Efficient and Scalable Vision Model Training is a benchmarking and workflow-oriented project designed to help users train deep vision models efficiently across multiple A100 or H100 GPUs using PyTorch’s Distributed Data-Parallel (DDP) on SLURM-managed clusters. Rather than being tied to a specific dataset or architecture, the project allows users to plug in available or their own model and dataset configurations. It demonstrates best practices for scalable training, including SLURM-native job management, environment setup, and optimized data loading strategies. The goal is to achieve strong performance and near-linear scaling while maintaining flexibility and reproducibility across a variety of vision workloads on the Kempner AI cluster.
| Workflow | Model | Dataset | Max Tested GPUs | Tags |
|---|---|---|---|---|
| imagenet1k_resnet50 | ResNet-50 | ImageNet-1k | 64 | A100, DDP |
| imagenet1k_alexnet | AlexNet | ImageNet-1k | 4 | A100, DDP |