Llama-2-13b quantized to 4-bit weights / 16-bit activations with KronQ (Kronecker-factored Hessian quantization). Weights are stored packed int4 (~6.9 GB vs 26 GB fp16); a fused dequant + bidirectional-incoherence (BiIP) CUDA kernel unpacks them on the fly.

Results (WikiText-2, seqlen 2048)

Perplexity: 4.96

Zero-shot accuracy:

PIQA	ARC-E	ARC-C	HellaSwag	WinoGrande	BoolQ	OBQA	Average
80.20	74.71	47.95	79.14	71.82	82.14	44.80	68.68

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

Usage

KronQ-packed checkpoint (model.safetensors carries biip_w_codes/scale/zero + BiIP buffers, see kronq_packed_config.json). Load with the KronQ runtime:

python eval_pretrained.py meta-llama/Llama-2-13b-hf donghyunli/Llama-2-13b-KronQ-W4A16 --ppl --zs

Recipe

Per-channel asymmetric W4, weight-only (a_bits=16), --alpha 0.25, bidirectional incoherence processing (BiIP), act_order, raw H_G. Calibrated on 128 WikiText-2 sequences.

License

Derivative of Llama-2-13b — subject to the Llama 2 Community License.