Padding
Filler tokens added to short sequences so a batch of them all share the same length.
Think of it like
Like adding packing peanuts so every box on the truck is the same size and stacks neatly.
Example
To process ten prompts of different lengths at once, the system pads the short ones with a special token until all match the longest.
How it actually works
Hardware wants uniform rectangles, so shorter sequences get pad tokens appended (or prepended). Those tokens carry no meaning and must be hidden from the model’s attention — that’s the attention mask’s job — otherwise the model would “read” the filler as content.
For product teams
Invisible plumbing that makes batching efficient; only surfaces as a bug when the mask is wrong.
For engineers
Appends pad tokens to reach a common sequence length; paired with an attention mask that zeroes them out.
Related
- Attention mask — What tells the model to ignore the filler.
- Truncation — The opposite operation: cutting to fit.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome