Post Training Quantization of Large Language Models with Microscaling Formats

Sharify, Saxena, Xu, Yazar, Soloveychik, Wang (NVIDIA/Vectorized). v3. Studies SmoothQuant + AWQ + GPTQ applied to MX formats; abstract claims 4-bit weights + 8-bit activations via MXINT with 'negligible accuracy loss compared to the uncompressed baseline'. Abstract-level read only; the per-benchmark tables were not transcribed, so no numeric delta is taken from this source.