Padding tokens fill sequences to uniform length for efficient batched processing. Why needed: Batched computation requires same sequence length, real sequences vary in length, padding fills the gap. Padding strategy: Right padding: Add PAD tokens at end (common for causal/decoder models). Left padding: Add PAD tokens at start (sometimes used for generation so outputs align). Attention mask: Critical companion to padding, tells model to ignore PAD tokens. Without mask, model would attend to meaningless PAD tokens. Token ID: Often 0, but varies by tokenizer. Should never contribute to loss or attention. Loss masking: Training loss excludes PAD positions, only compute loss on real tokens. Efficiency concern: Long padding wastes computation. Solutions include dynamic batching (group similar lengths), sequence packing. Memory: Padding inflates batch memory usage. Maximum sequence length should match data needs. Implementation: Tokenizer handles padding with padding=True, pad_to_max_length parameters. Always pair with attention_mask.