Update app.py

app.py

@@ -6,7 +6,12 @@ import httpx
 import asyncio
 import gradio as gr
 import os
-from spaces import GPU 
+from spaces import GPU
+from dotenv import load_dotenv
+
+load_dotenv()
+
+HUGGINGFACE_TOKEN = os.getenv("HUGGINGFACE_TOKEN")
 
 global_data = {
     'models': {},
@@ -51,7 +56,7 @@ class ModelManager:
     def load_model(self, model_config):
         if model_config['name'] not in self.models:
             try:
-                self.models[model_config['name']] = Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename'])
+                self.models[model_config['name']] = Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename'], use_auth_token=HUGGINGFACE_TOKEN)
             except Exception as e:
                 print(f"Error loading model {model_config['name']}: {e}")
 
@@ -83,7 +88,7 @@ def remove_duplicates(text):
             seen_lines.add(line)
     return '\n'.join(unique_lines)
 
-@GPU(duration=60) 
+@GPU(duration=60)
 def generate_model_response(model, inputs):
     try:
         response = model(inputs)