适配zeroGPU

app.py

@@ -53,7 +53,7 @@ TASK_PROMPT_MAPPING = {
     "STTC (Speech to Text Chat)": "首先将语音转录为文字，然后对语音内容进行回复，转录和文字之间使用<开始回答>分割。"
 }
 
-@spaces.GPU
+
 def init_model_my():
     logging.basicConfig(level=logging.DEBUG,
                         format='%(asctime)s %(levelname)s %(message)s')
@@ -71,7 +71,7 @@ def init_model_my():
     print(model)
     return model, tokenizer
 
-model, tokenizer = init_model_my()
+global_model, tokenizer = init_model_my()
 print("model init success")
 def do_resample(input_wav_path, output_wav_path):
     """"""
@@ -87,6 +87,7 @@ def do_resample(input_wav_path, output_wav_path):
     makedir_for_file(output_wav_path)
     torchaudio.save(output_wav_path, waveform, 16000)
 
+@spaces.GPU
 def true_decode_fuc(input_wav_path, input_prompt):
     # input_prompt = TASK_PROMPT_MAPPING.get(input_prompt, "未知任务类型")
     print(f"wav_path: {input_wav_path}, prompt:{input_prompt}")
@@ -120,11 +121,12 @@ def true_decode_fuc(input_wav_path, input_prompt):
     feat = feat.unsqueeze(0).cuda()
     # feat = feat.half()
     # feat_lens = feat_lens.half()
-    model = None
+    model = global_model.cuda()
+    model.eval()
     res_text = model.generate(wavs=feat, wavs_len=feat_lens, prompt=input_prompt)[0]
     print("耿雪龙哈哈：", res_text)
     return res_text
-@spaces.GPU
+
 def do_decode(input_wav_path, input_prompt):
     print(f'input_wav_path= {input_wav_path}, input_prompt= {input_prompt}')
     # 省略处理逻辑