Skip to content

Continuous Batching

PART 11 — Continuous Batching

Standard batching waits for a group of requests to finish together. Continuous batching injects them on the fly.

Example Trace:

Time T0: 
- Batch starts. 
- Req 1 (Prefill), Req 2 (Prefill)

Time T1 (Req 1 finished early!):
- Req 1 (Done. Evicted from GPU memory).
- Req 3 (NEW! Injected dynamically. Runs Prefill)
- Req 2 (Runs Decode token #2)

Time T2:
- Req 3 (Runs Decode token #1)
- Req 2 (Runs Decode token #3)
HLD impact: Continuous batching increases GPU throughput by 10x-20x compared to naive batching.