{
  "schema": "anvil-serving.qwen38-27b-qualification-summary/v1",
  "observed_date": "2026-08-14",
  "decision": "challenger; no-promotion",
  "hardware": {
    "product": "NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition",
    "cards": 2,
    "vram_gb_each": 96,
    "topology": "two equal independent PCIe split lanes; one TP=1 serve per card"
  },
  "runtime": {
    "image": "vllm/vllm-openai@sha256:4a2f33a884222f7049b983263ad9976f89452bb81affecf5b67d89ad35c1bc31",
    "engine_revision": "3a0914114705fa38d4c3171d0746c1a6b6f10209",
    "cuda": "13.0.1"
  },
  "checkpoints": {
    "bf16": {
      "repo": "Qwen/Qwen3.8-27B",
      "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
      "safetensors_files": 18,
      "safetensors_bytes": 55563006776
    },
    "official_fp8": {
      "repo": "Qwen/Qwen3.8-27B-FP8",
      "revision": "017b9c7af6b5689d5dd426a76e0bc077eb5ca20a",
      "safetensors_files": 66,
      "safetensors_bytes": 30866866928
    },
    "artifact_gate": "all local safetensors SHA-256 values matched pinned Hugging Face LFS identities; no executable, Python, pickle, PyTorch bin, native-library, auto_map, or trust_remote_code path was present"
  },
  "baseline": {
    "shared": {
      "context_tokens": 262144,
      "tensor_parallel_size": 1,
      "kv_cache_dtype": "fp8",
      "prefix_cache": false,
      "speculative_decoding": false,
      "thinking_disabled_control": "verified",
      "adaptive_reasoning_efforts_passed": ["low", "medium", "xhigh"]
    },
    "bf16_multimodal": {
      "max_num_seqs": 2,
      "gpu_memory_utilization": 0.9,
      "functional_checks": "all passed",
      "repeated_quality": "intelligence, session, tool, and 31225-token retrieval passed at 3/3 policy",
      "capacity_4k": {
        "c1_decode_tok_s_p50": 26.9,
        "c1_e2e_s_p50": 2.46,
        "c2_decode_tok_s_p50": 17.8,
        "c2_aggregate_output_tok_s": 27
      },
      "context": [
        {"actual_prompt_tokens": 31225, "ttft_s": 8.05, "status": "passed"},
        {"actual_prompt_tokens": 139428, "ttft_s": 53.44, "status": "passed"},
        {"actual_prompt_tokens": 241250, "ttft_s": 125.14, "status": "passed"}
      ],
      "multimodal": {
        "corpus_sha256": "ebff9dcc87a7fd13f801fc19eeea7271aec01a99fe560d721be99c1c9becad49",
        "attempts": 30,
        "passed": 30,
        "image": "12/12",
        "video": "14/14",
        "mixed": "4/4"
      }
    },
    "official_fp8_text": {
      "max_num_seqs": 5,
      "gpu_memory_utilization": 0.92,
      "functional_checks": "all passed before and after setting experiments",
      "repeated_quality": "intelligence, session, tool, and 31225-token retrieval passed at 3/3 policy",
      "capacity_4k": {
        "c1_decode_tok_s_p50": 47.9,
        "c1_e2e_s_p50": 1.69,
        "c5_decode_tok_s_p50": 16.2,
        "c5_aggregate_output_tok_s": 51
      },
      "context": [
        {"actual_prompt_tokens": 31225, "ttft_s": 5.88, "status": "passed"},
        {"actual_prompt_tokens": 139428, "ttft_s": 42.04, "status": "passed"},
        {"actual_prompt_tokens": 241250, "ttft_s": 104.71, "status": "passed"}
      ],
      "engine_kv_tokens": 1825809,
      "reported_full_window_concurrency": 6.96,
      "caveat": "startup warned that absent FP8 attention q/prob scaling factors defaulted to 1.0; quality equivalence to unquantized KV is not established"
