{
  "schema": "anvil-serving.sanitized-qualification-summary/v1",
  "observed_date": "2026-08-15",
  "source_base_revision": "6f3af00659441371b266b27bd910d932f428f277",
  "hardware": {
    "host_label": "Primary Node",
    "gpu": "NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition",
    "gpu_count": 2,
    "vram_gib_per_gpu": 96,
    "topology": "two concurrent TP=1 lanes, then model placement swapped across the equal PCIe cards"
  },
  "runtime": {
    "image": "lmsysorg/sglang@sha256:506525a5907ea22c9d445afb7c03603959b912de034d86915cf17da814f1a124",
    "image_label_engine_revision": "c4271c3fe1262fc2adbd162c33b25de5255251c5",
    "internal_build_version_revision": "561c8f3",
    "source_recipe_revision": "dd458f3212dd4ddf0e1a7907bbf539b660e70d21"
  },
  "common_recipe": {
    "tensor_parallel_size": 1,
    "context_tokens": 393216,
    "max_running_requests": 1,
    "language_only": true,
    "kv_cache_dtype": "fp8_e4m3",
    "attention_backend": "flashinfer",
    "chunked_prefill_tokens": 2048,
    "memory_fraction_static": 0.85,
    "prefix_cache": false,
    "max_mamba_cache_size": 1,
    "speculative_decoding": false,
    "thinking_default": "disabled"
  },
  "candidates": {
    "official_fp8": {
      "model": "Qwen/Qwen3.8-27B-FP8",
      "revision": "017b9c7af6b5689d5dd426a76e0bc077eb5ca20a",
      "cached_snapshot_bytes": 30890071640,
      "startup_weight_memory_gib": 28.47,
      "startup_kv_capacity_tokens": 1665740,
      "decision": ["qualified-control", "no-promotion"]
    },
    "inferact_nvfp4": {
      "model": "Inferact/Qwen3.8-27B-NVFP4",
      "revision": "6128240ebaf4eaa7bad2b3d1c72c37d677c5f462",
      "cached_snapshot_bytes": 26404422156,
      "startup_weight_memory_gib": 24.21,
      "startup_kv_capacity_tokens": 1805068,
      "decision": ["challenger", "no-promotion"]
    }
  },
  "artifact_safety": {
    "files": "documentation, tokenizer/config/template data, and seven Safetensors shards only",
    "tracked_executable_code": false,
    "auto_map_or_remote_code_hook": false,
    "pickle_or_pytorch_binary": false,
    "safetensors_files_verified": 7,
    "safetensors_tensors_verified": 2111,
    "headers_ranges_and_file_lengths_valid": true,
    "local_sha256_matches_hugging_face_lfs": true,
    "shard_sha256": [
      "4b71c5d9d5027c88d1df9f53b93ad77b8d5e428d5d6f75ae42063704d94930b9",
      "d0ceb042ef003ec63e00203823eb70d4815c9fc0430ba04a5442b6b8bde19093",
      "b633a26c1d3b97bf061585363b99d4c01650b4e3f6d0cbb18e7c7e03ca2de1e4",
      "485ea312626e2d1f3a820d9f7545809cef1e2e77414e3977a61d32dacd370bd0",
      "2aa61bab1ed252fe3cdb479286b5b7add7e8ad567501add5f6eeef3987b03d31",
      "8214cd21a46832380c06be552c42d9514c6cc9f27d98e09e2a09e36a8bd6e6ce",
      "c2b7cb6987e40f13b1ab6e552958afe47a759f86cac0c3c4a1c29f606dc080ef"
    ]
  },
  "functional": {
    "official_fp8": {
      "placements_passed": 2,
      "checks": ["coding", "structured-json", "128k-retrieval", "tools-20-of-20", "streaming-tools", "tool-result-continuation", "responses"],
      "reasoning_chars_with_thinking_disabled": 0
    },
    "inferact_nvfp4": {
      "placements_passed": 2,
      "checks": ["coding", "structured-json", "128k-retrieval", "tools-20-of-20", "streaming-tools", "tool-result-continuation", "responses"],
      "reasoning_chars_with_thinking_disabled": 0
    }
  },
  "bounded_quality": {
    "official_fp8": {"intelligence": "6/6", "session": "3/3", "tool": "3/3", "failures": 0},
    "inferact_nvfp4": {"intelligence": "6/6", "session": "3/3", "tool": "3/3", "failures": 0}
  },
  "capacity_4k_c1": {
    "protocol": "five 10-request runs per candidate; 4096 configured input tokens; 256-token output cap; concurrency one; values are means of each run median",
    "official_fp8": {
      "runs": 5,
      "ttft_p50_ms": 554.3,
      "e2e_p50_ms": 1450.9,
      "effective_prefill_tok_s_p50": 6512,
      "decode_tok_s_p50": 48.0,
      "aggregate_output_tok_s": 30.2
    },
    "inferact_nvfp4": {
      "runs": 5,
      "ttft_p50_ms": 429.3,
      "e2e_p50_ms": 1244.3,
      "effective_prefill_tok_s_p50": 8409,
      "decode_tok_s_p50": 57.9,
      "aggregate_output_tok_s": 38.4
    },
    "nvfp4_delta_percent": {
      "ttft": -22.6,
      "e2e": -14.2,
      "effective_prefill": 29.1,
      "decode": 20.6,
      "aggregate_output_throughput": 27.2
    }
  },
  "near_limit_context": {
    "actual_prompt_tokens": 388979,
    "official_fp8": {"status": "passed", "ttft_s": 258.13, "effective_prefill_tok_s": 1507, "decode_tok_s": 33.2},
    "inferact_nvfp4": {"status": "passed", "ttft_s": 248.75, "effective_prefill_tok_s": 1564, "decode_tok_s": 37.0}
  },
  "retained_failures_and_caveats": [
    "The first launch refused 393216 because it exceeds the advertised 262144 context until the explicit SGLang overwrite opt-in was added.",
    "The default multimodal CUDA IPC warmup failed on WSL2 with CUDA invalid resource handle; the matched text qualification uses language-only and does not qualify SGLang multimodal serving.",
    "The first otherwise-passing functional run leaked reasoning through the Responses adapter until a server-side thinking-disabled default was added.",
    "Both lanes warned that FP8 KV scaling factors were absent and defaulted to 1.0; bounded quality passed but equivalence to unquantized KV is not proven.",
    "The immutable image label and internal build-version string identify different SGLang revisions.",
    "The repository-provided crc32 file was stale and its md5 list empty, so immutable LFS identities and independently computed local SHA256 values were used.",
    "The generic evidence inspector flags unselected timing or suite fields in deterministic-quality and context-only artifacts; claims are limited to complete per-attempt and per-target records."
  ],
  "restoration": {
    "exact_original_split_restored": true,
    "official_fp8_direct_gate": "passed",
    "bf16_multimodal_direct_gate": "passed including image and OCR",
    "router_primary_identity_and_gate": "passed",
    "router_vision_general_gate": "passed",
    "router_vision_ocr_gate": "passed",
    "shared_memory_reclaimable_bytes": 0,
    "promotion_changed": false,
    "client_configuration_changed": false
  }
}