We have added support for Ubuntu 22.04 on first-gen Gaudi. Support for Ubuntu 18.04 is deprecated and will not be available from the next release.
We enabled DeepSpeed support for Model Pipeline Parallelism, Model Tensor Parallelism, and BF16Optimizer. See DeepSpeed Validated Configurations. In addition, DeepSpeed inference is now supported on Gaudi. See Inference Using DeepSpeed.
For Amazon EC2 DL1 Instance users, we have added EFA peer direct support, which enables significant performance improvements for distributed training across multiple DL1 instances.
We are introducing Gaudi2 Media Pipe APIs in SynapseAI 1.8.0. Users can now prepare batches of processed and augmented images and labels to be fed into training or inference. See Media Pipeline for more details.
Several reference models have been updated with instructions to train and infer on Habana Gaudi2 and Gaudi first-gen; this includes Wav2Vec training and inference, BLOOM 13B training, BLOOM 176B inference, ResNet-50 inference, ResNext101 inference, UNet2D inference, and Stable Diffusion training and inference.
You can find more information on SynapseAI 1.8.0 on Habana’s release notes page.