[V1] Refactor model executable interface for all text-only language models (#10374)

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

vllm/model_executor/models/stablelm.py

@@ -218,6 +218,9 @@ class StableLMEpochModel(nn.Module):
             make_empty_intermediate_tensors_factory(["hidden_states"],
                                                     config.hidden_size))
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -225,9 +228,13 @@ class StableLMEpochModel(nn.Module):
         kv_caches: List[torch.Tensor],
         attn_metadata: AttentionMetadata,
         intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, IntermediateTensors]:
         if get_pp_group().is_first_rank:
-            hidden_states = self.embed_tokens(input_ids)
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
@@ -265,6 +272,9 @@ class StablelmForCausalLM(nn.Module, SupportsPP):
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -272,9 +282,11 @@ class StablelmForCausalLM(nn.Module, SupportsPP):
         kv_caches: List[torch.Tensor],
         attn_metadata: AttentionMetadata,
         intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, IntermediateTensors]:
         hidden_states = self.model(input_ids, positions, kv_caches,
-                                   attn_metadata, intermediate_tensors)
+                                   attn_metadata, intermediate_tensors,
+                                   inputs_embeds)
         return hidden_states
 
     def compute_logits(

vllm/model_executor/models/starcoder2.py

@@ -221,6 +221,9 @@ class Starcoder2Model(nn.Module):
             make_empty_intermediate_tensors_factory(["hidden_states"],
                                                     config.hidden_size))
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -228,9 +231,13 @@ class Starcoder2Model(nn.Module):
         kv_caches: List[torch.Tensor],
         attn_metadata: AttentionMetadata,
         intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, IntermediateTensors]:
         if get_pp_group().is_first_rank:
-            hidden_states = self.embed_tokens(input_ids)
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
@@ -273,6 +280,9 @@ class Starcoder2ForCausalLM(nn.Module, SupportsPP):
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -280,9 +290,11 @@ class Starcoder2ForCausalLM(nn.Module, SupportsPP):
         kv_caches: List[torch.Tensor],
         attn_metadata: AttentionMetadata,
         intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, IntermediateTensors]:
         hidden_states = self.model(input_ids, positions, kv_caches,
-                                   attn_metadata, intermediate_tensors)
+                                   attn_metadata, intermediate_tensors,
+                                   inputs_embeds)
         return hidden_states
 
     def compute_logits(

vllm/model_executor/models/xverse.py

@@ -252,6 +252,9 @@ class XverseModel(nn.Module):
             make_empty_intermediate_tensors_factory(
                 ["hidden_states", "residual"], config.hidden_size))
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -259,9 +262,13 @@ class XverseModel(nn.Module):
         kv_caches: List[torch.Tensor],
         attn_metadata: AttentionMetadata,
         intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, IntermediateTensors]:
         if get_pp_group().is_first_rank:
-            hidden_states = self.embed_tokens(input_ids)
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
             residual = None
         else:
             hidden_states = intermediate_tensors["hidden_states"]
@@ -335,6 +342,9 @@ class XverseForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -342,9 +352,11 @@ class XverseForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
         kv_caches: List[torch.Tensor],
         attn_metadata: AttentionMetadata,
         intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, IntermediateTensors]:
         hidden_states = self.model(input_ids, positions, kv_caches,
-                                   attn_metadata, intermediate_tensors)
+                                   attn_metadata, intermediate_tensors,
+                                   inputs_embeds)
         return hidden_states
 
     def compute_logits(