Llama 3 8B quantized with signed Hadamard rotation, block-GPTQ, and Output-Fisher weak-column recovery. Projection weights use a 3.5-bpw INT3 base, selected columns are restored in BF16, and the embedding and output head use group-128 INT4.

Technical report · GitHub

Payload and evaluation

Format	Text bpw	Payload	WikiText-2 PPL ↓	ARC-299 ↑	MMLU-13,943 ↑
BF16	16.00	14.965 GiB	6.2559	50.50	41.04
Q4_K_M	4.89	4.583 GiB	6.4359	50.84	40.67
IQ3_S + imatrix	3.66	3.429 GiB	6.9929	44.15	39.87
WarpQuant Fisher R16E4H4	3.6256	3.389 GiB	7.3446	45.49	38.99

The repository stores the quantized values in BF16-compatible safetensors. The reported payload is the packed-equivalent analytical size including codes, scales, recovery values, and column indices.

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "HarimxChoi/WarpQuant-Llama-3-8B-R16E4H4"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Citation

bibtex

@misc{choi2026warpquant,
  author = {Harim Choi},
  title = {WarpQuant: Dual-Domain LLM Quantization via Hadamard Rotation and Output-Fisher Sensitivity},
  year = {2026},
  url = {https://harimxchoi.github.io/projects/warpquant/}
}