Fix FastAPI route addition - move outside Blocks context

app.py

@@ -411,43 +411,54 @@ with gr.Blocks(title="Router Model API - ZeroGPU") as gradio_app:
         inputs=[prompt_input, max_tokens_input, temp_input, top_p_input],
         outputs=output,
     )
-    
-    # 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:
-            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))
-    
-    @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
 
+# Add FastAPI routes directly to Gradio's app after Blocks context exits
+# Use load event to ensure app is ready
+def add_fastapi_routes():
+    """Add FastAPI routes to Gradio's underlying FastAPI app."""
+    try:
+        from fastapi.responses import JSONResponse
+        
+        @app.app.post("/v1/generate")
+        async def fastapi_generate_endpoint(request):
+            """FastAPI endpoint mounted in Gradio app."""
+            from fastapi import Request
+            try:
+                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))
+        
+        @app.app.get("/gradio")
+        async def fastapi_gradio_ui():
+            """FastAPI /gradio endpoint."""
+            return HTMLResponse(interactive_ui())
+        
+        @app.app.get("/")
+        async def fastapi_healthcheck():
+            """FastAPI healthcheck endpoint."""
+            return {
+                "status": "ok",
+                "model": MODEL_ID,
+                "strategy": ACTIVE_STRATEGY or "pending",
+            }
+        print("FastAPI routes added successfully")
+    except Exception as e:
+        print(f"Warning: Could not add FastAPI routes: {e}")
+
+# Use load event to add routes when app is ready
+app.load(add_fastapi_routes)
+
 if __name__ == "__main__":  # pragma: no cover
     app.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))