Check inputs before reserving the GPU, and reserve 120 s rather than 180.

app.py

@@ -50,7 +50,6 @@ def _audio_duration(path: str | None) -> float:
     return duration
 
 
-@spaces.GPU(duration=180)
 def process_fn(
     audio_path: str | None,
     prompt: str,
@@ -61,6 +60,8 @@ def process_fn(
     strength: float,
     seed: int,
 ) -> str:
+    # Checked before the GPU is reserved, so a request the model cannot take spends none of the
+    # caller's ZeroGPU allowance
     duration = _audio_duration(audio_path)
     prompt = (prompt or "").strip()
     if not prompt:
@@ -74,6 +75,14 @@ def process_fn(
             "with the end after the start."
         )
 
+    return run_edit(audio_path, prompt, mode, edit_start, edit_end, continuation_length,
+                    strength, seed)
+
+
+# A 6 s restyle measured about 24 s, and inputs are capped at 30 s
+@spaces.GPU(duration=120)
+def run_edit(audio_path, prompt, mode, edit_start, edit_end, continuation_length, strength,
+             seed) -> str:
     try:
         return edit_audio(
             audio_path=audio_path,