ardge/llama.cpp repository overview

Unofficial mirror of ghcr.io/ggml-org/llama.cpp, published for hosts that cannot reach ghcr.io. Not an official release.

• Upstream project: https://github.com/ggml-org/llama.cpp⁠ (the ggml authors)

• Mirrored tags: mirrored-server-cuda13-v<version> (NVIDIA), mirrored-server-vulkan-v<version> (any Vulkan GPU), mirrored-server-intel-v<version> (Intel, amd64 only) and mirrored-server-rocm-v<version> (AMD ROCm, amd64 only) — each one pins whatever upstream's moving tag of the same name, minus mirrored-, pointed at when that version was released, and is never repointed afterwards

• The images are byte-identical to upstream's, including the manifest digest. Each release records the upstream digest it copied; that digest is what a mirrored tag must still report:

  docker buildx imagetools inspect ardge/llama.cpp:mirrored-server-cuda13-v<version> --format '{{.Manifest.Digest}}'
  

  Copy

  Comparing against the upstream tag only matches until upstream moves it, which it does on every release.

llama.cpp is licensed under the MIT License, Copyright (c) 2023-2026 The ggml authors — https://github.com/ggml-org/llama.cpp/blob/master/LICENSE⁠

The images also bundle their base images' GPU runtimes, which keep their own licenses: NVIDIA CUDA (NVIDIA Deep Learning Container License), Intel oneAPI, AMD ROCm, and Mesa's Vulkan drivers.