Check inputs before reserving the GPU, so a request the model cannot take spends none of the caller's allowance.

app.py

@@ -102,9 +102,9 @@ if __name__ != "__main__":
         inference_pipeline.device = DEVICE
 
 
-@spaces.GPU(duration=60)
-@torch.inference_mode()
 def process_fn(vocal_filepath, n_timesteps):  # add , cfg_scale to re-enable
+    # Checked before the GPU is reserved, so a request the model cannot take spends none of the
+    # caller's ZeroGPU allowance
     if model_loading:
         raise gr.Error("Model is still loading, please wait a moment and try again.")
     if inference_pipeline is None:
@@ -115,11 +115,17 @@ def process_fn(vocal_filepath, n_timesteps):  # add , cfg_scale to re-enable
 
     try:
         duration = librosa.get_duration(path=vocal_filepath)
-        if not (3 <= duration <= 30):
-            raise gr.Error("Audio duration must be between 3 and 30 seconds.")
     except Exception as e:
-        raise gr.Error(f"Cannot read audio file or get duration: {e}")
+        raise gr.Error(f"Cannot read the audio file: {e}")
+    if not (3 <= duration <= 30):
+        raise gr.Error(f"Audio must be 3 to 30 seconds long, not {duration:.1f}.")
+
+    return accompany(vocal_filepath, n_timesteps)
 
+
+@spaces.GPU(duration=60)
+@torch.inference_mode()
+def accompany(vocal_filepath, n_timesteps):
     try:
         # The session's format, read before the downmix and resample the model
         # needs, and restored on both outputs