Estimate inference throughput in tokens per second from the generated token count and end-to-end latency, for capacity planning.
实际受批大小与排队影响。
| Item | Formula / Note | Source / Standard |
|---|---|---|
| Definition | throughput equals total generated tokens divided by total elapsed time | MLPerf Inference rules |
| Time to first token | first token latency is measured separately from subsequent inter-token latency | MLPerf Inference rules |
| Concurrency effect | batching raises aggregate throughput while raising per-request latency | NVIDIA Triton inference server docs |