[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 🔍

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 Error handling added for file opening failures Non-compliant requirements: None Requires further human verification: None 152 - Partially compliant Compliant requirements: Integer multiplication cast fixed in laya/scorer.cpp Workflow permissions declared in ci.yml Non-compliant requirements: None Requires further human verification: None 445 - Partially compliant Compliant requirements: The PR updates the submodule to the latest commit The test cases are updated to reflect the new dispatch behavior Non-compliant requirements: The PR does not address the root cause of the 49 failing test cases, but merely updates the submodule Requires further human verification: The classification of test cases (A-D) requires further human review to determine if the expectations need to be updated or if there are actual behavior changes that need to be addressed

⏱️ 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 update includes several fixes and changes related to HRX, including kernel name translation, norm input contiguity, and dispatch matching. While the update itself is straightforward, it's important to verify that all the changes are compatible with the current codebase and that no regressions are introduced. Subproject commit 40c4ccb76e48aafd44e024b0d2bb27b38bf54015 Architecture Registry Update The architecture registry is updated to reflect the new llama.cpp commit. This ensures that the registry is in sync with the latest changes in the submodule and that the correct architecture information is used for model loading and execution. "llama.cpp (hrx)": "40c4ccb76e48aafd44e024b0d2bb27b38bf54015"