[bong-water-water-bong]

One sentence under ## DFlash draft models (--dflash), after #558 rewrote that section with the 24 Oct HRX numbers:

Not on this pin. On llama.cpp pins before fork e91b608cd (this one, and the release candidate built on it) a
DFlash2 drafter fails to load with done_getting_tensors: wrong number of tensors; expected 81, got 58; fixed by
fork #139 + #140, lands with the next pin bump.

Without it the section reads as if --dflash works on the pinned llama.cpp. Docs only — no code, no config, nothing
that changes what is built. python3 tools/copyright.py --check rc=0.

🤖 Generated with Claude Code

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 6e6d6fa)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 558 - Partially compliant Compliant requirements: config/recipes.json updated with two new entries docs/serve.md updated with new measurements and guidance docs/recipes.md and docs/hrx.md updated with new recipe info All numbers are measured on HRX hardware Copyright check passes Non-compliant requirements: No leaderboard data added (as per instructions) Requires further human verification: Verification that the measurements are indeed from the specified HRX hardware Confirmation that the numbers in docs/serve.md are accurate and up-to-date 139 - Partially compliant Compliant requirements: Added documentation for --mmproj in docs/serve.md Updated registry/architectures.json with Zamba2_VLForConditionalGeneration mapping Vulkan pin updated to 2c9d3983 Non-compliant requirements: No actual code changes to implement --mmproj functionality (only documentation) Requires further human verification: Verification that --mmproj works correctly with Zamba2-VL models on Vulkan Confirmation that the Zamba2-VL converter PR is properly integrated 140 - Partially compliant Compliant requirements: No code changes in this PR Non-compliant requirements: The actual fix for the non-deterministic attention issue is not included in this PR Requires further human verification: Confirmation that the issue is resolved in the next pin bump

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Clarity The documentation states that DFlash2 drafter fails to load on the current llama.cpp pin, but doesn't clearly indicate that this is a temporary limitation that will be resolved in the next pin bump. This could lead to confusion for users who might expect DFlash2 to work with the current version. > **Not on this pin.** On llama.cpp pins before fork `e91b608cd` (this one, and the release candidate built on it) a DFlash2 drafter fails to load with `done_getting_tensors: wrong number of tensors; expected 81, got 58`; fixed by fork #139 + #140, lands with the next pin bump.

[github-actions]

Persistent review updated to latest commit 9105b72

[github-actions]

Persistent review updated to latest commit 04fe3e6

[github-actions]

Persistent review updated to latest commit 6e6d6fa