[mxj711]

feat(docker): add runtime image profiles and ROCm compose support

Summary

• Reorganized the Docker setup into runtime-specific image targets:
  • train-cuda
  • train-cuda-jax
  • train-rocm
  • deploy
  • unitree
• Kept train-cuda as the default Docker build target.
• Added a Docker publish workflow matrix for CUDA, CUDA+JAX, ROCm, and deployment images.
• Added separate Docker Compose files:
  • docker/docker-compose.yml for NVIDIA CUDA training
  • docker/docker-compose.rocm.yml for AMD ROCm training
• Added bilingual documentation for CUDA, ROCm, deployment, Unitree SDK handling, and Git LFS requirements.
• Removed the -dev image target from the public Docker setup.

Closes #154

Changes

Docker image profiles

Target	Image suffix	Purpose
train-cuda	-cuda	Default CUDA training stack with SKRL Torch, RSLRL, FastSAC, MotrixSim, and TensorBoard
train-cuda-jax	-cuda-jax	CUDA training stack with the additional SKRL JAX runtime
train-rocm	-rocm	ROCm PyTorch training stack
deploy	-deploy	ONNX deployment runtime with MotrixSim and MuJoCo backends
unitree	-unitree	Unitree SDK2/DDS deployment runtime

The CUDA and ROCm images remain separate because their locked PyTorch profiles are mutually exclusive.

The default CUDA image also receives:

• <version>-cuda
• <major.minor>-cuda
• <version>
• <major.minor>
• latest

Docker Compose

The CUDA Compose file uses an NVIDIA GPU reservation.

The ROCm Compose file uses AMD device passthrough:

devices:
  - ${MOTRIX_ROCM_COMPUTE_DEVICE:-/dev/kfd}
  - ${MOTRIX_ROCM_RENDER_DEVICE:-/dev/dri}
group_add:
  - video
  - render
security_opt:
  - seccomp=unconfined
shm_size: 1g

A specific AMD render node can be selected with:

MOTRIX_ROCM_RENDER_DEVICE=/dev/dri/renderD128

Dockerfile behavior

• Uses Python 3.10.
• Runs containers as a non-root motrix user.
• Installs the application under /app.
• Uses /workspace for user workloads and run output.
• Pre-creates /workspace/runs with the correct ownership.
• Installs locked third-party dependencies before copying workspace source.
• Excludes local caches, virtual environments, tests, build output, and run output from runtime images.
• Allows Unitree SDK2 native libraries to be injected through UNITREE_SDK_IMAGE.
• Does not provide or publish a -dev image.

Validation

Passed:

uv lock --check

docker compose -f docker/docker-compose.yml config
docker compose -f docker/docker-compose.rocm.yml config

dprint check \
  .dockerignore \
  docker/Dockerfile \
  docker/docker-compose.yml \
  docker/docker-compose.rocm.yml \
  docker/README.md \
  .github/workflows/docker-build.yml \
  .github/workflows/README.md \
  docs/source/zh_CN/user_guide/getting_started/container_deployment.md \
  docs/source/en/user_guide/getting_started/container_deployment.md

git diff --check

Also validated:

• The Docker workflow matrix contains only train-cuda, train-cuda-jax, train-rocm, and deploy.
• The Dockerfile does not contain a dev target.
• The default final stage is based on train-cuda.
• ROCm Compose correctly resolves /dev/kfd, /dev/dri, group access, seccomp settings, and shared memory.
• Dependency resolution was dry-run for the CUDA, CUDA+JAX, ROCm, deployment, Unitree, and full-workspace profiles.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.