Lightning-AI/pytorch-lightning

Choose upstream unless you specifically need this fork’s README/branding wording or want a very small downstream starting point. For production use, the 197-commit lag makes upstream the safer default.

Prefer this fork only if you need its older Lightning App/cloud/data behavior and can own the maintenance burden. For most adopters, upstream is the better default because it is much more current, actively maintained, and lower risk.

Choose this fork only if you need its specific customizations or removals; otherwise upstream is the safer default because it is far more current and retains more Lightning features.

Prefer upstream unless you specifically need a frozen snapshot; this fork adds no visible functionality and is materially behind on maintenance and fixes.

Choose this fork only if you specifically need the older Lightning internals or its experimental trainer/spawn refactors. For most adopters, upstream is the better choice because it is active, current, and far less risky to maintain.

Choose this fork only if you need historical Lightning behavior or a custom research baseline. If you want current training infrastructure, docs, Fabric, and ongoing compatibility work, upstream is the better choice.

Prefer upstream unless you specifically need this exact historical snapshot; this fork adds no visible features and is materially behind on maintenance and fixes.

Prefer this fork only if you need 2021-era Lightning behavior or specific legacy Lite/trainer semantics. For new work, upstream is the better choice because it is far more current, maintained, and compatible with the modern Lightning stack.

Choose this fork only if you specifically need the older Lightning behavior and are prepared to maintain it yourself. For new projects or teams wanting ongoing compatibility, upstream is the better default.