{
  "_comment": [
    "Pinned llama.cpp builds, fetched and verified by distinct_agent/runtime.py.",
    "A worker downloads the build recorded here and refuses to install it unless",
    "the bytes hash to `sha256`.",
    "",
    "KEYS. `windows-x64` is the build that runs on any Windows x64 machine, and",
    "is the fallback every machine can reach. A key with a suffix is an",
    "accelerated build offered only to a machine that has been checked for the",
    "hardware: `windows-x64-cuda` needs an NVIDIA driver advertising at least",
    "`requires_cuda`, which the worker reads from NVML. If it will not start,",
    "the worker steps down to the plain key rather than leaving a volunteer",
    "with a worker that does not boot.",
    "",
    "ARCHIVES. An accelerated build is not self-contained: llama.cpp ships the",
    "CUDA binaries in one archive and the CUDA runtime libraries they link",
    "against in another. Both are listed, both are verified, and either both",
    "install or neither does.",
    "",
    "A platform with no entry gets no automatic download. That is deliberate:",
    "this project already withholds a MODEL whose digest it cannot check, and",
    "the program that executes those models is held to the same bar. Operators",
    "on those platforms pass --llama-server, or run --demo-runner and offer the",
    "library's tools and skills with no model at all.",
    "",
    "TO ADD OR MOVE A PIN: download the asset, hash it, and record the result",
    "here. `scripts/record_llama_pin.py` does exactly that and prints the JSON;",
    "`scripts/apply_llama_pin.py` merges what it printed into this file. Never",
    "copy a digest from a release page or an API response: a digest that arrived",
    "with the download proves only that the bytes were not corrupted in transit,",
    "which TLS already said, and pinning it would dress trust-on-first-use up as",
    "verification."
  ],
  "builds": {
    "windows-x64": {
      "tag": "b10742",
      "accelerator": "cpu",
      "archives": [
        {
          "asset": "llama-b10742-bin-win-cpu-x64.zip",
          "url": "https://github.com/ggml-org/llama.cpp/releases/download/b10742/llama-b10742-bin-win-cpu-x64.zip",
          "sha256": "a923d80953d618335ae0073233fdcfb93760dfde646e957786894259aba87d72",
          "bytes": 18373032
        }
      ]
    },
    "windows-x64-cuda": {
      "tag": "b10742",
      "accelerator": "cuda",
      "requires_cuda": 12040,
      "archives": [
        {
          "asset": "llama-b10742-bin-win-cuda-12.4-x64.zip",
          "url": "https://github.com/ggml-org/llama.cpp/releases/download/b10742/llama-b10742-bin-win-cuda-12.4-x64.zip",
          "sha256": "2249c7e6337ea02a7f0c0832a2826e5a6de28b87f216427e746e50e86b87bba4",
          "bytes": 254220199
        },
        {
          "asset": "cudart-llama-bin-win-cuda-12.4-x64.zip",
          "url": "https://github.com/ggml-org/llama.cpp/releases/download/b10742/cudart-llama-bin-win-cuda-12.4-x64.zip",
          "sha256": "8c79a9b226de4b3cacfd1f83d24f962d0773be79f1e7b75c6af4ded7e32ae1d6",
          "bytes": 391443627
        }
      ]
    }
  }
}