AI

LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 01:18:25 UTC.

Briefings

LoGRA is a post-training RL framework that reduces memory usage by 45.7% on reasoning tasks by retaining learning signals in low-rank gradient sketches. It enables stable training of 27B-parameter models for over 1,100 steps on a single eight-GPU node, solving previous memory constraints for dense Adam. The method combines gradient compression with predicted-KL step control to prevent disruptive updates. Code is available via the Molt library.

Key points

• LoGRA reduces average training memory by up to 45.7% across reasoning tasks.
• It enables stable training of a 27B-parameter model for over 1,100 steps on a single eight-GPU node.

Context

Large language models benefit from reinforcement learning, but high memory demands limit adoption. LoGRA addresses this by compressing learning signals into low-rank sketches while maintaining performance.

Limitations

Memory reduction applies specifically to reasoning tasks; dense Adam training remains infeasible on single eight-GPU nodes for 27B models without LoGRA.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 24/10/2026, 18:28:06 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiehmy7xkgycejlj5tm4dttyixeuewmuzyohnwu2dsxdpnxlfymnr4