The Habana team is happy to announce the release of SynapseAI® version 1.6.0. In this release, we introduce preliminary inference capabilities on Gaudi. For further details, refer to the Inference on Gaudi guide. We have published a reference model, V-Diffusion to demonstrate the new inference capabilities.
New features with DeepSpeed are enabled, including support for ZeRO-2 and activation checkpointing.
In this release, we have made several version updates. We now support RHEL8 8.6 (previously 8.3), OpenShift 4.9, PyTorch 1.12.0 (previously 1.11.0), and Horovod 0.24.3 (previously 0.23.0.) With this release, we are dropping support for Kubernetes 1.19. Also, the Habana EKA AMI does not support docker as runtime; only containerD is supported.
New reference models include PyTorch implementations for YOLOX, Wav2Vec 2.0 pre-training, and v-diffusion inference. We have updated the DeepSpeed BERT reference model with LANS optimizer. For users interested in training models with PyTorch’s ZeRO optimizer, we have published an example of pre-training a 1.2B param BERT model. In addition, we validated SSD, Unet2D, Unet3D, and Transformer reference models on Gaudi2.
Conda is now pre-installed in all SynapseAI docker images. The Conda base environment is inactive by default, and needs to be activated manually. The Habana Media Loader now includes support for the COCO dataset. We have enabled PyTorch TensorBoard profiling. We are publishing pretrained models for TensorFlow and PyTorch ResNet-50 and BERT-Large; these can be found via the Habana Catalog page.
You can find more information on Habana’s release notes page.