Closed jiqing-feng closed 7 months ago
Hi, I think transformers==4.34.0 is ok. For newer transformers, because they changed their API, we need extra efforts to support them.
Our currently released version does not support flash_attn, so please do not load the model with flash-attn.
In terms of your problem, I think you can still use 4.34.0 and use pip install flash_attn
to solve it.
Hi, I think transformers==4.34.0 is ok. For newer transformers, because they changed their API, we need extra efforts to support them. Our currently released version does not support flash_attn, so please do not load the model with flash-attn. In terms of your problem, I think you can still use 4.34.0 and use
pip install flash_attn
to solve it.
Thx! I have fixed the problem with the newest version of transformers, see here. Lade should run well with this change.
BTW, have you ever tried Lade on CPU? It seems to have a large performance decay on the CPU. Do you have any clue about it?
Yes, running on the CPU should be very slow. But maybe you can set the hyper-parameter (LEVEL, WINDOW_SIZE, and GUESS_SET_SIZE in line config_lade) to a small number to see if it will have minor speedups. The main reason is that our method requires trading steps with flops. You can refer to our blog about this. CPU often has very few extra flops, so we can not expect a speedup.
It seem that the codes does not support the newest version of transformers, so I installed transformers==4.34.0 which is the version in requirements.txt. I got this error when I run
python minimal.py
When I use the newest version of transformers, the following error occurs: