Skip to content

Conversation

@fangfangssj
Copy link
Contributor

@fangfangssj fangfangssj commented Jan 30, 2026

PR Category

Bug Fix

Description

新增前向图捕捉部分,保存训练时前反向图
在sh脚本中新增device rewrite pass 来避免运行出错
参考https://github.com/PaddlePaddle/GraphNet/discussions/610#discussioncomment-15636620的设计

现有问题

  1. 在使用aot的时候,模型中的item会触发faketensor的同步,但是faketensor没有实际数据无法同步
  2. 提取出来的反向图,_scaled_dot_product_efficient_attention_backward会发生报错,无法运行

@paddle-bot
Copy link

paddle-bot bot commented Jan 30, 2026

Thanks for your contribution!

@lixinqi lixinqi merged commit b3b298b into PaddlePaddle:develop Jan 30, 2026
2 of 3 checks passed
@fangfangssj fangfangssj deleted the fix_backward branch January 31, 2026 01:49
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.

2 participants