replace skip_embed with input_embeds

python/sglang/srt/models/llama2.py

@@ -227,12 +227,12 @@ def forward(
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         input_metadata: InputMetadata,
-        skip_embed: bool = False,
+        input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
-        if not skip_embed:
+        if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)
         else:
-            hidden_states = input_ids
+            hidden_states = input_embeds
         residual = None
         for i in range(len(self.layers)):
             layer = self.layers[i]
@@ -264,9 +264,9 @@ def forward(
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         input_metadata: InputMetadata,
-        skip_embed: bool = False,
+        input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, input_metadata, skip_embed)
+        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
         return self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )

python/sglang/srt/models/llava.py

@@ -230,11 +230,11 @@ def forward(
                     pt += 1
 
             return self.language_model(
-                input_embeds, positions, input_metadata, skip_embed=True
+                input_ids, positions, input_metadata, input_embeds=input_embeds
             )
         elif input_metadata.forward_mode == ForwardMode.DECODE:
             return self.language_model(
-                input_ids, positions, input_metadata, skip_embed=False
+                input_ids, positions, input_metadata
             )
 
     def load_weights(

python/sglang/srt/models/mixtral.py

@@ -296,12 +296,12 @@ def forward(
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         input_metadata: InputMetadata,
-        skip_embed: bool = False,
+        input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
-        if not skip_embed:
+        if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)
         else:
-            hidden_states = input_ids
+            hidden_states = input_embeds
         residual = None
         for i in range(len(self.layers)):
             layer = self.layers[i]
@@ -330,9 +330,9 @@ def forward(
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         input_metadata: InputMetadata,
-        skip_embed: bool = False,
+        input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, input_metadata, skip_embed)
+        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
         return self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )

python/sglang/srt/models/qwen2.py

@@ -228,12 +228,12 @@ def forward(
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         input_metadata: InputMetadata,
-        skip_embed: bool = False,
+        input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
-        if not skip_embed:
+        if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)
         else:
-            hidden_states = input_ids
+            hidden_states = input_embeds
         residual = None
         for i in range(len(self.layers)):
             layer = self.layers[i]
@@ -265,9 +265,9 @@ def forward(
         input_ids: torch.Tensor,
         positions: torch.Tensor,
         input_metadata: InputMetadata,
-        skip_embed: bool = False,
+        input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, input_metadata, skip_embed)
+        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
         return self.logits_processor(
             input_ids, hidden_states, self.lm_head.weight, input_metadata
         )