{
  "schema": "anvil-serving.public-benchmark-summary/v1",
  "date": "2026-08-15",
  "decision": "no-promotion",
  "hardware": {
    "gpu": "NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition",
    "count": 2,
    "topology": "split TP=1 candidates followed by a cross-card swap"
  },
  "runtime": {
    "image": "lmsysorg/sglang@sha256:506525a5907ea22c9d445afb7c03603959b912de034d86915cf17da814f1a124",
    "image_label_revision": "c4271c3fe1262fc2adbd162c33b25de5255251c5",
    "cookbook_revision": "dd458f3212dd4ddf0e1a7907bbf539b660e70d21"
  },
  "models": [
    {
      "id": "Qwen/Qwen3.8-27B-FP8",
      "revision": "017b9c7af6b5689d5dd426a76e0bc077eb5ca20a",
      "quantization": "official FP8"
    },
    {
      "id": "Inferact/Qwen3.8-27B-NVFP4",
      "revision": "6128240ebaf4eaa7bad2b3d1c72c37d677c5f462",
      "quantization": "ModelOpt NVFP4 W4A4 with FP8 projections"
    }
  ],
  "common_configuration": {
    "tensor_parallel": 1,
    "context_tokens": 393216,
    "max_running_requests": 1,
    "kv_cache_dtype": "fp8_e4m3",
    "attention_backend": "flashinfer",
    "chunked_prefill_tokens": 2048,
    "radix_cache": false,
    "memory_fraction_static": 0.85,
    "thinking": "disabled"
  },
  "speculative_configuration": {
    "algorithm": "EAGLE",
    "steps": 3,
    "topk": 1,
    "draft_tokens": 4,
    "max_mamba_cache_size": 5
  },
  "performance_protocol": {
    "requests_per_run": 10,
    "runs_per_model": 5,
    "concurrency": 1,
    "configured_input_tokens": 4096,
    "max_output_tokens": 256,
    "aggregation": "mean of run medians across three first-placement and two swapped-placement runs"
  },
  "performance": [
    {
      "model": "official-fp8",
      "speculation": "off",
      "ttft_p50_ms": 554.3,
      "effective_prefill_tok_s_p50": 6512,
      "decode_tok_s_p50": 48.0,
      "e2e_p50_ms": 1450.9,
      "aggregate_output_tok_s": 30.2
    },
    {
      "model": "official-fp8",
      "speculation": "mtp3",
      "ttft_p50_ms": 569.3,
      "effective_prefill_tok_s_p50": 6341,
      "decode_tok_s_p50": 111.3,
      "e2e_p50_ms": 954.2,
      "aggregate_output_tok_s": 46.0,
      "decode_delta_percent": 131.9,
      "e2e_delta_percent": -34.2
    },
    {
      "model": "inferact-nvfp4",
      "speculation": "off",
      "ttft_p50_ms": 429.3,
      "effective_prefill_tok_s_p50": 8409,
      "decode_tok_s_p50": 57.9,
      "e2e_p50_ms": 1244.3,
      "aggregate_output_tok_s": 38.4
    },
    {
      "model": "inferact-nvfp4",
      "speculation": "mtp3",
      "ttft_p50_ms": 447.5,
      "effective_prefill_tok_s_p50": 8065,
      "decode_tok_s_p50": 98.1,
      "e2e_p50_ms": 913.7,
      "aggregate_output_tok_s": 50.9,
      "decode_delta_percent": 69.4,
      "e2e_delta_percent": -26.6
    }
  ],
  "sampled_speculative_telemetry": [
    {
      "model": "official-fp8",
      "accepted_length_range": [3.35, 3.52],
      "acceptance_rate_range": [0.78, 0.84]
    },
    {
      "model": "inferact-nvfp4",
      "accepted_length_range": [2.88, 3.12],
      "acceptance_rate_range": [0.62, 0.71]
    }
  ],
  "functional": {
    "both_models": {
      "coding": "pass",
      "json": "pass",
      "retrieval_131k": "pass",
      "tools": "20/20",
      "streaming_tools": "pass",
      "tool_result_recovery": "pass",
      "responses_subset": "pass",
      "thinking_disabled_reasoning_leakage": false
    },
    "quality": {
      "official_fp8": {"intelligence": "6/6", "session": "3/3", "tools": "3/3"},
      "inferact_nvfp4": {"intelligence": "6/6", "session": "3/3", "tools": "3/3"}
    },
    "near_limit_retrieval": {
      "requested_context_tokens": 389000,
      "official_fp8_elapsed_seconds": 193.2,
      "inferact_nvfp4_elapsed_seconds": 185.3,
      "official_fp8": "pass",
      "inferact_nvfp4": "pass"
    }
  },
  "multimodal": {
    "transport": "cpu",
    "mtp_enabled": true,
    "official_fp8": {
      "general_image_seconds": 9.3,
      "ocr_seconds": 0.7,
      "result": "pass"
    },
    "inferact_nvfp4": {
      "general_image_seconds": 7.7,
      "ocr_seconds": 0.7,
      "result": "pass"
    },
    "not_tested": ["video", "multi-image ordering", "32-image ceiling", "concurrency above one", "host-memory pressure"]
  },
  "retained_failures": [
    "The default CUDA-IPC multimodal feature transport failed with CUDA invalid-resource-handle errors in the exact WSL2/Docker/PyTorch runtime.",
    "The first 512-token image response found the required content but ended at the completion cap; the 1024-token rerun passed.",
    "The global editable CLI initially ignored the immutable recipe model_path; no measurement used those failed containers.",
    "The first final Primary readmission returned a transient 503; identity remained ready and the managed retry plus routed acceptance passed."
  ],
  "restoration": {
    "result": "pass",
    "primary": "official FP8 TP=1 393K MTP=3",
    "multimodal": "official BF16 TP=1 393K MTP=3 with existing 32-image ceiling",
    "router_aliases": ["llm.primary", "vision.general", "vision.ocr"],
    "all_tiers_admitting": true,
    "shared_memory_reclaimable_bytes": 0
  }
}