The Habana(R) Labs team is happy to announce the release of SynapseAI® version 1.3.0.
In this release, we have made several version updates. We now support PyTorch 1.10.1 (previously 1.10.0), PyTorch Lightning 1.5.8 (previously 1.5.0), TensorFlow 2.7.1 (previously 2.7.0). In addition, we are removing support for TensorFlow 2.6.2 and have added support for the recently released TensorFlow 2.8.0. Habana’s horovod is now based on v0.23.0 of horovod (previously v0.22.1.) We have also updated to Python version 3.8 for all supported OSes (previously python 3.7 for U18 and AL2) habana-tensorflow and habana-horovod Python packages are now available on the Python Package Index (PyPI). We plan to make the Python packages for PyTorch available on PyPI in a future release.
Additionally, we’re introducing support for OpenShift which provides an efficient and manageable way to orchestrate deep learning workloads at scale.
This release also includes HCCL demo which is published on GitHub and demonstrates HCCL usage and supports communication via Gaudi based NIC or host NIC (achieved using TCP and OFI.) The following list of collectives are currently supported: all_reduce, all_gather, eeduce_scatter, broadcast, and reduce.
Several new reference models have been enabled with the 1.3.0 release. This includes PyTorch implementations for SWIN Transformer, Electra Fine Tuning, Vision Transformer and MobileNetv2. TensorFlow models include Electra, Wide-and-Deep and DistilBert.
You can find more information on Habana’s release notes page.