Container Environments | NeMo Curator

Deploy NeMo Curator in containerized environments for reproducible, scalable data curation pipelines with pre-configured dependencies and optimized runtime settings.

Overview

NeMo Curator provides official Docker containers with all dependencies pre-installed and optimized for production workloads. Containers offer:

Reproducible Environments: Consistent software stack across development, testing, and production
Simplified Deployment: No manual dependency installation or environment configuration
GPU Acceleration: Pre-configured CUDA, cuDNN, and NVIDIA libraries for optimal performance
Multi-Modal Support: Built-in support for text, image, video, and audio curation
Cloud-Ready: Compatible with Kubernetes, Docker Swarm, and cloud container orchestries

When to use containers:

Production deployments requiring consistency and reliability
Multi-node cluster processing with identical environments
CI/CD pipelines for automated data curation workflows
Quick prototyping without local environment setup
GPU-accelerated processing in cloud environments

Available Containers

Main NeMo Curator Container

The primary container includes comprehensive support for all curation modalities:

Container registry: nvcr.io/nvidia/nemo-curator:{{ container_version }}

Supported modalities (amd64/x86_64):

✅ Text curation (CPU/GPU)
✅ Image curation (GPU required)
✅ Video curation (GPU required, FFmpeg included)
✅ Audio curation (GPU required for ASR; x86_64 Linux only)

arm64/aarch64 containers: Text, image, and video curation are supported on arm64 builds. Audio curation is not supported on arm64 because key dependencies (NeMo ASR, diarization, and related packages) are gated to x86_64 Linux in pyproject.toml. Missing audio packages on the arm64 image are expected, not a packaging defect. Use an amd64 container or host for audio workflows.

Pre-installed components:

NeMo Curator with all optional dependencies ([all] extras)
CUDA 12.8.1 with cuDNN
Python 3.12 with uv package manager
FFmpeg 8+ with NVENC support (for video processing)
Ray and its distributed computing libraries
NVIDIA optimized Python packages

Ray is the execution framework for all NeMo Curator pipelines. Dask is no longer used to run curation work; dask, dask-cuda, and dask-cudf are present in the image only as transitive dependencies of the RAPIDS raft-dask-cu12 package.

Curator Environment

Property	Value
Python Version	3.12
CUDA Version	12.8.1 (configurable)
Operating System	Ubuntu 24.04 (configurable)
Base Image	`nvidia/cuda:${CUDA_VER}-cudnn-devel-${LINUX_VER}`
Package Manager	uv (Ultrafast Python package installer)
Installation	NeMo Curator installed with all optional dependencies (`[all]` extras) using uv with NVIDIA index
Environment Path	Virtual environment at `/opt/venv`. Activate with `source /opt/venv/env.sh` after entering the container.

Security Hardening

The container build includes the following security measures:

ray_dist.jar removal: Ray’s Java support JAR is deleted during the build to remove a bundled jackson-core library affected by GHSA-72hv-8253-57qq (DoS via async JSON parser). NeMo Curator does not use Ray’s Java support, so this has no functional impact. A build-time verification guard fails the build if the JAR is not successfully removed.

Container Build Arguments

The main container accepts these build-time arguments for environment customization:

Argument	Default	Description
`CUDA_VER`	`12.8.1`	CUDA version
`LINUX_VER`	`ubuntu24.04`	Base OS version
`CURATOR_ENV`	`ci`	Curator environment type
`NVIDIA_BUILD_ID`	`<unknown>`	NVIDIA build identifier
`NVIDIA_BUILD_REF`	-	NVIDIA build reference

Environment Usage Examples

Text Curation

Uses the default container environment with CPU or GPU workers depending on the module.

Image Curation

Requires GPU-enabled workers in the container environment.