Skip to content

fix: align Qwen3 router precision and P2P staging (after cutoff) - #3697

Open
yueming-yuan wants to merge 2 commits into
mainfrom
yueming/qwen3-router-post-cutoff
Open

yueming-yuan wants to merge 2 commits into
mainfrom
yueming/qwen3-router-post-cutoff

Conversation

@yueming-yuan

@yueming-yuan yueming-yuan commented Sep 24, 2026 •

Copy link
Copy Markdown
Collaborator

Defer the Qwen3-MoE router precision change from #3584 until after the v0.1.1 cutoff. Requires the FP32 router support retained in SGLang #40997.

  • Match inference router dtype to the requested training dtype, preserving explicit inference overrides.
  • Construct P2P CPU staging models from engine configuration so their parameter dtypes and raw transfer layouts agree.

Two files, +7/-1; the moved patches are unchanged. This affects full-model Qwen3 as well as LoRA. The SGLang expert-index bounds and complete fused shared-expert fixes are in cutoff PR #41074.

Validation: full local pre-commit passed. Earlier Qwen full-model CI exercised these patches in #3584; the CPU P2P reproduction distinguishes the old BF16 staging mismatch from exact FP32 transfer. The new branch/dependency combination has not run GPU CI. P2P CI also logs transfer failures on unmodified main; final weight impact is not established.

Model tracking and release split.

ci-sglang-pr: 0a08c70dc89e196d0ebc59084af24c10999b6934
ci-megatron-pr: 651dd728e3c1f763b5ba3a8de4ac0d01f70e621c

@claude claude Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

Tip: disable this comment in your organization's Code Review settings.

This branch has not been deployed

No deployments
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