We are happy to announce the release of SynapseAI version 1.1.0. In addition to many improvements and bug fixes, the new version includes the following new capabilities:
We introduced support for PyTorch Lightning. You can find an example of Unet2D with PyTorch Lightning in our model references repository on Habana GitHub.
We now support PyTorch 1.9.1, TensorFlow 2.5.1, and TensorFlow 2.6.0. We have upgraded the Habana Horovod package to version 0.22.1. Also, we are introducing support for two new operating systems, RHEL 8.3 and Centos 8.3. With this release, we also enable installing PyTorch packages on a bare metal machine or a virtual machine.
We have enabled many new reference models for PyTorch: Unet2D, Transformer, ResNet152, DistilBERT, RoBERTa Base and Large. We improved the performance of ResNet50, ResNet152, ResNext101 by enabling the Habana dataloader. BERT-L for PyTorch can now be trained on up to 32 Gaudi devices. You can find all of them (and many other models) in our model reference here.
TensorFlow Distributed with HPUStrategy now uses HCCL API by default. We have enabled HCCL over TCP in PyTorch. With these changes, host-NIC based distributed training with TensorFlow and PyTorch can be enabled with TCP flows.
You can find more information on our release notes page.