데이터 병렬화와 모델 병렬화

One GPU is fine until your model or dataset gets big. Distributed training spreads the work across many GPUs so you finish hours faster.

There are two core strategies: split the data across GPUs, or split the model itself. Each solves a different bottleneck.

In data parallelism, every GPU holds a full copy of the model but trains on a different slice of the batch. It is the most common setup. 🚀

After each GPU computes gradients on its slice, they get averaged across all devices. Every model copy then takes the same step and stays identical.

Reach for data parallelism when your model fits on one GPU but training is just slow. You scale throughput by adding more devices.

In model parallelism, the network is too large for one GPU, so its layers are split across several devices, each holding only a piece.

With a split model, the output of one GPU becomes the input to the next. This handoff adds communication cost between devices.

A smarter model split is pipeline parallelism: micro-batches flow through the layer stages so no GPU sits idle waiting for the others.

Tensor parallelism goes finer, splitting a single big matrix multiply across GPUs. It powers training of the very largest language models.

Real frontier runs mix strategies: data parallel across nodes, model parallel inside each one. This hybrid approach trains models too big for any single machine.

Most projects only ever need data parallelism. Master that first, and reach for model splitting only when your network truly cannot fit.

You learned the two ways to scale: data parallelism copies the model and splits the batch, while model parallelism splits the network itself. Start with data parallel.