final try

app.py

@@ -478,6 +478,7 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
 
         # Model caching
         model_key = (pipeline_type, model_id)
+        model = None
         if model_key in loaded_models:
             model = loaded_models[model_key]
             logging.info("Loaded model from cache")
@@ -489,7 +490,7 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
                     remote_code=f"./Fun-ASR/model.py",
                     vad_model="fsmn-vad",
                     vad_kwargs={"max_single_segment_time": 30000},
-                    device=device,
+                    device='cpu', # 初始化在cpu，然后推理的时候移到GPU，保证利用好zeroGPU？
                     disable_update=True,
                     hub='ms',
                 )
@@ -499,7 +500,7 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
                     trust_remote_code=False,
                     vad_model="fsmn-vad",
                     vad_kwargs={"max_single_segment_time": 30000},
-                    device=device,
+                    device='cpu',
                     disable_update=True,
                     hub='ms',
                 )
@@ -509,6 +510,14 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
                 yield verbose_messages + error_msg, "", None
                 return
             loaded_models[model_key] = model
+        
+        try:
+            model.to(device)
+            logging.info(f"Model moved to device: {device}")
+        except Exception as e:
+            logging.error(f"Error moving model to device: {str(e)}")
+            yield verbose_messages + f"Error moving model to device: {str(e)}", "", None
+            return
 
         # Perform the transcription
         start_time_perf = time.time()
@@ -536,6 +545,8 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
                 merge_vad=True,
                 merge_length_s=15,
             )
+            
+        model.to('cpu')  # Move model back to CPU after inference to free GPU memory
 
         transcription = rich_transcription_postprocess(res[0]["text"])
         end_time_perf = time.time()