Using automatic deps management

This page explains how to use RAITAP's automatic dependency management. It lists hardware gotchas and useful flags.

Installing the deps

RAITAP will automatically analyse your config and install the deps when you run the job.

uv

uv run raitap --config-dir my-configs --config-name assessment

pip

raitap --config-dir my-configs --config-name assessment

Useful flags

Depending on your setup (uv, pip, ...) you might need to pass additional flags to make the install fully automatic.

Flag	When
--dry-run	Preview the inferred install
--allow-project-edit / -y	Let RAITAP modify your pyproject.toml and run the uv install
--exec-global	Install into the global Pip environment (not recommended)

Hardware gotchas

If you use a GPU, you will need to tweak your pyproject.toml. This is due to PyPI and ecosystem limitations.

Important

Replace BACKEND in the snippets below with torch or onnx depending on your setup.

CPU

On Linux, the plain torch package from PyPI is the CUDA build. Route it to the CPU index by adding the following to your pyproject.toml:

[project]
dependencies = [
  "raitap[BACKEND-cpu]>=0.9",
  "torch;       sys_platform == 'linux'", # redeclare to make the index below apply
  "torchvision; sys_platform == 'linux'", # redeclare to make the index below apply
]

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

[tool.uv.sources]
torch = { index = "pytorch-cpu" }
torchvision = { index = "pytorch-cpu" }

CUDA

torch does not publish separate CUDA builds on PyPI. You control the version via the index by adding the following to your pyproject.toml:

[project]
dependencies = [
  "raitap[BACKEND-cuda]>=0.9; sys_platform != 'darwin'",
  "torch;       sys_platform != 'darwin'", # redeclare to make the index below apply
  "torchvision; sys_platform != 'darwin'", # redeclare to make the index below apply
]

[[tool.uv.index]]
name = "pytorch-cuda"
url = "https://download.pytorch.org/whl/cu126" # CUDA 12.6, modern default; for older GPUs see below
explicit = true

[tool.uv.sources]
torch = { index = "pytorch-cuda" }
torchvision = { index = "pytorch-cuda" }

Builds for older NVIDIA GPUs (Volta / V100)

The cu126 wheels may not suit older cards.

1. Find your GPU generation on NVIDIA's current / legacy pages and a matching release on the PyTorch releases page.

2. Set the index url in your pyproject.toml to the matching family (see above), then run uv lock.

3. Run RAITAP.

Intel GPU

triton-xpu's PyPI releases are either yanked or pre-release. Route it to the Intel index by adding the following to your pyproject.toml:

[project]
dependencies = [
  "raitap[BACKEND-intel]>=0.9; sys_platform != 'darwin'",
  "torch;       sys_platform != 'darwin'", # redeclare to make the index below apply
  "torchvision; sys_platform != 'darwin'", # redeclare to make the index below apply
  "triton-xpu>=3.0.0rc0; sys_platform != 'darwin' and python_full_version < '3.14'",
]

[[tool.uv.index]]
name = "pytorch-intel"
url = "https://download.pytorch.org/whl/xpu"
explicit = true

[tool.uv.sources]
torch = { index = "pytorch-intel" }
torchvision = { index = "pytorch-intel" }
triton-xpu = { index = "pytorch-intel" }

Wheel availability

PyTorch indexes don't publish for every Python minor:

Extra	Index	Python	Platforms
torch-cpu / onnx-cpu	/whl/cpu	3.11–3.13	Linux, macOS, Windows
torch-cuda / onnx-cuda	/whl/cu126	3.11–3.13	Linux, Windows
torch-intel / onnx-intel	/whl/xpu	3.11–3.13	Linux, Windows

Faster locking on one OS

If you only target one OS, scope the resolver so it doesn't solve for all platforms at once:

[tool.uv]
environments = ["sys_platform == 'linux' and python_version >= '3.12'"]