Get Started

Install NeMo AutoModel

View as Markdown

This guide explains how to install NeMo AutoModel for LLM, VLM, and OMNI models on various platforms and environments. Depending on your use case, there are several ways to install it:

MethodDev ModeUse CaseRecommended For
📦 PyPI-Install stable release with minimal setupMost users, production usage
🐳 Docker-Use a preconfigured, isolated CUDA userspace environmentMulti-node deployments
🐍 Git Repo✅Install the latest base package directly from Git; select extras explicitlyPower users, testers
🧪 Editable Install✅Contribute to the codebase or make local modificationsContributors, researchers
🐳 Docker + Mount✅Develop against mounted source with container-provided CUDA dependenciesMulti-node deployments

Choose Your Installation Method

Pick the installation method that matches your needs and platform.

Decision Criteria

MethodBest ForProsCons
Docker ContainerProduction, multi-node, Debian-based systemsReproducible environment, preconfigured CUDA userspace and dependenciesRequires a compatible host NVIDIA driver; larger download size and container overhead
virtualenv (PyPI/Git)Local development, quick prototyping, macOSFast setup, lightweight, direct code accessManual dependency management, platform-specific issues

When to Use Docker Containers

Use Docker containers when you need:

  • Multi-node deployments: Containers ensure consistency across cluster nodes
  • Production environments: Reproducible builds with tested dependency versions
  • CUDA userspace isolation: Containers supply CUDA libraries and dependencies, but use the host NVIDIA driver; the driver must support the CUDA runtime in the container
  • Debian-based systems: Recommended for Ubuntu, Debian, and derivatives due to dependency complexity
  • Complex dependencies: Preconfigured environment with optimized CUDA packages such as TransformerEngine and DeepEP
  • Team consistency: Same environment across development, testing, and production

When to Use virtualenv

Use virtualenv (PyPI, Git, or editable install) when you need:

  • Local development: Fast iteration on code changes
  • Quick prototyping: Minimal setup for experimentation
  • macOS systems: Better native support without container overhead
  • Frequent code changes: Contributors working on the codebase (use editable install)
  • Compatible GPU drivers: The host driver supports the CUDA runtime selected by PyTorch
  • Lightweight setup: Minimal disk space and memory footprint

Platform-Specific Recommendations

Linux (Debian-based: Ubuntu, Debian)

Recommended: Docker Container

Debian-based systems can have dependency conflicts with system packages. Containers provide isolation and consistency.

$docker pull nvcr.io/nvidia/nemo-automodel:26.06.00
$docker run --gpus all -it --rm --shm-size=8g nvcr.io/nvidia/nemo-automodel:26.06.00

Alternative: virtualenv (if Docker is not available)

Ensure the host NVIDIA driver supports the CUDA runtime selected by your installation. The current source lock selects PyTorch cu130 wheels on Linux; prebuilt wheels do not require a separately installed CUDA toolkit.

$# Check the installed driver and its maximum supported CUDA version.
$nvidia-smi
$
$# Create an environment and install via PyPI.
$uv venv
$uv pip install "nemo-automodel"

Linux (RHEL, CentOS, Fedora)

Recommended: Docker Container

Containers avoid enterprise Linux package management complexity.

Follow the same Docker commands as Debian-based systems above.

macOS

Recommended: virtualenv

Docker on macOS has GPU limitations. Use native Python installation:

$# Create a local environment and install from PyPI.
$uv venv
$uv pip install "nemo-automodel"

GPU training on macOS is not supported. Use macOS for CPU-based experimentation or remote cluster submission.

Windows

Recommended: WSL2 + Docker

Run NeMo AutoModel in WSL2 with Docker Desktop:

  1. Install WSL2 and Docker Desktop.
  2. Use Docker container within WSL2 (follow Linux instructions).

Alternative: WSL2 and virtualenv

Install directly in WSL2 Ubuntu environment (follow Debian instructions).

