Fix research attention endpoint model compatibility

Make QKV hook registration more robust to support different model
architectures (CodeGen vs GPT-2 style attention modules)

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

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

backend/model_service.py

@@ -1327,10 +1327,20 @@ async def analyze_research_attention(request: Dict[str, Any], authenticated: boo
                 }
             return hook
 
-        # Register hooks on all qkv_proj modules
-        for layer_idx, layer in enumerate(manager.model.transformer.h):
-            hook = layer.attn.qkv_proj.register_forward_hook(make_qkv_hook(layer_idx))
-            hooks.append(hook)
+        # Register hooks on all qkv_proj modules (if available)
+        # This is model-specific - CodeGen uses different architecture
+        try:
+            if hasattr(manager.model, 'transformer') and hasattr(manager.model.transformer, 'h'):
+                for layer_idx, layer in enumerate(manager.model.transformer.h):
+                    if hasattr(layer, 'attn') and hasattr(layer.attn, 'qkv_proj'):
+                        hook = layer.attn.qkv_proj.register_forward_hook(make_qkv_hook(layer_idx))
+                        hooks.append(hook)
+                    elif hasattr(layer, 'attn') and hasattr(layer.attn, 'c_attn'):
+                        # GPT-2 style attention
+                        hook = layer.attn.c_attn.register_forward_hook(make_qkv_hook(layer_idx))
+                        hooks.append(hook)
+        except Exception as hook_error:
+            logger.warning(f"Could not register QKV hooks: {hook_error}")
 
         with torch.no_grad():
             current_ids = inputs["input_ids"]