Benchmarking Workload Profiles for Large Language Model Serving

This document defines standard workload profiles for benchmarking Large Language Model (LLM) inference serving systems. Each profile represents one class of production workload. A profile is defined by its input and output token distribution, output structure, latency sensitivity, concurrency pattern, and caching behavior. Profiles are organized into six groups: Non-Generative, Minimal Output, Interactive Streaming, Prefill-Heavy Generative, Decode-Heavy Generative, and Multi-Step Chained. This document works with "Benchmarking Methodology for Large Language Model Serving" [LLM-METHOD]. It specifies the workloads that the methodology's tests SHOULD be run against.