gradient accumulation

Gradient accumulation simulates larger batch sizes by accumulating gradients over multiple forward-backward passes before updating. **How it works**: Run forward and backward multiple times, sum gradients, then apply single optimizer step. Effective batch = micro-batch x accumulation steps. **Why useful**: GPU memory limits batch size. Want larger effective batch for training stability without more memory. **Implementation**: Call loss.backward() multiple times, then optimizer.step() and zero_grad(). Or use framework support. **Memory benefit**: Same memory as small batch, but large batch training dynamics. **Training dynamics**: Large batches often need learning rate scaling (linear scaling rule). May affect convergence. **Trade-off**: More forward/backward passes before update = slower wall-clock time. Worthwhile when batch size matters. **Common use cases**: Limited GPU memory, matching batch size across different hardware, very large batch training experiments. **Distributed training**: Accumulation within device, sync gradients after accumulation steps. Reduces communication frequency. **Best practices**: Scale learning rate appropriately, consider gradient normalization, validate against true large batch training.