Initialize ZeroGPU for the session-protected mounted Gradio app

app.py

@@ -278,6 +278,7 @@ def create_app(model, device='cuda:0', output_root=None, examples_manifest=None,
         return response
 
     access.middleware(app)
+    app.state.demo = demo
     return gr.mount_gradio_app(app, demo, path='/', allowed_paths=example_files, show_error=False,
         theme=theme, css=(Path(__file__).parent/'style.css').read_text(), footer_links=[])
 
@@ -296,5 +297,15 @@ if __name__ == '__main__':
         from zero_gpu import initialize, reconstruct_on_gpu
         initialize(args.model)
         inference_fn = reconstruct_on_gpu
-    uvicorn.run(create_app(args.model, args.device, args.output_root, args.examples,
-                           inference_fn=inference_fn), host=args.host, port=args.port)
+    app = create_app(args.model, args.device, args.output_root, args.examples,
+                     inference_fn=inference_fn)
+    server = uvicorn.Server(uvicorn.Config(app, host=args.host, port=args.port))
+    app.state.demo.server = server
+    if inference_fn is not None:
+        # Mounted Gradio uses Uvicorn rather than Blocks.launch. Execute the
+        # same startup hook from pinned spaces 0.51.3 to pack emulated CUDA.
+        from spaces.config import Config
+        if Config.zero_gpu:
+            from spaces.zero import startup
+            startup()
+    server.run()