AsyncGRPOTrainer adds LoRA support, validated on FSDP2 setup

QGallouedec · x · 2026-09-03

Developer @DirhousssiAmine announced a major update to AsyncGRPOTrainer: LoRA training is now supported. It was successfully tested on the sanity-dataset using an FSDP2 setup, meaning GRPO reinforcement-learning fine-tuning can now run with far fewer trainable parameters and lower memory requirements.

Original post →

More from coding & agent

coding & agent channel →