Why it matters

QPS math for LLM is nuanced. Understanding shapes capacity planning.

Advertisement

The architecture

QPS = throughput_tokens / avg_tokens_per_query.

Include prefill + decode.

QPS mathTokens/sec/GPUthroughputAvg tokens/queryinput + outputQPS/GPUthroughput / avgLLM QPS 10-100x lower than web QPS; long queries dominate cost
QPS.
Advertisement

How it works end to end

Query has input + output tokens.

Compute both.

Long tail dominates.