What is Data Parallelism?

Data parallelism is the most common way to train large models across multiple accelerators: each device holds a full copy of the model, processes a different slice of the batch, and gradients are synchronized between steps. It contrasts with model parallelism, which splits the model itself when it will not fit on one device, and the two are frequently combined in large runs. For an organization consuming models this is background — a fact about how the model was produced, with no bearing on how it behaves in deployment. Centralpoint's concerns begin after training: which model is selected at runtime, what it is permitted to retrieve, what rules govern its behaviour, and what is retained afterwards. That is why a change of model or provider requires no re-indexing and no rewriting of the organization's own logic.