Diffusion large language models (dLLMs) are emerging as a promising alternative to autoregressive models by generating tokens through bidirectional diffusion rather than left-to-right decoding. But their growing size and the repeated forward passes required for full-sequence denoising make deployment costly, and block-diffusion variants add a memory-heavy KV-cache on top of that.
Low-bit weight-activation quantization is an attractive remedy, yet existing post-training quantization methods for dLLMs depend on calibration data. That is a problem because activation distributions shift across masking states and denoising steps, undermining the assumptions behind calibration-based approaches.
Enter SoloQ, a calibration-free quantization framework from researchers publishing on Hugging Face. SoloQ maps weights and activations into a normalized rotated basis with a predictable marginal distribution, enabling data-independent quantization. It combines a structured K-RPBH rotation with a lightweight rescaling correction, and its predictable post-rotation distribution supports both distribution-matched codebooks and hardware-native NVFP4.
For block-diffusion models, SoloQ adds commit-time KV-cache quantization, compressing persistent states without disturbing the block currently being denoised.
Across full-sequence dLLMs (LLaDA and Dream) and block-diffusion dLLMs (Fast-dLLM v2 and Nemotron-Labs-Diffusion), SoloQ retains accuracy under 4-bit quantization and outperforms calibration-based baselines on knowledge- and reasoning-intensive benchmarks. With NVFP4, it reduces peak memory by up to 2.61x and accelerates end-to-end inference by up to 2.24x.