fix use flash attn

main.py

@@ -77,8 +77,8 @@ transcribe_pipeline = pipeline(
     torch_dtype=torch_dtype,
     #device=device,
 )
-if device == "cuda":
-    transcribe_pipeline.model = torch.compile(transcribe_pipeline.model, mode="max-autotune")
+#if device == "cuda":
+#    transcribe_pipeline.model = torch.compile(transcribe_pipeline.model, mode="max-autotune")
 
 # Warm up the model with empty audio
 logger.info("Warming up Whisper model with dummy input")
@@ -91,8 +91,8 @@ async def transcribe(audio: tuple[int, np.ndarray]):
     logger.info(f"Sample rate: {sample_rate}Hz, Shape: {audio_array.shape}")
     
     outputs = transcribe_pipeline(
-        #audio_to_bytes(audio),
-        audio_array,
+        #audio_to_bytes(audio), # pass bytes
+        audio_array, # pass numpy array
         chunk_length_s=3,
         batch_size=1,
         generate_kwargs={