Installation + Setup

Installing ART

The ART client can be installed into projects designed to run on any machine that runs python.

pip install openpipe-art

Running the server locally

The ART server can be run locally on any machine with a GPU. To install the backend dependencies required for training and inference, you can install the backend extra:

pip install --extra-index-url https://download.pytorch.org/whl/cu128 \
  "openpipe-art[backend]"

from art import TrainableModel, gather_trajectory_groups
from art.local.backend import LocalBackend

backend = LocalBackend()

model = TrainableModel(
    name="agent-001",
    project="my-agentic-task",
    base_model="OpenPipe/Qwen3-14B-Instruct",
)

await model.register(backend)

... the rest of your code ...

CUDA 13 hosts use openpipe-art[backend-cu130] and the PyTorch cu130 index.

Running Megatron

On a supported CUDA 12 trainer image, one install command provides ART’s controller, Monarch runtime, and the locked Megatron runtime contract:

pip install --extra-index-url https://download.pytorch.org/whl/cu128 \
  "openpipe-art[megatron]"

CUDA 13 hosts use openpipe-art[megatron-cu130] with https://download.pytorch.org/whl/cu130. Megatron currently requires Python 3.12. The first trainer launch materializes the exact trainer environment in a content-addressed node-local cache. No ART checkout or setup.sh invocation is required.

The image remains responsible for the NVIDIA driver and CUDA toolkit. For cross-host training it must also provide the NCCL network transport, MOFED/RDMA devices, and the kernel capabilities described in the multi-node deployment guide. ART validates these before allocating the model.

Tinker users install openpipe-art[tinker]; serverless users need only openpipe-art. These profiles do not install Megatron or vLLM dependencies.

Using a managed autoscaling backend

Instead of managing the GPUs and training processes yourself, you can optionally send inference and training requests to the W&B Training cluster, which autoscales to match your job’s demand. To do so, install openpipe-art without any extras and use ServerlessBackend:

pip install openpipe-art

from art import TrainableModel, gather_trajectory_groups
from art.serverless.backend import ServerlessBackend

backend = ServerlessBackend()

model = TrainableModel(
    name="agent-001",
    project="my-agentic-task",
    base_model="OpenPipe/Qwen3-14B-Instruct",
)

await model.register(backend)

... the rest of your code ...

To learn more about the ART client and server, see the docs below.

ART Client

The client is responsible for interfacing between your code and the ART backend.

ART Backend

The backend is responsible for generating tokens and training your models.