SoloQ: calibration-free 4-bit quantization cuts dLLM memory 2.61x, speeds inference 2.24x

UniversityofSouthernCalifornia · hf · 2026-10-12

USC researchers (with Yale and Korea University) released SoloQ, a calibration-free post-training quantization framework for diffusion LLMs (dLLMs).

Problem: dLLMs' full-sequence denoising incurs heavy forward passes and block-diffusion models add a memory-hungry KV-cache; existing PTQ methods need calibration data even as activation distributions shift across masking states and denoising steps
Method: SoloQ maps weights and activations into a normalized rotated basis via structured K-RPBH rotation plus a lightweight rescaling correction, enabling data-independent quantization; supports distribution-matched codebooks and hardware-native NVFP4, plus commit-time KV-cache quantization for block-diffusion models
Results: retains accuracy at 4-bit across LLaDA, Dream, Fast-dLLM v2 and Nemotron-Labs-Diffusion, outperforming calibration-based baselines on knowledge/reasoning benchmarks; NVFP4 cuts peak memory up to 2.61x and accelerates end-to-end inference up to 2.24x

More from Infra