Benchmarking and profiling inference
Benchmarking and profiling inference
Every optimization in the rest of this course is a claim that something got faster or cheaper. Two tools turn the claim into a number. A benchmark treats the server as a black box: it sends a realistic stream of requests and records time to first token, inter-token latency and throughput at the percentiles users feel. A profiler opens the box: it records every kernel the GPU ran during a few steps and tells you which ones took the time.
Use them in that order. The benchmark tells you whether there is a problem and how big it is. The profiler tells you why. Profiling a server that already meets its targets at the load you need is a pleasant afternoon with no payoff.
Continue reading
Pro unlocks every video lesson, the full notes, worked sizing and cost calculations, and runnable configs across the Inference Engineering course, from GPUs and kernels to quantization, speculative decoding, disaggregation and production autoscaling.