[`UltraVox`] Fix output type (#37224)

Signed-off-by: vasqu <antonprogamer@gmail.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

vllm/model_executor/models/ultravox.py

@@ -404,12 +404,14 @@ class UltravoxTransformerProjector(nn.Module, ModuleUtilsMixin):
             kwargs["layer_head_mask"] = None
 
         for layer in self.layers:
-            layer_outputs = layer(
+            hidden_states = layer(
                 hidden_states,
                 attention_mask=extended_attention_mask,
                 **kwargs,
             )
-            hidden_states = layer_outputs[0]
+            # BC version that allows for the old tupled output
+            if isinstance(hidden_states, tuple):
+                hidden_states = hidden_states[0]
 
         hidden_states = self.ln_post(hidden_states)
         hidden_states = self.linear_out(hidden_states)
@@ -509,13 +511,14 @@ class ModifiedWhisperEncoder(WhisperEncoder):
             kwargs["layer_head_mask"] = None
 
         for encoder_layer in self.layers:
-            layer_outputs = encoder_layer(
+            hidden_states = encoder_layer(
                 hidden_states,
                 attention_mask,
                 **kwargs,
             )
-
-            hidden_states = layer_outputs[0]
+            # BC version that allows for the old tupled output
+            if isinstance(hidden_states, tuple):
+                hidden_states = hidden_states[0]
 
         hidden_states = self.layer_norm(hidden_states)
         return hidden_states

vllm/transformers_utils/configs/ultravox.py

@@ -43,7 +43,6 @@ class UltravoxConfig(transformers.PretrainedConfig):
             use `False`, but v0.5 and above use `True`.
     """
 
-    wrapped_model_config: transformers.PretrainedConfig
     model_type = "ultravox"
     audio_token = "<|audio|>"
     is_composition = False
@@ -75,6 +74,7 @@ class UltravoxConfig(transformers.PretrainedConfig):
         self.num_projector_layers = num_projector_layers
 
         # N.B. May set the wrapped_model_config below.
+        self.wrapped_model_config: transformers.PretrainedConfig
         self.text_model_id = text_model_id
         if text_model_id is None:
             text_config = text_config or {}