-
Notifications
You must be signed in to change notification settings - Fork 3.6k
Description
Outline & Motivation
Firstly, thanks to Lightning team for your continued hard work and invaluable contributions to the open-source ecosystem!
As a follow-up to this issue, I’ve noticed that recent Lightning minor releases are no longer synchronized with PyTorch’s minor releases as they have been in the past:
PyTorch Version | Release Date | Release Artifact | Lightning Version | Release Date | Release Artifact | Lag (weeks: Lightning - PyTorch) |
---|---|---|---|---|---|---|
2.0 | Mar 15, 2023 | PyTorch 2.0.0 | 2.0.0 | Mar 15, 2023 | Lightning 2.0.0 | 0 |
2.1 | Oct 4, 2023 | PyTorch 2.1.0 | 2.1.0 | Oct 12, 2023 | Lightning 2.1.0 | 1 |
2.2 | Jan 30, 2024 | PyTorch 2.2.0 | 2.2.0 | Feb 7, 2024 | Lightning 2.2.0 | 1 |
2.3 | Apr 24, 2024 | PyTorch 2.3.0 | 2.3.0 | Jun 13, 2024 | Lightning 2.3.0 | 7 (50 days) |
2.4 | Jul 24, 2024 | PyTorch 2.4.0 | 2.4.0 | Aug 7, 2024 | Lightning 2.4.0 | 2 (14 days) |
2.5 | Oct 17, 2024 | PyTorch 2.5.0 | 2.5.0 | Dec 20, 2024 | Lightning 2.5.0 | 9 (64 days) |
2.6 | Jan 29, 2025 | PyTorch 2.6.0 | — | — | — | 28 (197 days) |
2.7 | Apr 23, 2025 | PyTorch 2.7.0 | — | — | — | 16 (113 days) |
2.8 | Aug 6, 2025 | PyTorch 2.8.0 | — | — | — | 1 (8 days) |
Pitch
I think both users and maintainers of downstream packages depending on Lightning (I'm the primary maintainer of a downstream Lightning package Fine-Tuning Scheduler) would benefit if this apparent versioning policy/cadence change were made explicit in the versioning documentation.
Can we update the versioning policy to clarify the current minor release policy/cadence for Lightning? E.g., 'minor releases are typically driven by key upstream PyTorch deprecations etc...' This would help the community better plan dependencies and upgrades.
Additional context
As a concrete example, Fine-Tuning Scheduler has a number of deprecations and enhancements aligned with various PyTorch minor versions that have been awaiting a Lightning minor release. If the release cadence of the last year is expected to be the new approach moving forward, it would be good to know so FTS can proceed with changing its release approach from being Lightning release synchronized to a PyTorch release alignment instead.
Thanks again!