{
  "schema_version": "infercrane.optimization-poc-receipt/v1",
  "recorded_at": "2026-09-20T00:31:00+02:00",
  "model": {
    "id": "Qwen/Qwen3-0.6B",
    "revision": "c1899de289a04d12100db370d81485cdf75e47ca",
    "hidden_size": 1024,
    "layers": 28,
    "rms_norm_epsilon": 0.000001
  },
  "software": {
    "torch": "2.11.0+cu130",
    "triton": "3.6.0",
    "vllm": "0.22.1",
    "aiperf": "0.12.0"
  },
  "custom_kernel": {
    "operator": "fused residual-add + RMSNorm",
    "correctness": "passed against a float32 PyTorch oracle on every measured shape",
    "h100": {
      "gpu": "NVIDIA H100 80GB HBM3",
      "compute_capability": "sm90",
      "app_id": "ap-dcovUBMC4LC8I95UFtzjll",
      "measurements": [
        {"rows": 1, "custom_ms": 0.005897, "vllm_ms": 0.006519, "speedup_vs_vllm": 1.1055, "speedup_vs_eager": 5.5209},
        {"rows": 8, "custom_ms": 0.005782, "vllm_ms": 0.006535, "speedup_vs_vllm": 1.1302, "speedup_vs_eager": 6.0210},
        {"rows": 32, "custom_ms": 0.005907, "vllm_ms": 0.006391, "speedup_vs_vllm": 1.0819, "speedup_vs_eager": 6.1948},
        {"rows": 128, "custom_ms": 0.006370, "vllm_ms": 0.006879, "speedup_vs_vllm": 1.0799, "speedup_vs_eager": 5.9084},
        {"rows": 512, "custom_ms": 0.007209, "vllm_ms": 0.007429, "speedup_vs_vllm": 1.0305, "speedup_vs_eager": 6.1187}
      ]
    },
    "l40s": {
      "gpu": "NVIDIA L40S",
      "compute_capability": "sm89",
      "app_id": "ap-XHRvkfzxANzUuxTWFr8VXn",
      "measurements": [
        {"rows": 1, "custom_ms": 0.004416, "vllm_ms": 0.004488, "speedup_vs_vllm": 1.0162},
        {"rows": 8, "custom_ms": 0.004422, "vllm_ms": 0.004919, "speedup_vs_vllm": 1.1123},
        {"rows": 32, "custom_ms": 0.004916, "vllm_ms": 0.005063, "speedup_vs_vllm": 1.0299},
        {"rows": 128, "custom_ms": 0.006382, "vllm_ms": 0.006829, "speedup_vs_vllm": 1.0700}
      ]
    },
    "decision": "retain as a correct kernel candidate, but do not integrate into serving yet",
    "reason": "the isolated win is only tens of nanoseconds per call and the candidate does not consistently beat torch.compile; the optimistic decode Amdahl ceiling is about 1.3-2.4 percent for the tested workload"
  },
  "paired_aiperf": {
    "gpu": "NVIDIA H100 80GB HBM3",
    "app_id": "ap-ojLEtM0RlbI2tFtvPy3vw3",
    "workload": {"input_tokens": 128, "output_tokens": 64, "streaming": true, "profile_runs": 3},
    "failures": 0,
    "measurements": [
      {"recipe": "eager", "concurrency": 1, "output_tokens_per_second": 88.9783, "ttft_p95_ms": 35.2737, "tpot_p95_ms": 11.8231, "latency_p95_ms": 779.8675},
      {"recipe": "compiled", "concurrency": 1, "output_tokens_per_second": 573.1606, "ttft_p95_ms": 15.5103, "tpot_p95_ms": 1.5907, "latency_p95_ms": 116.1183},
      {"recipe": "eager", "concurrency": 8, "output_tokens_per_second": 659.8918, "ttft_p95_ms": 62.3547, "tpot_p95_ms": 11.9683, "latency_p95_ms": 806.4452},
      {"recipe": "compiled", "concurrency": 8, "output_tokens_per_second": 3914.6829, "ttft_p95_ms": 34.1316, "tpot_p95_ms": 1.7069, "latency_p95_ms": 137.0689},
      {"recipe": "eager", "concurrency": 32, "output_tokens_per_second": 2375.5471, "ttft_p95_ms": 96.2253, "tpot_p95_ms": 12.3966, "latency_p95_ms": 867.8393},
      {"recipe": "compiled", "concurrency": 32, "output_tokens_per_second": 10485.4782, "ttft_p95_ms": 70.1075, "tpot_p95_ms": 2.1071, "latency_p95_ms": 200.0858}
    ],
    "compiled_vs_eager": [
      {"concurrency": 1, "throughput_speedup": 6.4416, "ttft_p95_reduction_percent": 56.03, "latency_p95_reduction_percent": 85.11},
      {"concurrency": 8, "throughput_speedup": 5.9323, "ttft_p95_reduction_percent": 45.26, "latency_p95_reduction_percent": 83.00},
      {"concurrency": 32, "throughput_speedup": 4.4139, "ttft_p95_reduction_percent": 27.14, "latency_p95_reduction_percent": 76.94}
    ],
    "decision": "accept compiled vLLM over eager for this workload"
  },
  "public_workload_reproduction": {
    "workload": {"requests": 1000, "concurrency": 100, "input_tokens": 1000, "output_tokens": 500, "streaming": true},
    "published_example": {"output_tokens_per_second": 22521.42, "request_throughput": 45.70, "hardware_and_runtime_disclosed": false},
    "balanced": {
      "app_id": "ap-5vEcwT7UJSD1wzoz4Xj7Xa",
      "output_tokens_per_second": 12950.2671,
      "request_throughput": 25.9005,
      "ttft_p95_ms": 393.7271,
      "tpot_p95_ms": 7.3180,
      "latency_p95_ms": 3961.5544,
      "failures": 0
    },
    "throughput_mode": {
      "app_id": "ap-XgXaR55SHroVxLIpNCwrDf",
      "output_tokens_per_second": 12880.3447,
      "request_throughput": 25.7607,
      "ttft_p95_ms": 407.4518,
      "tpot_p95_ms": 7.3601,
      "latency_p95_ms": 3951.0876,
      "failures": 0
    },
    "decision": "reject throughput mode for this workload; it reduced output throughput by 0.54 percent and worsened p95 TTFT",
    "comparison_boundary": "the public example omits hardware and runtime versions, so the 57.5 percent directional ratio is not an apples-to-apples product ranking"
  },
  "runtime_compatibility": {
    "finding": "vLLM 0.22.1 can mis-detect an incomplete vendored DeepGEMM backend on H100 for a BF16 model",
    "workaround": "VLLM_USE_DEEP_GEMM=0",
    "upstream_issue": "https://github.com/vllm-project/vllm/issues/41849"
  },
  "qualification_boundary": "synthetic performance evidence only; production trace replay, semantic quality, cost, and release gates remain required"
}