Skip to content

changelog and docs: four Unreleased lines corrected, a duplicate removed, and the speed-up and Flash Next recipe notes said for their setup - #2371

Merged
ashhart merged 3 commits into
mainfrom
lead/changelog-q32
Nov 7, 2026
Merged

ashhart merged 3 commits into
mainfrom
lead/changelog-q32

Conversation

@ashhart

@ashhart ashhart commented Nov 6, 2026 •

Copy link
Copy Markdown
Owner

All 55 bullets under ## Unreleased, checked against the code on main. 50 match it. This fixes four lines and removes one duplicate:

Line What was wrong Evidence
Clef-Flash on HIP Two bullets for one change. The earlier one uses the wording from before the decisions key and leaves #2048 out server/cli_capabilities.zig:30, docs/api.md:184; the later bullet stays
Metal prefill widths (#2293) "every Metal text family", but Flash Next's engine reports none no prefill_widths under zig/src/families/flashnext/
-Dnative-fp-decode "bit-equal to the default build" holds except for a weight carrying an e4m3 NaN code, which only a broken checkpoint has docs/kernel-contract.md:392-396
The cache tool's test (#2277) Its trigger list leaves out the tool's own test and build.zig.zon .github/workflows/no-gpu.yml:93-94

Credits, PR references and "part of" wording all match the house style, and nothing released sits under Unreleased. Docs only.

One more, from a second commit: the #2237 line now says "on one Mac". docs/speed-up-mode.md keeps its slowdown note, now as a measurement on two M5 Ultras in speed-up mode, which hasn't been rerun on a pair since. #2237 measured one Mac.

A third commit, on the Flash Next recipe: its EXL3 test and its draft-vocab build both say they run from a v0.6.6 checkout, and the draft-vocab pin is named as v0.6.6's tools/draft_vocab.py. The vocab command now puts its glob before --added-tokens, so the flag no longer reads as taking a value. It has the same arguments and gives the same output. Checked against the tag by the Flash Next owners.

@ashhart ashhart changed the title changelog: four Unreleased lines corrected against the code, and a duplicate removed changelog and docs: four Unreleased lines corrected, a duplicate removed, and the speed-up and Flash Next recipe notes said for their setup Nov 6, 2026
@ashhart

ashhart commented Nov 6, 2026

Copy link
Copy Markdown
Owner Author

/atc exchange 842bd3fb5f3a2a864c49f30cac939b32eb8edab8dc613bc22293177c1ee044c0

- drop the duplicate Clef-Flash on HIP bullet; keep the one carrying the
  current `decisions`-key wording and the full PR list
  (zig/src/server/cli_capabilities.zig:30, docs/api.md:184)
- Metal prefill widths: Flash Next's engine reports none, so 'every Metal
  text family' gains the exception (no prefill_widths reference under
  zig/src/families/flashnext/)
- native FP decode: bit-equality holds unless a weight carries an e4m3
  NaN code, which only a broken checkpoint has
  (docs/kernel-contract.md:367-370)
- cache-tool test trigger list: the tool's test and build.zig.zon also
  trigger it (.github/workflows/no-gpu.yml:93-94)
…they run from v0.6.6, and the vocab command puts its glob first
@ashhart
ashhart force-pushed the lead/changelog-q32 branch from 114dc8b to 230f36f Compare November 7, 2026 02:56
@ashhart
ashhart merged commit e41d195 into main Nov 7, 2026
4 of 7 checks 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