---
id: nvidia-resiliency-ext
version: "0.6.0"
license: Apache 2.0
license_treatment: permissive
maintenance: active
---

# nvidia-resiliency-ext — NVIDIA Resiliency Package

License: permissive · Maintenance: active · Downloads: 124.6K/mo

## What it is and what it does

NVIDIA Resiliency Extension (NVRx) is a modular toolkit for making large-scale PyTorch distributed training more resilient to hardware faults and performance degradation. It detects hung training ranks, identifies slower GPUs that slow down the entire job, and enables fast recovery without needing to reallocate compute nodes—critical for long-running training on shared infrastructure. The package provides asynchronous and local checkpointing frameworks to minimize work loss, health monitoring across GPU and CPU, and integration hooks for PyTorch Lightning and NeMo frameworks.

The package is actively maintained but explicitly experimental, with frequent breaking changes expected. It depends on a substantial stack including torch, grpcio, protobuf, numpy, and system-level NVIDIA monitoring tools. Installation requires specific Python, PyTorch, CUDA, and driver versions, making it suitable for managed training environments rather than casual experimentation.

Use it for:

- Detect and recover from hung ranks in multi-GPU training without manual job restart or SLURM reallocation.
- Identify straggler GPUs that slow down synchronous training and trigger corrective action.
- Implement frequent asynchronous checkpointing to minimize training loss when hardware fails mid-job.
- Monitor system-wide GPU and CPU health during long-running distributed training runs.
- Integrate fault tolerance into PyTorch Lightning training loops with minimal code changes.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Adds fault tolerance, checkpointing, and performance monitoring to PyTorch distributed training, detecting hung ranks, straggler GPUs, and enabling automatic recovery without reallocating compute resources.

Yes, if you run distributed PyTorch training at scale on shared infrastructure where hardware faults or performance variance are real concerns and you can meet the strict version requirements (Python >=3.10, <3.13; PyTorch >=2.5.1; CUDA >=12.8). The experimental status and frequent breaking changes mean you should pin versions carefully and monitor releases. Not recommended for single-GPU or development-only use cases.

## Install

pip install nvidia-resiliency-ext
uv add nvidia-resiliency-ext
poetry add nvidia-resiliency-ext

## Installing nvidia-resiliency-ext

Before you install: Medium friction: requires Python >=3.10, <3.13, PyTorch >=2.5.1, CUDA >=12.8, and NVML driver >=535. Wheels are available for x86_64 and arm64 on manylinux_2_39. Active maintenance with recent releases, though marked experimental and under active development with frequent breaking changes expected.

License in practice: Apache 2.0 permissive license allows commercial and private use with minimal restrictions, though the project is experimental and contributions are encouraged.

Quickstart:

pip install nvidia-resiliency-ext

import nvidia_resiliency_ext
from nvidia_resiliency_ext.fault_tolerance import FaultTolerance

# Integrate into PyTorch training loop
ft = FaultTolerance()

Requires PyTorch >=2.5.1, CUDA >=12.8, NVML driver >=535, and Python >=3.10, <3.13. NCCL version must be <2.28.3 or >=2.28.9 to avoid inprocess issues.

Verify before relying:

- Whether fault tolerance integration with PyTorch Lightning requires additional setup beyond importing the module.
- Whether straggler detection thresholds and GPU health check specifics are configurable for different workload types.
- Performance overhead of async and local checkpointing relative to standard PyTorch checkpointing.
- Compatibility with mixed-precision training and gradient accumulation patterns.

## Package facts

- License: Apache 2.0 (permissive)
- Python support: supports_current
- Install friction: medium
- Maintenance: active
- Downloads: 124.6K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

pytorch distributed training fault tolerance, gpu training resilience checkpointing, hung rank detection recovery, straggler detection gpu monitoring, async checkpointing pytorch, training job restart without reallocation, distributed training health checks, distributed-training, fault-tolerance, gpu-monitoring

[View on SkillFed](https://skillfed.io/packages/nvidia-resiliency-ext) · [View on PyPI](https://pypi.org/project/nvidia-resiliency-ext/)