32 ms·Faster RL Post-Training Rollouts via System-Integrated Speculative Decoding1 points by gmays 6mo ago