Fast reference benchmarks for training ML models with recipes. Designed to be easily forked and modified.
Figure 1: Comparison of MosaicML recipes against other results, all measured on 8x A100s on MosaicML Cloud.
Train the MosaicML ResNet, the fastest ResNet50 implementation that yields a โจ 7x โจ faster time-to-train compared to a strong baseline. See our blog for more details and recipes. Our recipes were also demonstrated at MLPerf, a cross industry ML benchmark.
๐ Get started with the code here.
Train the MosaicML DeepLabV3 that yields a โจ5xโจ faster time-to-train compared to a strong baseline. See our blog for more details and recipes.
๐ Get started with the code here.
![Training curves for various LLM sizes.](./llm/assets/loss-curve-light.png)
A simple yet feature complete implementation of GPT, that scales to 70B parameters while maintaining high performance on GPU clusters. Flexible code, written with vanilla PyTorch, that uses PyTorch FSDP and some recent efficiency improvements.
๐ Get started with the code here.