resiliency

Integrate nvidia-resiliency-ext fault tolerance into Megatron Bridge training workflows.

2.8k|332|Updated Feb 25, 2026
One-click install
npx skills add https://github.com/NVIDIA/skills --skill resiliency
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: resiliency
Source: https://github.com/NVIDIA/skills/tree/main/skills/Megatron-Bridge/resiliency
Command: npx skills add https://github.com/NVIDIA/skills --skill resiliency

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Megatron Bridge resiliency features provide fault tolerance, straggler detection, in-process restart, preemption, and a re-run state machine to improve stability of large-scale training workflows. Use when you need hang detection, automatic restart, or checkpoint recovery in distributed training.

Core Features & Use Cases

  • Fault tolerance using nvidia-resiliency-ext to automatically recover from failures during long runs.
  • Straggler detection and graceful handling to maximize resource utilization in multi-node training.
  • Experimental in-process restart and re-run state machine for rapid recovery without full job relaunch.
  • Preemption support and non-blocking checkpointing to minimize downtime during cluster pressure.
  • Async checkpointing and local checkpointing to accelerate recovery and reduce loss of progress.

Quick Start

Enable resiliency features by turning on the FT, NVRx straggler, IPR, and checkpointing options in Megatron Bridge.

Frequently Asked Questions about resiliency

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I add fault tolerance to distributed ML training?

Fault tolerance for distributed ML training is added by integrating nvidia-resiliency-ext into Megatron Bridge, enabling automatic recovery from failures during long runs. Turn on the FT options in Megatron Bridge to start.

What is straggler detection in multi-node training?

Straggler detection in multi-node training identifies slow nodes that hinder resource utilization, allowing graceful handling to maximize efficiency. Enable the NVRx straggler option in Megatron Bridge to activate this monitoring.

Can I restart a hung distributed training job without relaunching it?

In-process restart allows rapid recovery for a hung distributed training job without full job relaunch. Enable the experimental IPR and re-run state machine options in Megatron Bridge to trigger this quick recovery.

How do I minimize downtime during cluster preemption and checkpointing?

Downtime during cluster preemption and checkpointing is minimized using preemption support and non-blocking checkpointing. Enable async checkpointing and local checkpointing in Megatron Bridge to accelerate recovery and reduce progress loss.

Does Megatron Bridge support hang detection for large-scale model training?

Megatron Bridge supports hang detection for large-scale model training by integrating nvidia-resiliency-ext features. Use it when you need automatic restart or checkpoint recovery in distributed training workflows.

When should I enable resiliency features in my training workflow?

Enable resiliency features when your large-scale training workflow requires hang detection, automatic restart, straggler handling, or checkpoint recovery across distributed nodes. Turn on the FT, NVRx straggler, IPR, and checkpointing options in Megatron Bridge.