Update app.py

app.py

@@ -2,18 +2,6 @@ import os
 import spaces
 import torch
 import gradio as gr
-from huggingface_hub import snapshot_download, login
-from transformers.utils import move_cache
-
-
-
-
-LLM_MODEL_DIR = '/model'
-LLM_MODEL_ID = "mistral-community/Mistral-7B-v0.2"
-LLM_MODEL_REVISION = 'main'
-os.makedirs(LLM_MODEL_DIR, exist_ok=True)
-snapshot_download(LLM_MODEL_ID, revision=LLM_MODEL_REVISION, local_dir=LLM_MODEL_DIR) #, token=HF_TOKEN)
-move_cache()
 
 
 # cpu
@@ -28,6 +16,12 @@ print(zero.device) # <-- 'cpu' 🤔
 def greet(user):
     # print(zero.device) # <-- 'cuda:0' 🤗
     from vllm import SamplingParams, LLM
+    from transformers.utils import move_cache
+    from huggingface_hub import snapshot_download, login
+
+    LLM_MODEL_ID = "mistral-community/Mistral-7B-v0.2"
+    fp = snapshot_download(LLM_MODEL_ID)
+    move_cache()
 
     model = LLM(LLM_MODEL_DIR)
     sampling_params = dict(