QAM-W: Joint 2D Codebook Quantization for LLM Weights via Hadamard Rotation and Activation-Aware Scaling

In three linesQAM-W is a 2D quantization codec for LLM weights using Hadamard rotation and activation-aware scaling. Across 5 models (1.1B–13B), the activation-aware variant at ~5.5 bpw maintains ±0.4% BF16 perplexity, matching SmoothQuant W8A8 quality with 32% fewer weight bits. 2D coding outperforms polar coding by 2–15 pp.

Summary generated by Claude — human-verified