Skip to content

env_creation_compute_node.md

Latest commit

 

History

History
26 lines (21 loc) · 1.46 KB

env_creation_compute_node.md

File metadata and controls

26 lines (21 loc) · 1.46 KB
Raw

Compute-Node Environment Checklist

The complete offline installation procedure is in env_creation.md. Before launching Libra, validate the environment from an actual compute-node allocation rather than only on the login node.

srun --nodes=1 --ntasks=1 bash -lc '
  source "$HOME/rl_framework_env/bin/activate"
  set -e
  echo "Python: $(which python3)"
  python3 --version
  python3 -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
  python3 -c "import transformers, vllm; print(transformers.__version__, vllm.__version__)"
'

Check the following before submitting a long job:

  • The base Python interpreter and standard library are visible on every node.
  • The virtual environment is stored on a shared filesystem or reproduced consistently per node.
  • The CUDA runtime and driver are compatible with the installed PyTorch and vLLM builds.
  • python3, not an unrelated system python, is used consistently by the launchers.
  • The activation path in the Slurm script matches the environment where dependencies were installed.
  • Model, dataset, checkpoint, and log paths are visible from every participating node.
  • PYTHONPATH contains the parent directory of RL_Framework.

If a venv created on the login node reports No module named 'encodings' on a compute node, its base interpreter or standard library path is unavailable there. Recreate the venv from a Python installation visible to the compute nodes.