uv add --index fails to install PyTorch CUDA packages on Windows

[AbyssSkb]

Question

Is there a way to directly install the CUDA version of PyTorch via the command line on Windows and support project dependency management with uv?

I tried installing PyTorch using uv add torch --index on my computer, but it ultimately failed. The documentation doesn't seem to mention this. Why does uv add torch --index fail to install PyTorch?

Although I can install PyTorch by modifying the pyproject.toml file as mentioned in the documentation, I want to know if it is possible to install it directly through the command line (which is simpler and more familiar).

> uv add torch --index https://download.pytorch.org/whl/cu126
Resolved 16 packages in 4.52s
error: Distribution `markupsafe==3.0.2 @ registry+https://download.pytorch.org/whl/cu126` can't be installed because it doesn't have a source distribution or wheel for the current platform

hint: You're using CPython 3.10 (`cp310`), but `markupsafe` (v3.0.2) only has wheels with the following Python implementation tag: `cp313`

Platform

Windows 11 x86_64

Version

uv 0.6.0 (591f38c 2025-02-14)

[charliermarsh]

Unfortunately I don't think there's a way to avoid this right now due to how the PyTorch index is structured... They have a markupsafe wheel, but only for Linux. By default, if we see a package (like markupsafe) on an index that the user provided (like the PyTorch index), we avoid looking at other indexes for the same package, as a security concern.

If we supported a --constraints flag here, you could do --constraints markupsafe<3.0.2, but even that isn't great. Hmm.