ml-eng-lab Runtime Flow

How contributors reach notebook execution while preserving task-local paths.
Runtime flow from entry shell to notebook-local artifacts Diagram showing local virtual environment, Docker, Codespaces, and JupyterHub entry paths feeding active notebook task folders with local data and run directories. Local venv make install-torch-stack jupyter lab Docker image repo mounted at workdir shared memory sized for GNNs Codespaces postCreateCommand setup browser VS Code or JupyterLab JupyterHub wrapper genai-vanilla submodule bind-mounted persistence Runtime contract notebook CWD task directory Notebook top-to-bottom run seeded where needed ./data task-local inputs ./runs ignored run registry Validation pytest, ruff, verifier Legend: local execution cloud dev shell orchestration notebook-local data

Entry Choices

  • Local venv and Docker use the checked-out tree.
  • Codespaces provisions the same dependency order.

Path Contract

  • Notebook code resolves `./data` and `./runs` from the task folder.
  • Make targets preserve that CWD.

Validation

  • Verifier checks active notebook structure and stale output paths.
  • Tier targets replay papermill execution.