Skip to content

[Spec] Publish the final multi-layer EAGLE shared-read event - #36752

Merged
Qiaolin-Yu merged 3 commits into
sgl-project:mainfrom
paulzhang-tm:codex/mtp-final-shared-read-event
Sep 2, 2026
Merged

Qiaolin-Yu merged 3 commits into
sgl-project:mainfrom
paulzhang-tm:codex/mtp-final-shared-read-event

Conversation

@paulzhang-tm

@paulzhang-tm paulzhang-tm commented Aug 27, 2026

Copy link
Copy Markdown
Contributor

Motivation

Multi-layer EAGLE does not publish the final draft graph's shared-read completion event, so the overlap scheduler falls back to a broader forward-stream fence.

Modifications

  • Publish shared_read_done_event after the final draft step.
  • Cover both per-step and single-graph replay.
  • Use the final draft runner as the scheduler's shared-read event owner.

Tests

  • Added CPU tests for final-runner selection and final-step event publication.

CI States

Latest PR Test (Base): 🚫 Run #33448171148
Latest PR Test (Extra): ❌ Run #33448170895
Latest PR Test (AMD ROCm 7.2): ❌ Run #33448171383

Record shared-read completion after the final draft graph replay and expose the final draft runner as the scheduler WAR barrier owner. Cover both per-step and single-graph replay paths.
@Qiaolin-Yu
Qiaolin-Yu merged commit 3fa6b86 into sgl-project:main Sep 2, 2026
223 of 262 checks passed
StevenChenSE pushed a commit to StevenChenSE/sglang that referenced this pull request Sep 6, 2026
paulzhang-tm added a commit to paulzhang-tm/sglang that referenced this pull request Sep 9, 2026
paulzhang-tm added a commit to paulzhang-tm/sglang that referenced this pull request Sep 9, 2026
Resolve shared attention and conv-slot mappings into draft-owned buffers before target verification. Preserve the final-draft-runner fallback introduced in sgl-project#36752.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants