Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Dist dialect] Add pass for dist_reshape and nd_mesh_all2all in PIR #70027

Closed
wants to merge 7 commits into from

Conversation

pkuzyc
Copy link
Contributor

@pkuzyc pkuzyc commented Dec 6, 2024

PR Category

Auto Parallel

PR Types

New features

Description

Pcard-67164
Add pass for dist_reshape and nd_mesh_all2all in PIR.

Major changes:

  1. dist_reshape pass: replace the dist op "dist_op.dist_reshape" with reshape op for further execution.
  2. nd_mesh_all2all: use all-to-all communication when performing shard-to-shard reshard on multi-dimensional process mesh.
    2.1 This PR add a s_to_s reshard function in PIR.
    2.2 This PR modify the reshard pass, when resharding from [Replicate(), Shard(1)] to [Replicate(), Shard(0)], it will use all-to-all communication.

Minor:

  1. Fix the bug in setting no_need_buffer values.
  2. Fix the bug about getting wrong communication group in parsing communication context.
  3. Fix the bug in PIR sub_to_global mesh reshard function.

Copy link

paddle-bot bot commented Dec 6, 2024

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.

@pkuzyc pkuzyc force-pushed the pir_all2all_moe_pass branch from 95cc071 to a2a7c36 Compare December 6, 2024 13:45
Copy link

paddle-ci-bot bot commented Dec 14, 2024

Sorry to inform you that a2a7c36's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

@pkuzyc pkuzyc force-pushed the pir_all2all_moe_pass branch 4 times, most recently from f9a1b41 to d51cfe1 Compare December 20, 2024 07:56
@pkuzyc pkuzyc force-pushed the pir_all2all_moe_pass branch from d51cfe1 to d198142 Compare December 25, 2024 11:45
@pkuzyc pkuzyc marked this pull request as draft January 7, 2025 03:32
@pkuzyc pkuzyc force-pushed the pir_all2all_moe_pass branch from e60c1d9 to af7d730 Compare January 7, 2025 03:52
@pkuzyc pkuzyc removed the request for review from Xreki January 7, 2025 08:35
Copy link

paddle-ci-bot bot commented Feb 10, 2025

Sorry to inform you that 0953f87's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

@pkuzyc pkuzyc force-pushed the pir_all2all_moe_pass branch 5 times, most recently from 6926b44 to 606aaea Compare February 13, 2025 08:31
@pkuzyc pkuzyc force-pushed the pir_all2all_moe_pass branch from 606aaea to 8c22dfd Compare February 13, 2025 13:29
Copy link

paddle-ci-bot bot commented Feb 22, 2025

Sorry to inform you that 8c22dfd's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

@pkuzyc
Copy link
Contributor Author

pkuzyc commented Mar 8, 2025

all changes except the dist_reshape pass have been moved to #71264 and #71320

@pkuzyc pkuzyc closed this Mar 8, 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.

1 participant