Size the GPU reservation by measured cost.

app.py

@@ -120,9 +120,10 @@ def _resolve_instruments(text: str) -> tuple[list[str] | None, str]:
 
 
 # The audio is transcribed in five-second segments, so the cost grows with its length, and with the
-# size of the model
+# size of the model. Measured at 9 s for both 20 s and 60 s of audio on medium, which batches the
+# segments, and at 7 s and 14 s on large, padded, so ten minutes stays under 200 s on every size
 MAX_INPUT_SECONDS = 600
-SECONDS_PER_INPUT_SECOND = {"small": 0.5, "medium": 1.0, "large": 2.0}
+SECONDS_PER_INPUT_SECOND = {"small": 0.1, "medium": 0.15, "large": 0.3}
 
 
 def gpu_seconds(input_audio_path, variant, instruments_text, use_sampling, temperature):
@@ -150,11 +151,6 @@ def process_fn(
             f"The music is {duration:.1f} s long. Trim it to {MAX_INPUT_SECONDS // 60} minutes "
             "or less."
         )
-    if gpu_seconds(input_audio_path, variant, instruments_text, use_sampling, temperature) > 200:
-        raise gr.Error(
-            f"The {variant} model would need more GPU time for {duration:.0f} s of music than a "
-            "request can reserve. Choose a smaller model or trim the music."
-        )
     return transcribe(input_audio_path, variant, instruments_text, use_sampling, temperature)