Skip to content

[MoonEP] Log host capacity utilization - #35897

Draft
musclemuller wants to merge 1 commit into
sgl-project:moonepfrom
wirybeaver:codex/moonep-capacity-logging
Draft

musclemuller wants to merge 1 commit into
sgl-project:moonepfrom
wirybeaver:codex/moonep-capacity-logging

Conversation

@musclemuller

@musclemuller musclemuller commented Aug 21, 2026

Copy link
Copy Markdown

Summary

  • add per-dispatch DEBUG key=value logs for phase, rank, layer, token capacity, padding, utilization, and static padding ratio
  • thread MoeRunnerConfig.layer_id only into the MoonEP inner dispatcher and reject missing layer identity
  • preserve the existing over-capacity ValueError without emitting a normal capacity record

Scope

This is the independent host-side capacity-observability slice from #35783. It is based directly on the moonep branch and does not depend on #35834 or #34874.

Duplicated-expert statistics, prefetch volume, metrics aggregation, symmetric-pool lifecycle, and GPU/protocol work remain follow-ups.

Testing

  • 13 focused CPU tests in test_moonep_buffer.py
  • Ruff 0.15.1: F401, F821, UP037
  • isort 7.0.0
  • Black 26.1.0
  • registered-test registry check

Part of #35783.


CI States

Latest PR Test (Base): ❌ Run #32511192256
Latest PR Test (Extra): ❌ Run #32511192175
Latest PR Test (AMD ROCm 10): ❌ Run #32511192671

@musclemuller

Copy link
Copy Markdown
Author

CI is currently blocked before tests by the repository-wide MIN_BASE_SHA gate, which requires main@aa3f76679930. This PR intentionally targets the moonep staging branch at 1334810, as requested, so the branches are expected to diverge. Would you prefer to apply the bypass-maintenance label to this PR, or update the moonep branch? I have not rebased onto main because that would change the intended base and scope.

@Jiminator
Jiminator deleted the branch sgl-project:moonep September 14, 2026 04:42
@Jiminator Jiminator closed this Sep 14, 2026
@hnyls2002 hnyls2002 reopened this Sep 14, 2026
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.

3 participants