Running trained models to produce predictions or generated output in production. Covers latency and throughput tuning, batching and quantisation, GPU and accelerator utilisation, serving runtimes such as vLLM, TensorRT-LLM and Triton, and inference cost.