{
  "description": "Production vllm7 (amd/Qwen3.8-27B-Quark-AWQ-MXFP4 + DFlash2 k=7, LXC 2408). The gate reuses the live docker-vllm7.service arguments, so this profile only declares what to assert.",
  "service": "docker-vllm7.service",
  "model": "amd/Qwen3.8-27B-Quark-AWQ-MXFP4",
  "required_log_patterns": [
    [
      "quark mxfp4 W4A8 gemm",
      "\\[radiance\\.mxfp4\\] W4A8 fp8-WMMA GEMM ENABLED"
    ],
    [
      "mxfp4 decode kernel",
      "\\[radiance\\.mxfp4\\] decode kernel ON"
    ],
    [
      "all linear layers on our kernel",
      "\\[radiance\\.mxfp4\\] linear layers: (\\d+)/\\1 on our kernel, 0 FORCED"
    ],
    [
      "fused norm/quant armed",
      "\\[radiance\\.fused_norm\\] armed: add_rms_quant=1 silu_mul_quant=1 gdn_norm_quant=1"
    ],
    [
      "skinny gemm",
      "\\[radiance\\.gemm\\] skinny GEMM kernel ENABLED"
    ],
    [
      "gdn chunk scan",
      "\\[radiance\\.gdn\\] gdn_chunk_scan ENABLED"
    ],
    [
      "int4 lm_head",
      "lm_head quantised to int4 g128"
    ],
    [
      "int4 embedding",
      "embed_tokens quantised to int4"
    ],
    [
      "dflash drafter",
      "Resolved architecture: DFlash2DraftModel"
    ],
    [
      "dynamic draft controller",
      "RADIANCE_DYNAMIC_DRAFT"
    ]
  ],
  "forbidden_log_patterns": [
    [
      "python traceback",
      "Traceback \\(most recent call last\\)"
    ],
    [
      "engine died",
      "EngineDeadError|Engine core proc .* died"
    ],
    [
      "kernel fell back",
      "[1-9]\\d* FORCED ONTO AITER"
    ],
    [
      "cuda graph capture failed",
      "Capturing CUDA graphs? failed|CUDAGraph capture failed"
    ]
  ],
  "known_warnings": [
    "install_attn_config_hook failed",
    "triton_bundler.py",
    "Cubin file saved by TritonBundler not found",
    "Failed to JIT torch c dlpack extension",
    "torch/_inductor/"
  ],
  "smoke": [
    {
      "name": "arithmetic",
      "prompt": "What is 17*23? Reply with the number only.",
      "expect": "391",
      "max_tokens": 512
    },
    {
      "name": "reasoning parser",
      "prompt": "A train leaves at 14:35 and arrives 2h47m later. What time does it arrive? Answer with HH:MM only.",
      "expect": "17:22",
      "max_tokens": 512
    },
    {
      "name": "tool call",
      "prompt": "What is the weather in Berlin? Use the tool.",
      "expect_tool_call": "get_weather",
      "max_tokens": 512
    },
    {
      "name": "long prefill needle",
      "expect": "MARKER-7731",
      "max_tokens": 256,
      "prompt_needle": {
        "marker": "MARKER-7731",
        "filler_sentences": 4000
      },
      "timeout": 900
    }
  ],
  "thresholds": {
    "kv_tokens_min_ratio": 0.98,
    "aggregate_tps_min_ratio": 0.95,
    "acceptance_rate_min_delta": -0.02,
    "mean_tokens_per_step_min_delta": -0.3,
    "gsm8k_min_delta": -0.05,
    "startup_s_max_ratio": 1.5,
    "preemptions_max": 0
  },
  "smoke_note": "Every answer is additionally checked for leaked <think>/</think> tags (reasoning parser armed); an empty reasoning_content is normal for easy prompts."
}