New technique boosts accuracy of ultra-compressed AI model inference

Original: TORQ: Two-Level Orthogonal Rotation for MXFP4 Quantization

Source: arxiv.org ↗