Update app.py

app.py

@@ -2,6 +2,9 @@ import transformers
 import torch
 import os
 
+os.environ["HF_TOKEN"] = token
+os.environ["HUGGINGFACEHUB_API_TOKEN"] = token
+# os.environ["USE_FLASH_ATTENTION"] = "1"
 
 print(f"Device name: {torch.cuda.get_device_properties('cuda').name}")
 print(f"FlashAttention available: {torch.backends.cuda.flash_sdp_enabled()}")