[bong-water-water-bong]

Pins third_party/llama.cpp to 6617e78, which brings in two fork PRs.

llama.cpp #43: Q3_K and IQ3_S in the K-quant decode kernels.

• IQ3_S keeps its 512-entry grid in workgroup memory.
• With this, every weight type in the UD-Q4_K_XL 27B goes through these kernels.
• Qwen3.8-27B UD-Q4_K_XL decode: 11.60 → 12.15 tok/s, against 12.48 on Vulkan (97.3%).
• Decode-shaped wikitext PPL: 7.1595 (CPU 7.1475).
• test-backend-ops -o MUL_MAT -b HRX0 passes.

llama.cpp #44: HRX decode no longer spins a CPU core.

• The per-token stream wait sat in ROCr's signal wait, which returned from the KFD event ioctl on every completed command without sleeping.
• Long expected waits (over 20 ms) now sleep through 80% of the wait; shorter waits are unchanged.
• 27B decode now runs at 71–76 °C and about 125 W, where it was 83–93 °C and 133–137 W, and is 0.4% slower.
• Qwen3-0.6B and ZAYA1-8B decode are unchanged.

Also in this PR:

• registry/architectures.json is regenerated; --check-pins passes.
• docs/hrx.md is updated.

🤖 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 67a9015

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 43 - Partially compliant Compliant requirements: Performance improvements for Qwen3.8-27B UD-Q4_K_XL on HRX (12.15 tok/s vs 12.48 on Vulkan) Documentation updates in docs/hrx.md reflecting new kernel support and performance data Updated registry/architectures.json to reflect new llama.cpp commit Non-compliant requirements: No actual blog post content added to the repository No specific mention of the "hipBLAS trap" or sgemm silently returns wrong results on gfx1151 (Radeon 8060S): max rel err 1.7e4, no error raised ROCm/rocm-libraries#11530 issue Requires further human verification: Verification that the performance claims (e.g., 2.3-2.9x speedup) are accurate and measured on hardware Confirmation that the blog post content matches the project's voice and tone 44 - Partially compliant Compliant requirements: No changes to tools/site.py, README, or pages.yml found in the diff Non-compliant requirements: Site address change not implemented in this PR Requires further human verification: Need to verify that the DNS and GitHub Pages configuration changes are handled separately

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Performance Claim Verification The documentation claims a 12.15 tok/s decode speed for Qwen3.8-27B UD-Q4_K_XL on HRX, which is 97.3% of Vulkan performance. This specific performance number should be verified to ensure it matches the actual hardware measurements mentioned in the ticket. | Qwen3.8-27B UD-Q4_K_XL | 12.15 | 12.48 | 97.3% (was 9.0) | Thermal and Power Claim Verification The documentation states that 27B decode now runs at 71–76 °C and about 125 W, compared to 83–93 °C and 133–137 W previously. These thermal and power claims need to be confirmed as they represent significant hardware behavior changes. the wait; shorter ones are unchanged. 27B decode: 71-76 C and ~125 W instead of 83-93 C and 133-137 W, 0.4% slower; Qwen3-0.6B and ZAYA1-8B unchanged. `ONEBIT_HRX_BLOCKING_WAIT=1` turns it