Fix: Add routes directly in Blocks context instead of using load event

app.py

@@ -518,51 +518,46 @@ with gr.Blocks(
         outputs=[prompt_input, output],
     )
     
-    # Mount FastAPI routes onto Gradio's app using load event
-    # This must be done inside the Blocks context
-    def mount_fastapi_routes():
-        """Mount FastAPI routes onto Gradio's app after initialization."""
-        try:
-            from fastapi.responses import JSONResponse
-            from starlette.routing import Route
-            from starlette.requests import Request
-            
-            # Create route handlers that call the FastAPI endpoint functions
-            async def health_handler(request: Request):
-                """Handle GET /health requests."""
-                result = healthcheck()
-                return JSONResponse(content=result)
-            
-            async def generate_handler(request: Request):
-                """Handle POST /v1/generate requests."""
-                try:
-                    data = await request.json()
-                    payload = GeneratePayload(**data)
-                    # Call the FastAPI endpoint function directly
-                    result = generate_endpoint(payload)
-                    return JSONResponse(content={"text": result.text})
-                except Exception as exc:
-                    from fastapi import HTTPException
-                    raise HTTPException(status_code=500, detail=str(exc))
-            
-            async def console_handler(request: Request):
-                """Handle GET /console requests."""
-                html_content = interactive_ui()
-                return HTMLResponse(content=html_content)
-            
-            # Insert routes at the beginning of Gradio's router to ensure they're processed first
-            # This prevents Gradio's catch-all routes from intercepting our API endpoints
-            gradio_app.app.router.routes.insert(0, Route("/health", health_handler, methods=["GET"]))
-            gradio_app.app.router.routes.insert(0, Route("/v1/generate", generate_handler, methods=["POST"]))
-            gradio_app.app.router.routes.insert(0, Route("/console", console_handler, methods=["GET"]))
-            print(f"FastAPI routes mounted successfully on Gradio app. Total routes: {len(gradio_app.app.router.routes)}")
-        except Exception as e:
-            print(f"Warning: Could not mount FastAPI routes: {e}")
-            import traceback
-            traceback.print_exc()
-    
-    # Mount routes when Gradio app loads (must be inside Blocks context)
-    gradio_app.load(mount_fastapi_routes)
+    # Mount FastAPI routes directly onto Gradio's app (inside Blocks context)
+    # Add routes immediately so they're registered before Gradio processes requests
+    try:
+        from fastapi.responses import JSONResponse
+        from starlette.routing import Route
+        from starlette.requests import Request
+        
+        # Create route handlers that call the FastAPI endpoint functions
+        async def health_handler(request: Request):
+            """Handle GET /health requests."""
+            result = healthcheck()
+            return JSONResponse(content=result)
+        
+        async def generate_handler(request: Request):
+            """Handle POST /v1/generate requests."""
+            try:
+                data = await request.json()
+                payload = GeneratePayload(**data)
+                # Call the FastAPI endpoint function directly
+                result = generate_endpoint(payload)
+                return JSONResponse(content={"text": result.text})
+            except Exception as exc:
+                from fastapi import HTTPException
+                raise HTTPException(status_code=500, detail=str(exc))
+        
+        async def console_handler(request: Request):
+            """Handle GET /console requests."""
+            html_content = interactive_ui()
+            return HTMLResponse(content=html_content)
+        
+        # Insert routes at the beginning of Gradio's router to ensure they're processed first
+        # This prevents Gradio's catch-all routes from intercepting our API endpoints
+        gradio_app.app.router.routes.insert(0, Route("/health", health_handler, methods=["GET"]))
+        gradio_app.app.router.routes.insert(0, Route("/v1/generate", generate_handler, methods=["POST"]))
+        gradio_app.app.router.routes.insert(0, Route("/console", console_handler, methods=["GET"]))
+        print(f"FastAPI routes mounted successfully on Gradio app. Total routes: {len(gradio_app.app.router.routes)}")
+    except Exception as e:
+        print(f"Warning: Could not mount FastAPI routes: {e}")
+        import traceback
+        traceback.print_exc()
 
 # Enable queued execution so ZeroGPU can schedule GPU work reliably
 gradio_app.queue(max_size=8)