How to change options to decrease calculation consumption during evaluation_3d ?
Xiaolong-RRL opened this issue · 2 comments
Dear author:
Thanks for your interesting work.
I wonder how to change options as following to decrease calculation consumption during evaluation_3d, as inference_3d.py
does not have these args.
--cfg ./config/train_ds3.yaml # enable Deepspeed ZeRO stage3
--use_flash_attn # enable flash attention
--use_xformers # enable xformers
Best!
Xiaolong
Sorry for late reply. Recently we refactor our code and now we evaluate trained MLLMs via CheF.
To decrease computation cost during evaluation of LAMM, now we support LightLLM, which contains flash attention and other techniques. Therefore, there is no need to declare flash attention and xformer during inference.
In fact, flash-attn and xformers currently only works during training and only lightllm works in inference.
For your reference, you can add following line in evaluation configuration file:
use_lightllm: True
Note, lightllm mainly aims at inference, so during training, you might add these two lines to your config:
use_flash_attn: True
use_xformers: True
The flags are supported in LAMM models, and maybe prompted to other models accordingly.
This issue will be closed. Please reopen it if any further questions.