Why it matters

Chunked prefill improves latency mix. Understanding shapes serving.

Advertisement

The architecture

Chunk size C tokens per step.

Mix chunk prefill with decode tokens.

Chunked prefill flowLong prompt8k tokensChunk C = 512per stepMix with decodesin batchvLLM chunked prefill + TRT-LLM in-flight batching support this
Chunked prefill.
Advertisement

How it works end to end

Chunk size tunable.

Piggyback decodes on chunks.

Better ITL for mixed loads.