One of the main challenges in training Large Language Models (LLMs) is that they are often too large to fit on a single node or even if they fit, the training may be too slow. To address this issue, their training can be parallelized across multiple Gaudi accelerators (HPUs).