Fix: Keep Gradio as main app, mount FastAPI routes via load event, remove uvicorn.run()

app.py

@@ -517,14 +517,59 @@ with gr.Blocks(
         fn=lambda: ("", ""),
         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)
+            
+            # Add routes to Gradio's router
+            gradio_app.app.router.routes.append(Route("/health", health_handler, methods=["GET"]))
+            gradio_app.app.router.routes.append(Route("/v1/generate", generate_handler, methods=["POST"]))
+            gradio_app.app.router.routes.append(Route("/console", console_handler, methods=["GET"]))
+            print("FastAPI routes mounted successfully on Gradio app")
+        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)
 
 # Enable queued execution so ZeroGPU can schedule GPU work reliably
 gradio_app.queue(max_size=8)
 
-# Mount the Gradio UI on the FastAPI app so both live side-by-side
-app = gr.mount_gradio_app(fastapi_app, gradio_app, path="/gradio")
+# Set app to Gradio for Spaces compatibility (sdk: gradio requires Gradio app)
+# Spaces will handle running the server automatically
+app = gradio_app
 
 if __name__ == "__main__":  # pragma: no cover
-    import uvicorn
-
-    uvicorn.run(app, host="0.0.0.0", port=int(os.environ.get("PORT", 7860)))
+    # For local testing only - Spaces handles server startup
+    app.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))