GPU/CUDA support - Windows solution

[archasek]

If anybody needs to test it on Windows, this is what worked for me.

1. pyproject.toml:

[[tool.poetry.source]]
name = "pytorch-cuda"
url = "https://download.pytorch.org/whl/cu121"
priority = "explicit"

[tool.poetry.group.windows-gpu]
optional = true

[tool.poetry.group.windows-gpu.dependencies]
torch = [
    {version = ">=2.5.1+cu121", source = "pytorch-cuda", markers = 'sys_platform == "win32"'},
]
torchvision = [
    {version = ">=0.20.1+cu121", source = "pytorch-cuda", markers = 'sys_platform == "win32"'},
]

2. poetry install --with windows-gpu --extras ui

[archasek]

@dolfim-ibm do you think it's worth adding to the repo?

[dolfim-ibm]

Managing the extra indexes (specially the torch versions) with poetry seems to create more issues. This is why in #48 I'm actually migrating to uv.

We could add the Windows extra there, but I think this won't needed anymore once we make the project consumable directly via pip install docling-serve (see #47 ).
The target is to simplify and don't require the actual clone of the repo, if not for developments.