The Habana® Labs team is happy to announce the release of SynapseAI® version 1.2.0. In addition to many improvements and bug fixes, the new version includes the following new capabilities.
SynapseAI software now supports PyTorch 1.10.0, PyTorch Lightning 1.5.0, TensorFlow 2.7.0, and TensorFlow 2.6.2.
We’ve added Containerd support to habanalabs-container-runtime, support for debugging and profiling through profiler configuration, and support for PyTorch Custom Ops is now available.
SynapseAI and Habana Gaudi® now enables many new reference models for training, including for PyTorch: SSD ResNet34, Unet3D, MobileNet_V2, GoogleNet, Albert Large/XXLarge, and BART. For the TensorFlow framework, we have enabled: VisionTransformer and RetinaNet (model evaluation).
As part of our commitment to scale-up/scale-out, we added support for distributed scale-out using HCCL (Habana Collective Communications Library) based Host-NIC over OFI libFabric. Now, the TensorFlow CycleGAN reference model supports up to 8 Gaudi processors, while ResNext101 and BERT-Large PreTraining support up to 64 Gaudis. On PyTorch, we extended the support of ResNet50 to up to 32 Gaudis.
You can find more information on Habana’s release notes page.