Common Issues and Solutions

GPU driver compatibility errors

  • Problem: CUDA version mismatch between host and application
  • Solution: Update the host NVIDIA driver to support the CUDA runtime in the selected wheel or container, or select a runtime compatible with the installed driver. Docker does not replace or isolate the host GPU driver

Dependency conflicts on Debian/Ubuntu

  • Problem: System packages conflict with Python packages
  • Solution: Use Docker container or create isolated virtualenv with uv

Out of memory during container startup

  • Problem: Insufficient shared memory for PyTorch data loading
  • Solution: Increase --shm-size parameter (e.g., --shm-size=16g)

TransformerEngine import failures

  • Problem: Incorrect CUDA toolkit or missing dependencies
  • Solution: Use pre-configured Docker container

Prerequisites

System Requirements

  • Python: 3.10 or higher
  • GPU runtime: The current source lock selects PyTorch cu130 wheels on Linux, while the current container build uses a CUDA 13.2 development image. These are separate installation artifacts, not a blanket CUDA 11.8+ requirement
  • NVIDIA driver: For GPU execution, the host driver must support the CUDA runtime shipped by the selected PyTorch wheel or container
  • CUDA toolkit: A host toolkit is not required for prebuilt wheels or containers. Install one only when building CUDA extensions from source, and match it to the selected PyTorch or container runtime
  • Memory: Minimum 16GB RAM, 32GB+ recommended
  • Storage: At least 50GB free space for models and datasets

Hardware Requirements

  • GPU: NVIDIA GPU with 8GB+ VRAM (16GB+ recommended)
  • CPU: Multi-core processor (8+ cores recommended)
  • Network: Stable internet connection for downloading models

Installation Options for Non-Developers

This section explains the easiest installation options for non-developers, including using uv with PyPI or leveraging a preconfigured NVIDIA NeMo Docker container. Both methods offer quick access to a stable NeMo AutoModel installation and its base dependency set.

For most users, the easiest way to get started is using uv:

$uv venv
$uv pip install "nemo-automodel"

This installs the latest stable release of NeMo AutoModel from PyPI.

To verify the install in the environment:

$uv run --no-project python -c \
>  "import nemo_automodel; print(nemo_automodel.__version__)"

See nemo-automodel on PyPI.

Install with NeMo Docker Container

You can use NeMo AutoModel with the NeMo Docker container. Pull the container by running:

$docker pull nvcr.io/nvidia/nemo-automodel:26.06.00

The above docker command uses the 26.06.00 container. Use the most recent container version to ensure you get the latest version of AutoModel and its dependencies like PyTorch, Transformers, etc.

Then you can enter the container using:

$docker run --gpus all -it --rm \
>  --shm-size=8g \
>  -v "$(pwd)"/checkpoints:/opt/Automodel/checkpoints \
>  nvcr.io/nvidia/nemo-automodel:26.06.00

Persist your checkpoints. By default, checkpoints are written to checkpoints/ inside the container. Because --rm destroys the container on exit, any data stored only inside the container is lost. Always bind-mount a host directory for the checkpoint path (as shown with -v above) so that your trained weights survive after the container stops. You can also mount additional directories for datasets and Hugging Face cache:

$docker run --gpus all -it --rm \
>  --shm-size=8g \
>  -v /path/to/your/checkpoints:/opt/Automodel/checkpoints \
>  -v /path/to/your/datasets:/datasets \
>  -v /path/to/your/hf_cache:/root/.cache/huggingface \
>  nvcr.io/nvidia/nemo-automodel:26.06.00

Models that require CUDA-specific packages (e.g., Nemotron). Some model families—such as Nemotron Nano and Nemotron Flash—depend on packages like mamba-ssm and causal-conv1d. If these packages are built from source, the build requires a CUDA toolkit that matches the selected PyTorch runtime. The NeMo Automodel Docker container ships with these dependencies pre-built, so using the container is the recommended approach for fine-tuning Nemotron and other models with similar requirements.

