Case Study: Verda achieves scalable, reproducible GPU cluster workflows with DataCrunch's Pyxis and Enroot integration

Verda’s 16xH200 cluster integration with DataCrunch powers multi-node TorchTitan training

Verda, a company specializing in AI cloud services, faced the challenge of configuring and deploying large-scale, multi-node distributed AI training workloads on their DataCrunch Instant Clusters. They needed a solution to manage containerized environments efficiently while ensuring reproducibility and avoiding performance issues related to shared storage and permissions.

DataCrunch provided a solution by guiding Verda through the integration of Pyxis, a Slurm plugin, and Enroot, a lightweight container runtime. This setup enabled Verda to successfully run a production-ready distributed training of a Llama 70B model, achieving a structured and reproducible workflow for their machine learning workloads on the HPC environment.

Verda