Handle tensor outputs from Qwen decoder layers

src/model_adapters.py

@@ -107,14 +107,16 @@ class LlamaStyleAdapter(ModelAdapter):
             hidden_states,
             attention_mask=attention_mask,
             position_ids=position_ids,
-            past_key_values=past_key_values,
+            past_key_value=past_key_values,
             use_cache=use_cache,
             position_embeddings=position_embeddings,
             cache_position=cache_position,
         )
-        hidden_states = layer_outputs[0]
-        new_kv = layer_outputs[1] if len(layer_outputs) > 1 else None
-        return hidden_states, new_kv
+        # Recent Transformers decoder layers return a tensor directly. Indexing
+        # it like a tuple would drop the batch dimension after the first layer.
+        if isinstance(layer_outputs, torch.Tensor):
+            return layer_outputs, None
+        return layer_outputs[0], layer_outputs[1] if len(layer_outputs) > 1 else None
 
     def apply_final_norm(self, hidden_states: Tensor) -> Tensor:
         return self._norm(hidden_states)