Dao-AILab/flash-attention

How to specify cuda 1 when finetuning?

karry5921 opened this issue · 1 comments

The cuda 0 I am using has no memory, so I need to specify cuda 1. I tried two ways: CUDA_VISIBLE_DEVICES=1 deepspeed ${deepspeed_args} and deepspeed --include="localhost:1" ${deepspeed_args} , but it seemed that both ways didn't work out .
The log error is as below:
截屏2024-03-25 17 40 41
Could you please tell me how to specify cuda 1?

I think you meant to post this to another repo?