Skip to content
This repository was archived by the owner on Sep 8, 2026. It is now read-only.

STRUCTURE.md

Latest commit

 

History

History
45 lines (33 loc) · 2.43 KB

STRUCTURE.md

File metadata and controls

45 lines (33 loc) · 2.43 KB
Raw

How everything is structured now (2026-07-24)

The one engine

poisonxa16/pxq_llama.cpp (public, standalone) — THE engine and THE product. One codebase, one name, in the llama.cpp -> ik_llama.cpp -> pxq_llama.cpp lineage. Contains everything:

  • Engine + all architectures (Laguna, Cohere2-MoE/North, Gemma-4, GLM, qwen35moe, deepseek, ...)
  • PXQ codecs (PXQ1/2/4/6) + PXQU universal mixed-tier maps
  • All speed levers (spec1row, router-fuse, enhance, spec-smalln, cublas-eager, volta-cublas, mtp-lazy-warmup, fuse-deltanet)
  • All correctness ports (WMMA-K6, dmmv-OOB, fused-MoE indices, norm-BF16, graph-v2, + 26 upstream)
  • Release packaging, docs, community credits (bradrlaw, Last-Guitar-5924)

Box working tree: the working tree (one pxq tree per machine). Branch: main. Binary: build-unified. Everything — dev, bench, serve, gauntlet — runs from this one build. Build with --runtime=nvidia (needs libcuda for the CUDA-driver-API link).

The private mirror

poisonxa16/pxq_llama.cpp-private (private) — same code + the secret universal-quant recipe files (.tiers, tier-maps.json, pxqu_wrel.py, pxqu_golden.py, pxa-bench/pxq-universal/). The recipes are .gitignored so they never reach the public repo; on the private remote they are force-added. One tree -> two remotes: origin = public clean, private = dirty backup.

Upstream

ikawrakow/ik_llama.cpp (external) — the upstream we pull improvements from. Add as a remote to pxq_llama.cpp when cherry-picking upstream PRs; fold them straight into main. No staging fork.

Retired / legacy

  • poisonxa16/ik_llama.cpp (the old fork) — DELETED 2026-07-24. Content fully folded into pxq_llama.cpp. Its campaign branches are bundled at ik_llama-campaign-branches.bundle in the offline archive.
  • A local legacy ik_llama tree — kept ONLY because the live brains still run its build-mmfast. It retires the day the brains move onto the pxq_llama.cpp unified build (a validated brain-engine swap; owner go required). After that, one engine, period.
  • Historical builds live on the array: <archive>/ (old worktrees, experiment binaries, the campaign bundle, MANIFEST.md). Never on the cards' cache.

The law

One working tree, one build, everything runs from it. Archive, don't fork. See ONE-BUILD-POLICY.md. The 8+ worktree / 2-repo sprawl was the root cause of the cross-history merge pain; it does not recur.