b70-optimization-lab

2026-06-29 Current llama.cpp Gemma Record Worktree Capture

This records the exact dirty source state of:

/home/steve/src/llama.cpp-gemma-record-repro-c926

Base upstream commit:

c926ad098 vulkan: link ggml-cpu when GGML_VULKAN_CHECK_RESULTS / RUN_TESTS are enabled (#24444)

Patch artifact:

patches/gemma4-26b-a4b-q8-b70/20260629-current-llamacpp-gemma-record-worktree.patch

The patch is a cumulative research stack, not a minimal upstream patch. It preserves both promoted/default-off code paths and later negative experiments so the local state can be reconstructed before switching model efforts.

Current promoted Gemma Q8 result tied to this source family:

High-value enabled flags in the promoted lane include:

Reapply from a clean llama.cpp checkout at c926ad098:

cd /home/steve/src/llama.cpp-gemma-record-repro-c926
git checkout c926ad098
git apply /home/steve/qwen36-results-main/patches/gemma4-26b-a4b-q8-b70/20260629-current-llamacpp-gemma-record-worktree.patch

Then rebuild with the VDR2 command in results/gemma4-26b-a4b-q8-b70/reproduce.md.

Do not treat every code path in this patch as promoted. See the adjacent negative patch notes and sweep notes for failed flags, including BF16 direct mulmat-id, rowpack/pairslot variants, selected-softmax split-N, no-bonus verifier variants, stage-split MTP, skip-identity out-ids, and dFlash PR 22105.