Meta-Llama-3-8B — KronQ W2A16 g128 (packed int2)
Paper: arXiv:2607.07964 · Code: GitHub
Meta-Llama-3-8B quantized to 2-bit weights / 16-bit activations, group size 128, with KronQ (Kronecker-factored Hessian quantization). Packed int2 (~3.9 GB); a fused dequant + bidirectional-incoherence (BiIP) CUDA kernel with per-group scales unpacks on the fly.
Results (WikiText-2, seqlen 2048)
Perplexity: 10.756
Zero-shot accuracy:
| PIQA | ARC-E | ARC-C | HellaSwag | WinoGrande | BoolQ | OBQA | Average |
|---|---|---|---|---|---|---|---|
| 68.99 | 56.31 | 34.39 | 53.34 | 63.30 | 65.93 | 34.20 | 53.78 |
(lm-evaluation-harness 0-shot; acc_norm for PIQA/HellaSwag/ARC, acc for WinoGrande/BoolQ.)
Usage
python eval_pretrained.py meta-llama/Meta-Llama-3-8B donghyunli/Meta-Llama-3-8B-KronQ-W2A16-g128 --ppl --zs
Recipe
Group-128 asymmetric W2, weight-only (a_bits=16), --alpha 0.25, --act_order, BiIP, raw H_G. 128 WikiText-2 calibration sequences.
License
Derivative of Meta-Llama-3-8B — llama3 license.