The company says Zephon preserves the same sequence of global training batches when GPU count, operator parallelism, or execution backend changes, including pipelines that tokenize, pack, and mix data during training.
The release targets resumed jobs and cluster resizing, where a conventional loader can change the samples seen after a checkpoint. DatologyAI reports an experiment training a 1B-parameter model for 20B tokens on 8 to 64 GPUs: FineWeb evaluation scores varied by up to 0.64 points with a standard loader and 0.011 with Zephon. Reported DCLM Core v1 variation was 0.82 and 0.05 points, respectively.
DatologyAI also reports 16,000 samples per second for fetch-and-batch, versus 5,800 for Grain, and provides a paper, documentation, and integrations for TorchTitan and Megatron-LM. The source does not establish independent reproduction of the performance or evaluation results, nor does it show how Zephon performs across other training stacks or workloads.