Llama-2-7b-KronQ-W3A16-fake/README.md

base_model, language, license, pipeline_tag, library_name, tags

base_model	language	license	pipeline_tag	library_name	tags
meta-llama/Llama-2-7b-hf	en	llama2	text-generation	transformers	kronq quantization fake-quant fp16

Llama-2-7b — KronQ W3A16 (fake-quant fp16)

Paper: arXiv:2607.07964 · Code: GitHub

⚠️ Fake-quant fp16 checkpoint. The 3-bit weights are stored in fp16 (KronQ does not pack int3), so this repo is the same size as bf16 — no compression or speedup, for PPL / accuracy reproduction only. For deployable low-bit, see the W4A16 / W2A16 (packed int4/int2) repos.

Llama-2-7b quantized to 3-bit weights with KronQ (Kronecker-factored Hessian quantization), exported as a standard fp16 model.

Results (WikiText-2, seqlen 2048)

Perplexity: 5.84

Zero-shot accuracy:

PIQA	ARC-E	ARC-C	HellaSwag	WinoGrande	BoolQ	OBQA	Average
77.09	72.39	42.58	72.04	67.64	75.35	41.60	64.10

(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 needed:

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("donghyunli/Llama-2-7b-KronQ-W3A16-fake", torch_dtype="float16").cuda()
tok = AutoTokenizer.from_pretrained("donghyunli/Llama-2-7b-KronQ-W3A16-fake")

Recipe

Per-channel asymmetric W3, 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-2-7b — subject to the Llama 2 Community License.