Upload app.py

app.py

@@ -5,6 +5,7 @@ from dotenv import load_dotenv
 load_dotenv()
 import os
 hf_token = os.getenv("HUGGINGFACEHUB_API_TOKEN")
+# os.environ['HUGGINGFACE_TOKEN'] = os.getenv("HUGGINGFACEHUB_API_TOKEN")
 
 
 # [선택1] 거대모델 랭체인 Custom LLM (HF InferenceClient) - 70B가 무료!!!, openai보다 성능 안떨어짐 (스트리밍은 아직 안됨)
@@ -32,7 +33,7 @@ class CustomInferenceClient(LLM, KwArgsModel):
 
     def __init__(self, model_name: str, hf_token: str, kwargs: Optional[Dict[str, Any]] = None):
         # inference_client = InferenceClient(model=model_name, token=hf_token)
-        inference_client = InferenceClient(model=model_name, token=os.getenv("HUGGINGFACEHUB_API_TOKEN"))
+        inference_client = InferenceClient(model=model_name, token=hf_token)
         super().__init__(
             model_name=model_name,
             hf_token=hf_token,