Skip to navigation

Install

View as MarkdownOpen in Claude

Prerequisites

Before installing NVIDIA AITune, make sure your system meets these requirements:

  • Operating System: Linux (Ubuntu 22.04+ recommended)
  • Python: Version 3.11 or newer
  • PyTorch: Version 2.8 or newer
  • TensorRT: Version 10.3 or higher (for TensorRT backend)
  • NVIDIA GPU: Required for GPU-accelerated tuning

You can install NVIDIA AITune from PyPI.

pip install --extra-index-url https://pypi.nvidia.com aitune

Pinning to a specific PyTorch version

If you need to pin AITune’s dependencies (torch, torch-tensorrt, torchao) to a specific PyTorch minor version, use one of the version extras:

ExtraPyTorch version
torch282.8.x
torch292.9.x
torch2102.10.x
torch2112.11.x
torch2122.12.x
pip install --extra-index-url https://pypi.nvidia.com "aitune[torch211]"

Selecting a CUDA version

To target a specific CUDA version, add the matching PyTorch index:

pip install --extra-index-url https://pypi.nvidia.com --extra-index-url https://download.pytorch.org/whl/cu130 "aitune[torch211]"

ONNX Runtime CUDA version

Select the CUDA 13 extra when installing from a source checkout with UV:

uv pip install --torch-backend=cu130 ".[torch210,onnxruntime-gpu-cuda13]"

Use the onnxruntime-gpu-cuda12 extra for the CUDA 12 PyPI build:

uv pip install --no-sources --torch-backend=cu128 ".[torch29,onnxruntime-gpu-cuda12]"

pip uses PyPI and therefore installs CUDA 12. Use the same extra explicitly:

pip install --extra-index-url https://pypi.nvidia.com "aitune[torch29,onnxruntime-gpu-cuda12]"

CUDA 13 with pip needs an explicit ORT reinstall from the CUDA 13 feed:

pip install onnxruntime-gpu==1.26.0 --no-deps --force-reinstall --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-13/pypi/simple/

If ORT is already installed, use UV’s targeted reinstall option with the CUDA 13 extra:

uv pip install --torch-backend=cu130 --reinstall-package onnxruntime-gpu ".[torch210,onnxruntime-gpu-cuda13]"

NGC container install

You can use NGC PyTorch containers, which include the required NVIDIA dependencies:

NGC PyTorch containerTorch version
nvcr.io/nvidia/pytorch:25.08-py32.8a0
nvcr.io/nvidia/pytorch:25.10-py32.9a0
nvcr.io/nvidia/pytorch:26.01-py32.10a0
nvcr.io/nvidia/pytorch:26.03-py32.11a0
nvcr.io/nvidia/pytorch:26.05-py32.12a0

Inside an NGC container, use pip and the package name:

pip install --extra-index-url https://pypi.nvidia.com aitune

Installing from Source

# Clone the repository
git clone https://github.com/ai-dynamo/aitune
cd aitune
pip install --extra-index-url https://pypi.nvidia.com .

or use editable mode for development:

pip install --extra-index-url https://pypi.nvidia.com -e .