Skip to content
Trending storyDeveloping

NVIDIA's LoGRA trains all LLM weights with RL using less memory than LoRA

1 article1 sourcesince Oct 12Last article 2h ago ·

Overview

AISummary of 1 article

NVIDIA proposes LoGRA, a low-rank gradient sketch method for reinforcement learning post-training of LLMs that updates all of a model's weights.

Elvis Saravia reports that at 1.5B parameters, LoGRA averages 7.2 GiB of training memory versus 13.2 GiB for rank-256 LoRA, with Pass@4 about equal (81.8% vs. 81.4%) and Pass@1 about 2 points lower. At 7B, average training memory per GPU drops 45.7%, from 31.8 to 17.3 GiB, with peak accuracy close to dense Adam (72.3% and 72.5% Pass@1).

The report says LoGRA trains a 27B model for over 1,100 steps on one eight-GPU node, where dense Adam runs out of memory.

Written by AI from the articles below · updated Oct 12, 6:48 PM ET

Check the sources:

Latest developmentOct 12, 6:20 PM ET
NVIDIA's LoGRA cuts RL training memory below LoRA for LLMs

Article timeline

The articles in this story. Times are ET.

Oct 12
  1. elvisX
    NVIDIA's LoGRA cuts RL training memory below LoRA for LLMs

    AINVIDIA's LoGRA trains all of a model's weights with reinforcement learning using less memory than rank-256 LoRA. At 1.5B, it averages 7.2 GiB of training memory versus 13.2 GiB for LoRA, with Pass@4 about equal (81.8% vs. 81.4%) and Pass@1 about 2 points lower. At 7B, average training memory per GPU drops 45.7%, from 31.8 to 17.3 GiB, and the method trains a 27B model for over 1,100 steps on one eight-GPU node where dense Adam runs out of memory.

    Image from @omarsar0's post

Heat trend

Current comparable heat
9
Peak comparable heat
10Oct 12, 7:00 PM ET
Change in past 24 hours
–

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.