Skip to content

Conversation

@jeff41404
Copy link
Contributor

@jeff41404 jeff41404 commented Feb 20, 2025

PR Category

Auto Parallel

PR Types

Bug fixes

Description

pcard-89456

During the validation of the multimodal Qwen2VL, the dynamic to static encountered an error when running operator of flash_attn_unpadded in pir mode. The reason was that the input of operator of flash_attn_unpadded lacked distributed attributes. Tracing along the calculation graph, it was found that the distributed attributes would be lost after pd_op.expand. Further examination of the code reveals that ExpandOp (ExpandOp::InferMeta, etc.) under PIR is handwritten, and there is no processing for distributed attributes in the handwritten logic (regular OP generated code will have relevant processing). The reason for handwriting is described in PR #58767 , so this PR will add processing for distributed attributes in ExpandOp::InferMeta.

image
image
image

@paddle-bot
Copy link

paddle-bot bot commented Feb 20, 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.

Copy link
Contributor

@From00 From00 left a comment

Choose a reason for hiding this comment

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

LGTM

@jeff41404 jeff41404 merged commit 496947d into PaddlePaddle:develop Feb 21, 2025
35 checks passed
@jeff41404 jeff41404 deleted the add_call_to_spmd_in_expand_infermeta_when_use_pir branch February 21, 2025 03:15
Enigmatisms pushed a commit to Enigmatisms/Paddle that referenced this pull request Mar 6, 2025
YqGe585 pushed a commit to YqGe585/Paddle that referenced this pull request May 7, 2025
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