{
  "schema": "kira-benchmark-plan/v1",
  "status": "planned",
  "campaign": "MI350X production-inference qualification",
  "run_schema": "https://kiraruntime.cloud/benchmark-run.schema.json",
  "publication_policy": "No performance claim is promoted until the exact run is reproducible and its manifest, errors, and raw measurements are retained.",
  "target_hardware": {
    "accelerator": "AMD Instinct MI350X",
    "shape": "8 GPU node",
    "memory": "288 GB HBM3E per GPU",
    "topology": "record from runtime inventory; never infer"
  },
  "software_manifest": [
    "OS and kernel",
    "AMD driver and ROCm versions",
    "container image digest",
    "runtime and framework git revision",
    "attention backend",
    "graph/compile settings",
    "scheduler settings",
    "KV-cache dtype and allocation",
    "GPU clocks/power mode when available"
  ],
  "workload_matrix": {
    "runtimes": [
      "SGLang",
      "vLLM",
      "Kira-qualified profile"
    ],
    "concurrency": [
      1,
      8,
      32,
      64,
      128
    ],
    "input_tokens": [
      512,
      4096,
      16384
    ],
    "output_tokens": [
      256,
      512
    ],
    "repetitions_per_point": 3,
    "warmup": "continue until compilation/graph capture and allocator behavior are stable; record warmup separately"
  },
  "metrics": [
    "request throughput",
    "output tokens per second",
    "TTFT p50/p95/p99",
    "TPOT/ITL p50/p95/p99",
    "end-to-end latency p50/p95/p99",
    "GPU memory utilization",
    "KV-cache capacity/pressure",
    "GPU utilization",
    "request errors and OOMs"
  ],
  "evidence_rules": [
    "Never discard failed or OOM points from the report.",
    "Keep identical prompt/output accounting across compared runtimes.",
    "Pin model revision and weight precision.",
    "Retain raw per-request measurements, not only aggregates.",
    "Repeat each promoted point at least three times.",
    "Separate benchmark results from public playground latency."
  ]
}