{
  "schema_version": 1,
  "image_digest": "sha256:0f1cdcc8891f1cc3a444121eb61d366289a1cbba285f0892dcbb24bc94961692",
  "hardware": {
    "gpus": "2x RTX PRO 6000 Blackwell Workstation Edition 96 GB",
    "gpu_ids": [1, 3],
    "power_limit_w_each": 300
  },
  "kv_cache_dtype": "nvfp4_ds_mla",
  "profiles": [
    {
      "name": "multimodal_dflash2_7",
      "vision_enabled": true,
      "speculator": "DFlash2-7 external draft",
      "max_model_len": 98304,
      "available_kv_cache_gib": 5.68,
      "gpu_kv_cache_tokens": 129473,
      "max_concurrency": 1.32
    },
    {
      "name": "language_only_dflash2_7",
      "vision_enabled": false,
      "speculator": "DFlash2-7 external draft",
      "max_model_len": 98304,
      "available_kv_cache_gib": 8.14,
      "gpu_kv_cache_tokens": 184619,
      "max_concurrency": 1.88
    },
    {
      "name": "language_only_mtp3",
      "vision_enabled": false,
      "speculator": "built-in MTP3",
      "max_model_len": 131072,
      "available_kv_cache_gib": 6.49,
      "gpu_kv_cache_tokens": 1376256,
      "max_concurrency": 10.50
    }
  ],
  "interpretation": {
    "vision_off_dflash_token_gain_pct": 42.59,
    "mtp3_vs_dflash_language_only_token_ratio": 7.45,
    "note": "NVFP4 MLA physical layouts differ between the external DFlash draft and built-in MTP profile. Compare supported token capacity, not GiB alone."
  }
}