New Data Loader Makes AI Training More Reliable and Cheaper

When training AI models, the data they see is crucial. But preparing that data can be expensive and time-consuming. Typically, each experiment requires creating a preprocessed copy of the dataset, which takes storage and compute. Since most experiments fail, this wastes resources. Zephon, a new data loader for TorchTitan (a framework for training large AI models), solves this by preparing data on the fly during training, using idle CPUs. This means no preprocessed copies are needed, saving time and storage.

Another problem Zephon addresses is determinism. With standard data loaders, the order of data samples depends on the number of GPUs used. So if you resume training on a different number of GPUs, the model sees a different sequence, leading to different results. This makes it hard to compare experiments. Zephon ensures the same data order regardless of GPU count, so results are consistent. In tests, models trained with Zephon had nearly identical scores across different GPU counts, while standard loaders showed variations up to 0.82 points—enough to mislead researchers.

Zephon asks users to define datasets, a work source (which mixes and orders data), and a pipeline of operators (which process samples). This separation allows easy changes to the data mixture without affecting processing. Zephon also includes a token estimation utility to ensure correct mixing proportions. Overall, Zephon makes AI training more efficient and reliable, which could lead to faster improvements in AI models and lower costs.