Skip to content

docs: register and task doc — gemma4 image chunk vs. batch fixed in the fork (ADR 0036) - #322

Merged
glennneuber merged 4 commits into
mainfrom
docs/adr0036-register
Sep 19, 2026
Merged

glennneuber merged 4 commits into
mainfrom
docs/adr0036-register

Conversation

@glennneuber

@glennneuber glennneuber commented Sep 18, 2026 •

Copy link
Copy Markdown

The register row and the task-doc paragraph that still read "measured, decision pending" now point at ADR 0036 / #320, merged tonight.

Second commit — the deploy record. ADR 0036 went live on :11497 at 22:30:18 as 0.34.1-dynres-16-g16649e8 (the v0.34.1 release image with a Go-only binary swap from main 16649e8; no native input changed). The README's Deployed line moves to that stamp and the task doc's deploy section records the swap, its rollback, and the verification against production itself: gemma4:31b launched at -b 2048 -ub 2048 and decoded its 1100-token image as batch 1/1 (scene IoU 0.963), and the MLX runner answered a plain request after its cold JIT. The preflight matrix keeps the tag's full-run stamp, since the swap changes nothing the run measures natively.

🤖 Generated with Claude Code

glennneuber and others added 4 commits September 18, 2026 21:59
…ed in the fork (ADR 0036, #320)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ME Deployed line, task doc deploy record

The Go-only swap on the v0.34.1 payload went live 2026-09-18 22:30:18 and was
verified on production itself: gemma4:31b launched at -b 2048 -ub 2048 and
decoded its 1100-token image in one batch (scene IoU 0.963); the MLX runner
still serves (gemma4:12b-nvfp4 answered OK). Old container kept for rollback.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…he mlx-metal promotion

#324 added the Apple Silicon deployment to the same block this branch was
updating with the ADR 0036 rebuild stamp. Both are true and neither replaces
the other: the CUDA host runs 0.34.1-dynres-16-g16649e8 with the drafting knob,
the Apple Silicon host runs 0.34.0-maxusai-8a7ba949 without it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…line

#328 amended the Apple Silicon sentence — the stamp equivalence under ADR 0032,
and the launchd environment does set OLLAMA_MLX_DRAFT_UNDER_GRAMMAR=0 after all.
That correction is kept verbatim; this branch contributes only the CUDA line,
which moves to the ADR 0036 rebuild stamp.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@glennneuber
glennneuber merged commit 6df3ee9 into main Sep 19, 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.

2 participants