seperate .to(device) ?

app.py

@@ -1,6 +1,6 @@
 import os
 import spaces
-
+# only debug for hf now
 REPO_TYPE = "hf" 
 if REPO_TYPE not in ["hf", "ms"]:
     raise ValueError("REPO_TYPE must be either 'hf' for Hugging Face or 'ms' for ModelScope.")
@@ -13,40 +13,48 @@ else:
 
 
 # 1. 定义本地路径和远程仓库ID
-FUN_ASR_NANO_LOCAL_PATH = "./Fun-ASR/model"
-SENSE_VOICE_SMALL_LOCAL_PATH = "./Fun-ASR/model/SenseVoiceSmall"
+MODEL_CACHE_DIR = "./models"
+FUN_ASR_NANO_LOCAL_PATH = os.path.join(MODEL_CACHE_DIR, "Fun-ASR-Nano")
+SENSE_VOICE_SMALL_LOCAL_PATH = os.path.join(MODEL_CACHE_DIR, "SenseVoiceSmall")
+VAD_MODEL_LOCAL_PATH = os.path.join(MODEL_CACHE_DIR, "fsmn-vad")
+
+# 创建模型缓存目录
+os.makedirs(MODEL_CACHE_DIR, exist_ok=True)
 
+# 设置ModelScope环境变量以使用本地缓存
+os.environ['MODELSCOPE_CACHE'] = MODEL_CACHE_DIR
+# 禁用远程下载，强制使用本地模型（可选，如果想要确保只使用本地模型）
+# os.environ['MODELSCOPE_DISABLE_REMOTE'] = '1'
+
+print(f"ModelScope缓存目录设置为: {MODEL_CACHE_DIR}")
 
 if REPO_TYPE == "ms":
     FUN_ASR_NANO_REPO_ID = "FunAudioLLM/Fun-ASR-Nano-2512"
     SENSE_VOICE_SMALL_REPO_ID = "iic/SenseVoiceSmall"
+    VAD_MODEL_REPO_ID = "iic/speech_fsmn_vad_zh-cn-16k-common-pytorch"
 else:
     FUN_ASR_NANO_REPO_ID = "FunAudioLLM/Fun-ASR-Nano-2512"
     SENSE_VOICE_SMALL_REPO_ID = "FunAudioLLM/SenseVoiceSmall"
+    VAD_MODEL_REPO_ID = "funasr/fsmn-vad"
 
 # 2. 检查本地是否存在，不存在则下载
-if not os.path.exists(FUN_ASR_NANO_LOCAL_PATH):
-    print(f"正在下载模型 Fun-ASR-Nano 到 {FUN_ASR_NANO_LOCAL_PATH} ...")
-    snapshot_download(
-        repo_id=FUN_ASR_NANO_REPO_ID,
-        local_dir=FUN_ASR_NANO_LOCAL_PATH,
-        ignore_patterns=["*.onnx"], # 如果你不需要onnx文件，可以过滤掉以节省时间和空间
-    )
-    print("模型下载完毕！")
-else:
-    print("检测到本地模型文件，跳过下载。")
-
+def download_model_if_not_exists(repo_id, local_path, model_name):
+    """如果本地模型不存在，则下载模型"""
+    if not os.path.exists(local_path):
+        print(f"正在下载模型 {model_name} 到 {local_path} ...")
+        snapshot_download(
+            repo_id=repo_id,
+            local_dir=local_path,
+            ignore_patterns=["*.onnx"], # 如果你不需要onnx文件，可以过滤掉以节省时间和空间
+        )
+        print(f"{model_name} 模型下载完毕！")
+    else:
+        print(f"检测到本地 {model_name} 模型文件，跳过下载。")
 
-if not os.path.exists(SENSE_VOICE_SMALL_LOCAL_PATH):
-    print(f"正在下载模型 {SENSE_VOICE_SMALL_REPO_ID} 到 {SENSE_VOICE_SMALL_LOCAL_PATH} ...")
-    snapshot_download(
-        repo_id=SENSE_VOICE_SMALL_REPO_ID,
-        local_dir=SENSE_VOICE_SMALL_LOCAL_PATH,
-        ignore_patterns=["*.onnx"], # 如果你不需要onnx文件，可以过滤掉以节省时间和空间
-    )
-    print("模型下载完毕！")
-else:
-    print("检测到本地模型文件，跳过下载。")
+# 下载所有需要的模型
+download_model_if_not_exists(FUN_ASR_NANO_REPO_ID, FUN_ASR_NANO_LOCAL_PATH, "Fun-ASR-Nano")
+download_model_if_not_exists(SENSE_VOICE_SMALL_REPO_ID, SENSE_VOICE_SMALL_LOCAL_PATH, "SenseVoiceSmall")
+download_model_if_not_exists(VAD_MODEL_REPO_ID, VAD_MODEL_LOCAL_PATH, "VAD Model")
 
 
 
@@ -66,13 +74,13 @@ import importlib
 from funasr import AutoModel
 from funasr.utils.postprocess_utils import rich_transcription_postprocess
 
-# Model configurations for Hugging Face deployment
+# Model configurations for local deployment
 FUN_ASR_NANO_MODEL_PATH_LIST = [
-    "Fun-ASR/model", # local path, ms
+    FUN_ASR_NANO_LOCAL_PATH, # local path
 ]
 
 SENSEVOICE_MODEL_PATH_LIST = [
-    "Fun-ASR/model/SenseVoiceSmall", # local path together with this hf space
+    SENSE_VOICE_SMALL_LOCAL_PATH, # local path
 ]
 
 class LogCapture(io.StringIO):
@@ -406,7 +414,7 @@ def get_model_options(pipeline_type):
 # Dictionary to store loaded models
 loaded_models = {}
 
-@spaces.GPU()
+@spaces.GPU(duration=40)
 def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_password, pipeline_type, model_id, download_method, start_time=None, end_time=None, verbose=False):
     """
     Transcribes audio from a given source using SenseVoice.
@@ -478,7 +486,6 @@ 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")
@@ -488,9 +495,9 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
                     model=model_id,
                     trust_remote_code=True,
                     remote_code=f"./Fun-ASR/model.py",
-                    vad_model="fsmn-vad",
+                    vad_model=VAD_MODEL_LOCAL_PATH,  # Use local VAD model path
                     vad_kwargs={"max_single_segment_time": 30000},
-                    device='cpu', # 初始化在cpu，然后推理的时候移到GPU，保证利用好zeroGPU？
+                    device=device,
                     disable_update=True,
                     hub='ms',
                 )
@@ -498,9 +505,9 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
                 model = AutoModel(
                     model=model_id,
                     trust_remote_code=False,
-                    vad_model="fsmn-vad",
+                    vad_model=VAD_MODEL_LOCAL_PATH,  # Use local VAD model path
                     vad_kwargs={"max_single_segment_time": 30000},
-                    device='cpu',
+                    device=device,
                     disable_update=True,
                     hub='ms',
                 )
@@ -511,14 +518,9 @@ def transcribe_audio(audio_input, audio_url, proxy_url, proxy_username, proxy_pa
                 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
-
+        # move seperately?
+        model.model.to(device)
+        model.vad_model.to(device)
         # Perform the transcription
         start_time_perf = time.time()
         
@@ -545,8 +547,9 @@ 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
+        
+        model.model.to("cpu")
+        model.vad_model.to("cpu")
 
         transcription = rich_transcription_postprocess(res[0]["text"])
         end_time_perf = time.time()