Size the GPU reservation by the mixture's length and check requests before the GPU is reserved.

app.py

@@ -40,6 +40,7 @@ import contextlib
 import tempfile
 from pathlib import Path
 
+import soundfile as sf
 import torch
 import torchaudio
 from huggingface_hub import snapshot_download
@@ -93,8 +94,34 @@ model_card = ModelCard(
 )
 
 
-@spaces.GPU
+# The whole mixture is separated, so the cost grows with its length
+MAX_INPUT_SECONDS = 600
+
+
+def gpu_seconds(input_audio, model_name, niter, wiener_win_len):
+    # Measured at about 10 s for a minute of audio, building the separator included, padded
+    return 15 + 0.25 * sf.info(input_audio).duration
+
+
 def process_fn(input_audio, model_name, niter, wiener_win_len):
+    # 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:
+        raise gr.Error("Provide the mixture to separate.")
+    try:
+        duration = sf.info(input_audio).duration
+    except Exception as exc:
+        raise gr.Error(f"Could not read the audio file: {exc}") from exc
+    if duration > MAX_INPUT_SECONDS:
+        raise gr.Error(
+            f"The mixture is {duration:.1f} s long. Trim it to {MAX_INPUT_SECONDS // 60} minutes "
+            "or less."
+        )
+    return separate(input_audio, model_name, niter, wiener_win_len)
+
+
+@spaces.GPU(duration=gpu_seconds)
+def separate(input_audio, model_name, niter, wiener_win_len):
     audio, rate = torchaudio.load(input_audio)
 
     # Open-Unmix separates at 44.1 kHz, so the session's rate is kept here and