Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training
NVIDIA researchers detail an end-to-end system for online draft co-training that speeds speculative decoding in large-scale long-context RL post-training.