Upload app.py with huggingface_hub

app.py

@@ -18,12 +18,13 @@ if hf_token:
 # every call into a GPU function burns quota. So call this exactly ONCE at boot
 # with the smallest budget (1s), and keep all real generation on CPU below
 # (NOT decorated). That way quota is only touched here, never per request.
-@spaces.GPU(duration=1)
-def _gpu_startup_touch():
-    return "ok"
-
-
-_gpu_startup_touch()
+try:
+    @spaces.GPU(duration=1)
+    def _gpu_startup_touch():
+        return "ok"
+    _gpu_startup_touch()
+except Exception as e:
+    print(f"Warning: ZeroGPU touch failed (running CPU-only): {e}")
 
 
 # Model cache
@@ -43,11 +44,17 @@ def load_model(model_name):
     return MODEL_CACHE[model_name]
 
 
-# Pre-warm both models at startup so the first user request is fast
-print("Pre-warming models...")
-for _name in ["small-music", "small-sfx"]:
-    load_model(_name)
-print("All models ready!")
+# Pre-warm both models in background so the first user request is fast
+# but the Gradio server starts immediately without waiting for downloads
+import threading
+
+def _prewarm():
+    print("Pre-warming models...")
+    for _name in ["small-music", "small-sfx"]:
+        load_model(_name)
+    print("All models ready!")
+
+threading.Thread(target=_prewarm, daemon=True).start()
 
 
 def generate_audio(prompt, duration, steps, cfg_scale, seed, model_name):