Skip to content

[vLLM][MoE] Avoid top-k boundary ties causing test failures#7580

Draft
sarah12121212 wants to merge 5 commits into
moe2-workspacemanager-xpu-initfrom
moe2-avoid-topk-boundary-ties
Draft

[vLLM][MoE] Avoid top-k boundary ties causing test failures#7580
sarah12121212 wants to merge 5 commits into
moe2-workspacemanager-xpu-initfrom
moe2-avoid-topk-boundary-ties

Conversation

@sarah12121212

Copy link
Copy Markdown
Contributor

Summary

test_fused_moe generates random BF16 routing scores for top-k selection. Different routing implementations may resolve scores tied at the selection boundary using different expert IDs, causing numerical failures unrelated to the correctness of the MoE kernels.

This PR:

  • Detects if the kth and k+1th routing scores are equal.
  • Rerolls rows containing a boundary tie.
  • Fails if a tie-free input cannot be generated after 10 attempts.
  • Unskips 30 affected xe2 cases and 28 corresponding default cases.

Validation

Full vLLM MoE suite on B580 (Tested locally):

926 passed
2,061 skipped
0 failed

Full vLLM MoE suite on max 1100 (Tested locally):

Blocked by #7462 (it needs to be merged first)

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