You already know: training FLOPs (6 × params × tokens) and mixed-precision training (D5 LLM pretraining); the training loop in PyTorch (D4); tensor parallelism and MoE at inference (D12 Advanced optimization); stragglers and lockstep work (D14 Distributed compute); the D14 bridge's memory and time arithmetic.
Why this matters for Atlas. No single GPU can hold or train a frontier model, so training splits the work along one or more dimensions: the data (each GPU sees different examples), the tensors inside each layer, the layers themselves (pipeline), the sequence, or the experts of an MoE model. Each split saves memory or time in a different way and costs a different kind of communication. Real training runs combine them ("3D" or "4D" parallelism), chosen to fit the model in memory and keep the expensive communication on the fastest links. This lesson builds each strategy in NumPy, proves it computes the same thing as one big GPU would, and quantifies what it costs.
1. Data parallelism: same model, different data
The question. The model fits on one GPU but training is too slow. What's the simplest way to use 8 GPUs?
Start from nothing. Give every GPU a full copy of the model and a different slice of each batch. Each computes gradients on its slice; the gradients are averaged across GPUs (an all-reduce, next lesson); every GPU applies the same averaged gradient, so the copies stay identical. Mathematically it's the same as training on the whole batch on one GPU, because the gradient of a mean loss is the mean of the per-slice gradients. PyTorch DDP does exactly this, overlapping the all-reduce with the backward pass.