post-training quantization

FAIR-Calib: Frontier-Aware Instability-Reweighted Calibration for Post-Training Quantization of Diffusion Large Language Models

Diffusion Large Language Models (dLLMs) face a "stability lag" due to irreversible token commitment, a problem exacerbated by Post-Training Quantization (PTQ) errors. FAIR-Calib proposes a two-stage PTQ framework that uses a position prior and layer-wise calibration to protect fragile frontier states, enhancing quantization for dLLMs.