mlsys-workflow
GitHub用于规划MLSys会议提交全流程的项目管理技能，涵盖从选题、实验、撰写到反辩及最终发布的端到端流程。重点解决GPU资源稀缺问题，提供基于2026年日程的倒推计划与团队协作机制。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill mlsys-workflow -g -y
SKILL.md
Frontmatter
{
"name": "mlsys-workflow",
"description": "Use when planning an MLSys submission cycle end to end, from venue-fit and track choice through evaluation freeze, the October deadline, the four-day January response, notification, camera-ready, the March artifact-evaluation submission, and the May conference, with backward planning built around scarce GPU time and team ownership."
}
MLSys Workflow
Use this as the project-management skill for a Conference on Machine Learning and Systems submission. MLSys runs one cycle per year, so a missed gate costs twelve months or a reroute to another venue. All dates below are the verified 2026-cycle timetable (access 2026-07-08) — treat them as shape, and replace every one with the current mlsys.org dates page before committing a plan.
MLSys is a conference with rotating per-edition leadership (2026: General Chair Luis Ceze; Program Chairs Zhihao Jia and Aakanksha Chowdhery) and no APC — proceedings publish open on proceedings.mlsys.org, and the money side is registration, where accepted-paper authors get a two-week reserved-ticket window.
The 2026-cycle shape
| Gate | 2026 anchor | What it demands |
|---|---|---|
| CFP and track choice | Summer 2025 (CFP publicized ~August) | Research vs industrial routing decided |
| Paper + appendix deadline | October 30, 2025 | Frozen evaluation, blinded PDF, separate appendix upload |
| Reviews released | January 12, 2026 | All authors on deck immediately |
| Author response due | January 16, 2026 | Four days — pre-planned, not improvised |
| Notification | January 25-26, 2026 | Branch: camera-ready+AE, or postmortem+reroute |
| Artifact evaluation submission | March 8, 2026 | Badge-ready package + Artifact Appendix |
| AE window | March 8 - April 8, 2026 | One author on call for evaluator questions |
| Conference | May 18-22, 2026 (Bellevue) | Registration inside reserved window, talk, travel |
Backward plan from the paper deadline
| Weeks out | Systems-paper milestone |
|---|---|
| 12+ | Bottleneck measurement done; mechanism designed; venue/track locked |
| 10 | Baselines installed, version-pinned, tuning protocol agreed |
| 8 | End-to-end prototype produces its first honest comparison |
| 6 | Full experiment matrix launched — GPU allocation is the critical path from here |
| 4 | Ablations and sensitivity sweeps done; figures generated from logs |
| 3 | Complete draft in the official two-column style; internal review by one systems and one ML reader |
| 2 | Evaluation freeze; only reviewer-anticipation runs remain |
| 1 | Anonymity sweep, appendix assembly as separate file, bibliography full-author check |
| 0 | Upload paper + appendix; verify both files from a logged-out view |
The distinctive constraint versus ML-venue planning: the evaluation competes for shared hardware. Reserve cluster time for weeks 6-4 when the matrix runs, and again for the January response window, where two rebuttal experiments may need machines on two days' notice.
Standing team rituals
Weekly (weeks 12->2):
- experiment-matrix review: rows launched / landed / failed, GPU budget burned
- claims board: abstract claim -> supporting figure -> status (red/yellow/green)
- baseline watch: new releases of compared systems since last week?
Once (week 3): mock review — one systems-culture and one ML-culture reader,
scored on workload realism, baseline fairness, attribution, and tails.
Response-week protocol (pre-agreed in December):
- who reads reviews hour 0, who owns machines, who drafts, who signs off
Failure modes by stage
- Week 6 matrix slip: the sweep launches late, so ablations get cut — and ablation absence is precisely what MLSys reviews punish. Cut breadth (hardware variants) before cutting attribution (ablations).
- Baseline rot: a compared system ships a major release in September; deciding whether to re-run everything at week 4 needs a pre-agreed rule ("re-run headline comparisons if the release notes claim >10% on our metric").
- Response-window surprise: the four-day window found nobody available in January. It is on the calendar from the day of submission.
- Post-acceptance pile-up: camera-ready, AE package, and conference logistics all land in February-March; assign three different owners at notification time.
- Reroute paralysis after rejection: the systems-venue calendar (OSDI/SOSP/NSDI/ ASPLOS/ATC deadlines) should be listed in the postmortem doc before the decision arrives.
The reroute decision, prepared in advance
Because the cycle is annual, "revise and resubmit to next MLSys" costs a year. Write the decision rule before the notification arrives:
| Rejection pattern | Indicated move |
|---|---|
| Workload/baseline objections, mechanism praised | Strengthen evaluation, resubmit to next MLSys — the reviews are a work list |
| "Insufficient novelty, solid engineering" | Industrial track next cycle, or a systems venue that weighs deployment insight |
| "Wrong audience" signals from both cultures | Re-read mlsys-topic-selection; the project may be OSDI/ASPLOS/NSDI-shaped |
| Split reviews, one culture convinced | Rewrite for the unconvinced culture; the paper, not the project, missed |
Maintain the alternative-deadline list (OSDI, SOSP, NSDI, ASPLOS, ATC, EuroSys, SC) with dates in the project doc from week 12, so a January rejection flows into a concrete February plan instead of a stall.
Ownership map (minimum viable)
- Evaluation owner: matrix, logs, figure regeneration.
- Baseline owner: versions, tuning parity, release watch.
- Writing owner: 10-page budget, style-kit compliance, claims board.
- Compliance owner: anonymity, appendix upload, OpenReview fields, deadlines.
- (Post-acceptance) AE owner: package, badges, evaluator on-call.
One person may hold two hats; no hat may be unassigned. Industrial-track submissions add a sixth hat that academic teams forget: the internal-approval owner, who clears company names, trace releases, and benchmark numbers with legal/PR — a process whose latency routinely exceeds the writing itself, so it starts at week 10, not week 1.
Cycle-volatility warning
Every date above, the track structure, the response-window length, and the AE mechanics are 2026 facts. The 2027 CFP was not yet visible at access time (待核实, all of it) — rebuild this table from mlsys.org before planning a real cycle.
Output format
[Current stage] fit / building / evaluating / writing / submitted / response / decided / AE
[Next official gate] <date + source URL>
[Critical path] <three tasks, with GPU-time dependencies flagged>
[Claims board] <claim -> figure -> status>
[Ownership] <evaluation/baseline/writing/compliance/AE -> person>
[Risk register] <matrix slip / baseline rot / response availability / logistics pile-up>
Version History
- 9f86f09 Current 2026-07-19 17:00
