Installation

This page explains how to install RAITAP from PyPI and how to select the optional dependency groups you need.

Use uv add if you are working inside a managed Python project (pyproject.toml exists). Use uv pip install or pip install otherwise.

1. Install RAITAP

uv

uv add raitap

pip

pip install raitap

2. Install optional dependencies

Execution dependencies

RAITAP supports both PyTorch and ONNX models, and both CPU and GPU execution. To avoid conflicts, only install the dependencies that match your setup.

First, choose the right group from this table:

	CPU	CUDA	Intel GPU
Torch	torch-cpu	torch-cuda	torch-intel
ONNX	onnx-cpu	onnx-cuda	onnx-intel

Then, adapt the following example command and run it:

uv

uv add "raitap[onnx-cpu]" # replace `onnx-cpu` with your group

pip

pip install "raitap[onnx-cpu]" # replace `onnx-cpu` with your group

Note

• CUDA corresponds to NVIDIA GPUs.

• torch-intel uses the Intel XPU API directly.

• onnx-intel uses the OpenVINO ONNX Runtime.

• Apple MPS support is coming soon.

Assessment dependencies

You can then install the dependencies for the assessment modules you want to use.

For instance, if you want to assess the model’s transparency, run:

uv

uv add "raitap[transparency]"

pip

pip install "raitap[transparency]"

If you plan to use a single underlying framework (here Captum), you can run the following instead:

uv

uv add "raitap[captum]"

pip

pip install "raitap[captum]"

Note

Alibi requires dependency overrides before it can be installed. See Alibi (transparency) for the exact steps.

Combine multiple extras

Of course, such optional dependency groups can be combined. For instance:

uv

uv add "raitap[onnx-cpu,transparency]"

pip

pip install "raitap[onnx-cpu,transparency]"