[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)	851fa6c8d8f2	f08bd76f672d
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:

• f08bd76f6 jinja: accept \xNN escapes in the lexer (weekly: run Lemonade's suite against a running lemond; clearer held reasons #135)
• c3eef46c0 models: add Flint (FlintForCausalLM), ported from upstream #30928 (weekly: force the checkout of main in the workspace #134)
• 830e2c161 tests: fix test-hrx-ops from the first device run (engine HRX: hrx-backend-test and test-hrx-ops abort on stale kernel-routing expectations #444) (Bump ZINC: zolotukhin/zinc 29bc350ac4cb #132)
• f1670b43c hrx: the absent bias/residual slots of mul_mat_id_postops bind a read-only value (engine hrx-backend-test: two fused plans fail the command-program verifier (transient read before write) #533 verify) (docs: ZAYA1-base and ZAYA1-reasoning-base checked; the ZAYA multi-sequence fix #131)
• 8509e42c3 tests: re-read the open rows at the current pin (engine hrx-backend-test: 49 host cases red after the AMD-core re-head — classification (stale expectation vs needs a ruling) #445, fork 7ca8ba5a) (Pin llama.cpp 706ad1a: ZAYA with several sequences per ubatch #130)
• 7ca8ba5a4 tests: follow the pin's routes and report verifier failures (engine hrx-backend-test: 49 host cases red after the AMD-core re-head — classification (stale expectation vs needs a ruling) #445) (ZAYA1 legacy checkpoints and remote chat templates: pin llama.cpp 567d2dc; ZAYA1-74B docs #129)
• 6e57798e5 tests: name the kquant decode route for the few-token matmuls in test-hrx-ops (engine HRX: hrx-backend-test and test-hrx-ops abort on stale kernel-routing expectations #444, item 2) (tools/gguf_to_q4nx.py: repack a GGUF into a Q4NX model directory #128)
• 5fd1d30fa ggml-hrx: fix the PAD param dump so the tree compiles, and run test-hrx-ops case by case (engine HRX: hrx-backend-test and test-hrx-ops abort on stale kernel-routing expectations #444) (Pin llama.cpp 4380daf: ZAYA1-74B (sliding-window layers) on Vulkan #127)
• 9b846c173 tests: assert the publication this line makes, not the fused producers f2099e9b removed (engine HRX: restore consumer-driven activation publication (dropped whole in f2099e9b) #449) (site: meta descriptions and titles sized for search results #126)

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 🔶 135 - Partially compliant Compliant requirements: The bump check now starts the build's lemond on its own port (13399, LEMONADE_TEST_PORT) with this build's 1bit, and stops it afterwards The suite needs numpy. The venv now installs Lemonade's own test/requirements.txt The held reason quoted the passing ctest log. It now names the step that failed and the last lines of that step's log Non-compliant requirements: Lemonade's test/server_llm.py expects a lemond already listening Requires further human verification: Lemonade's test/server_llm.py expects a lemond already listening 134 - Partially compliant Compliant requirements: The workspace belongs to the run, so the checkout is now forced (-f) Build outputs inside submodules are untouched, since submodule update handles those The strixhalo cron driver got the same fix Non-compliant requirements: A dry-run rehearsal of the Sunday run stopped at its first step: a file edited by hand in the workspace (~/.cache/1bit-weekly/src) blocked git checkout Requires further human verification: A dry-run rehearsal of the Sunday run stopped at its first step: a file edited by hand in the workspace (~/.cache/1bit-weekly/src) blocked git checkout 444 - Partially compliant Compliant requirements: The routing history mixes our commits with f2099e9b and 420b4e6fe ("take AMD's corpus wholesale"), so each expectation needs checking against the routing we actually intend Update the test where the current route is intended; fix the routing where it isn't Non-compliant requirements: Two binaries abort on their first failed REQUIRE, so the rest of their cases never run Move both binaries to the testing_suite.h runner, so one failure doesn't hide every later case and a run gives real pass/fail totals Requires further human verification: Two binaries abort on their first failed REQUIRE, so the rest of their cases never run Move both binaries to the testing_suite.h runner, so one failure doesn't hide every later case and a run gives real pass/fail totals

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Architecture Mapping Inconsistency The addition of FlintForCausalLM to the architectures registry may introduce an inconsistency if the corresponding GGUF architecture (flint) is not fully supported or tested on the HRX backend. This could lead to runtime errors or unexpected behavior when models using this architecture are loaded. "FlintForCausalLM": { "gguf": "flint", "backends": [ "hrx" ] },

[bong-water-water-bong]

Superseded by #565 (same hrx-system hold, llama.cpp moved further along 1bit/hrx-vulkan-patched to e7e5cd08f9bb: DFlash2 port + strict selector, Flash-Next, IQ2_S routing, Bonsai Hadamard fix). Closing.