[V1] Refactor model executable interface for multimodal models (#10570)

Signed-off-by: Roger Wang <ywang@roblox.com>

vllm/model_executor/models/fuyu.py

@@ -35,6 +35,7 @@ from vllm.model_executor.models.persimmon import PersimmonForCausalLM
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
 from vllm.multimodal.image import cached_get_image_processor
+from vllm.multimodal.inputs import NestedTensors
 from vllm.multimodal.utils import (cached_get_tokenizer,
                                    consecutive_placeholder_ranges)
 from vllm.sequence import (VLLM_TOKEN_ID_ARRAY_TYPE, IntermediateTensors,
@@ -302,6 +303,25 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
         vision_embeddings, _ = self.vision_embed_tokens(image_input["data"])
         return vision_embeddings
 
+    def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
+        image_input = self._parse_and_validate_image_input(**kwargs)
+        if image_input is None:
+            return None
+        vision_embeddings = self._process_image_input(image_input)
+        return vision_embeddings
+
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: Optional[NestedTensors] = None,
+    ) -> torch.Tensor:
+        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
+        if multimodal_embeddings is not None:
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids, inputs_embeds, multimodal_embeddings,
+                _IMAGE_TOKEN_ID)
+        return inputs_embeds
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -309,24 +329,19 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
         kv_caches: List[torch.Tensor],
         attn_metadata: AttentionMetadata,
         intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
         **kwargs: object,
     ):
         if intermediate_tensors is not None:
-            input_ids = None
             inputs_embeds = None
-        else:
-            image_input = self._parse_and_validate_image_input(**kwargs)
 
-            if image_input is not None:
-                vision_embeddings = self._process_image_input(image_input)
-                inputs_embeds = self.language_model.model.embed_tokens(
-                    input_ids)
-                inputs_embeds = merge_multimodal_embeddings(
-                    input_ids, inputs_embeds, vision_embeddings,
-                    self.image_token_id)
-
-            else:
-                inputs_embeds = None
+        # NOTE: In v1, inputs_embeds is always generated at model runner, this
+        # condition is for v0 compatibility.
+        elif inputs_embeds is None:
+            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
+            inputs_embeds = self.get_input_embeddings(input_ids,
+                                                      vision_embeddings)
+            input_ids = None
 
         hidden_states = self.language_model(
             input_ids=input_ids,