Alexandre L.-Piché

🚀 The RL community keeps pushing boundaries — from better on-policy data and partial rollouts to in-flight weight updates that mix KV caches across models during inference. Continuing inference while weights change and KV states stay stale sounds wild — but that’s exactly what PipelineRL makes work. vLLM is proud to power this kind of modular, cutting-edge RL innovation. Give it a try and share your thoughts!

Don't sleep on PipelineRL -- this is one of the biggest jumps in compute efficiency of RL setups that we found in the ScaleRL paper (also validated by Magistral & others before)! What's the problem PipelineRL solves? In RL for LLMs, we need to send weight updates from trainer to generator (to generate data from our latest policy being trained). (Conventional PPO-off-policy) A naive approach would be to "start generators on a batch, wait for all sequences to complete, update the model weights for both trainers and generators, and repeat. Unfortunately, this approach leads to idle generators and low pipeline efficiency due to heterogeneous completion times. (Pipeline-RL) Instead, we simply let the generators continue generating tokens without discarding or finishing ongoing generations in-flight whenever we need to do a weight update -- doing an "in-flight" weight update. As such our KV caches for these generations would be stale, as they would come from LLM with earlier copy(ies) of the weights) but this is ok (see below).

If you’re like Yuki and you are tired of slow RL. Join us in 15 minutes to learn more about Pipeline RL in room 523AB.