{
  "schema": "anvil-serving.sanitized-benchmark-summary/v1",
  "date": "2026-08-15",
  "subject": "Qwen3.8 27B SGLang BF16 versus quantized multimodal consolidation A/B",
  "decision": {
    "current": "unchanged official FP8 text plus official BF16 multimodal vLLM split",
    "preferred_single_service_candidate": "official-fp8",
    "lowest_latency_candidate": "inferact-nvfp4",
    "promotion": "no-promotion"
  },
  "hardware": {
    "product": "NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition",
    "count": 2,
    "vram_gb_each": 96,
    "topology": "split TP=1; BF16 fixed on one card and quantized candidates sequential on the other"
  },
  "runtime": {
    "image": "lmsysorg/sglang@sha256:506525a5907ea22c9d445afb7c03603959b912de034d86915cf17da814f1a124",
    "engine_revision": "c4271c3fe1262fc2adbd162c33b25de5255251c5",
    "cookbook_revision": "dd458f3212dd4ddf0e1a7907bbf539b660e70d21"
  },
  "configuration": {
    "tensor_parallel": 1,
    "context_tokens": 393216,
    "max_running_requests": 1,
    "kv_cache": "fp8_e4m3",
    "memory_fraction_static": 0.85,
    "attention_backend": "flashinfer",
    "chunked_prefill_tokens": 2048,
    "radix_cache": false,
    "mtp": {
      "algorithm": "EAGLE",
      "steps": 3,
      "top_k": 1,
      "draft_tokens": 4,
      "gdn_state_slots": 5
    },
    "multimodal_feature_transport": "cpu",
    "thinking_mode": "disabled"
  },
  "multimodal_protocol": {
    "corpus_sha256": "ac06eb4c295e1fbd91d582bb1330ea55a70ea775f291e2c323fd4fc4f00c45db",
    "cases": 6,
    "repetitions_per_case": 3,
    "attempts_per_model": 18,
    "concurrency": 1,
    "max_images_per_request": 2,
    "max_tokens": 1024,
    "case_types": [
      "scene",
      "ocr",
      "chart",
      "ui",
      "spatial-count",
      "two-image-comparison"
    ]
  },
  "text_protocol": {
    "runs_per_model": 3,
    "requests_per_run": 10,
    "concurrency": 1,
    "configured_input_tokens": 4096,
    "max_output_tokens": 256
  },
  "results": [
    {
      "id": "official-bf16",
      "model": "Qwen/Qwen3.8-27B",
      "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
      "multimodal_passed": 18,
      "multimodal_attempts": 18,
      "multimodal_latency_p50_ms": 915.3,
      "multimodal_latency_p95_ms": 2722.9,
      "ttft_p50_ms": 910.4,
      "effective_prefill_p50_tok_s": 3966,
      "decode_p50_tok_s": 62.7,
      "e2e_p50_ms": 1596.5,
      "aggregate_output_tok_s": 27.0,
      "observed_device_memory_mib": "81729-82089"
    },
    {
      "id": "official-fp8",
      "model": "Qwen/Qwen3.8-27B-FP8",
      "revision": "017b9c7af6b5689d5dd426a76e0bc077eb5ca20a",
      "multimodal_passed": 18,
      "multimodal_attempts": 18,
      "multimodal_latency_p50_ms": 587.5,
      "multimodal_latency_p95_ms": 1885.5,
      "ttft_p50_ms": 576.8,
      "effective_prefill_p50_tok_s": 6261,
      "decode_p50_tok_s": 111.4,
      "e2e_p50_ms": 962.0,
      "aggregate_output_tok_s": 45.8,
      "observed_device_memory_mib": 83203
    },
    {
      "id": "inferact-nvfp4",
      "model": "Inferact/Qwen3.8-27B-NVFP4",
      "revision": "6128240ebaf4eaa7bad2b3d1c72c37d677c5f462",
      "multimodal_passed": 18,
      "multimodal_attempts": 18,
      "multimodal_latency_p50_ms": 447.9,
      "multimodal_latency_p95_ms": 1727.0,
      "ttft_p50_ms": 452.6,
      "effective_prefill_p50_tok_s": 7975,
      "decode_p50_tok_s": 97.7,
      "e2e_p50_ms": 918.3,
      "aggregate_output_tok_s": 50.8,
      "observed_device_memory_mib": 83537
    }
  ],
  "restoration": {
    "shared_memory_files": 0,
    "shared_memory_reclaimable_bytes": 0,
    "primary_identity": "passed",
    "omni_identity": "passed",
    "router_tiers_admitting": true,
    "routed_primary_sentinel": "passed",
    "routed_ocr_sentinel": "passed"
  },
  "limits": [
    "concurrency one only",
    "maximum two images per request",
    "no video",
    "no 32-image request",
    "no host-memory-pressure soak",
    "no broad vision-quality benchmark",
    "BF16 was not swapped across cards"
  ]
}