feat: dynamic zero gpu time

model.py

@@ -1,6 +1,7 @@
 from transformers import pipeline
 from accelerate import Accelerator
 import spaces
+import librosa
 
 model_id = "JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW"
 
@@ -13,7 +14,15 @@ def load_model():
     pipe = pipeline("automatic-speech-recognition", model=model_id, device=device)
 
 
-@spaces.GPU()
+def get_gpu_duration(audio: str) -> int:
+    y, sr = librosa.load(audio)
+    duration = librosa.get_duration(y=y, sr=sr) / 60.0
+    gpu_duration = (duration + 59.0) // 60.0
+    print(f"{duration=}, {gpu_duration=}")
+    return max(1, int(gpu_duration))
+
+
+@spaces.GPU(duration=get_gpu_duration)
 def transcribe_audio_local(audio: str) -> str:
     print(f"{audio=}")