ContextQMD
Back to Eliza

llama.cpp patches — DEPRECATED archival drop (2026-05-09)

packages/app-core/scripts/aosp/llama-cpp-patches/README.md

2.0.12.3 KB
Original Source

llama.cpp patches — DEPRECATED archival drop (2026-05-09)

Superseded by elizaOS/llama.cpp @ v0.1.0-eliza. These patch series are no longer applied on AOSP builds. compile-libllama.mjs now points at the combined Eliza fork, which has TBQ + QJL + Q4_POLAR baked in (plus the Metal kernel sources). See the fork consolidation strategy doc for the migration story.

The directory and apply-patches.mjs script are kept in-tree for one release as a rollback path. To re-enable the legacy flow:

  1. In compile-libllama.mjs, restore the prior pin:
    js
    export const LLAMA_CPP_TAG    = "main-b8198-b2b5273";
export const LLAMA_CPP_COMMIT = "b2b5273e8b275bb96362fe844a5202632eb3e52b";
export const LLAMA_CPP_REMOTE = "https://github.com/Apothic-AI/llama.cpp-1bit-turboquant.git";
  2. Re-add the applyVendoredPatches({ srcDir: cacheDir, log }) call after patchLlamaCppSourceForMusl in ensureLlamaCppCheckout.

The two flows produce equivalent libraries on AOSP today (W1-A and W1-B patches were lifted directly into the fork tree). The fork path:

  • Eliminates the quadratic merge-conflict cost between QJL and Polar in ggml-common.h / ggml-cpu.c (both modified the same type_traits table; in-tree the conflict is resolved once, in patches it's resolved on every cherry-pick from a moving base).
  • Adds the Metal kernel sources alongside the C/NEON sources, so a future bun run build:metal doesn't need to apply runtime patches.
  • Standardizes on a single LLAMA_CPP_TAG across the host (DFlash) and AOSP build paths (next agent: align build-llama-cpp-dflash.mjs once DFlash is ported into the fork).

Once a follow-up release verifies no consumer pins the patches directly, this directory can be deleted.

Historical content (frozen)

qjl/

Five patches that added GGML_TYPE_QJL1_256 = 46 + GGML_OP_ATTN_SCORE_QJL to apothic@b2b5273. Now landed on elizaOS/llama.cpp branch eliza/qjl and merged into eliza/integration (slot 46 unchanged).

polarquant/

Four patches that added GGML_TYPE_Q4_POLAR = 45 to apothic@b2b5273. Now landed on elizaOS/llama.cpp branch eliza/polarquant — slot bumped from 45 to 47 so QJL (46) and Polar (47) coexist without colliding on the reserved hole at 45 (which was GGML_TYPE_COUNT in the TBQ-only build).