ternary-QAT 0.1.2

Ternary QAT for transformers, built for Ternary-Bonsai by Prism-ML

ternary_QAT

Lightweight ternary QAT for Ternary-Bonsai unpacked text models.

Qwen 3.5 is untested right now!

This is designed from the start to be Unsloth compatible.

Weights are ternarized to {-1, 0, 1} per group of (128/64/user-defined) consecutive weights along the last dim, matching the Bonsai on-disk format (verified bit-exact). Embeddings + all nn.Linear modules (attn, MLP, lm_head) are ternarized; norms stay FP.

Based on Prism-ML's whitepaper: https://github.com/PrismML-Eng/Bonsai-demo/blob/main/ternary-bonsai-8b-whitepaper.pdf

Install

Have your preferred torch version installed first so that this doesn't install the CPU version (which you probably don't want)

pip install ternary_QAT                 # core (torch only)
pip install "ternary_QAT[peft,transformers]"  # + LoRA / model loading

if you use uv:
uv pip install ternary_QAT --torch-backend=auto

Use

Full-finetune

from transformers import AutoModelForCausalLM
from ternary import swap_linear, TernaryConfig

model = AutoModelForCausalLM.from_pretrained("prism-ml/Ternary-Bonsai-1.7B-unpacked")
swap_linear(model, TernaryConfig(group_size=128))
# ... train normally; ternarize fires in every Linear.forward

LoRA (ternary frozen base + FP adapters)

from peft import LoraConfig, get_peft_model
from ternary import swap_linear, TernaryConfig, ternarize_lora_params, reternarize_merged_linears

model = ...  # load model
swap_linear(model, TernaryConfig(group_size=128))
model = get_peft_model(model, LoraConfig(r=128, lora_alpha=128, ...))

# ... train ...

# at save: ternarize adapter, merge, re-ternarize merged linears
ternarize_lora_params(model)
model = model.merge_and_unload()
reternarize_merged_linears(model)
model.save_pretrained("./out")

See examples/ for LoRA, FFT, and Unsloth examples.

Learning rate

Ternary QAT needs 10-50x higher LR than standard fine tuning.

The lowest usable LR I've found so far is around 7e-4, so experiment in that range up to the e-3s, depending on rank and dataset size.