Refuse a recording over 10 minutes before reserving the GPU, and size each reservation

The reservation is computed from the length (300 s of piano measured at about 15 s), and unreadable inputs are refused before the GPU is reserved.

app.py

@@ -8,6 +8,8 @@ import librosa
 import tempfile
 import os
 
+import soundfile as sf
+
 # setuptools dropped this module in version 81, and it is only needed here
 # to locate data files inside the installed transkun package, which
 # importlib.resources does from the standard library.
@@ -41,7 +43,7 @@ def load_model(device: str):
     checkpoint = torch.load(defaultWeight, map_location=device)
     model = TransKun(conf=conf).to(device)
 
-    # Mirrors your checkpoint loading logic :contentReference[oaicite:3]{index=3}
+    # Checkpoints keep their weights under either key
     if "best_state_dict" in checkpoint:
         model.load_state_dict(checkpoint["best_state_dict"], strict=False)
     else:
@@ -50,7 +52,16 @@ def load_model(device: str):
     model.eval()
     return model
 
-@spaces.GPU
+# The whole recording is transcribed in one pass, so the cost grows with its length
+MAX_INPUT_SECONDS = 600
+
+
+def gpu_seconds(input_file):
+    # Measured at about 15 s for 300 s of piano, building the model included, padded
+    return 10 + 0.1 * sf.info(input_file).duration
+
+
+@spaces.GPU(duration=gpu_seconds)
 def transcribe(input_file):
     device = "cuda" if torch.cuda.is_available() else "cpu"
     model = load_model(device)
@@ -84,6 +95,19 @@ def transcribe(input_file):
     return out_path
 
 def process_fn(input_audio_path: str) -> 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 gr.Error("Provide the piano recording to transcribe.")
+    try:
+        duration = sf.info(input_audio_path).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 recording is {duration:.1f} s long. Trim it to {MAX_INPUT_SECONDS // 60} "
+            "minutes or less."
+        )
     midi_path = transcribe(input_audio_path)
     return midi_path