Reference · Glossary

Batching (inference)

Last updated

Running **multiple model requests together** in one forward pass to improve GPU throughput — often at the cost of some latency.

When to use

High-volume serving where maximizing tokens/sec matters more than instant first token for every user.

When not to

Interactive chat where each user expects immediate streaming — tiny batches or single requests may feel snappier.

Example

A server queues 8 embedding requests for 20ms, runs one GPU batch, returns all eight vectors together.