[bong-water-water-bong]

Moves the HRX submodules along our line. AMD's tested pair (ROCm/ggml-staging-automation) is llama.cpp dbe1d6e1a27e + hrx-system 61c152ac4b40; 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)	8d9ce2c1aa60	0d3ed9e3ff21
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 61c152ac4b40).

llama.cpp commits new to the engine:

• 0d3ed9e3f qwen: make rmsnorm q8 publication consumer-driven again (engine HRX: restore consumer-driven activation publication (dropped whole in f2099e9b) #449) (HRX: dense on HRX0 with experts on another device fails for Qwen MoE models (fused router) #108)
• c26c071a0 tests: retire the HRX: symmetric-i4 consumer route re-enabled by AMD aadd6025 after f2099e9b removed it #450 known-fail guard, the symmetric-i4 route is not a regression (engine HRX: symmetric-i4 consumer route re-enabled by AMD aadd6025 after f2099e9b removed it #450) (ONNX on the GPU on Linux: ONNX Runtime's WebGPU provider (ONNX_WEBGPU=1) #109)
• 29dd88200 qwen: publish the next-RMSNorm q8 alternate from the fused routed-down kernel (engine HRX: restore consumer-driven activation publication (dropped whole in f2099e9b) #449) (ZAYA1 (Zyphra) on the GPU: pin llama.cpp 5556bf2, serve routes fork-only architectures #107)
• 953b6397b tests: retire the HRX: restore consumer-driven activation publication (dropped whole in f2099e9b) #449 guard for gdn rmsnorm-gate publication (engine HRX: restore consumer-driven activation publication (dropped whole in f2099e9b) #449) (hrx: scalar-scatter SET_ROWS for the non-FA path (-fa off) (#95) #106)
• 7227a7711 tests: per-case runner and the hrx-backend-test: 49 host cases red after the AMD-core re-head — classification (stale expectation vs needs a ruling) #445 ACCEPT rows (registry: causes of the two ZINC failures (Qwen2.5-7B load crash, MiniCPM5 empty reply) #105)
• 8dafc4a05 tests: build the HRX tests against the ggml-hrx we ship (engine#422) (HRX: pin llama.cpp 96f6b89, fixing #95 (GLM-4.7-Flash, Qwen3-Coder at full context) #104)

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 57a1160)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 449 - Partially compliant Compliant requirements: The PR restores the gdn_rmsnorm_gate_f32_publish kernel which was previously unreachable The PR updates the llama.cpp submodule to include the necessary changes for re-enabling publication kernels Non-compliant requirements: The PR does not explicitly show the re-registration of all the *_publish_* kernels in dispatch_registration/ The PR does not show the updated test cases that would validate the restored functionality Requires further human verification: Verification that all the *_publish_* kernels are correctly registered and functional Confirmation that the test cases are updated to reflect the restored behavior 108 - Partially compliant Compliant requirements: The PR updates the llama.cpp submodule which includes fixes for binding length and tensor materialization Non-compliant requirements: The PR does not explicitly show the implementation of the binding length fix or the tensor materialization fix Requires further human verification: Verification that the binding length fix works correctly in practice Confirmation that the tensor materialization is correct when experts are on another device 450 - Partially compliant Compliant requirements: The PR updates the llama.cpp submodule which includes the removal of the known-fail guard for the symmetric-i4 route Non-compliant requirements: The PR does not explicitly show the measurement or analysis of the symmetric-i4 route performance Requires further human verification: Verification that the symmetric-i4 route is indeed not a regression anymore Confirmation that the test case reflects the current behavior correctly

⏱️ 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 commit in the architectures.json file. While this is a necessary update, it's important to ensure that all the changes in the new commit are compatible with the existing codebase and that no regressions are introduced. "llama.cpp (hrx)": "0d3ed9e3ff210145fbf2f80f4de2e93e6fb05cf6"

[github-actions]

Persistent review updated to latest commit 57a1160