Language

Docker¶

The checked-in Dockerfile is the Linux NVIDIA/CUDA container path. It installs UniLab runtime dependencies, the Motrix extra, and dev/test tooling. macOS Docker is not a primary path, and ROCm uses a separate image (see below) rather than this CUDA Dockerfile.

Build¶

docker build -t unilab:latest .

Quickly check the image entrypoint (it defaults to uv run train --help):

docker run --rm unilab:latest

Check CUDA visibility inside the container:

docker run --rm --gpus all unilab:latest \
  uv run python -c "import torch; print(torch.cuda.is_available())"

Run With A Mounted Checkout¶

docker run --rm --gpus all -it \
  -v "$(pwd):/workspace/UniLab" \
  -v unilab-venv:/workspace/UniLab/.venv \
  -w /workspace/UniLab \
  unilab:latest bash

Putting .venv in a named volume keeps the container’s virtual environment from overwriting the host repository directory, and leaves the host uv workflow clean when you switch back to it.

Inside the container, use the same commands as the host workflow:

uv run train --algo ppo --task go2_joystick_flat --sim mujoco
uv run train --algo sac --task g1_walk_flat --sim mujoco

Evaluation and playback work the same way inside the container; --load-run -1 picks the latest run:

uv run eval --algo ppo --task go2_joystick_flat --sim mujoco --load-run -1

Because the checkout is bind-mounted, training artifacts still write back to the mounted repository’s logs/ directory.

ROCm Containers¶

ROCm containers use AMD’s ROCm PyTorch image and the make sync-rocm workflow instead of the repository’s CUDA Dockerfile. Inside the container, run make sync-rocm first to activate the ROCm profile, then use uv run ... for training. The container typically needs the AMD device mounts /dev/kfd and /dev/dri, plus --group-add=video and --ipc=host.

A minimal example using AMD’s ROCm PyTorch image:

docker run --rm -it --network=host --ipc=host \
  --device=/dev/kfd --device=/dev/dri --group-add=video \
  -v "$(pwd):/workspace/UniLab" -w /workspace/UniLab \
  rocm/pytorch:latest bash

Use the ROCm PyTorch image tag that matches your installed ROCm version; see AMD’s ROCm PyTorch images for the available tags.