[bong-water-water-bong]

Moves third_party/llama.cpp e44c9d01a4d5 -> 0e255d774032 (fork tip after #98, #99 via #100, #100). third_party/hrx-system stays at 98d05d94: 4ba76c18 returns end-of-text on long prompts (#369). Registry regenerated (hrx 292, npu 7, mapped 292; tools/registry_build.py --check-pins clean).

Verified on strixhalo (performance mode, HRX0 visible via IREE_HAL_AMDGPU_LIBHSA_PATH, every GPU step under box.lock; logs ~/wt/pinverify-0e255d7/):

• ctest: 25/25 passed.
• tests/serve_e2e.sh hrx and cpu on Qwen3-0.6B: PASS, "Paris.", 32 chunks each.
• 4,732-token needle (-np 1 --jinja, temperature 0, cache_prompt false): 20/20 answer "7341", 5 completion tokens, same hash on all 20; probe: finish=stop, 0 NaN lines.
• test-backend-ops -b HRX0 twice: 1141/1141 both, Backend HRX0 OK (the HRX: intermittent MUL_MAT f32 m=1 n=1 k=2048 error on common.mul_mat.f32_f32_decode (about 1 in 40 runs at fork tip e44c9d0) #370 flake did not appear).
• Format matrix through 1bit serve --device hrx (one Paris request each, 0 NaN lines): Qwen3-0.6B-NVFP4, Qwen3.8-27B-iMatrix-NVFP4-MTP, qwen38-27b-qat-q2_0, Qwen3-0.6B-emb-iq3_s, Qwen3-0.6B-emb-iq4_nl all answer Paris.
• llama-bench (-fa 1 -p 512 -n 128 -r 3): 27B NVFP4 pp512 117.9 / tg128 9.16 (PR Registry: reviewed architecture gaps, and a guard against faking support with an alias #98: 115.6 / 9.14); 27B Q2_0 QAT 163.6 / 20.42 (Windows: 1bit.exe and a Vulkan llama-server.exe, cross-built on Linux #99: 164.1 / 20.76).
• Regression guard on the shared dequantizer: gpt-oss-20b MXFP4 Paris (36 tokens, 33.1 tok/s); Qwen3.8-27B UD-Q4_K_XL Paris; its llama-bench pp512 368.4 / tg128 12.13 (earlier pins ~335 / ~9.8; a gain if anything, not confirmed by an interleaved A/B).

CI here builds without HRX; the above is the Strix Halo gate from docs/hrx.md.

🤖 Generated with Claude Code

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 98 - Partially compliant Compliant requirements: Add docs/arch-gaps.md with evidence for eleven HF architecture classes Add registry/significant.json with the eleven classes Add tools/registry_build.py --check-gaps functionality Update docs/registry.md to link the gap review page Non-compliant requirements: The alias table and its 100% coverage claim were not included (as noted in the ticket description) Requires further human verification: Verification that the gap review is accurate and complete 99 - Partially compliant Compliant requirements: Cross-build engine for Windows 10+ x64 on Linux Produce two static executables: 1bit.exe and llama-server.exe 1bit.exe finds llama-server.exe beside it Use CreateProcess with kill-on-close job object for backend processes Use Winsock for free_port and std::signal for signals Maintain POSIX paths unchanged Build with Windows 10 API level for cpp-httplib Static linking with MinGW Fetch pinned inputs with sha256 checks Non-compliant requirements: Not yet tested on a real Windows machine Requires further human verification: Testing on a real Windows machine 100 - Partially compliant Compliant requirements: Support ONNX Runtime GenAI models via 1bit serve --device onnx Use third_party/ryzenai-server as the backend server Build ryzenai-server against ONNX Runtime GenAI 0.11.2 and ONNX Runtime 1.23.2 Support CPU execution with optional Ryzen AI Software for NPU/hybrid Include proper NOTICE file entries Ensure binary carries DT_RPATH $ORIGIN for runtime dlopen Non-compliant requirements: ryzenai-server.exe in scripts/build-windows.sh not yet implemented (mentioned as next step) Requires further human verification: Verification of ONNX backend functionality on target hardware

⏱️ Estimated effort to review: 4 🔵🔵🔵🔵⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Subproject Update The subproject commit for llama.cpp has been updated from e44c9d01a4d5110cecaf45472a717f46cd1abbf9 to 0e255d774032dc6b4b06a157d2feb6be2175d3ab. This update includes changes from forks #98 through #100, which may introduce new functionality or behavior changes that should be verified for compatibility and correctness. Subproject commit 0e255d774032dc6b4b06a157d2feb6be2175d3ab Registry Source Update The registry source for llama.cpp (hrx) has been updated to reflect the new commit hash. This change aligns the registry with the updated llama.cpp version, ensuring that the architecture mappings are consistent with the latest codebase changes. "llama.cpp (hrx)": "0e255d774032dc6b4b06a157d2feb6be2175d3ab"