Installation

AReno has native CUDA and MLX installation paths. Linux and WSL2 use the CUDA installer; Apple Silicon uses a normal pip source install with platform-marked MLX dependencies.

The Backends guide compares both native runtimes. Its platform pages describe the complete, parallel train/serve setup:

• Linux and CUDA for Linux, WSL2, NVIDIA CUDA, distributed topology, and CUDA checkpoints.

• Apple Silicon and MLX for Apple Silicon, unified memory, MLX checkpoints, and the single-process runtime.

Install CUDA AReno

Clone the repository and run one command:

git clone https://github.com/inclusionAI/AReno.git
cd AReno
bash scripts/install.sh

Before changing Python, the installer checks required system tools, rejects WSL1, and verifies that nvidia-smi can see a GPU. It then uses an active virtualenv or conda environment when available, reuses the repository’s .venv when it is ready, or creates .venv automatically. If an IDE does not expose environment activation metadata, the installer detects and reuses a Python interpreter that already provides PyTorch instead of creating an empty .venv. Finally, it checks for CUDA-enabled PyTorch 2.6 or newer, detects CUDA build support, installs AReno’s remaining dependencies, selects the attention setup, builds the CUDA extension, and runs areno check. The installer never installs or upgrades PyTorch because the correct build depends on the machine’s CUDA platform. If PyTorch is missing or incompatible, it stops with guidance for the selected Python environment. Other packages that already satisfy AReno’s requirements are reused; only missing or incompatible packages are installed or updated.

Successful installation ends with AReno is ready and the exact command to start using AReno. If installation stops, the same script reports the failed stage, explains the immediate reason, prints targeted suggestions, and preserves complete command output in the user state directory, usually ~/.local/state/areno/install.log.

To preview the plan without changing the environment:

bash scripts/install.sh --dry-run

Continue with Linux and CUDA for CUDA training, serving, memory controls, checkpoints, model support, and SDK configuration.

Compatibility matrix

Environment	Status	Notes
Linux x86_64 + NVIDIA GPU	Supported	Primary training/serving target. Use CUDA-enabled PyTorch >= 2.6 and build areno_accel.
Linux aarch64 / Grace-Blackwell	Supported	Start from a compatible CUDA-enabled aarch64 PyTorch environment, such as a current NVIDIA NGC PyTorch development container. The installer validates it and builds AReno against it.
Windows WSL2 + NVIDIA GPU	Supported	Follow the Linux install path inside WSL2. Native Windows is not supported.
macOS Apple Silicon	Supported	Use native arm64 Python and the MLX pip installation below. The CUDA installer does not apply.
CPU-only environments	Not supported	Training and serving require either NVIDIA CUDA or Apple Silicon MLX.

Install on Apple Silicon

The repository installer currently prepares the CUDA toolchain, so do not run scripts/install.sh on macOS. Use a native arm64 virtual environment:

git clone https://github.com/inclusionAI/AReno.git
cd AReno
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .

The project metadata installs MLX dependencies only on Apple Silicon and CUDA dependencies only on Linux. Verify that Python is not running through Rosetta:

python -c "import platform; from areno.api import DefaultBackend; print(platform.machine(), DefaultBackend)"

Expected output contains arm64 and BackendType.MLX. Continue with Apple Silicon and MLX for training, serving, checkpoints, memory controls, and multimodal support.

Docker

Docker is the setup escape hatch when you want to verify AReno before debugging local Python, PyTorch, or CUDA build state. Build the CUDA runtime image from the repository root, then run the same readiness check used by local installs:

docker build -t areno .
docker run --gpus all --rm -it areno areno check

Use --build-arg PIP_INDEX_URL=... if your environment requires a package mirror.

If you need local project files, model files, or a Hugging Face cache inside the container, mount them explicitly:

docker run --gpus all --rm -it \
  -v $PWD:/workspace \
  -v $HOME/.cache/huggingface:/root/.cache/huggingface \
  areno \
  areno check

Host checklist:

nvidia-smi
docker run --gpus all --rm nvidia/cuda:12.4.1-base-ubuntu22.04 nvidia-smi
docker run --gpus all --rm areno areno check

Docker gives you a known-good Python/PyTorch/CUDA user-space environment. It does not fix host-side requirements: the host still needs a working NVIDIA driver, NVIDIA Container Toolkit support for --gpus all, and a driver new enough for the container CUDA runtime. Model downloads, Hugging Face tokens, cache paths, network access, disk space, and multi-node or custom networking remain user environment concerns and are outside the first Docker setup path.

Post-install checklist

The installer runs the readiness check automatically. You can rerun it at any time:

areno check

For setup reports, also collect a machine-readable environment bundle:

areno env --json

On CUDA, areno check reports common build-time and runtime setup problems such as missing or CPU-only PyTorch, unsupported PyTorch versions, missing CUDA_HOME or nvcc, and missing build dependencies. On Apple Silicon, verify the backend and MLX device from the active environment before loading a checkpoint.