Reproducibility contract

Claims require a workload.

TokenLabs publishes inference results as bounded engineering evidence—not universal framework, model, or quantization rankings. Every conclusion applies only to the disclosed hardware, software, traffic model, and SLO.

Publication rule: a result is “published” only when the run is complete, the comparison controls are disclosed, request failures are accounted for, machine-readable evidence is linked, and the recommendation includes limitations. Everything else remains in the lab notebook.

Metric definitions

TTFTTime from request submission to the first output token. Report p50, p95, and p99 when the sample count supports them.
ITL / TPOTDelay between generated tokens / average time per output token. The exact aggregation used by the load generator is disclosed per study.
Output throughputTotal generated output tokens divided by measured wall-clock duration. It is paired with per-user latency, not treated as a standalone winner metric.
Request throughputCompleted requests per second under the stated arrival model, concurrency, and output distribution.
GoodputRequests or tokens completed while satisfying the study's explicit latency and error SLO.
ISL / OSLInput and requested output sequence lengths. Actual tokenizer output and stop-token behavior are disclosed when they differ.

Controls for comparative studies

Model and quality

  • Same checkpoint, tokenizer, prompt set, and generation settings
  • Same quantization when comparing frameworks
  • Quality evaluation when comparing quantizations

Serving environment

  • GPU, topology, driver, CUDA/ROCm, container, and framework version
  • KV dtype, tensor parallelism, memory limits, caching, scheduler, graphs, and token budget
  • Warm-up and exclusion policy

Traffic

  • Exact or distributed ISL/OSL, concurrency, request count, and seed
  • Open-loop or closed-loop arrival model
  • Streaming mode, timeouts, retries, and client location

Run validity

A run is valid when

  • The intended model and configuration are confirmed from server logs.
  • Warm-up is separated from measurement.
  • Errors, timeouts, OOMs, and restarts are reported.
  • Measured samples and percentile method are documented.
  • Raw request metrics and configuration metadata are preserved.

A comparison is not generalized when

  • Models or quantizations differ across arms.
  • One arm uses caching or scheduler features the other does not.
  • Sample count is too small for stable tail percentiles.
  • Capacity interruption leaves unmatched cells.
  • Network or client placement materially differs.

Reproduce a run

Each published study links its raw CSV/JSON and names the exact workload. Deployment manifests, benchmark harnesses, result artifacts, and experiment notes live in the TokenLabs repository. Study-specific commands take precedence over this general contract.