NVIDIA's LoGRA Cuts RL Training Memory by 45.7%

NVIDIA's new LoGRA method accumulates low-rank gradients via rank-256 random projections during backpropagation, avoiding full gradient construction. It cuts RL training memory by 45.7%—lower than LoRA—and enables training a 27B model on 8 GPUs.

2026-10-12 ~ 2026-10-13 · 2 related posts

• LoGRA cuts RL training memory by 45.7%, enabling 27B RL on a single 8-GPU node — 青稞AI · 2026-10-12
• NVIDIA's LoGRA cuts RL training memory by 45.7%, using less than LoRA — omarsar0 · 2026-10-13