rl.def

Bootstrap: docker
# Ubuntu 24.04 provides glibc 2.39. This is new enough for the driver-side
# EGL/GL libraries injected by the B200 cluster (which require GLIBC_2.38) and
# remains backward-compatible with driver libraries from the V100/A100 hosts.
From: nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04

%files
    pyproject.toml /opt/vbrl-src/pyproject.toml
    uv.lock /opt/vbrl-src/uv.lock
    .python-version /opt/vbrl-src/.python-version
    README.md /opt/vbrl-src/README.md
    src/vbrl /opt/vbrl-src/src/vbrl

%environment
    export PYTHONDONTWRITEBYTECODE=1
    export PYTHONUNBUFFERED=1
    export PYTHONNOUSERSITE=1
    export PATH=/opt/venv/bin:$PATH
    export MUJOCO_GL=egl
    export MPLBACKEND=Agg
    # $HOME is read-only during the build and shared or quota-limited on compute
    # nodes; without this matplotlib warns and re-scans its font cache per job.
    export MPLCONFIGDIR=/tmp/matplotlib
    export PYOPENGL_PLATFORM=egl
    export HF_HUB_OFFLINE=1
    export TRANSFORMERS_OFFLINE=1

%post
    set -e
    # uv creates and populates /opt/venv, so pip and venv are not needed; git is,
    # for the pinned R3M source dependency.
    apt-get update && apt-get install -y --no-install-recommends \
        python3 python3-dev \
        git curl ca-certificates ffmpeg \
        libegl1 libgl1 libgles2 libglvnd0 libglx0 libopengl0 \
        libosmesa6 libglfw3 \
        && rm -rf /var/lib/apt/lists/*

    # pyproject.toml plus uv.lock is the single environment specification. uv
    # installs the exact transitive closure the lockfile pins, with hashes, and a
    # development host runs `uv sync` against the same file -- so the image and
    # the workstation cannot drift. Bump a version by editing pyproject.toml,
    # running `uv lock`, and rebuilding this image.
    curl -LsSf https://astral.sh/uv/0.12.5/install.sh \
        | env UV_INSTALL_DIR=/usr/local/bin INSTALLER_NO_MODIFY_PATH=1 sh

    # --frozen refuses to re-resolve, so the build fails loudly if uv.lock is
    # stale rather than quietly installing something else. --no-dev keeps ruff,
    # mypy, and build out of the image. --no-editable installs vbrl as a regular
    # wheel, which is what supplies the mjlab.tasks entry point that fresh
    # TorchrunX workers discover; a bound checkout still wins via PYTHONPATH.
    cd /opt/vbrl-src
    UV_PROJECT_ENVIRONMENT=/opt/venv uv sync --frozen --no-dev --no-editable
    export PATH="/opt/venv/bin:${PATH}"

    # Bake the backbones in so a cluster job never touches the network. Each
    # upstream caches in its own layout under one root; a development host runs
    # the same command. %environment is not applied during %post, but set these
    # explicitly so the build cannot be broken by a future export of them.
    HF_HUB_OFFLINE=0 TRANSFORMERS_OFFLINE=0 vbrl-fetch-backbones /opt/vbrl-models
    vbrl-fetch-backbones /opt/vbrl-models --verify
    chmod -R a+rX /opt/vbrl-models

%test
    # Validate the vendor-neutral graphics ABI without requiring a GPU during
    # the build. Apptainer --nv supplies only the host-specific NVIDIA side.
    python - <<'PY'
import ctypes

for library in ("libEGL.so.1", "libOpenGL.so.0", "libGL.so.1"):
    ctypes.CDLL(library)

from OpenGL import EGL
import mjlab
import mujoco

assert EGL.eglQueryString is not None
# MJLab registers its own Mjlab-* tasks, so ownership comes from the
# registration delta vbrl.tasks records, never from a name prefix.
from vbrl.tasks import vbrl_task_ids

# One known ID per task family rather than a count, so registering a new
# architecture does not require editing this file. tests/test_task_registration
# pins the exact set.
task_ids = set(vbrl_task_ids())
expected = {
    "Mjlab-LiftCube-CollisionCam-DinoV2ViTS14-LocalGrid7-Trossen",
    "Mjlab-LiftCube-RealTexture-DinoV2ViTS14-LocalGrid7-Trossen",
    "Mjlab-PushT-State-TrossenIdentified",
}
missing = expected - task_ids
assert not missing, (
    "Installed VBRL task entry point did not populate the task-local "
    f"registrations; missing: {sorted(missing)}"
)
print(f"MuJoCo {mujoco.__version__}: EGL/OpenGL available, {len(task_ids)} tasks")
PY

%runscript
    exec "$@"