Add FastAPI routes directly to Gradio app instead of mounting

app.py

@@ -412,28 +412,39 @@ with gr.Blocks(title="Router Model API - ZeroGPU") as gradio_app:
         outputs=output,
     )
     
-    # Mount FastAPI routes using Gradio's load event
-    def mount_fastapi_routes():
-        """Mount FastAPI routes when Gradio app loads."""
+    # Add FastAPI routes directly to Gradio's app after it's created
+    @gradio_app.app.post("/v1/generate")
+    async def fastapi_generate_endpoint(request):
+        """FastAPI endpoint mounted in Gradio app."""
+        from fastapi import Request
+        from fastapi.responses import JSONResponse
         try:
-            # Access Gradio's underlying FastAPI app
-            gradio_app.app.mount("/v1", fastapi_app)
-            gradio_app.app.mount("/gradio", fastapi_app)
-            print("FastAPI routes mounted successfully")
-        except (AttributeError, Exception) as e:
-            print(f"FastAPI routes mounting note: {e}")
-            # Try alternative mounting approach
-            try:
-                import starlette
-                # Mount using Starlette's mount
-                if hasattr(gradio_app, 'app') and hasattr(gradio_app.app, 'mount'):
-                    gradio_app.app.mount("/v1", fastapi_app)
-                    gradio_app.app.mount("/gradio", fastapi_app)
-            except Exception as e2:
-                print(f"Alternative mounting also failed: {e2}")
+            data = await request.json()
+            payload = GeneratePayload(**data)
+            text = _generate_with_gpu(
+                prompt=payload.prompt,
+                max_new_tokens=payload.max_new_tokens or MAX_NEW_TOKENS,
+                temperature=payload.temperature or DEFAULT_TEMPERATURE,
+                top_p=payload.top_p or DEFAULT_TOP_P,
+            )
+            return JSONResponse(content={"text": text})
+        except Exception as exc:
+            from fastapi import HTTPException
+            raise HTTPException(status_code=500, detail=str(exc))
     
-    # Use load event to mount routes when app is ready
-    gradio_app.load(mount_fastapi_routes)
+    @gradio_app.app.get("/gradio")
+    async def fastapi_gradio_ui():
+        """FastAPI /gradio endpoint."""
+        return HTMLResponse(interactive_ui())
+    
+    @gradio_app.app.get("/")
+    async def fastapi_healthcheck():
+        """FastAPI healthcheck endpoint."""
+        return {
+            "status": "ok",
+            "model": MODEL_ID,
+            "strategy": ACTIVE_STRATEGY or "pending",
+        }
 
 # Set app to Gradio Blocks for Spaces - ZeroGPU requires Gradio SDK
 app = gradio_app