Skip to content

[DO NOT MERGE] Add MegaMoE engineering skill - #1111

Draft
GwilliamHu wants to merge 1 commit into
mainfrom
ghu/megamoe-engineering-skill
Draft

GwilliamHu wants to merge 1 commit into
mainfrom
ghu/megamoe-engineering-skill

Conversation

@GwilliamHu

Copy link
Copy Markdown
Member

Important

Personal reference PR for agent reuse. Please do not merge.

Motivation

Preserve the MegaMoE architecture, validation, debugging, performance-tuning,
environment, and historical-evidence knowledge needed to work safely across
FlyDSL, AITER, MORI, ATOM, SGLang, and the customer FlyDSL_universe branch.

Technical Details

  • Add a progressively disclosed megamoe-engineering skill with task-based
    routing to focused references.
  • Capture protocol invariants for fixed/compact dispatch, quant/prepare,
    Stage1/Stage2, fanout/dedup, readiness, AOT/preload, CUDA Graph, and capacity.
  • Add trace-led tuning and root-cause workflows, 8-GPU/AOT/serving validation
    gates, model geometry/onboarding guidance, and a dated evidence index.
  • Add a read-only preflight helper that reports source/runtime/GPU ownership
    provenance without initializing the GPU runtime.
  • Link the skill from CLAUDE.md for repository-local discovery.

Test Plan

  • Run the repository agent-document/API checks.
  • Run the skill validator.
  • Syntax-check and execute the read-only preflight helper.

Test Result

  • python3 scripts/check_repo.py: PASS
  • python3 scripts/check_docs_api.py --all: PASS
  • quick_validate.py .claude/skills/megamoe-engineering: PASS
  • bash -n .../scripts/preflight.sh: PASS
  • Preflight end-to-end smoke: PASS

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