Skip to content

ci: add cancel pr workflow - #1069

Merged
zhyncs merged 1 commit into
sgl-project:mainfrom
zhyncs:cancel
Aug 13, 2024
Merged

zhyncs merged 1 commit into
sgl-project:mainfrom
zhyncs:cancel

Conversation

@zhyncs

@zhyncs zhyncs commented Aug 13, 2024

Copy link
Copy Markdown
Contributor

Motivation

When a PR has already been approved and merged, it is unnecessary for the workflow that is still running on this PR.

Modification

Briefly describe the changes made in this PR.

Checklist

  1. Ensure pre-commit pre-commit run --all-files or other linting tools are used to fix potential lint issues.
  2. Confirm that modifications are covered by complete unit tests. If not, please add more unit tests for correctness.
  3. Modify documentation as needed, such as docstrings or example tutorials.

@zhyncs
zhyncs enabled auto-merge (squash) August 13, 2024 06:36
@zhyncs
zhyncs disabled auto-merge August 13, 2024 06:36
@zhyncs
zhyncs merged this pull request into sgl-project:main Aug 13, 2024
@zhyncs
zhyncs deleted the cancel branch August 13, 2024 06:37
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 1, 2026
…neither idle nor a stall

1068cap (2026-09-01): 4 group flip pairs, the SAME 3-request cohort seam-
re-prefilled from zero in every TP window (442840 tok total), zero decode
batches ever built. Two arm faces of one booking blind spot -- the policy
has no notion of 're-prefilling toward decode':
(A) windows 1+4: 'pending prefill > 0 ... nothing decoding' armed ~13 s in,
    while the cohort sat queued (bundle_is_mid_flight correctly answers
    False for queued-retracted per W37-E -- but not-decoding is not
    not-owed-this-layout's-service).
(B) windows 2+3: 'decode bundle STALLED, set has not shrunk for 51.8s'
    fired at a member whose prefix was actively growing -- the progress
    clock knew only bs-shrink (:4155).

One mechanism, two sites, no new ledger:
- observe_idle: sinking seam_cohort_pending_tokens credits the bundle
  progress clock (state marker last_seam_cohort_pending, phase-entry reset).
- the purity/pending tp_to_pp arm holds while seam_cohort_dwell_active()
  (same sgl-project#1032 FIX 1 bound: lapses after SEAM_COHORT_DWELL_ROUNDS, so it can
  never hold forever); named holding line is the metal instrument.

Desk-proven by execution: clock clause truth table (sink=progress,
flat/rise=not), dwell predicate arms both ways. Metal falsifier: driver
load must reach Decode batch > 0 with flips still reconciling.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 1, 2026
…s the pp-ward arm

Pins for 9295258, red-by-name on revert of either half:
- sinking seam_cohort_pending_tokens stamps last_bundle_progress_at with
  running_bs flat; flat and RISING do not (truth table executed, the 1068cap
  windows-2+3 false-STALL inverted).
- the cohort marker resets at phase entry, so a cutover's own zeroing can
  never read as service progress.
- _decide_from_load consults seam_cohort_dwell_active() on the pp-ward path
  and refuses with the named line 'seam cohort in service (sgl-project#1069)'; the
  dwell predicate lapses at SEAM_COHORT_DWELL_ROUNDS (bounded, behavioral).

Suite: 19/19 green (PYTHONPATH=<worktree>/python, htsglang-gpu venv).
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