Skip to content

install_O2.md

Latest commit

 

History

History
37 lines (33 loc) · 2.14 KB

install_O2.md

File metadata and controls

37 lines (33 loc) · 2.14 KB
Raw

Install on O2 at Harvard Medical School

O2 is a high-performance computing platform based on Linux, located at Harvard Medical School. The platform is managed by the Research Computing Group, part of HMS IT. See the O2 documentation website for platform-wide details (accounts, SLURM scheduling, storage quotas).

O2 does not support Docker, so this repository must be installed with the non-Docker uv workflow described in install_local.md, adapted to O2's module system and job scheduler:

  1. Log in to O2 and start an interactive session with a GPU if you need one for training (see the O2 docs for current srun/sbatch GPU-partition syntax).
  2. Load a Python 3.12-compatible toolchain via module load, then install uv per the uv installation guide.
  3. Clone the repository and sync dependencies: 
    git clone git@github.com:ccb-hms/computervision.git
cd computervision
uv sync --frozen
  4. Install Detectron2 the same way as in install_local.md, from source: 
    uv run python -m pip install --no-build-isolation "git+https://github.com/facebookresearch/detectron2.git"
    This requires a working CUDA toolchain in your loaded modules; consult O2's GPU documentation if the build fails.
  5. Copy env to .env and set DATA_DIR to a path on O2 storage you have write access to and sufficient quota for (dental X-ray datasets and model checkpoints are large) — see environment-variables.md.
  6. Run Jupyter Lab through O2's Open OnDemand portal, or start it manually in your interactive session with uv run jupyter lab and tunnel the port over SSH, per O2's Jupyter documentation.

Because module names, GPU partitions, and portal URLs change over time, treat the O2 platform docs linked above as authoritative for anything not specific to this repository.