[bong-water-water-bong]

Moves third_party/llama.cpp from 851fa6c8d8f2 to the fork tip f08bd76f672d (branch 1bit/hrx-vulkan-patched).

Fork PRs #126-#131: hrx-backend-test and test-hrx-ops on the suite runner with route expectations re-ruled at this pin (engine #444/#445/#449), the PadParams op-params dump branch (the tree did not compile with GGML_HRX=ON since fork #114), and #533 case 1 (the fused MoE postops kernel's absent bias/residual slots bind a read-only value; verifier passes; per-token logprobs bit-identical on Qwen3-Coder-30B).

Fork PR #132 (tests only): require_matmul_route() accepts the table's kernel or the registered kquant decode family and prints the scheduled sequence when neither is there; stale decode/prefill sibling names and three routed-MoE names replaced from the first device run's log. The hw-gate ops phase at this pin is the device re-run engine #444 waits for.

Fork PR #135: the vendored jinja lexer accepts \xNN escapes, which Flint's chat template needs before the model even loads (five lexer tests).

Fork PR #134: the Flint architecture (RecursiveSI/flint, 40B total / 2.5B active MoE coder, Apache-2.0), ported from upstream llama.cpp #30928 with three small prerequisites (per-layer RoPE flags, a K-shift guard for NoPE layers, saver opt-out). CPU-verified on ryzen with unsloth's UD-Q4_K_XL (coherent greedy output on three prompts, PPL 5.56 on 10 chunks). This gate is the first HRX build of that code; the device run (serve chat, llama-bench, KLD vs CPU) and the Lemonade catalog entry follow once the leaderboard releases the GPU lane.

Merge by hand after hw_gate passes (auto-merge off); the hw-gate label goes back on once the leaderboard releases the GPU. Not part of the rc-2026-10-23 packages (the owner tags 74f64c7).

🤖 Generated with Claude Code

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit c7bfb06)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 126 - Partially compliant Compliant requirements: Meta description and title formatting rules implemented README page description updated to "Meet the engine" Blog index description updated to a real description All pages under character limits No duplicate titles Visible headings unchanged Non-compliant requirements: None Requires further human verification: None 131 - Partially compliant Compliant requirements: ZAYA models are convertible and their perplexity values are documented Decode performance for ZAYA1-reasoning-base is documented ZAYA1-8B comparison perplexity is documented Multi-sequence fix is implemented Non-compliant requirements: None Requires further human verification: None 444 - Partially compliant Compliant requirements: Test expectations updated to reflect current routing Test binaries are moved to testing_suite.h runner Non-compliant requirements: None Requires further human verification: None

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

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Subproject Commit Update The PR updates the subproject commit for llama.cpp from 851fa6c8d8f220074ba73c7450ffb2219b27180b to f08bd76f672d06509e053bcdf68bb8fbadc0fb4f. This is a standard update of the third-party dependency, but it's important to verify that all the changes introduced by this commit align with the intended behavior and do not introduce regressions, especially in the context of HRX and Vulkan support. Subproject commit f08bd76f672d06509e053bcdf68bb8fbadc0fb4f

[github-actions]

Persistent review updated to latest commit 7b6fbbe

[github-actions]

Persistent review updated to latest commit 03a06ab

[github-actions]

Persistent review updated to latest commit c7bfb06

[bong-water-water-bong]

Superseded by #565 (Bump HRX: llama.cpp e7e5cd08f9bb + hrx-system 82804197bcbe), which moves the same
third_party/llama.cpp pin past f08bd76f672d and carries the hrx-system hold-file checks with it. Closing this one
so there is a single llama.cpp bump in flight; nothing here is lost, the commits it pointed at are inside #565's pin.

Closing per the release queue (the pin only moves once #565's hardware gate is green).