Can someone merge A-DMA to accelerate finetuning?

[ILG2021]

Checks

• This template is only for feature request.
• I have thoroughly reviewed the project documentation but couldn't find any relevant information that meets my needs.
• I have searched for existing issues, including closed ones, and found not discussion yet.
• I am using English to submit this issue to facilitate community communication.

1. Is this request related to a challenge you're experiencing? Tell us your story.

The finetune is very slow.

2. What is your suggested solution?

The code is here:
https://github.com/ZhikangNiu/A-DMA

3. Additional context or comments

@ZhikangNiu

4. Can you help us with this feature?

• I am interested in contributing to this feature.

[SWivid]

We may not have time recently, this work is as you said open source, if you are interested you can try it directly @ILG2021