X-ScaleAI provides an optimized and integrated software stack for high-performance distributed pre-training, fine-tuning, and inference. We support any model defined in PyTorch or HuggingFace, including Large Language Models such as Llama-3, OLMo, Pythia, and BERT and Vision Models such as ResNet, U-Net, ViT and Stable Diffusion. You can also define your own model and your own dataset. Leave the headache of scaling up your AI workloads to us, and focus on your domain-specific strengths.

The end-to-end optimized software stack in X-ScaleAI provides out-of-the-box optimal performance for distributed AI workloads. It has a proprietary MVAPICH MPI implementation that has been tuned for the Elastic Fabric Adapter (EFA) instances on AWS. It comes with a very simple one-command launcher, xscale-ai-run, that significantly simplifies launching a distributed training workload on a AWS parallel cluster. No more complex commands and suboptimal performance.

X-ScaleAI also provides an easy-to-use API to easily scale up your AI workloads. The API automatically applies various optimizations pertaining to distributed data loading and training. It provides scalable model checkpoint and restart support for long-running training and fine-tuning applications. Use X-ScaleAI and save time and effort involved in getting a distributed AI set up and optimized. Reduce your carbon footprint and time to solution by using our optimized stacks.