{
  "schema_version": "infercrane.runtime-compatibility-screen/v1",
  "recorded_at": "2026-09-20T02:20:00+02:00",
  "evidence_class": "real-gpu-negative-compatibility-evidence",
  "model": {
    "id": "Qwen/Qwen3-0.6B",
    "revision": "c1899de289a04d12100db370d81485cdf75e47ca"
  },
  "hardware": {
    "gpu_request": "H100!",
    "gpu": "NVIDIA H100 80GB HBM3",
    "compute_capability": "sm90"
  },
  "candidate": {
    "runtime": "sglang==0.5.10.post1",
    "benchmark": "aiperf==0.12.0",
    "image": "nvidia/cuda:12.8.1-devel-ubuntu22.04 with Python 3.12 and libnuma1",
    "stable_recipe": [
      "--disable-piecewise-cuda-graph"
    ],
    "workload": {
      "input_tokens": 128,
      "output_tokens": 64,
      "streaming": true,
      "concurrency_lanes": [1, 8, 32]
    }
  },
  "attempts": [
    {
      "app_id": "ap-EApdiR1fr0n4XGkJjPn4QO",
      "outcome": "rejected",
      "finding": "runtime-only image could not import optional DeepGEMM because no CUDA toolkit was present"
    },
    {
      "app_id": "ap-yw5ZkSPC8LPkthEtdkU0jS",
      "outcome": "rejected",
      "finding": "Qwen3 RoPE JIT path required CUDA_HOME and nvcc"
    },
    {
      "app_id": "ap-qjrEgjTuFZMVuVnX98S496",
      "outcome": "rejected",
      "finding": "native SM90 extension required the libnuma runtime"
    },
    {
      "app_id": "ap-qsZiNWXy1TZsDg2O2QlJD0",
      "outcome": "rejected",
      "finding": "default experimental piecewise CUDA graph warmup caused an illegal memory access after ordinary CUDA graph capture succeeded"
    },
    {
      "app_id": "ap-01FRg3pRRCCh4slj4xjpFm",
      "outcome": "aborted",
      "finding": "stable fallback server started, but the AIPerf streaming campaign did not complete promptly"
    },
    {
      "app_id": "ap-wMaUiacP5zPw2vD7eZCNfP",
      "outcome": "aborted",
      "finding": "engine-neutral local token counting did not make the standardized AIPerf campaign complete within the operator's cost bound"
    }
  ],
  "decision": "reject this SGLang recipe for the MVP qualification set",
  "reason": "the pinned runtime did not produce a complete, zero-failure AIPerf receipt under the bounded H100 campaign; startup compatibility fixes alone are not performance evidence",
  "qualified_alternative": "vLLM 0.22.1 compiled recipe in qwen3-0.6b-modal-2026-09-20.json",
  "qualification_boundary": "this rejects only the exact runtime, model, image, accelerator, and workload tuple; it is not a general claim about SGLang"
}