Size the reservation from current timings

The estimate was measured when the model was built inside every call. It is built when the worker starts now, and a call measures about 2 to 8 s for 30 s of piano and 10 to 15 s for 300 s, so the reservation is 8 s plus 0.05 s per second of audio, down from 10 s plus 0.1 s.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

app.py

@@ -67,8 +67,9 @@ MAX_GPU_SECONDS = 900
 
 
 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
+    # Measured at about 2 to 8 s for 30 s of piano and 10 to 15 s for 300 s, upload included,
+    # padded. The model is built when the worker starts, not inside the call.
+    return 8 + 0.05 * sf.info(input_file).duration
 
 
 @spaces.GPU(duration=gpu_seconds)