Open Models Almanac
Methodsoptimizationtraining parallelism

single mention · optimization

Cross-device average token-count loss normalization

Each device divides token-level losses by the average number of loss tokens across devices rather than its local token count.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See training parallelism for the methods filed beside it.

Filed under optimization › training parallelism