Installation Options for Developers

This section provides installation options for developers, including pulling the latest source from GitHub, using editable mode, or mounting the repo inside a NeMo Docker container.

Install from GitHub (Source)

If you want the latest features from the main branch or want to contribute:

Use uv With Git Repo

$# Create the environment before using uv pip in a clean directory.
$uv venv
$
$# Install the base package.
$uv pip install "nemo-automodel @ git+https://github.com/NVIDIA-NeMo/Automodel.git"
$
$# Or request the standard extras bundle in the requirement spec.
$uv pip install "nemo-automodel[all] @ git+https://github.com/NVIDIA-NeMo/Automodel.git"

uv venv creates .venv; subsequent uv pip commands in this directory use that environment. The first Git requirement installs only the base dependencies; optional extras are installed only when the requirement includes an extras list.

Install in Developer Mode (Editable Install)

To contribute or modify the code:

$git clone https://github.com/NVIDIA-NeMo/Automodel.git
$cd Automodel
$uv sync --locked --all-groups --extra all

uv sync installs AutoModel in editable mode, so changes to the code are immediately reflected in Python.

Mount the Repo into a NeMo Docker Container

To develop against your local checkout while reusing the container’s pre-built dependencies, bind-mount your local Automodel directory over /opt/Automodel, overriding the source installed in the image:

$# Step 1: Clone the AutoModel repository.
$git clone https://github.com/NVIDIA-NeMo/Automodel.git
$cd Automodel
$
$# Step 2: Pull a compatible container image (replace the tag as needed).
$docker pull nvcr.io/nvidia/nemo-automodel:26.06.00
$
$# Step 3: Run the container, sync the mounted checkout, and run a usage example.
$docker run --gpus all --network=host -it --rm \
>  --shm-size=32g \
>  -v "$(pwd)":/opt/Automodel \
>  nvcr.io/nvidia/nemo-automodel:26.06.00 /bin/bash -c "\
>    cd /opt/Automodel && \
>    bash docker/common/update_pyproject_pytorch.sh /opt/Automodel && \
>    uv sync --locked --extra all --all-groups && \
>    automodel examples/llm_finetune/llama3_2/llama3_2_1b_squad.yaml"

The update_pyproject_pytorch.sh step is required. Without it, uv sync tries to reinstall torch, which causes CUDA version mismatches and TransformerEngine import failures, because uv cannot recognize the torch baked into the PyTorch base container.

Install Profiles

NeMo AutoModel provides several install extras for different use cases.

Base Install (default)

Installs the base library dependencies, including PyTorch. It does not install the optional TransformerEngine, Mamba, and other packages in the cuda extra:

$uv venv
$uv pip install "nemo-automodel"

Extras Reference

The following table covers every optional profile declared by the package. “Included in [all]” means the complete profile is pulled in by that bundle.

ExtraPurposeIncluded in [all]
cliNeMo-Run launcher supportYes
cudaTransformerEngine, Mamba, causal convolution, grouped GEMM, and TileLang CUDA packagesYes
cuda_sourcebitsandbytesNo
delta-databricksDelta Lake and Databricks accessYes
diffusionDiffusion modelsYes
diffusion_kernelsDiffusion plus the kernels packageNo
extraPerceptron and SentencePieceYes
faFlashAttentionNo
ffpaFFPA attentionNo
flaFlash Linear AttentionYes
mediaUnion of vlm-media and diffusion-mediaNo
moecuda, fla, and DeepEPNo; DeepEP is omitted
mscMulti-Storage ClientNo
s3Amazon S3 supportYes
vlmCore vision-language model packagesYes
vlm-mediaVLM media decoding and preprocessingNo
diffusion-mediaDiffusion media preprocessing and exportNo
allStandard bundle described belowBundle

CLI and NeMo-Run Extra (Login Nodes)

If your login node or CI environment needs the AutoModel CLI together with the NeMo-Run launcher, install the cli extra:

$uv venv
$uv pip install "nemo-automodel[cli]"

The cli extra is additive: it installs the normal nemo-automodel base dependencies, including PyTorch, and adds nemo-run. PyYAML is already a base dependency and is also listed in the extra. The automodel and am entry points come from the base package, and no separate NeMo-Run installation step is required. This extra is not a reduced or CLI-only package profile.

