Pin llama.cpp c9283cf: HRX ZAYA decode 67 -> ~90 tok/s (fork #36-#40)

[bong-water-water-bong]

Moves third_party/llama.cpp 45c1770 -> c9283cf (fork #36-#40). Regenerates the registry, and adds a docs/hrx.md entry.

ZAYA1-8B HRX0 decode 67 -> ~90 tok/s. Each step was measured in interleaved A/B runs on strixhalo:

• site: the 1bit.MONSTER template, a cleaned-up blog, three new posts #36: skips CONTs of fresh contiguous results in zaya.cpp. Bitwise the same output, +2.4%.
• Bump ZINC: zolotukhin/zinc b9123c649f81 #39: the CCA convolution (~35 nodes) as one dispatch: 72 -> 81 tok/s.
• Bump Laya: source 23a17522, model 55cf4c4e #40: the CCA Q/K mixing and per-head norms (~30 nodes) as one dispatch: 81 -> 90 tok/s.
• Bump ZINC: zolotukhin/zinc b9123c649f81 #39 and Bump Laya: source 23a17522, model 55cf4c4e #40 correctness: teacher-forced top-1 vs FP32 stays 69/96. Batch-1 PPL goes 33.31 -> 33.49 -> 33.45 (CPU 34.30).
• Lean option: ROCmFP4 (Vulkan) and ROCmI4 (ROCm, W4A4) behind 1bit serve --lean #37, Bump XDNA: amd/xdna-driver efddccd30568 #38: remove runtime integer division from earlier HRX kernels, which newer Loom rejects.

Each new .loom file carries a reference kernel and check.case differentials that pass under iree-test-loom --sanitizer=access.

RFC #213 decode gate (tg128, HRX = this pin, Vulkan = weekly build, interleaved):

model	HRX	Vulkan	HRX/Vulkan
Qwen3-0.6B Q4_K_M	320.8	356.4	90.0% (at the edge)
Qwen3-Coder-30B-A3B Q4_K_M	90.5	93.8	96.5%
ZAYA1-8B Q4_K_M	89.5	93 (fork Vulkan build)	~96%
Qwen3.8-27B UD-Q4_K_XL	7.25	12.48	58%, far off

The 27B is the open gap. HRX also draws 150 W and peaks at 91 °C, against Vulkan's 91 W and 59 °C. This UD file is mostly Q5_K and IQ4_XS, which miss HRX's Q4_K/Q6_K token-1 fast paths. Routing the generic WMMA matmuls to the decode GEMV only reaches 7.55 tok/s, so it needs its own profile (it's a hybrid with gated delta-net layers). Separately, the HRX IQ4 decode fix (fork branch 1bit/hrx-iq4-fix) isn't merged yet.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 901446d

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 36 - Partially compliant Compliant requirements: The site template and layout are implemented New blog posts are added Blog cleanup with archive.tsv is included copyright.py and pages.yml changes are included Non-compliant requirements: The specific implementation details of the site's CSS, header, and footer are not visible in the diff The exact structure of the pages (home, overview, docs, blog, etc.) is not visible in the diff Requires further human verification: Visual verification of the site's appearance and layout Confirmation that all pages are correctly implemented and styled 40 - Partially compliant Compliant requirements: The Laya commit has been updated in the diff Non-compliant requirements: The scripts/fetch-laya.sh verification and Laya tests are not visible in the diff Requires further human verification: Confirmation that scripts/fetch-laya.sh was run and verified Confirmation that Laya tests were run and passed 39 - Partially compliant Compliant requirements: The zinc commit has been updated in the diff Non-compliant requirements: The build-zinc.sh and test commands are not visible in the diff Requires further human verification: Confirmation that build-zinc.sh was run and tested on Strix Halo Confirmation that the specific model and prompt test was performed

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Clarity The documentation for the ZAYA decode fusions is detailed but could benefit from clearer explanation of how the performance improvements were measured and validated. Specifically, the mention of interleaved A/B runs and the reference to iree-test-loom --sanitizer=access should be clarified to ensure reproducibility. - **ZAYA decode fusions (llama.cpp #36-#40).** ZAYA1-8B HRX0 decode 67 -> ~90 tok/s (Vulkan 93), each step measured in interleaved A/B runs: - #36 skips CONTs of fresh contiguous results in `src/models/zaya.cpp` (views keep theirs: on HRX their consumers take slower strided paths), bitwise the same output: +2.4%. - #39 runs the CCA convolution (Q/K concat, conv-state update, depthwise conv, grouped 2-tap conv and biases, ~35 graph nodes) as one dispatch (`zaya.cca_conv.decode_f32`): 72 -> 81 tok/s. - #40 runs the CCA query/key mixing and per-head norms (~30 nodes) as one dispatch (`zaya.cca_qk_norm.decode_f32`): 81 -> 90 tok/s. #39 and #40 accumulate in their own order, so output is not bitwise the same; teacher-forced top-1 against the FP32 model stays 69/96 and batch-1 perplexity moves 33.31 -> 33.49 -> 33.45 (CPU backend 34.30). Both match decode only (one token, one sequence). #37 and #38 remove runtime integer division from #35's and #32's kernels, which newer Loom rejects. Each new `.loom` file carries a reference kernel and `check.case` differentials that pass under `iree-test-loom --sanitizer=access`. `GGML_HRX_DISABLE_DISPATCH=zaya.cca` turns the two ZAYA kernels off for comparison.