    }
  },
  "setting_experiments": {
    "mtp3": {
      "only_change": "speculative_config method=mtp, num_speculative_tokens=3",
      "functional_checks": "all passed",
      "repeated_quality": "all passed",
      "c1_decode_tok_s_p50": 94.8,
      "c5_aggregate_output_tok_s": 54,
      "quality_context_decode_tok_s": 100.4,
      "warning": "vLLM reported the inherited 4096 max batched tokens may be suboptimal for MTP3; observed draft acceptance varied by workload",
      "retained_failure": "first managed start failed before load because registry quoting stripped JSON quotes; the same value succeeded after serialization-only correction"
    },
    "prefix_cache": {
      "only_change": "prefix caching enabled",
      "workload": "five concurrent requests with a 30000-token shared prefix",
      "enabled_cold": {"ttft_s_p50": 9.18, "aggregate_output_tok_s": 23},
      "enabled_warm": {"ttft_s_p50": 0.41, "aggregate_output_tok_s": 142},
      "disabled_first": {"ttft_s_p50": 16.59, "aggregate_output_tok_s": 9},
      "disabled_repeat": {"ttft_s_p50": 16.39, "aggregate_output_tok_s": 9},
      "caveat": "the endpoint omitted prompt_tokens_details.cached_tokens, so the result is timing-based reuse evidence"
    },
    "unquantized_kv": {
      "only_change": "kv_cache_dtype fp8 to auto, resolved with bfloat16 model dtype",
      "functional_checks": "all passed",
      "c1_decode_tok_s_p50": 47.8,
      "c5_aggregate_output_tok_s": 51,
      "actual_prompt_tokens_passed": 244573,
      "engine_kv_tokens": 929913,
      "reported_full_window_concurrency": 3.55,
      "decision": "viable accuracy-oriented control; no short-context speed gain and about half the full-window KV-token capacity"
    }
  },
  "gaps": [
    "No candidate router alias, promotion, or routed client acceptance was attempted.",
    "The durable separate-worker context, agentic, and SWE harness was not submitted because the candidate had no approved router alias; direct API tool and coding checks are not SWE-bench evidence.",
    "Adaptive reasoning effort runs prove control plumbing and visible/reasoning separation, not a quality ranking among effort levels.",
    "The official FP8 attention-scaling warning requires a matched independent quality A/B before treating FP8 KV as quality-equivalent to unquantized KV."
  ],
  "private_artifact_digests": {
    "bf16_preflight": "cdc8e0a45eef62ad384829a548c2b5ef041d8841e5e5f2b1737d56520ecfb917",
    "fp8_preflight": "038c48969b5e20e902ab954c866d686e14823128d418480ecc6d46c56556aa7e",
    "bf16_quality": "3198ea2251b7579a24cfedb89bccbb640783cd3b655923e343570b2e3aeb46e9",
    "fp8_quality": "c84fe8ec26055327dcc0d88866e6f7a247051e7cea59073dd84f8e77a1b3d0e9",
    "bf16_context": "e85eef2eae82f0831eb3dddaac81540dc7c45a7130ca1343f32eab9148deff06",
    "fp8_context": "02c7c0a0f57d0b6b2e0d861d0356ce3c2a4d04b27e83e2d478617fa1442fe4c8",
    "bf16_multimodal": "d8540307e0ddafe247357cec2b18b6eef3537b193c135abaa94a02ab473861cc",
    "mtp3_quality": "6a531da51412f1887b0c4875de7f4f503c183c96f9debdab05407bae097a1b33",
    "prefix_enabled_warm": "adf9a600061738ebdf6c61aeb7cb99c6d155cc62ea4855d11f330beb863c2f01",
    "prefix_disabled_repeat": "16b3a5a95677ddad43cfce3567a85849ef3d9c804d24dc1ee55b8e726cbc6016",
    "unquantized_kv_context": "72c7db0160a0f39c816ac93f6f16039c50019aeb3212a1a1226b3c609b9a90c1"
  }
}