VLM Dependencies

For vision-language model training, add the VLM extras:

$uv venv
$uv pip install "nemo-automodel[vlm]"

[vlm] installs the core VLM stack (image processors, tokenizers). Models that decode video or use OpenCV / Qwen vision preprocessing — Qwen2.5-VL, Qwen3-VL, Qwen3-Omni, Mistral VLMs, and Nemotron-Omni — additionally require the vlm-media extra (see Media Extras):

$uv venv
$uv pip install "nemo-automodel[vlm,vlm-media]"

Media Extras (Video & Image Decode)

Media decode dependencies (OpenCV, decord, PyAV via the Qwen utils, and imageio-ffmpeg) bundle FFmpeg, which is no longer packaged in the container. They are opt-in and split by domain:

ExtraPulls inNeeded for
vlm-mediaOpenCV, decord, qwen-vl-utils, qwen-omni-utilsVLM video reading, Qwen vision preprocessing, Mistral image tokenization
diffusion-mediaimageio-ffmpeg, OpenCVDiffusion image/video preprocessing and T2V export
mediaboth of the aboveConvenience union
$uv venv
$uv pip install "nemo-automodel[vlm,vlm-media]"              # VLM + media
$uv pip install "nemo-automodel[diffusion,diffusion-media]"  # diffusion + media

[all] does not include the media extras, and they are not installed in the Docker container by default — FFmpeg is no longer packaged in the container. To install the standard bundle together with media in a fresh environment:

$uv venv
$uv pip install "nemo-automodel[all,media]"

CUDA-Specific Packages

For models requiring TransformerEngine, Mamba, or the other CUDA-compiled packages in the cuda extra:

$uv venv
$uv pip install "nemo-automodel[cuda]"

bitsandbytes is provided by the separate cuda_source extra:

$uv venv
$uv pip install "nemo-automodel[cuda_source]"

Standard Extras Bundle

The standard bundle includes exactly cli, cuda, delta-databricks, diffusion, extra, fla, s3, and vlm:

$uv venv
$uv pip install "nemo-automodel[all]"

[all] excludes cuda_source, diffusion_kernels, fa, ffpa, moe, msc, vlm-media, diffusion-media, and their media union. Add the extras you need explicitly. The media example above installs [all] and [media] together.

You can combine extras:

$uv venv
$uv pip install "nemo-automodel[vlm,cuda]"

Summary

GoalCommand or Method
Stable install (PyPI)uv venv, then uv pip install "nemo-automodel"
Base package + NeMo-Run extrauv venv, then uv pip install "nemo-automodel[cli]"
CUDA packages (TransformerEngine, Mamba, etc.)uv venv, then uv pip install "nemo-automodel[cuda]"
bitsandbytesuv venv, then uv pip install "nemo-automodel[cuda_source]"
VLM + media (Qwen/Mistral/video)uv venv, then uv pip install "nemo-automodel[vlm,vlm-media]"
Diffusion + mediauv venv, then uv pip install "nemo-automodel[diffusion,diffusion-media]"
Latest base package from GitHubuv venv, then uv pip install "nemo-automodel @ git+https://github.com/NVIDIA-NeMo/Automodel.git"
Editable install (dev mode)uv sync --locked --all-groups --extra all after cloning
Use mounted source in DockerRun update_pyproject_pytorch.sh, then uv sync --locked --all-groups --extra all