[Copilot]

The reported D=256 compile failure involved subgroup-vote results feeding conditional branches. Later runtime testing compiled and dispatched the kernel successfully; the failure was observed with a separate CLI compiler.

• Kernel: Pin llama.cpp to the always-stage candidate. Replace both subgroup votes with constants, forcing the existing masked-value staging path and removing the collectives:

  %key_tile_all_visible = scalar.constant false : i1
  %key_tile_any_visible = scalar.constant true : i1
  

  This stages every key tile, trading performance for a simpler path.
• Registry: Update the recorded llama.cpp pin to match the submodule.
• Review note: The candidate is on 1bit/fa-always-stage; .gitmodules tracks 1bit/hrx-vulkan-patched. The candidate’s masked-KV check was not the authoritative private test, and the original compile failure is no longer reproducible on demand.

• Fixes HRX Loom decode-split FA fails to compile for D=256, 24 q heads / 4 kv heads after fork #72 (hrx-system 51b1739) #314

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 314 - Partially compliant Compliant requirements: Fix the compile failure of the decode-split flash-attention kernel for D=256 with 24 query heads and 4 KV heads on HRX Loom Update the llama.cpp submodule to a version that resolves the compilation issue Update the registry to reflect the new llama.cpp commit Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Subproject Commit Mismatch The PR updates the llama.cpp submodule commit, but does not verify that the new commit actually resolves the compile issue described in the ticket. It's possible that the new commit may not fully address the problem or may introduce new issues. Subproject commit cf4dfd801e6ac7cd06a42ff012ab453c9bd086da Registry Update Consistency The registry file is updated to reflect the new llama.cpp commit, but there is no verification that the commit hash in the registry matches the actual pinned commit in the submodule. This could lead to inconsistencies if the submodule and registry are not in sync. "llama.cpp (hrx)": "cf4dfd801e6ac7cd06a42ff012ab453c9bd086da"