2.6.4
Breaking Changes📦 pytorch-lightningView on GitHub →
⚠ 1 breaking🐛 5 fixes🔧 8 symbols
Summary
This patch release removes deprecated Neptune logger support due to service sunsetting and includes several bug fixes related to FSDP mixed precision initialization, validation interval checks, and profiler precision.
⚠️ Breaking Changes
- Support for Neptune logger has been removed. Users relying on this logger must migrate to an alternative solution, as the public service for Neptune is being sunsetted.
Migration Steps
- If you were using the Neptune logger, you must switch to a different logging integration as Neptune support has been removed.
🐛 Bug Fixes
- Fixed `val_check_interval` raising `ValueError` when `limit_val_batches=0` and the interval exceeds training batches.
- Fixed FSDP mixed precision settings (`bf16-mixed`, `16-mixed`) from incorrectly initializing model parameters in half precision instead of fp32.
- Fixed `device_mesh` type hint in `FSDPStrategy` to correctly accept a 2-element tuple when configured via the CLI.
- Fixed model size display formatting in `RichModelSummary`.
- Fixed floating-point precision in `SimpleProfiler` duration aggregation by utilizing `math.fsum`.