Skip to content

Releases

v0.3.0

Choose a tag to compare

@oguzeldereli oguzeldereli released this 09 Nov 13:24

Convert full-precision Qwen3.5 / 3.8 models into ternary networks that TTF
(ttf-ternary v0.3.0+) keeps training.

  • The streaming conversion (python -m ttfconv.convert): one decoder layer at a time from the safetensors.
    • Centred ternary GPTQ.
    • Plug-ins:
      • norm-gain fold + Hadamard rotation;
      • coordinate descent on the trits;
      • ternary low-rank adapters (prop, gmetric, randomized SVD);
      • 4-bit embedding / head.
    • Float adapters fitted to the rounding error (--svd-rank), and the fine residual (extra digits).
  • ttfconv.q38: Qwen3.8-Flash-Next conversion.
    • Every routed expert and dense projection to trits, layer by layer on one GPU.
    • The KD teacher cache (top 64 + 64 sampled tail).
  • scripts/q38: the PLE n-gram table's pruning + product quantisation (7.95 GB instead of 51.2 GB), slim checkpoints.
  • Recipes: recipes/qwen-0.8b.toml (the best measured 0.8B recovery) and recipes/qwen-27b.toml;
    docs/CONVERT_27B.md.
  • A Qwen3.5-0.8B quickstart in the README: text to converted, recovered model on one 16 GB GPU.
  • The released models made with it: ttf-Qwen3.8-27B-GGUF,
    ttf-Qwen3.8-Flash-Next-GGUF.

Install: pip install "ttf-ternary[qwen] @ git+https://github.com/oguzeldereli/ttf-ternary@v0.3.0" "ttf-conversion @ git+https://github.com/oguzeldereli/ttf-conversion@v0.3.0".

License: PolyForm Noncommercial 1.0.0.