{
  "model": {
    "repository": "satgeze/GLM-5.3-Flash-EXL3-TR3-3.5bpw",
    "revision": "62587d6015184a26773a4ed1b751a9c5fa469cd6"
  },
  "served_model": "glm53-mixed35-v84-nospec-scout",
  "runtime": {
    "image_digest": "sha256:0403cbbc83450b2ca7d4fdeabe58bc8daf785872b12ebda1f1fb32a1bfba2fb4",
    "base_image_digest": "sha256:0f1cdcc8891f1cc3a444121eb61d366289a1cbba285f0892dcbb24bc94961692",
    "overlay_revision": "8fc95f0da72072a49a697f2164410b851a4e7377",
    "reported_version": "0.1.dev20111+g7f1e92bec.d20260827",
    "overlay_files": [
      "w4a16/host.py",
      "w4a16/kernel.py",
      "w4a16/mixed_trellis.py",
      "fused_moe/_impl.py",
      "vllm/exl3.py"
    ]
  },
  "hardware": {
    "host": "Primary Node",
    "os": "native Linux",
    "gpus": "2x RTX PRO 6000 Blackwell Max-Q 96 GB",
    "host_ram": "96 GB installed; approximately 89.7 GiB OS usable",
    "swap_gib": 8
  },
  "configuration": {
    "tp": 2,
    "ep": false,
    "dcp": 1,
    "quantization": "mixed K3/K4 EXL3",
    "kv": "nvfp4_ds_mla",
    "attention": "B12X_MLA_SPARSE",
    "context_tokens": 327680,
    "scheduler_concurrency": 1,
    "speculation": false,
    "prefill_block_m": 64,
    "max_num_batched_tokens": 2048,
    "gpu_memory_utilization": 0.97,
    "prefix_caching": false,
    "nccl_p2p_disabled": true,
    "host_memory_limit": null
  },
  "scope": "single startup attempt; no inference requests; co-resident desktop and benchmark client",
  "launcher": {
    "repository_revision": "fb30405d81272b13b57f62db7a537bafb2372b2d",
    "version": "1.1.0",
    "source": "anvil_serving/cli.py",
    "dirty_state": "clean at candidate startup; evidence authored afterwards"
  },
  "launch": {
    "managed_recipe": "candidate-recipe.toml",
    "captured_recipe_digest": "b42652030185356b3590fab3d6c24ebc9ddce8e12ddb9b82b5fe42a25351e4bb",
    "captured_registry_digest": "c0330d2dda535ddb327c121aefc37c8c8355e04a1a91064a9100eaa6dab43236",
    "endpoint": "http://127.0.0.1:30001/v1",
    "gpu_selection": [
      "GPU-example-1",
      "GPU-example-2"
    ],
    "redaction": "Endpoint port and GPU identities are synthetic; captured digests fingerprint the original private recipe, not the redacted TOML."
  }
}