Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training
An end-to-end system for large-scale online draft co-training that achieves strong scaling at 256K tokens with significant memory savings over prior work, and the PP transport incurs modest overhead.
Zi-Li Wang, Zhao-Peng Qiu, Yuekai Zhang et al.
· 3 citations
· ⚡1