Split the GPU work from process_fn, and import spaces first

process_fn carried the GPU decorator itself, so nothing was checked before a GPU would be reserved on ZeroGPU. It now checks the inputs and hands the model work to a decorated function, so the Space runs unchanged on any hardware. spaces is now imported before torch and PyHARP, as ZeroGPU requires.

app.py

@@ -1,3 +1,5 @@
+import spaces  # imported first, as ZeroGPU requires
+
 import os
 import tempfile
 from pathlib import Path  
@@ -11,7 +13,6 @@ import torch
 from pyharp import ModelCard, build_endpoint, SampleRate, Channels
 from pyharp.tags import Subcategory
 from df import enhance, init_df
-import spaces
 
 # -----------------------------
 # Model metadata for HARP
@@ -84,12 +85,22 @@ def apply_attenuation_db(noisy: np.ndarray, enhanced: np.ndarray, noise_atten_db
     return out.astype(np.float32)
 
 
-@torch.inference_mode()
-@spaces.GPU(duration=120)
 def process_fn(input_audio_path: str, noise_atten_db: float) -> str:
+    # Checked before the GPU is reserved, so a request the model cannot take spends none of the
+    # caller's ZeroGPU allowance
     if not input_audio_path:
-        raise ValueError("No input audio provided.")
+        raise gr.Error("Provide the speech to denoise.")
+    try:
+        sf.info(input_audio_path)
+    except Exception as exc:
+        raise gr.Error(f"Could not read the audio file: {exc}") from exc
 
+    return denoise(input_audio_path, noise_atten_db)
+
+
+@spaces.GPU(duration=120)
+@torch.inference_mode()
+def denoise(input_audio_path: str, noise_atten_db: float) -> str:
     # The session's format, read before the downmix and resample that
     # DeepFilterNet needs, and restored before saving
     info = sf.info(input_audio_path)