Skip to content

Plan TP=1 with EP>1 simulation fix - #303

Draft
tianhao909 with Copilot wants to merge 1 commit into
masterfrom
copilot/fix-simulation-failure-tp-1-ep-greater-than-1
Draft

tianhao909 with Copilot wants to merge 1 commit into
masterfrom
copilot/fix-simulation-failure-tp-1-ep-greater-than-1

Conversation

Copilot AI commented Aug 27, 2026 •

Copy link
Copy Markdown

Simulations with TP=1 and EP>1 lose EP/DP_EP group membership and suppress forward/input-gradient EP collectives.

  • Root cause
    • SimCCL derives EP/DP_EP groups from TP groups, but omits singleton TP groups.
    • SimAI applies the empty TP dimension mask to EP collectives.
  • Proposed changes
    • Represent TP=1 ranks as singleton TP groups in SimCCL.
    • Use data-parallel dimensions for EP forward/input-gradient collectives.
    • Preserve existing TP=1 communicator suppression and TP>1 behavior.
  • Scope
    • Analysis and implementation plan only; this PR currently contains no code changes.

@CLAassistant

Copy link
Copy Markdown

CLA assistant check
Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
You have signed the CLA already but the status is still pending? Let us recheck it.

Copilot AI linked an issue Aug 27, 2026 that may be closed by this pull request
Copilot AI changed the title [WIP] Fix simulation failure when TP=1 and EP>1 Plan TP=1 with EP>1 simulation fix Aug 27, 2026
Copilot AI requested a review from tianhao909 August 27, 2026 10:58
@gabrielecastellano

Copy link
Copy Markdown
Contributor

The required fix is in the submodule, not in this repository. I have opened PR aliyun/SimCCL#5 with the fix, and it is currently awaiting merge.

Please do not implement an alternative fix in the main repository. Once that PR is merged, update the submodule pointer in this PR to the merged commit containing the fix.

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.

TP=1,EP>1时仿真失败

4 participants