Paper Detail
Lorenz K. Muller, Philippe Bich, Chiara Boretti, Hyun-Min Chang, Jiawei Zhuang, Lukas Cavigelli
Weight-activation quantization can reduce the memory and compute cost of neural-network inference, but recent methods often require expensive optimization and architecture-specific propagation of basis transforms across layers. We introduce MACaQ, an MXFP4 W4A4 post-training quantization method that treats every linear layer independently. MACaQ applies a per-channel scale and a block-Hadamard transform online to each layer's activations, while absorbing the corresponding inverse transforms into its weights. Retaining these transforms online avoids cross-layer dependencies and enables a simple streaming implementation of the quantization process, in which only one layer must reside in memory at a time. To improve over round-to-nearest quantization, we introduce Flip-Refinement, a Hessian-aware coordinate-descent procedure that greedily moves individual FP4 codes to adjacent representable values. Despite executing the activation transforms at inference time, MACaQ adds only 2.3% end-to-end overhead over pure MXFP4 on Qwen3-32B in vLLM. MACaQ reaches lower perplexity than DuQuant++ and MR-GPTQ, and already matches their best perplexity on Llama-3-8B in 11 times less quantization time. We further apply MACaQ without architecture-specific modifications to an image diffusion model, illustrating the generality of its layer-local design and show quantization of a 122B parameter model in 64 minutes, using only 32.5GB GPU RAM.
No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.
No ranking explanation is available yet.
No tags.
@misc{muller2026macaq,
title = {MACaQ: Fast, Model Agnostic Calibrated Quantization for Low-Precision Weights and Activations},
author = {Lorenz K. Muller and Philippe Bich and Chiara Boretti and Hyun-Min Chang and Jiawei Zhuang and Lukas Cavigelli},
year = {2026},
abstract = {Weight-activation quantization can reduce the memory and compute cost of neural-network inference, but recent methods often require expensive optimization and architecture-specific propagation of basis transforms across layers. We introduce MACaQ, an MXFP4 W4A4 post-training quantization method that treats every linear layer independently. MACaQ applies a per-channel scale and a block-Hadamard transform online to each layer's activations, while absorbing the corresponding inverse transforms into},
url = {https://huggingface.co/papers/2611.10649},
keywords = {weight-activation quantization, post-training quantization, MXFP4, W4A4, per-channel scale, block-Hadamard transform, activation transforms, inverse transforms, linear layer, cross-layer dependencies, streaming implementation, round-to-nearest quantization, Flip-Refinement, Hessian-aware coordinate-descent, FP4 codes, perplexity, Qwen3-32B, vLLM, DuQuant++, MR-GPTQ, Llama-3-8B, image diffusion model, layer-local design, code available, huggingface daily},
eprint = {2611.10649},
archiveprefix = {arXiv},
}
{}