A live demo of Stable Diffusion was presented by Pat Gelsinger in Intel Innovation in September and there has been a lot of interest from our users since then. With this release, we are publishing the model code and scripts on the HabanaAI Model-References GitHub repository.
We have upgrade versions of several libraries with SynapseAI 1.7.0, including DeepSpeed 0.7.0, PyTorch Lightning 1.7.7, TensorFlow 2.10.0 & 2.8.3, horovod 0.25.0, libfabric 1.16.1, EKS 1.23, and Open Shift 4.11.
SynapseAI installation processes have been simplified with a single automated installation script. We have also simplified PyTorch Habana Mixed Precision by reducing the number of configuration options and improving the default behavior. The Habana Media Loader now includes support for the Medical Segmentation Decathlon (BraTS) dataset.
Several reference models have been updated with instructions to train on Habana Gaudi2 for both PyTorch (ResNet-50 with LARS optimizer, SSD and transformer) and TensorFlow (SSD, transformer and UNET2D/3D). In addition, we have inference enabled on PyTorch BERT-large, BLOOM-7B and Stable Diffusion. You can find more information on Habana’s release notes page.