Fix startup on ZeroGPU hardware

- Declare a placeholder @spaces.GPU function: ZeroGPU refuses to start
without one. Inference stays on the CPU.
- Mount Gradio without SSR: Spaces enable it, which starts a Node server
on port 7860 and prevented uvicorn from binding it.
- Point the page config at this Space.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

app.py

@@ -5,11 +5,23 @@ A Gradio Space runs `python app.py` and expects a server on port 7860: this serv
 
 On the free tier a Gradio Space gets ZeroGPU hardware. The GPU is deliberately not used: it is
 metered per visitor (a couple of minutes a day) and queued, which does not suit a score
-recomputed on every keystroke, and Laya answers in a fraction of a second on CPU anyway. So the
-`spaces` package is never imported and inference is pinned to the CPU.
+recomputed on every keystroke, and Laya answers in a fraction of a second on CPU anyway. So
+inference is pinned to the CPU.
 """
 import os
 
+# ZeroGPU refuses to start a Space that declares no @spaces.GPU function, so one is declared but
+# never called. `spaces` must be imported before anything that imports torch (it patches CUDA
+# initialisation); outside Hugging Face the package is absent or has no effect.
+try:
+    import spaces
+
+    @spaces.GPU
+    def _zerogpu_placeholder():
+        """Never called: only satisfies ZeroGPU's startup check."""
+except Exception as error:  # absent locally; any other failure shows in the Space logs
+    print(f"spaces unavailable ({type(error).__name__}: {error})", flush=True)
+
 # One model on CPU keeps memory and latency reasonable; override in the Space settings.
 os.environ.setdefault("LAYA_BACKEND", "torch")
 os.environ.setdefault("LAYA_MODELS", "typed")
@@ -27,7 +39,9 @@ with gr.Blocks(title="Laya Demo API") as info:
         "La démo est sur `/`, l'état de l'API sur `/health`."
     )
 
-app = gr.mount_gradio_app(app, info, path="/gradio")
+# Spaces set GRADIO_SSR_MODE, which starts a Node server on port 7860 and leaves uvicorn unable
+# to bind it ("address already in use"); this page needs no server-side rendering.
+app = gr.mount_gradio_app(app, info, path="/gradio", ssr_mode=False)
 
 if __name__ == "__main__":
     uvicorn.run(app, host="0.0.0.0", port=7860)

requirements.txt

@@ -1,7 +1,7 @@
-# Hugging Face Space (Gradio SDK, which installs gradio itself). CPU build of PyTorch, pinned to a
-# version ZeroGPU hardware supports.
---extra-index-url https://download.pytorch.org/whl/cpu
+# Hugging Face Space (Gradio SDK, which installs gradio itself). Standard PyTorch build, as the
+# `spaces` package (needed by ZeroGPU hardware) expects it; inference is still forced onto the CPU.
 torch==2.13.0
 laya==0.3.20
 fastapi
 uvicorn
+spaces

web/config.js

@@ -1,6 +1,6 @@
 // URL of the Laya API used by the static deployment (e.g. on Vercel).
 // Replace it with your Hugging Face Space URL, without a trailing slash.
-window.LAYA_API = "https://YOUR-USER-laya-demo.hf.space";
+window.LAYA_API = "https://robynsd-laya-demo.hf.space";
 
 // Pause after the last keystroke before calling the API (ms). A CPU server takes
 // about 1 to 2 s per request, so it should not be called on every keystroke.