Repository navigation
Bump gpu-kernels: 744bd713d968 (top-1 batched window from 5 tokens) - #526
Merged
Merged
Conversation
…om 5 tokens gpu-kernels #26 (pi bw220): top-1 MoE models (ZAYA1-8B) take the batched MoE decode kernels only on 5..8-token steps. ZAYA1-8B aggregate decode (90 W cap, PCM pad): N=8 152.8 -> 209.7 (+37%) kept, N=2 136.5 -> 135.6 (-0.7%, was -2.2% with the full window), N=4 unchanged; Coder/27B/4B never present route_count 1, so #25 numbers are untouched. test-backend-ops MUL_MAT_ID 132/132. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
bong-water-water-bong
added
the
hw-gate
PR Reviewer Guide 🔍(Review updated until commit f061e5d)Here are some key observations to aid the review process:
|
|
Persistent review updated to latest commit f061e5d |
bong-water-water-bong
pushed a commit
that referenced
this pull request
Oct 23, 2026
Two review fixes on the candidate section: - `gpu-kernels` pin `abaa6bc484d5` -> `744bd713d968` (#526, the top-1 batched window from 5 tokens), with the batched cells noted as measured one pin earlier, and the engine row now names `7871c71d` — main when the last release-blocking PR merged. Adds the `npu-runtime` pin (`ce9fa59adf95`) so every row the release run records has a row here. - the candidate paragraph pointed at a file that only exists on the release box; public docs name files in this repository only, so it says "the release run" and what that run records. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
bong-water-water-bong
added a commit
that referenced
this pull request
Oct 23, 2026
* docs: the v2026.43 candidate section and the README status table (#525 follow-up) Data for the owner's release today, no prose for its own sake: - `docs/releases.md` gains a **v2026.43 (candidate, 23 Oct)** section: the release sha and every pin, the four artifacts and the six checks as tables (marked *pending* until the release box's `RELEASE-CANDIDATE.md` exists), and a grouped changelog of everything merged since the 18 Oct rehearsal - HRX speed (verify kernels +31-39 % on 27B MTP, split-K decode 23.7 -> 47.1 on Qwen3.5-4B 32K, Flash-Next MoE 5.1 -> 26 tok/s, prompt FA +45 % on Qwen3-4B, warm-up TTFT 2.64 -> 0.80 s), the batched N=8 cells (50.7 / 138.1 / 194.3 aggregate tok/s), correctness (strict placement by default, 43/43 census, #315 resolved on a 37-arm soak, #433 closed on 120 loads), the sealed NPU packaging, the Lemonade one-hop + catalog proofs (22 HRX entries), the OS image's signed UKI, and the tooling (gpu-hold queue, census, kernel runbook). - `README.md` gets the same status as a table under the opening paragraph, with each number's standard (lemonade bench standard or ours) and the power state where it matters. Every line names its PR; every number comes from the auto-opt database rows, the catalog proofs or a closed issue, and carries the standard it was measured under. Nothing is claimed as shipped: the candidate's artifacts and checks stay *pending* until the release run writes them. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * docs: v2026.43 pins after #526, and no internal file references Two review fixes on the candidate section: - `gpu-kernels` pin `abaa6bc484d5` -> `744bd713d968` (#526, the top-1 batched window from 5 tokens), with the batched cells noted as measured one pin earlier, and the engine row now names `7871c71d` — main when the last release-blocking PR merged. Adds the `npu-runtime` pin (`ce9fa59adf95`) so every row the release run records has a row here. - the candidate paragraph pointed at a file that only exists on the release box; public docs name files in this repository only, so it says "the release run" and what that run records. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> --------- Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Moves config/gpu-kernels.config from
abaa6bc484d5to744bd713d968.gpu-kernels #26: top-1 MoE models (ZAYA1-8B) use the batched MoE decode kernels only on 5..8-token steps; 2..4-token top-1 steps return to the previous path. ZAYA1-8B aggregate decode (strixhalo, 90 W cap, PCM pad): N=8 152.8 -> 209.7 tok/s (+37%), N=2 136.5 -> 135.6 (-0.7%), N=4 unchanged, N=1 unchanged; Coder-30B / 27B / 4B untouched (never route_count 1). test-backend-ops MUL_MAT_ID 132/132 vs CPU.
Merge by hand after hw_gate passes (auto-merge off).
🤖 Generated with Claude Code