The pyproject.toml file does not include an extra optional dependency for torch+CUDA.

You can replace cu128 in the --index-url to whatever version you like and that is also supported.
RTX 50 series / Blackwell requires at least cu128 (I think?) and torch 2.7.x seems to be the earliest pytorch version to come with cuda 12.8 builds

pip install git+https://github.com/DraconicDragon/vibe \
  "torch>=2.7.1" \
  "safetensors>=0.6.2" \
  "timm>=1.0.22" \
  "transformers>=5.0.0" \
  "einops" \
  --index-url https://download.pytorch.org/whl/cu128 \
  --extra-index-url https://pypi.org/simple

Note: For Maxwell/Pascal GPUs you made need to replace cu128 with cu126 or cu124

Note: Don't install more than one onnxruntime variant (onnxruntime, onnxruntime-gpu, onnxruntime-rocm, etc.) in the same environment. They overwrite each other's entry points and the conflict won't be reported as an error and will just silently break (for example, if you had onnxruntime-gpu installed at first, then install onnxruntime - you will only have access to the execution providers from onnxruntime and need to uninstall and install onnxruntime-gpu again to be able to access CUDA EP and TensorRT EP) If you encounter any weirdness, try to uninstall any onnxruntime package you have and then just install the one you want to have

Windows:

pip install "git+https://github.com/DraconicDragon/vibe[onnx-cuda]"

Linux + System level CUDA/cuDNN / PyTorch + CUDA:

This is a bit weird for me but as far as I understood CUDA in relation to linux and onnxruntime:
If torch+cuda is installed it should already come with the required nvidia packages and this lib will try to search for those in the same environemnt (works if torch's cuda version is the one that onnxruntime expects, usually cu12x). Otherwise you will need to install CUDA/cuDNN libraries from your system package manager, or through pip packages - info for that further below.

pip install "git+https://github.com/DraconicDragon/vibe[onnx-cuda]"

Note: If your CUDA/cuDNN is version 13.x, you need onnxruntime-gpu to be v1.24.1 or higher - pyproject.toml pins the version to '>=1.17.3' so it should automatically install 1.24.1 or higher, but to be sure you can do pip install "git+https://github.com/DraconicDragon/vibe[onnx-cuda-cu13]"

Linux, no torch/system CUDA:

Pulls CUDA libraries from PyPI
You will likely need this if you are on linux, and do not have torch+cuda installed or CUDA/cuDNN installed through your package manager.

pip install "git+https://github.com/DraconicDragon/vibe[onnx-cuda-linux-bundled-cu12]"

Note: If you want to use Cuda 13.x, do pip install "git+https://github.com/DraconicDragon/vibe[onnx-cuda-linux-bundled-cu13]"