activation-recompute

Official

Reduce GPU memory with selective recompute.

AuthorNVIDIA
Version1.0.0
Installs0

System Documentation

What problem does it solve?

Activation recompute trades GPU memory for compute by discarding intermediate activations during the forward pass and recomputing them during backward in Megatron Bridge.

Core Features & Use Cases

  • Supports selective recompute (per-submodule) and full-layer recompute (per transformer layer) to tune memory-compute tradeoffs.
  • Provides clear enablement guidelines and compatibility notes for CUDA graphs, CPU offloading, and model configurations.
  • Applies to large transformer training workflows where memory is a bottleneck, enabling deeper or larger models without hardware upgrades.

Quick Start

Configure recompute_granularity to selective and set recompute_modules to your target submodules to begin trading memory for compute in Megatron Bridge.

Dependency Matrix

Required Modules

None required

Components

Standard package

💻 Claude Code Installation

Recommended: Let Claude install automatically. Simply copy and paste the text below to Claude Code.

Please help me install this Skill:
Name: activation-recompute
Download link: https://github.com/NVIDIA/skills/archive/main.zip#activation-recompute

Please download this .zip file, extract it, and install it in the .claude/skills/ directory.
View Source Repository

Agent Skills Search Helper

Install a tiny helper to your Agent, search and equip skill from 471,000+ vetted skills library on demand.