v23.06.00 (v1.0)

What’s Changed

New Features

• Added support for row-slicing as a parallel strategy
• Added support for data-parallel as a parallel strategy
• Allow mix-matching data-parallel, table-parallel, row-slicing, and column-slicing. Refer to User Guide for more details.
• Added IntegerLookup layer that supports on-the-fly vocabulary building, on both CPU and GPU

Breaking Changes

• Added NVIDIA cuCollections as submodule for GPU hash map support
• Now support TensorFlow 2.12. Note that this change breaks the build with TF 2.09 and earlier.

Improvements

• Improved package import

Bug Fixes

• fixes input offset overflow due to automatic table concatenating
• fixes potential graph mismatching problems in broadcast

Full Changelog: v23.03.00...v23.06.00

v23.03.00

What’s Changed

New Features

• NVIDIA Hopper™ architecture families support (compute capability 9.0).
• Added support for Keras Model fit api.
• Added support for Horovod callbacks in case of hybrid data/model parallel.

Breaking Changes

• Now support TensorFlow 2.12. Note that this change breaks build with TF 2.09 and earlier.
• Now require horovod version 0.27 or later.

Improvements

• Improved unit tests

Bug Fixes

• Use tf.shape for graph mode support by @edknv in #6

New Contributors

• @edknv made their first contribution in #6

Full Changelog: v0.3...v23.03.00

v0.3

What’s Changed

New Features

• CUDA 12 support
• Automatic concatenation of multiple embedding tables for greatly improved speed
• Support model parallel with user-defined custom keras layer through DistributedEmbedding wrapper

Improvements

• Support cases where number of workers is greater than number of tables.
• For corner cases where diffrerent slices of a table are placed onto same worker, they will be merged into single slice now.

Breaking Changes

• move submodule from CUB to NVIDIA Thrust for better compatibilities

Bug Fixes

• Better error handling in set_weight() when weights are not initialized
• Better error handling when global batchsize is not divisible by number of workers

Full Changelog: v0.2...v0.3

v0.2

What’s Changed

Breaking Changes

• added new dependency NVIDIA CUB as submodule

New Features

• SparseTensor is supported as embedding input, in addition to Dense and Ragged Tensor.
• Add support and example for keras model.fit() api through custom train_step() function

Improvements

• Improved embedding lookup speed when input is multi-hot with combiner.
• Improved embedding lookup speed when input is one-hot, regardless of its combiner and format(Tensor, SparseTensor or RaggedTensor)
• Add support for data parallel input, cpu embedding and TF native embedding api as options in benchmark

Bug Fixes

• fix build with tensorflow 2.10+
• fix a bug where batch dimension could be None at early stage in graph mode

Full Changelog: v0.1...v0.2

v0.1

Initial release