Fix: Use load event to add routes, insert at position 0 for priority

app.py

@@ -481,56 +481,41 @@ with gr.Blocks(
         outputs=[prompt_input, output, status_display],
     )
     
-    # Note: API routes will be added after Blocks context to avoid interfering with Gradio's static assets
-
-# Attach API routes directly onto Gradio's FastAPI instance
-# Use include_router which is the proper FastAPI way to add routes
-api_router = APIRouter()
-
-
-@api_router.get("/health")
-def api_health() -> dict[str, str]:
-    """Health check endpoint."""
-    return healthcheck()
-
-
-@api_router.post("/v1/generate", response_model=GenerateResponse)
-def api_generate(payload: GeneratePayload) -> GenerateResponse:
-    """Generate endpoint."""
-    return generate_endpoint(payload)
-
-
-# Include router on Gradio's FastAPI app
-try:
-    gradio_app.app.include_router(api_router)
-    print(f"API router included successfully. Total routes: {len(gradio_app.app.routes)}")
-except AttributeError:
-    # Fallback: add routes directly if include_router not available
-    print("include_router not available, adding routes directly...")
-    from fastapi.responses import JSONResponse
-    from starlette.routing import Route
-    from starlette.requests import Request
-    
-    async def health_handler(request: Request):
-        return JSONResponse(content=healthcheck())
-    
-    async def generate_handler(request: Request):
+    # Add API routes using Gradio's load event to ensure they're registered correctly
+    def add_api_routes():
+        """Add API routes after Gradio is fully initialized."""
         try:
-            data = await request.json()
-            payload = GeneratePayload(**data)
-            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))
+            from fastapi.responses import JSONResponse
+            from starlette.routing import Route
+            from starlette.requests import Request
+            
+            async def health_handler(request: Request):
+                """Handle GET /health requests."""
+                return JSONResponse(content=healthcheck())
+            
+            async def generate_handler(request: Request):
+                """Handle POST /v1/generate requests."""
+                try:
+                    data = await request.json()
+                    payload = GeneratePayload(**data)
+                    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))
+            
+            # Add routes directly to Gradio's router
+            # Insert at position 0 to ensure they're checked before Gradio's catch-all routes
+            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"]))
+            print(f"API routes added via load event. Total routes: {len(gradio_app.app.routes)}")
+        except Exception as e:
+            print(f"Warning: Could not add API routes: {e}")
+            import traceback
+            traceback.print_exc()
     
-    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"]))
-    print(f"Routes added directly. Total routes: {len(gradio_app.app.routes)}")
-except Exception as e:
-    print(f"Warning: Could not add API routes: {e}")
-    import traceback
-    traceback.print_exc()
+    # Register the load event handler
+    gradio_app.load(add_api_routes)
 
 warm_start()