{
  "schema": "anvil-serving.campaign-configuration/v1",
  "classification": "configured values; startup memory and correctness live in separate native evidence",
  "source": "selected managed recipes; public-safe field extraction omits operator paths, endpoint addresses and device identifiers",
  "hardware": {
    "cpu": "Intel Xeon 654",
    "physical_cores": 18,
    "threads": 36,
    "installed_ram_gib": 128,
    "gpu_count": 2,
    "gpu": "RTX PRO 6000 Blackwell Max-Q",
    "vram_mib_per_gpu": 97887,
    "loaded_link_observation": "hardware-loaded-link.json"
  },
  "cells": [
    {
      "id": "baseline-r11",
      "model_repository": "brandonmusic/GLM-5.3-Flash-tr3-4bpw",
      "model_revision": "a5fee929cf4888b1824323e33e8a19b60129e025",
      "runtime_image": "verdictai/glm53-flash-exl3-k4:r19-sm120-tp2-ep2-dcp2-v84-language-only@sha256:0f1cdcc8891f1cc3a444121eb61d366289a1cbba285f0892dcbb24bc94961692",
      "context_tokens": 327680,
      "quantization": "uniform K4 EXL3/TR3 routed experts; native-precision non-routed tensors; FP8 DS-MLA KV; speculation disabled; vision enabled",
      "host_memory_limit_mib": 53248,
      "host_ram_plus_swap_limit_mib": 53249,
      "host_memory_reserve_mib": 12288,
      "prefix_caching": "enabled in managed recipe",
      "speculation": "disabled",
      "tensor_parallel": 2,
      "decode_context_parallel": 1,
      "max_sequences": 4,
      "batch_tokens": 2048,
      "configured_output_tokens": 65536,
      "image_limit": 8,
      "video_limit": 0,
      "kv_bytes_per_rank": null
    },
    {
      "id": "qwen-step5500",
      "model_repository": "local-inference-lab/Qwen3.8-Flash-Next-NVFP4",
      "model_revision": "6909a5bed089a48fa07e956d3915af2537de9368",
      "runtime_image": "ghcr.io/local-inference-lab/vllm:karmic-kraken-beta-20261005-e1ccbaf0fc0f4f4c@sha256:6008f020c23065bef71390194f968af9f5ebbedf57807b1659f876632e4a1067",
      "context_tokens": 262144,
      "quantization": "QAD NVFP4 routed experts and PLE; MXFP8 attention/shared experts; native profile KV to verify; FP32 recurrent state; no speculation; GPU PLE",
      "host_memory_limit_mib": 53248,
      "host_ram_plus_swap_limit_mib": 53249,
      "host_memory_reserve_mib": 12288,
      "prefix_caching": "enabled in managed recipe",
      "speculation": "disabled",
      "tensor_parallel": 2,
      "decode_context_parallel": 1,
      "max_sequences": 4,
      "batch_tokens": 6019,
      "configured_output_tokens": 65536,
      "image_limit": 8,
      "video_limit": 0,
      "kv_bytes_per_rank": 8589934592
    },
    {
      "id": "glm-csf",
      "model_repository": "local-inference-lab/GLM-5.3-Flash-NVFP4-MXFP8-CSF-QAD",
      "model_revision": "4f90b74165cb1542cca35730b7dfa4d550bcf0ce",
      "runtime_image": "ghcr.io/local-inference-lab/vllm:karmic-kraken-beta-20261005-e1ccbaf0fc0f4f4c@sha256:6008f020c23065bef71390194f968af9f5ebbedf57807b1659f876632e4a1067",
      "context_tokens": 327680,
      "quantization": "QAD NVFP4-CSF routed experts with MXFP8 attention/shared experts; FP8 KV; VRAM-only prefix cache; MTP disabled",
      "host_memory_limit_mib": 53248,
      "host_ram_plus_swap_limit_mib": 53249,
      "host_memory_reserve_mib": 12288,
      "prefix_caching": "enabled in managed recipe",
      "speculation": "disabled",
      "tensor_parallel": 2,
      "decode_context_parallel": 2,
      "max_sequences": 4,
      "batch_tokens": 4096,
      "configured_output_tokens": 65536,
      "image_limit": 8,
      "video_limit": 0,
      "kv_bytes_per_rank": 7650410496
    }
  ],
  "request_controls": "See native artifacts and workload-manifest.json; model-family reasoning profiles differ.",
  "baseline_output_source": "retained selected primary contract; baseline model generation-config is mounted separately"
}