Fix: Correct layer hook output format for layer_norm compatibility

- Fixed "layer_norm() argument 'input' must be Tensor, not tuple" error
- Hook now properly handles different input/output formats
- Maintains compatibility with subsequent layers expecting tensors

The previous fix was returning tuples when tensors were expected,
breaking layer_norm operations after skipped layers.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

backend/model_service.py

@@ -298,17 +298,23 @@ class ModelManager:
             
             def create_layer_hook():
                 def hook(module, input, output):
-                    # Pass through input unchanged (skip layer)
-                    # Handle different input formats
-                    if isinstance(input, tuple) and len(input) > 0:
-                        input_tensor = input[0]
-                    else:
-                        input_tensor = input
+                    # Skip layer by passing through input unchanged
+                    # The input to a transformer layer is (hidden_states, optional_attention_mask, ...)
+                    # The output is (hidden_states, optional_attention_weights, ...)
+                    # We want to pass the input hidden states as if the layer did nothing
                     
-                    # Return input with same format as output
-                    if isinstance(output, tuple):
-                        return (input_tensor,) + output[1:]
-                    return input_tensor
+                    if isinstance(input, tuple):
+                        # Return input hidden states, but keep any additional outputs from the layer
+                        # This maintains compatibility with the expected output format
+                        if isinstance(output, tuple) and len(output) > 1:
+                            # Keep attention weights etc. from output if present
+                            return (input[0],) + output[1:]
+                        else:
+                            # Just return the input hidden states
+                            return input[0]
+                    else:
+                        # Simple tensor input/output
+                        return input
                 return hook
             
             # Apply hooks and log what's being disabled