Drop the length cap on the music, keeping the reservation sized to it

The model works through it in segments or chunks, so its length costs time rather than memory. ZeroGPU already refuses a request that the caller's tier or remaining allowance cannot cover, before reserving anything, and tells a free account how to get more. A cap set so that a free account could afford every request therefore only turned away callers who could. The only request refused here now is one whose reservation would outlast PyHARP's 15 minute timeout, which would stop it anyway.

app.py

@@ -121,8 +121,11 @@ 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. 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
+# segments, and at 7 s and 14 s on large, and at 58 s for 600 s on medium, padded
+# PyHARP stops a job that runs past its timeout_s, 15 minutes by default, so a request whose
+# reservation would exceed that is refused rather than started. A shorter one is left to
+# ZeroGPU, which admits it when the caller's own allowance covers it.
+MAX_GPU_SECONDS = 900
 SECONDS_PER_INPUT_SECOND = {"small": 0.1, "medium": 0.15, "large": 0.3}
 
 
@@ -146,10 +149,11 @@ def process_fn(
         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:
+    if gpu_seconds(input_audio_path, variant, instruments_text, use_sampling,
+                   temperature) > MAX_GPU_SECONDS:
         raise gr.Error(
-            f"The music is {duration:.1f} s long. Trim it to {MAX_INPUT_SECONDS // 60} minutes "
-            "or less."
+            f"The music is {duration:.0f} s long, more than one request can transcribe in the "
+            f"{MAX_GPU_SECONDS // 60} minutes it may run. Trim it, or use a smaller model."
         )
     return transcribe(input_audio_path, variant, instruments_text, use_sampling, temperature)