SoloQ: Calibration-Free Quantization for Diffusion Language Models

SoloQ. Calibration-Free Quantization for Diffusion Language Models. By Donghyun Lee and seven others, from the University of Southern California, Yale University, and one other institution. Published on the fifth of October. It links to a code repository that is no longer there.

Diffusion large language models generate tokens by bidirectional diffusion rather than one after another, and they are a promising alternative to autoregressive models. They are also expensive to deploy. Full-sequence denoising calls compute-heavy forward passes over and over, and block-diffusion models add a memory-heavy KV-cache on top. Quantizing both weights and activations to low bit widths is the obvious remedy. The catch is that existing post-training quantization methods for these models depend on calibration data, while activation distributions keep shifting across masking states and denoising steps. Calibrating against a target that moves is the problem.

SoloQ removes calibration altogether. It maps weights and activations into a normalized, rotated basis whose marginal distribution is predictable. If you know in advance what the distribution will look like, you do not need data to discover it, and quantization becomes data-independent. That sidesteps the moving target entirely. There is no calibration set to choose, and so no calibration set that might fail to represent a masking state or a denoising step the model meets later.

Two pieces do the mapping: a structured K-RPBH rotation, and a lightweight rescaling correction. Because the distribution after rotation is predictable, SoloQ can use codebooks matched to that distribution, and it also supports the hardware-native NVFP4 format. Block-diffusion models get one more component. SoloQ quantizes the KV-cache at commit time, compressing the persistent states without disturbing the block that is still being actively denoised.

Tests cover full-sequence models, LLaDA and Dream, and block-diffusion models, Fast-dLLM v2 and Nemotron-Labs-Diffusion. Under four-bit quantization SoloQ retains accuracy, and it outperforms calibration-based baselines on benchmarks that lean on knowledge and reasoning. Those baselines had data to tune against; SoloQ had none, and still came out ahead.

The efficiency numbers come from the hardware-native path. With NVFP4, peak memory falls by up to two point six one times, and end-to-end inference runs up to two point two four times faster.

The move is to make the quantity being quantized predictable before trying to quantize it. Once the rotation fixes what the distribution looks like, the shifting activations of a diffusion model stop being a calibration problem, and the same predictability opens the door to a hardware-native format.