> ## Documentation Index
> Fetch the complete documentation index at: https://koharu.rs/llms.txt
> Use this file to discover all available pages before exploring further.

# Hardware and runtimes

> Check GPU support and understand what Koharu downloads and stores.

Koharu runs every model on one GPU and draws pages with it. It downloads the native libraries and model files that GPU needs; you do not need a CUDA or ROCm SDK installation to run a release.

## Device selection

The first time a model runs, Koharu picks the best GPU on the machine and keeps it for the session:

| Rank | GPU |
| - | - |
| 1 | Metal, or a discrete NVIDIA GPU |
| 2 | A discrete AMD GPU |
| 3 | An integrated NVIDIA GPU |
| 4 | An integrated AMD GPU |
| 5 | Any other GPU through Vulkan |

Within a rank, more memory wins, then a newer CUDA architecture. If the best GPU cannot be used, the run fails rather than falling back to the next one, so a model never runs somewhere you did not expect. There is no CPU-only mode: without a supported GPU, models do not run.

> [tab] NVIDIA
>
> CUDA 13.3 support is available on Windows and Linux. Koharu uses a CUDA device only when the
> driver reports CUDA 13.3 or newer and the GPU has compute capability 7.5 or newer, which means a
> Turing-class or newer GPU with an R610-series or newer driver. See NVIDIA's [driver and
> architecture
> matrix](https://docs.nvidia.com/datacenter/tesla/drivers/cuda-toolkit-driver-and-architecture-matrix.html)
> and [driver downloads](https://www.nvidia.com/en-us/drivers/).

> [tab] AMD
>
> ROCm 10.1 support is available on x86-64 Windows and Linux and depends on the exact GPU,
> operating system, and driver. Install a compatible [AMD driver](https://www.amd.com/en/support).
> Koharu downloads the ROCm runtime it needs. Check AMD's [compatibility
> matrix](https://rocm.docs.amd.com/en/docs-10.1.0/compatibility/compatibility-matrix.html) for
> platform requirements.

> [tab] Apple silicon
>
> Koharu uses Metal on Apple-silicon macOS.

> [tab] Vulkan
>
> Vulkan is the remaining option on Windows and Linux. Local language models and the diffusion
> inpainters run on the Vulkan GPU. libtorch has no Vulkan build, so detection, OCR, and LaMa run
> on the CPU on a Vulkan machine, which is slower.

> [!NOTE]
>
> The canvas, exports, and page thumbnails are drawn on the GPU as well, so Koharu needs a working
> graphics driver even to open a project.

## Downloads and storage

| Data | When it is downloaded |
| - | - |
| libtorch, llama.cpp, stable-diffusion.cpp, and CUDA or ROCm | The first time any model runs |
| Model weights, pinned to a Hugging Face commit | The first time that model loads |
| Fonts from Koharu's font dataset | When a text needs a face your system lacks |

Everything goes into one store: the `store` folder in Koharu's installation folder on Windows, and `koharu/store` under the operating system's cache folder on macOS and Linux. Downloads are staged and validated before they move into place, so an interrupted download is retried on the next use. After an update, older runtime versions are removed in the background; model files stay. See [data locations](https://koharu.rs/reference/data.md).

## Memory and model residency

The activity center shows downloads and processing runs. A model loads the first time a stage needs it and stays loaded for later runs. It unloads when the settings switch to another model.

Nothing is unloaded to make room: a model that does not fit fails, and the error names the free memory. Local language models shrink their context and move layers to system memory to fit; the diffusion inpainters place their weights in free GPU memory, then system memory.

Smaller local models and quantizations need less memory, with a possible quality trade-off. A model that fits on disk can still exceed device memory.

If a run fails at the first model, check network access to GitHub, PyPI, AMD's repository, and Hugging Face, update the driver, and retry once. Capture the full error if it repeats and follow [troubleshooting](https://koharu.rs/reference/troubleshooting.md).


## Related topics

- [Install Koharu](https://koharu.rs/installation.md)
- [Koharu documentation](https://koharu.rs/index.md)
- [Architecture](https://koharu.rs/development/architecture.md)
- [Formats and data](https://koharu.rs/reference/data.md)
- [Troubleshooting](https://koharu.rs/reference/troubleshooting.md)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.