@whitecircle open-sourced Halo today - their in-house training framework.
Halo delivers 2.3–2.8× the throughput of stock TRL at lower peak memory. It also scales past fine-tuning to agentic and asynchronous RL, with SGLang handling rollout generation.
Halo delivers 2.3–2.8× the throughput of stock TRL at lower peak memory. It also scales past fine-tuning to agentic and asynchronous RL, with SGLang handling rollout generation.
0:42