Release History

Changelog

0.0.0 (2020-05-07)

  • Package initialized.

0.0.1 (2020-06-09)

  • Initial release.

0.2.0 (2020-08-18)

  • Dramatically improved documentation.

  • Added channel-distributed convolutional layer.

  • Abstracted convolutional layer interface. It now auto-selects implementation.

  • Added pre-forward hooks so that communication buffers are only allocated when the shape of the input tensor changes.

  • Improved general consistency of layer structure and member names.

0.3.0 (2020-12-01)

  • Corrected use of dtype in internal buffers.

  • Cleaned up partition API.

  • Fixed a bug where MPI resources were not released.

  • Removed assumption that transpose requires load-balanced input.

  • Added smarter buffer re-use.

  • Added distributed batch normalization layer.

  • Added distributed upsampling interpolation layer.

0.4.0 (2021-09-01)

  • Reorganized code to follow standard PyTorch naming

  • Fixed bugs related to invalid convolution arguments

  • Improved convolution and pooling implementations to reduce constraints on inputs

  • Added all-sum-reduce

  • Added distributed loss functions

  • Added initial GPU support for MPI backend (experimental)

  • Moved from Travis-CI to GitHub Actions

  • Multiple documentation fixes