Update modeling_llama3.py

modeling_llama3.py

@@ -548,20 +548,10 @@ class Llama3ForConditionalGeneration(Llama3PreTrainedModel, GenerationMixin):
         if cross_attention_mask is not None and cache_position is not None:
             cross_attention_mask = cross_attention_mask[:, :, cache_position]
             full_text_row_masked_out_mask = full_text_row_masked_out_mask[:, :, cache_position]
-
-        if audio_features is not None:
-            if input_ids is None:
-                raise ValueError("You must provide `input_ids` if you pass `audio_features`.")
-            
-            inputs_embeds = self.audio_model(
-                audio_features=audio_features,
-                input_ids=input_ids,
-                return_dict=False,
-            )
-            input_ids = None
             
         outputs = self.language_model(
             input_ids=input_ids,
+            audio_features=audio_features,
             attention_mask=attention_mask,
             position_ids=position_ids,
             cross_attention_states=cross_attention_states,