⚠️ Fake-quant fp16 checkpoint. 3-bit weights stored in fp16 (KronQ does not pack int3) — same size as bf16, for PPL/accuracy reproduction only. For deployable low-bit see the W4A16 / W2A16 (packed) repos.

Llama-2-70b-hf quantized to 3-bit weights with KronQ, exported as a standard fp16 model.

Results (WikiText-2, seqlen 2048)

Perplexity: 3.66

Zero-shot accuracy:

PIQA	ARC-E	ARC-C	HellaSwag	WinoGrande	BoolQ	OBQA	Average
82.26	81.86	58.28	82.45	78.53	82.97	48.80	73.59

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

Usage

Loads as a standard fp16 model (no KronQ code):

python

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("donghyunli/Llama-2-70b-KronQ-W3A16-fake", torch_dtype="float16", device_map="auto")

Recipe

Per-channel asymmetric W3, weight-only (a_bits=16), --alpha 0.25, BiIP, act_order, raw H_G.

License

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