changed to uvicorn setup for HF v13

main/api.py

@@ -51,7 +51,7 @@ class InferenceApi(LitAPI):
             stream: bool = False
     ) -> Any:
         """Make an authenticated request to the LLM Server."""
-        access_token = os.environ.get("InferenceAPI")
+        access_token = os.environ.get("InfAPITokenWrite")
         headers = {"Authorization": f"Bearer {access_token}"} if access_token else {}
         base_url = self.llm_config.get('base_url', 'http://localhost:8002')
         full_endpoint = f"{base_url.rstrip('/')}/{self._get_endpoint(endpoint).lstrip('/')}"

main/main.py

@@ -36,7 +36,7 @@ def create_app():
     logger = setup_logging()
 
     # Log into Hugging Face Hub
-    access_token = os.environ.get("InferenceAPI")
+    access_token = os.environ.get("InfAPITokenWrite")
     if access_token:
         try:
             login(token=access_token)