Fixed State, Long Reach: What a Constant-Size Cache Buys Block Diffusion at Scale
This work pretraining three 3B block-diffusion denoisers (attention, mamba, and hybrid) on 300B tokens under one single-frontier objective and decoding all three through a single cached interface finds the Mamba cache delivers 4.3x lower latency, 11x less memory, and 2.6x higher single-stream throughput.
Vaibhav Singh, Pierre-André Noël, Torsten Scholak et al.
· 1 citation