[bong-water-water-bong]

Moves third_party/llama.cpp from dd74f6b to 6e42b51, which is llama.cpp fork #55 squashed on top of #53/#54.

Fork #55 routes Q4_K/Q5_K/IQ4_XS prompt matmuls to HRX's existing q8_1 x4 prefill kernel:

• Q5_K and IQ4_XS now get Q4_K's activation/GLU policy.
• The generic fused SwiGLU (priority 290) leaves 256-2048-token prompt chunks to it.

On Qwen3.8-27B UD-Q4_K_XL those projections had been taking 94% of HRX prompt time in the generic F32 WMMA kernels.

Measured on this pin:

Qwen3.8-27B UD-Q4_K_XL, HRX0	old routing	this pin
pp512 (llama-bench -b 512 -ub 512)	98.5	334.6
pp2048	98.1	309.5
1bit serve --device hrx, 14,435-token prompt	90.7 tok/s	264.7 tok/s (same text)
KLD vs BF16 logits (wikitext-2, 20 x 512)	0.00717	0.00712

test-backend-ops -o MUL_MAT -b HRX0: 287/287. GGML_HRX_Q8_PREFILL_RELAX=0 restores the old routing.

Docs:

• docs/hrx.md gets a section and an "Our patches" entry.
• docs/serve.md's auto-routing note, which said "97-99 tok/s until fork q4nx: document the two Q8 code layouts behind one chunk size #55 is pinned" (added in serve: --device auto keeps Vulkan for architectures HRX does not map, --mtp and --mmproj #271), now carries the measured numbers.
• So does the docs/laya.md policy note.

🤖 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 3247a23

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 3247a23)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 55 - Partially compliant Compliant requirements: Added documentation in docs/npu.md about the two Q8 code layouts Updated registry/architectures.json to reflect the new llama.cpp commit Non-compliant requirements: No code change was implemented, as stated in the ticket description Requires further human verification: None 53 - Partially compliant Compliant requirements: Updated Discord invite link to a permanent one Added Community link to README Non-compliant requirements: None Requires further human verification: None 54 - Partially compliant Compliant requirements: Added canonical URLs Added link previews with og:url, og:type, og:image, and twitter:card Added structured data (JSON-LD) Added robots.txt and sitemap.xml Implemented IndexNow Non-compliant requirements: None Requires further human verification: Verification of search engine indexing and link preview functionality may require manual checks

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Clarity The documentation in docs/hrx.md mentions performance improvements but does not specify the exact hardware or conditions under which these improvements were measured, which could be important for reproducibility. Measured on Strix Halo, llama.cpp `6e42b51`, Qwen3.8-27B UD-Q4_K_XL, 2026-10-01: | metric | old routing | this pin | |---|---|---| | pp512 (llama-bench, `-b 512 -ub 512`) | 98.5 | **334.6** | | pp2048 | 98.1 | **309.5** | | `1bit serve --device hrx`, 14,435-token prompt | 90.7 tok/s | **264.7 tok/s** (same text) | | KLD vs BF16 logits (wikitext-2, 20 x 512) | 0.00717 | 0.00712 | | same top token | 96.13% | 96.27% | `test-backend-ops -o MUL_MAT -b HRX0`: 287/287. Decode does not change, since both rules apply only to chunks of at least 256 tokens. A chunk's tokens beyond the last multiple of 256 still go to the generic kernel. Performance Claim Verification The performance claims in docs/serve.md (e.g., pp512 335 tok/s) should be verified to ensure they are accurate and reproducible under the specified conditions. | `.gguf` | `auto`, `hrx` | the HRX build's llama-server on `HRX0` (docs/hrx.md). In a build without HRX, `auto` still means Vulkan, and so does `auto` for what HRX does not run yet, or runs well below Vulkan: an architecture the HRX build does not map (registry/architectures.json; among them Qwen3.8-Flash-Next and Zyphra Zamba, Zamba2 and BlackMamba), `--moe-slots`, `--parallel` on a gated delta-net model (Qwen3.5, Qwen3.8, Qwen3-Next), `--mtp` (HRX drafts at 65-78% of Vulkan speed) and `--mmproj` (an open RFC #213 gate). Since the HRX prompt-matmul routing (llama.cpp fork #55), prompts on HRX run at pp512 335 tok/s on Qwen3.8-27B UD-Q4_K_XL (was 97-99; ROCm 365-368), and a 14K-token prompt through `serve` at 265 tok/s (was 91) |

[bong-water-water-bong]

Review (PR-Agent duty): pin 6e42b51 is the tip of 1bit/hrx-vulkan-patched (checked with ls-remote) and is fork #55, which I reviewed. The registry change is the pin line only, the numbers match #55, and the docs have the before/after table. Merging.

[github-actions]

Persistent review updated to latest commit f0aea87

[github-actions]

Persistent review updated to latest commit 3247a23