| --- |
| pretty_name: Mirid Runtime Packs |
| license: other |
| task_categories: |
| - text-generation |
| - text-to-image |
| tags: |
| - mirid |
| - llama.cpp |
| - stable-diffusion.cpp |
| - local-inference |
| - vulkan |
| - rocm |
| - metal |
| --- |
| |
| # Mirid Runtime Packs |
|
|
| Verified, independently updateable local-inference packages used by [Mirid](https://mirid.ai). |
|
|
| Each entry in `runtime-packages.manifest.json` records the upstream project and release, original download URL, byte size and SHA-256 digest. Files mirrored from upstream are stored without modification. Mirid-built Python wheels additionally record their source archive, build arguments and validation result. |
|
|
| ## Layout |
|
|
| - `runners/text/`: private GGUF inference runners from `llama.cpp`. |
| - `runners/image/`: private image inference runners from `stable-diffusion.cpp`. |
| - `bindings/`: Python wheels retained for Mirid's embedded compatibility path. |
|
|
| CPU packages are the universal baseline. Vulkan, HIP, ROCm and Metal packages are optional acceleration paths selected only after a local device probe succeeds. |
|
|
| The Apple Silicon binding set includes `mlx`, `mlx-metal` and `mlx-lm` for Python 3.12 on macOS 14 or newer. Their ordinary pure-Python dependencies should still be resolved by the prepared Mirid environment. |
|
|
| ## Trust and licences |
|
|
| These packages are redistributions or builds of the named upstream projects. Their original licences remain authoritative: |
|
|
| - [`ggml-org/llama.cpp`](https://github.com/ggml-org/llama.cpp) |
| - [`abetlen/llama-cpp-python`](https://github.com/abetlen/llama-cpp-python) |
| - [`leejet/stable-diffusion.cpp`](https://github.com/leejet/stable-diffusion.cpp) |
| - [`william-murray1204/stable-diffusion-cpp-python`](https://github.com/william-murray1204/stable-diffusion-cpp-python) |
|
|
| `source-sha256-verified` means Mirid matched the published size and SHA-256 after downloading the file. `import-passed` means a Mirid-built wheel was also installed into a clean environment and imported successfully. Neither label claims that every accelerator has been exercised on physical hardware; Mirid's release notes record those tests separately. |
|
|
