add HF token

app.py

@@ -1,3 +1,5 @@
+import os
+
 import gradio as gr
 from huggingface_hub import InferenceClient
 
@@ -7,10 +9,12 @@ from transformers import AutoTokenizer
 from model.modeling_llamask import LlamaskForCausalLM
 from model.tokenizer_utils import generate_custom_mask, prepare_tokenizer
 
+
+access_token = os.getenv("HF_TOKEN")
 model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
 device = 'cpu'
 
-model = LlamaskForCausalLM.from_pretrained(model_id, torch_dtype= torch.bfloat16)
+model = LlamaskForCausalLM.from_pretrained(model_id, torch_dtype= torch.bfloat16, access_token=access_token)
 model = model.to(device)
 tokenizer = AutoTokenizer.from_pretrained(model_id, padding_side="left")