SoloQ: Calibration-Free Quantization for Diffusion Language Models
SoloQ claims 4-bit diffusion LLM inference without calibration data, while preserving accuracy and cutting memory and latency.
The paper targets diffusion language models, where activation distributions shift across masking states and denoising steps. SoloQ rotates and rescales weights and activations into a normalized basis so quantization can be done data-independently. For block-diffusion models, it also quantizes committed KV-cache states without disturbing the block still being denoised. The authors report up to 2.61x peak-memory reduction and 2.24x end-to-end inference speedup with NVFP4. HF Daily Papers' note
The paper targets diffusion language models, where activation distributions shift across masking states and denoising steps. SoloQ rotates and rescales weights and activations into a normalized basis so quantization can be done data-independently. For block-diffusion models, it also quantizes committed KV-cache states without disturbing the block still being denoised. The authors report up to 2.61x peak-memory reduction and 2.24x end-to-end inference speedup with NVFP4. HF Daily Papers' note
score 4