Llama-2-70b quantized to 2-bit weights / 16-bit activations, group size 128, with KronQ (Kronecker-factored Hessian quantization). Packed int2 (~19 GB); fused dequant + BiIP CUDA kernel with per-group scales.

Results (WikiText-2, seqlen 2048)

Perplexity: 4.846

Zero-shot accuracy:

PIQA	ARC-E	ARC-C	HellaSwag	WinoGrande	BoolQ	OBQA	Average
79.22	75.76	49.83	74.51	76.64	78.13	44.20	68.33

(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-W2A16-g128 --ppl --zs --distribute

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

Recipe

Group-128 asymmetric W2, weight-only, --alpha 0.25, --act_order, BiIP, raw H_G.

License

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