DEV Community

Cover image for MACaQ Uses Layer-Local Calibration to Quantize Weights and Activations to MXFP4
aimodels-fyi
aimodels-fyi

Posted on Originally published at aimodels.fyi

MACaQ Uses Layer-Local Calibration to Quantize Weights and Activations to MXFP4

This is a Plain English Papers summary of a research paper called MACaQ Uses Layer-Local Calibration to Quantize Weights and Activations to MXFP4. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.

MACaQ makes four-bit weights and activations practical without layer-by-layer engineering

MACaQ is a post-training method that quantizes both weights and activations to MXFP4, a four-bit microscaling format. On Llama-3-8B it matches the best perplexity of DuQuant++ and MR-GPTQ in 11 times less quantization time; on Qwen3-32B it runs 3.06 times faster than FP16 at batch 1, with only 2.3% decode overhead from its added transforms. It also quantizes a 122-billion-parameter Qwen3.5 model in 64 minutes using 32.5 GB of GPU memory.

The central design choice is to handle each linear layer independently. MACaQ leaves small activation transforms online instead of propagating them through neighboring layers, avoiding architecture-specific changes and allowing quantization to stream one layer at a time.

Scaling, rotation, and local code refinement reduce quantization error

For each layer, MACaQ first rescales input channels to balance activation and weight magnitudes, then applies a block-Hadamard rotation. A Hadamard transform mixes values within groups of 32 channels, spreading concentrated activation energy; the inverse effect on the weights is absorbed before quantization, so the unquantized matrix product is unchanged....


Continue reading the full paper summary on AIModels.fyi →

Top comments (0)