Researchers at the University of Southern California introduced SoloQ, a calibration-free quantization framework for diffusion language models, in a paper published on October 5 and submitted to Hugging Face on October 12. Diffusion language models generate tokens through bidirectional diffusion, but repeated full-sequence denoising requires compute-intensive forward passes; block-diffusion models also use memory-intensive KV caches. Existing post-training quantization methods rely on calibration data, even though activation distributions change across masking states and denoising steps.

SoloQ maps weights and activations into a normalized rotated basis with a predictable marginal distribution, allowing quantization without calibration data. It combines structured K-RPBH rotation with a lightweight rescaling correction. The resulting post-rotation distributions support distribution-matched codebooks and hardware-native NVFP4. For block-diffusion models, SoloQ also applies commit-time KV-cache quantization to compress persistent states without perturbing the block being actively denoised.

The researchers tested SoloQ on full-sequence models LLaDA and Dream, and block-diffusion models Fast-dLLM v2 and Nemotron-Labs-Diffusion. Under 4-bit quantization, it retained accuracy and outperformed calibration-based baselines on knowledge- and reasoning-intensive benchmarks. With NVFP4, SoloQ reduced peak memory by up to 2.61X and accelerated end-to-end inference by up to 2.24X. The paper lists Donghyun Lee, Arkapravo Ghosh, Varun Manjunath, Bumjoon Kyle Rhee, Hyunho Kook, Shiting Xiao, Youngeun Kim and Priyadarshini Panda as authors.

Researchers at the University of Southern California introduced SoloQ, a calibration-free quantization framework for diffusion language models, in a paper published on October 5 and submitted to Hugging Face on October 12. Diffusion language models generate tokens through bidirectional diffusion, but repeated full-sequence denoising requires compute-intensive forward passes; block-diffusion models also use memory-intensive KV caches. Existing post-training quantization methods rely on calibration data, even though activation distributions change across masking states and denoising steps.

SoloQ maps weights and activations into a normalized rotated basis with a predictable marginal distribution, allowing quantization without calibration data. It combines structured K-RPBH rotation with a lightweight rescaling correction. The resulting post-rotation distributions support distribution-matched codebooks and hardware-native NVFP4. For block-diffusion models, SoloQ also applies commit-time KV-cache quantization to compress persistent states without perturbing the block being actively denoised.

The researchers tested SoloQ on full-sequence models LLaDA and Dream, and block-diffusion models Fast-dLLM v2 and Nemotron-Labs-Diffusion. Under 4-bit quantization, it retained accuracy and outperformed calibration-based baselines on knowledge- and reasoning-intensive benchmarks. With NVFP4, SoloQ reduced peak memory by up to 2.61X and accelerated end-to-end inference by up to 2.24X. The paper lists Donghyun Lee, Arkapravo Ghosh, Varun Manjunath, Bumjoon Kyle Rhee, Hyunho Kook, Shiting Xiao, Youngeun Kim and Priyadarshini Panda as authors.