# transcribe-cpp-native-cu12

CUDA 12 native provider for
[`transcribe-cpp`](https://pypi.org/project/transcribe-cpp/), the Python
bindings for [transcribe.cpp](https://github.com/handy-computer/transcribe.cpp).

Install via the extra — it is **additive**, alongside the default provider:

```sh
pip install "transcribe-cpp[cu12]"
```

At runtime the best installed provider wins (CUDA outranks Vulkan/CPU); force
one with `TRANSCRIBE_NATIVE_PROVIDER=cu12` or per-model
`Model(..., backend="cuda")`.

What's inside: `libtranscribe` + ggml backend modules (conservative CPU +
CUDA with fatbins for sm 75/80/86/89/90 + Vulkan). Vulkan is bundled on
purpose: this provider outranks the default one when both are installed, so
without it a `[cu12]` install deployed onto a non-NVIDIA machine (AMD/Intel
GPU) would silently lose Vulkan acceleration. With it, this artifact is a
strict superset of the default provider — CUDA wins on NVIDIA boxes, Vulkan
picks up everywhere else, CPU is the floor.

The CUDA toolkit is not vendored — cudart/cublas come from the
`nvidia-*-cu12` runtime wheels this package depends on, and `libcuda.so.1`
is your system driver. On a machine without an NVIDIA driver the CUDA module
simply doesn't load and Vulkan/CPU keep working.

Linux x86_64 and Windows x64 wheels. Building from source with CUDA goes
through the `transcribe-cpp-native` sdist:

```sh
CMAKE_ARGS="-DTRANSCRIBE_CUDA=ON -DTRANSCRIBE_VULKAN=ON -DTRANSCRIBE_GGML_BACKEND_DL=ON" \
    pip install --no-binary :all: transcribe-cpp-native
```