---
license: other
license_name: nvidia-cuda-toolkit-eula
license_link: https://docs.nvidia.com/cuda/eula/index.html
tags:
  - cuda
  - cudnn
  - windows
  - runtime
  - animhaven
---

# AnimHaven CUDA Runtime

NVIDIA **CUDA Toolkit + cuDNN runtime DLLs** for the AnimHaven app
family ([AnimStart](https://github.com/Initnatsnok/animstart), AnimRoom).
Distributed separately from the application installers so the base
download stays small (~150 MB) and only NVIDIA-GPU users pull these
~2.85 GB of native libraries on demand.

---

## What's in this repo

A flat directory of Windows `.dll` files comprising:

| Family | Files | Approx. size |
|---|---|---:|
| **CUDA Toolkit** | `cublas64_12.dll`, `cublasLt64_12.dll`, `cudart64_12.dll` | ~548 MB |
| **cuDNN** | `cudnn*.dll` (engines, graph, ops, heuristic) | ~947 MB |
| **CUDA Math** | `cufft64_11.dll`, `cufftw64_11.dll`, `curand64_10.dll` | ~350 MB |
| **NVRTC** | `nvJitLink_120_0.dll`, `nvrtc*.dll` | ~262 MB |
| **CUDA engine bindings** | `ggml-cuda.dll` (from llama.cpp build), `onnxruntime_providers_cuda.dll` | ~741 MB |
| **Total** | ~20 files | **~2.85 GB** |

All files are NVIDIA's redistributable binaries vendored from the CUDA
Toolkit installer + cuDNN release. No modifications.

---

## Usage

**This repo is consumed by the in-app installer** — you typically don't
download it directly. AnimStart's SetupWizard ("GPU acceleration"
step) and Settings → Advanced → "GPU acceleration" both pull these
files via HuggingFace LFS into `<APPDATA>/AnimHaven/cuda_runtime/`,
verify SHA256, and register the directory with the Windows DLL loader
via `AddDllDirectoryW`.

The manifest (file paths + sizes + SHA256) is **baked into the
AnimStart binary** at compile time via
`include_str!("anim-shared/cuda-runtime-manifest.json")`, so the
installer has zero network round-trips on top of the file downloads
themselves.

### Direct download (advanced)

If you need to manually populate `cuda_runtime/`, e.g. for offline
provisioning of a kiosk machine, clone the repo with Git LFS:

```bash
git lfs install
git clone https://huggingface.co/Animhaven/cuda-runtime
# Copy contents of cuda-runtime/ to %APPDATA%/AnimHaven/cuda_runtime/
```

---

## License

The DLLs in this repo are NVIDIA software, redistributed under their
respective EULAs:

- **CUDA Toolkit components** (`cublas*.dll`, `cudart*.dll`, `cufft*.dll`,
  `curand*.dll`, `nvJitLink*.dll`, `nvrtc*.dll`) — covered by the
  [NVIDIA CUDA Toolkit End User License Agreement](https://docs.nvidia.com/cuda/eula/index.html).
- **cuDNN components** (`cudnn*.dll`) — covered by the
  [NVIDIA cuDNN Software License Agreement](https://docs.nvidia.com/deeplearning/cudnn/sla/index.html).
- **`ggml-cuda.dll`** — derived from llama.cpp (MIT) compiled against
  the CUDA Toolkit.
- **`onnxruntime_providers_cuda.dll`** — derived from ONNX Runtime
  (MIT) compiled against CUDA.

The AnimStart in-app installer surfaces both NVIDIA EULAs as direct
links and treats a click on "Install" as user acceptance (Soft-EULA
model). The acceptance is persisted per-installation so the dialog
doesn't re-prompt.

---

## Versioning

The repo tracks a single rolling pack on `main`. When the CUDA Toolkit
or cuDNN version changes, the entire pack is re-uploaded — git history
preserves prior versions for any downstream clients that pinned a SHA
in their manifest.

Current version is encoded in the `manifest.json` baked into the
AnimStart binary (field `version`, e.g. `v12.4-cudnn9`).

---

## Why this pack instead of asking users to install CUDA?

1. **One-click UX** — typical users don't know what CUDA Toolkit is or
   which version AnimStart wants. Installing the full toolkit (~3 GB
   developer tools they don't need) is wasteful.
2. **Version isolation** — different apps using different CUDA versions
   conflict in `PATH`. By keeping our copy in `<APPDATA>/AnimHaven/`
   and registering only for our process, we don't pollute the
   system-wide DLL search order.
3. **Exact ABI match** — the `ggml-cuda.dll` and
   `onnxruntime_providers_cuda.dll` shipped here are compiled against
   the specific CUDA Toolkit version in this repo. A user-supplied
   different version would either fail to link or produce subtle
   inference bugs.