Reference serving workloads

These optional, versioned profiles make independently published runs easier to compare without preventing custom workloads. They describe online serving—not isolated prefill or decode microbenchmarks.

Machine-readable profile catalog

ProfileInput tokensOutput tokensRecommended concurrencyIntent
serve-fixed-1k-256-v11,0242561, 4, 8, 16, 32, 64Balanced serving
serve-fixed-8k-256-v18,1922561, 2, 4, 8, 16, 32Long prompt
serve-fixed-256-1k-v12561,0241, 2, 4, 8, 16, 32Long generation

Copy the selected manifest_fragment, set one recommended concurrency value, and record the profile ID in workload.extensions["runpile.profile_id"]. Publish each concurrency as a separate run and include every completed measured point—not only winners or representative points. The profile fixes endpoint, streaming, token shapes, synthetic generator semantics, sampling, EOS behavior, cache/prefix state, client placement, warm-up, and failure treatment.

Harness and runtime versions, model identity, hardware, launch command, benchmark command, and material runtime parameters remain facts of the individual run. Do not claim a profile when the executed command differs from its fragment.