iteration / step

An iteration or step is one update of model weights after processing one batch, the atomic unit of training. **Definition**: Forward pass on batch, compute loss, backward pass, optimizer step = one iteration. **Relationship to epochs**: steps_per_epoch = dataset_size / batch_size. Total steps = epochs x steps_per_epoch. **LLM training**: Often measured in steps rather than epochs. Millions of steps for large models. **What happens each step**: Load batch, forward pass, compute loss, backward pass (gradients), optimizer update, (optional logging). **With gradient accumulation**: Logical step may span multiple forward-backward passes before optimizer update. **Logging frequency**: Log every N steps (e.g., 100). Too frequent is expensive, too infrequent misses issues. **Checkpointing**: Save model every N steps or epochs. Balance between safety and storage. **Learning rate per step**: Most schedulers update LR per step, not per epoch. Smoother adaptation. **Steps vs samples**: Sometimes report samples (steps x batch size) for comparisons across batch sizes. **Progress tracking**: Steps are wall-clock-neutral metric. Epochs depend on dataset size.