Docker image tarball (
docker save | zstd -6, 78.6GB uncompressed) with a fully built
NVIDIA-NeMo/RL @
5f89b3ae training environment:
Contains only public/redistributable content (NVIDIA CUDA redistributables per the CUDA
EULA, Apache-2.0 NeMo-RL, PyPI wheels). No model weights, no tokens, no private code.
1huggingface-cli download JWei05/nemorl-gemma4-cu130-env rl-distill-nemorl-env-cu130-20260729.docker.tar.zst --local-dir .
2zstd -d -c rl-distill-nemorl-env-cu130-20260729.docker.tar.zst | docker load
3# then materialize your training repo around the baked checkout, e.g.:
4# docker run ... rl-distill-nemorl-env:cu130-20260729 bash -c # 'cd /workspace/rl-distill && git init && git remote add origin <your-fork> && # git fetch --depth 1 origin main && git checkout -f FETCH_HEAD && <run>'
