Skip to content

Conversation

@DongBaiYue
Copy link
Contributor

@DongBaiYue DongBaiYue commented Nov 10, 2025

  • Add global param to refined_recompute
  • Optimize ERNIE-4.5-VL SFT config for P800 96GB

refined_recompute: "global:x"的含义:n层ErnieDecoderLayer中有x层不重计算,n-x层重计算。默认所有层都会进行重计算。
n由模型结构确定,在4.5 VL中,n=14。P800上设置global:3时可以平衡显存和性能。

@paddle-bot
Copy link

paddle-bot bot commented Nov 10, 2025

Thanks for your contribution!

Copy link
Contributor

@yongqiangma yongqiangma left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

Copy link
Collaborator

@nepeplwu nepeplwu left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@PaddlePaddle PaddlePaddle locked and limited conversation to collaborators Nov 18, 2025
@PaddlePaddle PaddlePaddle unlocked this conversation Nov 18, 2025
@zjjlivein zjjlivein closed this Nov 18, 2025
@zjjlivein zjjlivein reopened this Nov 18, 2025
@DongBaiYue DongBaiYue closed this Nov 18, 2025
@DongBaiYue DongBaiYue reopened this Nov 18, 2025
@DongBaiYue DongBaiYue closed this Nov 18, 2025
@DongBaiYue DongBaiYue reopened this Nov 18, 2025
* Add `global` param to `refined_recompute`
* Optimize ERNIE-4.5-VL SFT config for P800 96GB
@DongBaiYue DongBaiYue force-pushed the global_refined_recompute branch from 72de648 to 72fb925 Compare November 18, 2025 07:10
@PaddlePaddle PaddlePaddle locked and limited conversation to collaborators Nov 27, 2025
@PaddlePaddle PaddlePaddle unlocked this conversation Nov 27, 2025
@DongBaiYue DongBaiYue marked this pull request as draft November 27, 2025 07:26
@DongBaiYue DongBaiYue marked this pull request as ready for review November 27, 2025 07:26
@DongBaiYue DongBaiYue merged commit 2590403 into PaddlePaddle:develop Dec 1, 2025
9 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants