Environments
Environment notes connect software stacks, dependency boundaries, and public-safe run records. Reusable definitions live under AI Systems; infra pages focus on scheduler, GPU runtime, storage mount, or server-operation boundaries.
Environment는 run identity의 일부입니다.
Compiler, CUDA stack, Python package set, container image, runtime mount가 바뀌면 model code가 그대로여도 결과가 달라질 수 있습니다.
Scope
- Software stacks used for training, inference, preprocessing, and analysis.
- Environment modules, containers, lockfiles, image digests, and runtime wrappers.
- Driver/runtime boundaries such as host driver versus container user space.
- Public-safe environment records that explain reproducibility without exposing private machines.
Notes
Checks
- Can the environment be reconstructed from public-safe metadata?
- Are CUDA, compiler, Python, package, and container versions recorded at the right granularity?
- Is the host-specific part separated from portable user-space dependencies?
- Are private registry names, mount paths, credentials, and hostnames excluded?
Routing
| Question | Go To |
|---|---|
| What are modules and containers? | Environment modules and containers |
| How should a run environment be captured? | Environment management, Reproducibility |
| Is the issue a scheduler job script or cluster launch? | HPC, Slurm job script |
| Is the issue an operational failure? | Server operations, Incident response |