Skip to content- Sources: primary
- Summary: HiLo keeps replicated optimizer states on rollout workers, so the learner sends compressed gradients and each worker replays the optimizer step to reach the same policy version. Its Spar compressor combines spectral truncation, adaptive rank allocation and error feedback so training rank no longer sets communication cost. Across on-policy distillation and reward-guided runs it stays close to full synchronization while sending as little as 1.6% of adapter bytes and speeds end-to-end updates up to 2.2 times under bandwidth limits.
- Why it matters: Teams running RL or distillation across machines with slow links can cut weight sync traffic without lowering LoRA rank.
- Background: In on-policy post-training, rollout workers must generate with the latest policy, so every update has to reach them before the next batch of samples. With LoRA only small adapter matrices change, but adaptive optimizers such as Adam produce updates that compress poorly, which is why sending gradients and replaying the optimizer locally helps.