donghyunli/Llama-2-70b-KronQ-W4A16-g128

Llama-2-70b — KronQ W4A16 g128 (packed int4)

Paper: arXiv:2607.07964 · Code: GitHub

Llama-2-70b quantized to 4-bit weights / 16-bit activations, group size 128, with KronQ (Kronecker-factored Hessian quantization). Packed int4 (~35 GB); fused dequant + BiIP CUDA kernel with per-group scales.

Results (WikiText-2, seqlen 2048)

Perplexity: 3.380

Zero-shot accuracy:

PIQA	ARC-E	ARC-C	HellaSwag	WinoGrande	BoolQ	OBQA	Average
82.70	79.21	56.23	84.05	79.79	84.95	48.40	73.62

(lm-evaluation-harness 0-shot; acc_norm for PIQA/HellaSwag/ARC, acc for WinoGrande/BoolQ.)

Usage

python eval_pretrained.py meta-llama/Llama-2-70b-hf donghyunli/Llama-2-70b-KronQ-W4A16-g128 --ppl --zs --distribute

(70B: --distribute spreads across GPUs.)

Recipe

Group-128 asymmetric W4, weight-only, --alpha 0.25, --act_order, BiIP, raw H_G.

License

Derivative of Llama-2-70b — llama2 license.