Coming from PyTorch

If you already know PyTorch, most ESPnet3 pieces are still familiar. The main shift is that ESPnet3 organizes them as recipes, Systems, and stage-specific configs.

Use these pages to map the usual PyTorch questions:

Data and dataloader

Use normal PyTorch datasets, then wire them through `dataset:` and `dataloader:`.

Model and system

Split model code and workflow code across `src/model.py` and `System` methods.

Multi-GPU

See how Lightning training parallelism and provider-runner parallelism differ.

Logging and debug

Use built-in logging, logger config, profiler config, and fast debug knobs.

Customize the training loop

Move to recipe-local LightningModule, trainer, or System code when the default loop is not enough.