens(input_ids)
         if multimodal_embeddings is not None: