Update inference.py

inference.py

@@ -13,6 +13,7 @@ class LiveInfer:
     def __init__(self, ) -> None:
         args = parse_args()
         args.resume_from_checkpoint = 'live1+_aug_2e/'
+        args.attn_implementation = 'sdpa'
         self.model, self.tokenizer = build_model_and_tokenizer(is_training=False, set_vision_inside=True, **asdict(args))
         self.model.to('cuda')