Skip to content

fix: do not default fair mixed-prefill on TP3 or TP4 - #187

Merged
MiaAI-Lab merged 2 commits into
mainfrom
fix/tp34-mixed-prefill-off
Sep 15, 2026
Merged

MiaAI-Lab merged 2 commits into
mainfrom
fix/tp34-mixed-prefill-off

Conversation

@MiaAI-Lab

Copy link
Copy Markdown
Owner

Summary

  • PR feat(scheduler): fair v5 mixed-prefill as the TP2 default #186 merged fair as the default on every launcher. TP=3 then crashed during post-ready warmup on this kit; fair is unmeasured on TP=3/TP=4.
  • Restore TP=3 0 (start-tp3.sh, .env.tp3.example) and TP=4 skip (start-tp4.sh, .env.tp4.example). Both examples set the value explicitly so they override a fair .env.
  • TP=2 stays fair. Opt in on 3/4 with GLM53_MIXED_PREFILL_CHUNK=fair.

Test plan

  • python3 tests/test_numeric_config.py
  • Confirm a fresh .env.tp3 / .env.tp4 copy from the examples does not start with fair

Made with Cursor

MiaAI-Lab and others added 2 commits September 15, 2026 13:28
fair v5 stays the TP=2/TP=4 default. TP=3 goes back to CHUNK=0 in
start-tp3.sh and .env.tp3.example so a copied example overrides a fair
.env. Opt in with GLM53_MIXED_PREFILL_CHUNK=fair.

Co-authored-by: Cursor <cursoragent@cursor.com>
fair v5 stays the TP=2 default only. TP=4 goes back to skip in
start-tp4.sh and .env.tp4.example so a copied example overrides a fair
.env. Opt in with GLM53_MIXED_PREFILL_CHUNK=fair.

Co-authored-by: Cursor <cursoragent@cursor.com>
@MiaAI-Lab
MiaAI-Lab merged commit b517eef into main Sep 15, 2026
plotarmordev added a commit to nood-co1/GLM-5.3-Flash-EXL3-2x-DGX-Sparks that referenced this pull request Sep 15, 2026
…ault

Carries the reviewed gate-v2 direction from PR MiaAI-Lab#80 onto current main's versioned installer; fair v5 (the TP=2 default since MiaAI-Lab#186/MiaAI-Lab#187) is untouched.

- overlay/patch_scheduler_decode_floor.py: installer v6. skip/cap only, both features disabled at 0. WARM_TOKENS admits a request whose uncached remainder fits instead of holding it; MAX_WAIT_MS releases a request the skip hold starved, under LATE_CAP tokens/step (time-to-first-service bound, not TTFT). Explicit cap keeps its cap; v1-v5 images are unpatched and re-patched with the same fail-closed drift refusal, and a v6 image is verified by round-trip before its marker is trusted.
- start.sh: knobs default 0/0/512, validated (0..1000000, 0..600000, 64..8192), forwarded on every rank and logged. No hand-rolled _cli_* capture: main's generic _caller_overrides already gives caller exports precedence over .env, including explicitly empty ones.
- .env.example / README.md / CHANGELOG.md: documented as opt-in with the crawl caveat.
- tests: gate behavior (default-off, zero-disables, both transitions, requeue survival, explicit-cap precedence, fair/off isolation, bounds and invalid-value fallback), knob bounds, and the v0-v5 migration matrix plus marker-only refusal. The live validator refuses to score a disabled gate instead of reporting a false pass or failure.
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