MACaQ: Fast, Model Agnostic Calibrated Quantization for Low-Precision Weights and Activations

Researchers introduced MACaQ, an MXFP4 W4A4 post-training quantization method for weight-activation quantization that reduces the memory and compute cost of neural-network inference without the expensive optimization and architecture-specific propagation of basis transforms across layers required by recent methods. The work targets low-precision weights and activations and is described as fast and model agnostic.

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