Llama-2-7b quantized to 2-bit weights / 16-bit activations with KronQ (Kronecker-factored Hessian quantization). Weights are stored packed int2 (~2.3 GB vs 13 GB fp16); a fused dequant + bidirectional-incoherence (BiIP) CUDA kernel unpacks them on the fly at inference.

Results (WikiText-2, seqlen 2048)

Perplexity: 8.19

Zero-shot accuracy:

PIQA	ARC-E	ARC-C	HellaSwag	WinoGrande	BoolQ	OBQA	Average
67.95	54.34	31.31	51.16	60.77	64.25	34.80	52.08

(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:

bash

# clone https://github.com/<...>/KronQ and build the CUDA kernels, then:
python eval_pretrained.py meta-llama/Llama-2-7b-hf donghyunli/Llama-2-7b-KronQ-W2A16 --ppl --zs

Recipe

Per-channel asymmetric W2, weight-only (a_bits=16), --alpha 0.25, bidirectional incoherence processing (BiIP, Hadamard kernel), act_order. Calibrated on 128 WikiText-2 sequences with raw H_G. To reproduce instead of using these weights, see the KronQ repo's "Reproduce" path.

License

Derivative of Llama-2-7b — subject to the Llama 2 Community License.