Abstract
Post-training quantization (PTQ) is a widely adopted technique for compressing large language models (LLMs) without retraining. Existing second-order PTQ methods, including GPTQ, construct quantization objectives exclusively from input activation statistics, effectively assuming that all output channels contribute equally to the layer-wise reconstruction objective. We propose KronQ, a PTQ framework that challenges this assumption by introducing the gradient covariance into the quantization pipeline. Under the Kronecker-factored Hessian approximation, the quantization loss depends jointly on both the activation and gradient covariances, and KronQ exploits this at two complementary levels. (1) KronQ introduces bidirectional incoherence processing, extending the existing input-side random rotation to the output dimension using the gradient covariance, reducing weight magnitude variance across both input and output dimensions. (2) KronQ derives a new sensitivity metric for inter-layer mixed-precision allocation, driven by the gradient and activation Hessian traces. Notably, in the case of 2-bit weight-only quantization on LLaMA-3-70B, while GPTQ and GPTAQ diverge or produce degenerate quantizations (>2000 perplexity on WikiText-2), KronQ achieves 7.93 perplexity.
Community
What got me was the gradient covariance angle. I've run GPTQ enough times to know it treats every output channel like it matters equally, and that's always felt like a shortcut. KronQ adds the gradient covariance into the quantization objective via a Kronecker-factored Hessian approximation, which means it actually weighs which output channels matter more for the reconstruction loss. Makes intuitive sense — not all weights pull their weight in the forward pass.
The practical question is whether the extra compute at quantization time pays off. Second-order methods already cost more than RTN or AWQ, and KronQ adds another matrix factorization step on top. For a one-time quant on a model you'll serve millions of times, sure. For iterating on quant recipes during development, the overhead might sting.
I'd be curious to see perplexity numbers at 3-bit and 2-bit specifically. That's where GPTQ tends to fall apart and where a better loss objective would show its teeth. If KronQ holds at 3-bit where GPTQ starts hallucinating, that's the real win.
Models citing this paper 30
donghyunli/Llama-2-7b-KronQ-W2A16
Datasets citing this paper 5
donghyunli/Llama-2-70b-KronQ-HG
donghyunli/Meta-Llama-3-70B-KronQ-HG
Spaces citing this paper 0
No Space linking this paper