NVIDIA/Atex: A TensorFlow Extension

This repository holds NVIDIA-maintained utilities to improve GPU performance and usability for Tensorflow training and inference. The intent of Atex is to make up-to-date utilities available to users as quickly as possible. As such, these utilities are experimental. Some we will upstream and support permanently in projects such as TensorFlow or Keras. Others will eventually be discontinued.

Contents

1. Fused Layer/Instance Normalization

nv_norms.LayerNormalization and nv_norms.InstanceNormalization are fused implementations designed to replace the tf.keras.layers.LayerNormalization and tfa.layers.InstanceNormalization. The Keras and Addons implementations compose many ops, such as Mul, Add, etc., to compute the stats and then scale/offset the results. In contrast, the NVIDIA fused implementation provided in nv_norms compute the norms in a single operation, eliminating many expensive round-trips to device memory and significantly improving performance. More in nv_norms.

2. Structured Sparsity

This is a project for Tensorflow on supporting fine-grained structured sparsity for the NVIDIA Ampere GPU architecture. We only need users to add a couple lines to their python script and then the pretrained model can be automatically pruned to benefit from the sparse Tensor Cores (available from Ampere GPUs) to achieve faster inference speed after deployment. More in structured_sparsity.

Installation from source

To build this package from source, run the following command in the root directory of this package.

pip install .

Contribution guidelines

Please review the Contribution Guidelines.

GitHub issues will be used for tracking requests and bugs.

License

BSD License 2.0