Skip to content

Fix DSV3 speed degradation after adapting to Paddle 3.3.0#3823

Open
lshpku wants to merge 1 commit intoPaddlePaddle:release/v1.0from
lshpku:fix-dsv3-speed-issue
Open

Fix DSV3 speed degradation after adapting to Paddle 3.3.0#3823
lshpku wants to merge 1 commit intoPaddlePaddle:release/v1.0from
lshpku:fix-dsv3-speed-issue

Conversation

@lshpku
Copy link
Contributor

@lshpku lshpku commented Feb 4, 2026

PR types

Performance optimization

PR changes

Models

Description

修复 DeepseekV3 在升级到 Paddle 3.3.0 后出现的性能问题,包括:

  1. 关闭默认的 barrier_ep,这个功能会阻塞 overlap
  2. fused_rms_norm_ext_grad 的 amp 有问题,需要设置 auto_cast(False)
  3. rearrange_kv 没有开启动转静,需要手动开启

@paddle-bot
Copy link

paddle-bot bot commented Feb 4, 2026

Thanks for your contribution!

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.

1 participant