Skip to content

Refresh existing serving skills and fix profiling tools - #141

Merged
BBuf merged 56 commits into
mainfrom
codex/refresh-skills-2026-11-08
Nov 9, 2026
Merged

BBuf merged 56 commits into
mainfrom
codex/refresh-skills-2026-11-08

Conversation

@BBuf

@BBuf BBuf commented Nov 8, 2026 •

Copy link
Copy Markdown
Owner

Update the existing serving skills against pinned November 8 source, preserving their workflows and reference structure. Refresh profiler APIs, DS V4.1 fusion/overlap conditions, vLLM compiler dispatch, source paths and KV-memory accounting in the existing documents.

  • Fix capture controls and artifact isolation: support current SGLang Rust ingress, negotiate vLLM session parameters, validate labels/steps, and stop recording after failed probes.
  • Fix target/draft KV accounting per rank and resize round; a 30 GiB target plus a 2 GiB draft now reports 32 GiB.
  • Correct stale PR states and issue-as-PR references in existing catalogs/history. Fix pagination and cache freshness in the maintenance tools.
  • Remove the unused profiling wrapper, low-value tests and duplicated guidance. Research notes, standalone PR-review documents and bulk status inventories are excluded from the final diff.

Validation: 273 CPU tests and 241 subtests passed, one optional test skipped; pre-commit, offline link checks, skill validation and diff whitespace checks passed. No GPU capture, kernel benchmark or model-accuracy rerun was performed.

@BBuf BBuf changed the title Refresh profiler and serving skills from November upstream source Refresh serving skills, profiler capture and PR evidence from current source Nov 8, 2026
BBuf added 27 commits November 8, 2026 22:27
BBuf added 26 commits November 9, 2026 02:25
@BBuf BBuf changed the title Refresh serving skills, profiler capture and PR evidence from current source Refresh existing serving skills and fix profiling tools Nov 9, 2026
- Restore the `#paged-kv-storage-contract` anchor (the dated heading broke four links).
- Rust ingress at the SGLang pin mounts no `/ready` route; `/ready*` is only auth-exempt.
- vLLM #60721 moved `csrc/libtorch_stable/*` back under `csrc/`: update current-state
  rows (status-audited PR rows keep their PR-head paths) and the compile guide.
- Mark the AsyncTP / sequence-parallel rows and `enable_sp` / `fuse_gemm_comms`
  toggles as removed by vLLM #60517.
- Point the refresh commands at the November pins and FlashInfer v0.7.3; record the
  October pins in the October refresh note.
- Restore the guard against retired vLLM partial-prefill flags.
- Open-PR watch REST fallback: send GH_TOKEN/GITHUB_TOKEN when set and explain
  anonymous rate-limit failures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@BBuf
BBuf merged commit a569b78 into main Nov 9, 2026
1 check passed
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