TRL Adds Tree Packing, Cutting Agentic RL Compute Over 26x

Hugging Face TRL has merged tree packing into AsyncGRPO, exploiting shared rollout trajectories in GRPO to cut pre-training forward/backward compute by over 26x for agentic RL, delivering the same reward at a fraction of the training compute.

2026-10-15 ~ 2026-10-16 · 2 related posts

• Tree packing lands in TRL AsyncGRPO, cutting agentic RL trainer compute 26x — SergioPaniego · 2026-10-15
• Tree packing lands in TRL's AsyncGRPO, cutting agentic RL training compute by over 26x — joecole · 2026-10-16