NVIDIA Unveils LoGRA, Cutting RL Training Memory by Up to 45.7%

NVIDIA introduced LoGRA, a memory optimization method for LLM reinforcement-learning post-training that compresses gradients into low-rank gradient sketches, cutting training memory by up to 45.7% while keeping updates stable.

2026-10-07 ~ 2026-10-08 · 2 related posts

• NVIDIA's LoGRA cuts LLM RL training memory by up to 45.7%, enables 27B RL on a single 8-GPU node — nvidia · 2026-10-07
• NVIDIA's LoGRA cuts RL training memory by up to 45.7%, trains 27B model where Adam OOMs — mark_k · 2026-10-08