Skip to content

Conversation

@Xing-lil
Copy link
Contributor

@Xing-lil Xing-lil commented Jun 6, 2025

PR Category

Auto Parallel

PR Types

Bug fixes

Description

  • Fix pp_degree > 1 bug for sharding overlap in auto dy.
  • Accuracy alignment for FLAGS_enable_sharding_overlap = true/false in pp_degree > 1.
  • [TODO] Add ut to check sharding overlap always occurs at the last backward of each step.

Pcard-70448

@paddle-bot
Copy link

paddle-bot bot commented Jun 6, 2025

你的PR提交成功,感谢你对开源项目的贡献!
请关注后续CI自动化测试结果,详情请参考Paddle-CI手册
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

if "pp" in fleet.auto.get_mesh().dim_names:
param_group_len = (
param_group_len * fleet.auto.get_mesh().get_dim_size("pp")
)
Copy link
Contributor

Choose a reason for hiding this comment

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

流水线并行的 dim_name 不一定是 pp

Copy link
Contributor Author

Choose a reason for hiding this comment

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

这里是指还有 vpp_degree 吗?

Copy link
Contributor

Choose a reason for hiding this comment

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

不是,pp 是自定义的一个名称,虽然默认用 "pp" 来命名 流水线并行,但也可以用其它名称。因为框架其它多处使用该命名,可先保持一致,后续再统一。

Copy link
Contributor Author

Choose a reason for hiding this comment

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

好的 感谢

@Xing-lil Xing-lil merged commit ca1aa82 into PaddlePaddle:develop Jun 9, 2025
49 of 50 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

2 participants