Build And Dependency

Build and Dependency

Quick Start

Clone and install:

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

Or use the NeMo-AutoModel container from NVIDIA NGC (pick a published tag from the NGC catalog — e.g. 26.04):

docker pull nvcr.io/nvidia/nemo-automodel:26.04
docker run --gpus all -it nvcr.io/nvidia/nemo-automodel:26.04

Installation Options

Option 1: NeMo-AutoModel Container (NGC)

The container ships with all dependencies pre-installed at /opt/Automodel (WORKDIR) with the venv at /opt/venv. Run as-is:

docker run --gpus all --network=host -it --rm --shm-size=32g \
    nvcr.io/nvidia/nemo-automodel:26.04 /bin/bash

Mounting your local checkout into the container

To develop against your host checkout, bind-mount it over /opt/Automodel to override the installed source:

docker run --gpus all --network=host -it --rm --shm-size=32g \
    -v <local-Automodel-path>:/opt/Automodel \
    nvcr.io/nvidia/nemo-automodel:26.04 /bin/bash

Inside the container, patch pyproject.toml / uv.lock for the PyTorch base image, then re-sync:

cd /opt/Automodel
bash docker/common/update_pyproject_pytorch.sh /opt/Automodel
uv sync --locked --all-groups --extra all

Warning: the update_pyproject_pytorch.sh step is required. Without it, uv sync will try to reinstall torch, which leads to CUDA version mismatches and TE import failures — uv cannot recognize the torch baked into the PyTorch base container.

Option 2: uv (Recommended for Local Development)

--all-groups pulls the build, docs, and test dev groups (defined in pyproject.toml); drop it for a runtime-only install.

uv sync --locked --all-groups                          # base + dev groups
uv sync --locked --all-groups --extra cuda             # CUDA support
uv sync --locked --all-groups --extra fa               # flash-attention
uv sync --locked --all-groups --extra moe              # mixture-of-experts
uv sync --locked --all-groups --extra vlm              # vision-language models (core)
uv sync --locked --all-groups --extra vlm-media        # + video/Qwen/Mistral decode (opencv, decord, qwen-utils; FFmpeg-bearing)
uv sync --locked --all-groups --extra diffusion        # diffusion models
uv sync --locked --all-groups --extra diffusion-media  # + diffusion preprocessing/export (imageio-ffmpeg, opencv)
uv sync --locked --all-groups --extra media            # vlm-media + diffusion-media (union)
uv sync --locked --all-groups --extra delta-databricks # Delta Lake / Databricks
uv sync --locked --all-groups --extra all              # all standard extras (EXCLUDES media — FFmpeg kept opt-in)

The media extras (vlm-media, diffusion-media, media) bundle FFmpeg and are deliberately excluded from all and from the container image — add them explicitly for video/image decode.

Option 3: pip

Full install (matches uv sync --extra all):

pip install -e ".[all]"

Login-node / submitter-only install — lightweight package for SLURM, k8s, or NeMo-Run job submission without local CUDA deps:

pip install nemo-automodel[cli]

Package Management

Always use uv. Do not introduce pip install commands in scripts or docs.

Environment Variables

export HF_TOKEN="hf_..."           # Hugging Face token for gated models
export WANDB_API_KEY="..."         # Weights & Biases logging
export HF_HOME="/path/to/hf_cache" # Hugging Face cache directory

CLI Usage

The entry point is automodel (defined at nemo_automodel._cli.app:main).

Pattern: automodel <command> <domain> -c <config.yaml>

# LLM
automodel finetune llm -c examples/llm_finetune/llama3_2/llama3_2_1b_squad.yaml
automodel pretrain llm -c config.yaml
automodel kd llm -c config.yaml
automodel benchmark llm -c config.yaml

# VLM
automodel finetune vlm -c config.yaml

# Diffusion
automodel finetune diffusion -c config.yaml

# Retrieval
automodel finetune retrieval -c config.yaml

Override any config value from the CLI:

automodel finetune llm -c config.yaml --model.name_or_path meta-llama/Llama-3.2-1B

Common Pitfalls

Source: NVIDIA-NeMo/Automodel — distributed by TomeVault.