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

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

vllm/model_executor/models/blip2.py

@@ -16,6 +16,7 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.inputs import NestedTensors
 from vllm.multimodal.utils import consecutive_placeholder_ranges
 from vllm.sequence import IntermediateTensors, SequenceData
 
@@ -609,6 +610,25 @@ class Blip2ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
 
         return self.language_projection(query_output)
 
+    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,
+                BLIP2_IMAGE_TOKEN_ID)
+        return inputs_embeds
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -616,6 +636,7 @@ class Blip2ForConditionalGeneration(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,
     ) -> Union[SamplerOutput, IntermediateTensors]:
         """Run forward pass for BLIP-2.
@@ -648,32 +669,24 @@ class Blip2ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
         See also:
             :class:`Blip2ImageInputs`
         """
+
         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.get_input_embeddings(
-                    input_ids)
+        # 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
 
-                inputs_embeds = merge_multimodal_embeddings(
-                    input_ids, inputs_embeds, vision_embeddings,
-                    BLIP2_IMAGE_TOKEN_ID)
-
-                input_ids = None
-            else:
-                inputs_embeds = None
-
-        hidden_states = self.language_model.model(
-            input_ids,
-            positions,
-            kv_caches,
-            attn_metadata,
-            intermediate_tensors=intermediate_tensors,
-            inputs_embeds=inputs_embeds)
+        hidden_states = self.language_model.model(input_ids,
+                                                  positions,
+                                                  kv_caches,
+                                                  attn_metadata,
+                                                  intermediate_tensors,
+                                                  inputs_embeds=inputs_embeds)
 
         return hidden_states