Disaggregated Quantization Optimizes LLM Prefill and Decode Phases

Disaggregated Quantization (DQ) uses specialized compute formats and storage placement to treat prefill and decode separately. On Qwen3 and Gemma 3, training NVFP4 prefillers for 27B models improves 1-bit accuracy by 32.5 points while maintaining high performance at 2-3-bit decode.