[bong-water-water-bong]

Moves the HRX submodules along our line. AMD's tested pair (ROCm/ggml-staging-automation) is llama.cpp 1ec5a7a69da5 + hrx-system 5f4dfc8731a1; the workflow merges it into our branches (a conflict fails the run and names the files).

	from	to
third_party/llama.cpp (1bit-MONSTER/llama.cpp 1bit/hrx-vulkan-patched)	e7e5cd08f9bb	40c4ccb76e48
third_party/hrx-system (1bit-MONSTER/hrx-system 1bit/main)	82804197bcbe	82804197bcbe

hrx-system HELD at 82804197bcbe by config/hrx-system-hold.config (AMD's tested pin is 5f4dfc8731a1).

llama.cpp commits new to the engine:

• 40c4ccb76 hrx: translate add-on kernel names through a weak hook before scanning a code object (moe_trace: create the trace file 0644 and fail if it cannot be written #153)
• 15a46404e llama-graph: make norm inputs contiguous for qwen4exp only (qwen35 one-token decode back on the fused GDN kernel) (Fix CodeQL alerts: size_t index math in the Laya scorer, read-only token for ci #152)
• 8c0ab23a6 tests: qwen_matmul_dispatch runs to the end at the pin (engine hrx-backend-test: 49 host cases red after the AMD-core re-head — classification (stale expectation vs needs a ruling) #445) (third_party: bump llama.cpp pin for the HRX decode-split multipass fix (engine#115) #151)
• 6b397ee90 tests: the packing case pins the guarded packer, both arms measured (engine hrx: disjoint transients no longer share an arena slot (arena 256 -> 512 on the same plan) #534) (Zyphra ZAYA1-VL-8B (pin llama.cpp f30cc43), Zamba2-VL template fix (pin llama.cpp-vulkan 4a2c0656) #149)
• 8ee20a088 hrx: refuse a dispatch match that reads a value before it is written (engine hrx-backend-test: two fused plans fail the command-program verifier (transient read before write) #533, hrx: fusion must withdraw dispatches it makes redundant (reader emitted before the fused writer, engine #533) #570) (serve: --moe-slots streams MoE experts from the drive (Vulkan) #150)
• adef66c17 tests: keep the FILL constant in snapshots, and bring both HRX suites to the pin (engine HRX: hrx-backend-test and test-hrx-ops abort on stale kernel-routing expectations #444, hrx-backend-test: 49 host cases red after the AMD-core re-head — classification (stale expectation vs needs a ruling) #445) (serve: --device hrx decodes without decode-split flash attention by default (#140) #148)
• 6ded110b5 qwen4exp/ternary: TQ2_0 load-time repack onto the fast k128 ternary decode kernel (BitNet b1.58 on HRX) (Discussions are open #147)
• 21a2822e4 cli: -no-cnv means one turn, and a non-TTY stdin ends the interactive loop (engine#561) (actions: bump actions/checkout from 4.4.0 to 7.0.1 #145)

CI here builds without HRX. Before merging, on Strix Halo run the long-prompt correctness
probe and test-backend-ops -b HRX0; record the exact validated hrx-system pin in
config/hrx-long-prompt-validated-pin.config:
python3 tests/hrx_long_prompt.py
cmake -B build -G Ninja -DONEBIT_HRX=ON && cmake --build build --target onebit && for d in hrx cpu; do tests/serve_e2e.sh build/1bit <Qwen3-0.6B Q4_K_M .gguf> $d; done

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit bc8ee93)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 153 - Partially compliant Compliant requirements: The trace file is now opened with explicit 0644 mode The tool will now fail and exit if it cannot open the trace file path Non-compliant requirements: None Requires further human verification: None 152 - Partially compliant Compliant requirements: The integer multiplication cast issue in laya/scorer.cpp has been addressed The ci.yml workflow now declares the required permissions Non-compliant requirements: None Requires further human verification: None 445 - Partially compliant Compliant requirements: The PR updates the submodule to the new commit, which should address the issues Non-compliant requirements: The classification of the 49 failing cases is not fully addressed in this PR diff Requires further human verification: The actual test results and case classifications need to be verified manually

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Submodule Update The PR updates the llama.cpp submodule to commit 40c4ccb76e48aafd44e024b0d2bb27b38bf54015. This is a major change that could introduce compatibility issues or behavioral changes. Ensure that all functionality remains intact and that no regressions are introduced. Subproject commit 40c4ccb76e48aafd44e024b0d2bb27b38bf54015 Architecture Registry Update The architecture registry has been updated to reflect the new llama.cpp commit. Verify that this update correctly reflects the supported architectures and that no architecture support has been inadvertently removed or altered. "llama.cpp (hrx)": "40c4ccb76e48aafd44e024b0d2bb27b38bf54015"

[bong-water-water-bong]

Triage (CPU box ryzen, no GPU work):

What the diff is — two hunks: third_party/llama.cpp e7e5cd08 -> 40c4ccb76e48, and the llama.cpp (hrx) source-SHA line in registry/architectures.json. No row or count changes, which is right: the 8 upstream commits added (#145 -no-cnv, #147 TQ2_0 load-time repack, #148/#149/#151 tests, #150 refuse a dispatch match that reads a value before it is written, #152 norm inputs contiguous for qwen4exp only, #153 translate add-on kernel names through a weak hook) register no new architecture. e7e5cd08 is a strict ancestor of 40c4ccb76e48, so this is a fast-forward of the pin.

Checks — build, Analyze (c-cpp/python/rust), CodeQL pass; the hrx-system gitlink agrees with the hold file guard passes (config/hrx-system-hold.config, gitlink still 82804197bcbe); hw_gate pass 30m28s — the only one of the five open engine PRs where hw_gate actually ran rather than skipped. pr_agent fails on infrastructure alone: the self-hosted runner's model (coder-llm on :8090) does not answer while the measurement pass owns the box.

Risk, one sentence: it moves the kernel tree the leaderboard cites — it adds #152 (qwen4exp norm contiguity, which puts qwen35 one-token decode back on the fused GDN kernel), #150 (dispatch refuses a match that reads a value before it is written), #147 (TQ2_0 load-time repack) and #153 (add-on kernel-name hook) to the tree the rows come from, so a regression would surface as a leaderboard row that no longer reproduces — and hw_gate has already passed on exactly this tree.

Not merging this one myself — it moves the pin, which is your call. Recommendation in the triage brief: merge after the current ~/lb pass closes its row.

Provenance gap worth fixing in the same motion (found while checking the pin): bench/npu-leaderboard/*.json and docs/npu-leaderboard.md:27-29 record engine=0d0c846, and docs/hrx.md:400 records the rows as 0d0c846 + llama.cpp 8ee20a0. But git ls-tree 0d0c846 third_party/ says e7e5cd08 — and e7e5cd08 is an ancestor of 8ee20a0, i.e. the 27 measured rows ran a tree 4 commits ahead of what the pin's own submodule record says. That matches ~/briefs/LB-AUDIT-2026-10-28.md. Whichever way it goes, docs/hrx.md:400 and the submodule record should agree.

[github-actions]

Persistent review updated to latest commit 8e47d7b

[github-actions]

Persistent review updated to latest commit 7ad3446

[github-actions]

Persistent review updated to latest commit 1fc4dfe

[github-actions]

Persistent review updated to latest commit bc8ee93