uv selects an unavailable onnxruntime version (1.20.1) despite it not existing on PyPI for the given platform

[markovejnovic]

Summary

Preface

Note: I have observed the following issue with poetry too. Considering I believe uv shares this issue with poetry, I am opening a ticket with uv too. This ticket will resemble the aforementioned python-poetry/poetry#10151

uv appears to select an onnxruntime version which is incompatible with my system. With a very minimal pyproject.toml:

[project]
name = "bugs"
version = "0.1.0"
description = ""
authors = []
requires-python = ">=3.9,<4.0"
dependencies = [
    "onnxruntime",
]

uv selects the 1.20.1 onnxruntime package. This is, however, incompatible with my system, as evidenced by pip index and I would expect a 1.19.2:

$ uv run pip index versions onnxruntime
WARNING: pip index is currently an experimental command. It may be removed/changed in a future release without prior warning.
onnxruntime (1.19.2)
Available versions: 1.19.2, 1.19.0, 1.18.1, 1.18.0, 1.17.3, 1.17.1, 1.17.0, 1.16.3, 1.16.2, 1.16.1, 1.16.0, 1.15.1, 1.15.0, 1.14.1, 1.14.0, 1.13.1, 1.12.1, 1.12.0, 1.11.1, 1.11.0, 1.10.0, 1.9.0, 1.8.1, 1.8.0, 1.7.0

I have attached the uv.lock file. Notice

[[package]]
name = "onnxruntime"
version = "1.20.1"

I've installed uv as per the recommended way via the install script.

Platform

Rocky Linux 9.4 (Blue Onyx)

Version

uv 0.5.29

Python version

Python 3.9.18

[charliermarsh]

You can solve this with:

[project]
name = "bugs"
version = "0.1.0"
description = ""
authors = []
requires-python = "==3.9.*"
dependencies = [
    "onnxruntime",
]

[markovejnovic]

Perhaps, I'll give it a shot when I'm at my computer. That being said, it does feel wrong that uv picked a version of the package that is incompatible with the current working interpreter.

Also I would like to avoid pinning the python version in my project.

[charliermarsh]

Of course, but it's inherently a fairly hard problem. The package doesn't publish source distributions and has limited platform support. It's impossible to come up with a resolution that covers all possible platforms. This is a more narrow case, but that's why it isn't working here. We see a version that has support for at least one of the platforms you care about (i.e., platform support for Python 3.10 onward), and we choose that.

[charliermarsh]

(I agree the experience here should be better; just trying to explain what's going on and why it's difficult. We also don't privilege the currently-running interpreter at all, which I think is correct -- we're solving generically, ignoring your platform, Python version, etc., but also contributes here.)

[charliermarsh]

I kind of expected this to work:

[project]
name = "bugs"
version = "0.1.0"
description = ""
authors = []
requires-python = ">=3.9,<4.0"
dependencies = [
    "onnxruntime",
]

[tool.uv]
environments = [
    "python_version == '3.9'",
    "python_version > '3.9'",
]

Which is a nicer solution in that you don't have to change your supported Python range. But it's not working as expected, so looking into that.

[markovejnovic]

Okay so I don't think this is a uv issue. after discussing with the maintainers of poetry, I've learned that poetry relies on the python-requires value in onnxruntime's project declaration rather than which packages are actually available for my python distribution.

uv likely does the same and that seems like a natural and good behavior.

I'm closing this ticket consequently. Please refer to the poetry ticket for the exact discussion.

[markovejnovic]

@dimbleby has kindly opened a PR to amend the issue on onnxruntime's side microsoft/onnxruntime#23604