uv sync and uv run --no-project end up in different version with same requirements

[ThomasHezard]

Hello

I had an issue with torch dependency. I am pretty sure the problem comes from torch and I opened an issue on that matter: pytorch/pytorch#141781

However, while investigating my issue I encountered a uv behaviour I do not fully understand.

The issue can be reproduced with the following very simple pyproject.toml:

[project]
name = "uv-torch-test"
version = "0.1.0"
requires-python = ">=3.8"
dependencies = [
    "torch",
]

If you uv sync with Python 3.8 (either using .python-version or --python 3.8 option), it fails with the following error

error: Distribution `torch==2.5.1 @ registry+https://pypi.org/simple` can't be installed because it doesn't have a source distribution or wheel for the current platform

I think this is expected given the issue described here: pytorch/pytorch#141781.

However, the following command succeeds and indicates that uv ends up installing torch 2.4.1 in this case

uv run --no-project --python 3.8 --with "torch" python -c "import torch; print(torch.__version__)"

In conclusion:

• One feature request:

  • Having an option to print the package versions in uv run --no-project the same way they are printed in uv sync would have helped me a lot. The only solution to print them today is to use the --verbose option which is extremely verbose. I use a lot the "uv run --no-project" command in my unit testing CI jobs, and adding the --verbose would make the test results and logs difficult to read, whereas I just need the package versions.

• Two questions:

  • Why does uv run --no-project --python 3.8 --with "torch" dependency resolution ends up choosing 2.4.1 and not 2.5.1 when 2.5.1 is identified as compatible with Python 3.8 by uv sync?
  • Would Introduce a --multi-version preference mode #8686 help avoid this behaviour? (really waiting for this one, it would help on several issues I encountered 😉 )

Thank you 🙏

[ThomasHezard]

Hello

Quick update after some additional testing.

pip install torch in a blank Python 3.8 environment installs 2.4.1, which indeed is the real last Python 3.8 compatible torch release, despite what the metadata of the version says. This leads me to two conclusions:

• pip does not respect its own interface, in the sense that it does not fail if a package metadata is not properly written (in that torch 2.5.0/2.5.1 case, the package does not provide the wheels or sdist for all supposedly compatible Python versions), which bugs me a lot as a developer (as well as the fact that this has been pushed on PyPI without any problem) 😱
• I am not sure about that, but I think what pip does under the hood is that it looks at the available wheels or sdist rather than looking at the requires_python specifications (I looked a bit at the pip sources but I did not find where this is done yet).

On the uv side:

• uv sync seems to resolve the dependencies by looking at the packages metadata, which totally makes sense to me, and the error I got definitely comes from a problem in the PyTorch releases and is therefore fully legit I think it is a good think uv reports it as an error.
• I still do not understand why uv run --no-project --python 3.8 --with "torch" does not fail. Does it use pip resolver under the hood?

In general, I wanted to say that uv sync resolver seems really excellent to me as its rigorous "fail-fast" behaviour helped me spot errors in packages metadata (mine or others) really easily, and write perfectly reliable pyproject.toml and lock files.

[charliermarsh]

So, I think uv run --no-project --python 3.8 --with "torch" will look for the first version of torch that has a wheel for Python 3.8, whereas uv sync will look for the first version of torch that's compatible with your project, and attempt to use that for all supported Python versions. In this case, it picks 2.5.1, since it has wheels for some supported Python versions. It's related to #5182.

[charliermarsh]

The difference is that uv sync is doing a "universal" solve, whereas uv run --with only has to solve for your current platform / environment, since it's not creating a universal lockfile.

[charliermarsh]

I think #8686 would indeed help though it's definitely a tricky case.

[charliermarsh]

(Happy to answer any further questions.)

[ThomasHezard]

The difference is that uv sync is doing a "universal" solve, whereas uv run --with only has to solve for your current platform / environment, since it's not creating a universal lockfile.

One more question to be sure I understand well :)

My understanding was that, whether you try to resolve the torch for Python>=3.8 (universal) or Python==3.8(current environment), the obvious resolution seems to be

• 2.5.1 if you look at the metadata (the requires_python field),
• 2.4.1 if you look at the available wheels,

because torch released the 2.5.0 and 2.5.1 versions with a requires_python at 3.8 while only providing wheels for Python 3.9 to 3.13 (see pytorch/pytorch#141781).

In a sense my question is: Do uv sync and uv run --no-project look at the requires_python or the available wheels to resolve dependencies?

If both do look at the available wheels, I do not understand why the universal resolution (with uv sync or uv lock) chooses 2.5.1 (without forking for Python 3.8) and not 2.4.1 because 2.4.1 if the most recent version compatible with Python >= 3.8, not 2.5.1.

Is there something I misunderstood there in this "universal" resolution concept?

[ThomasHezard]

I'm not fluent enough in rust to look at the uv code, but I tried to interpret the uv verbose logs.
I'm not sure what to interpret from these 😅

--

Here is (part of) the output of uv lock --python 3.8 --verbose with the pyproject.toml of my initial message:

DEBUG uv 0.5.5 (Homebrew 2024-11-27)
[...]
DEBUG Using Python request `3.8` from explicit request
[...]
Using CPython 3.8.20
[...]
DEBUG Solving with installed Python version: 3.8.20
DEBUG Solving with target Python version: >=3.8
[...]
DEBUG Adding transitive dependency for uv-torch-test==0.1.0: torch*
DEBUG Found fresh response for: https://pypi.org/simple/torch/
DEBUG Searching for a compatible version of torch (*)
DEBUG Selecting: torch==2.5.1 [compatible] (torch-2.5.1-cp310-cp310-manylinux1_x86_64.whl)

--

Here is (part of) the output of uv run --no-project --verbose --python 3.8 --with "torch" python:

DEBUG uv 0.5.5 (Homebrew 2024-11-27)
[...]
DEBUG No project found; searching for Python interpreter
DEBUG Searching for Python 3.8 in virtual environments, managed installations, or search path
[...]
DEBUG Using Python 3.8.20 interpreter at: [...]/.local/share/uv/python/cpython-3.8.20-macos-aarch64-none/bin/python3.8
[...]
DEBUG Solving with installed Python version: 3.8.20
DEBUG Solving with target Python version: >=3.8.20
DEBUG Adding direct dependency: torch*
DEBUG Found fresh response for: https://pypi.org/simple/torch/
DEBUG Searching for a compatible version of torch (*)
DEBUG Searching for a compatible version of torch (<2.5.1 | >2.5.1)
DEBUG Searching for a compatible version of torch (<2.5.0 | >2.5.0, <2.5.1 | >2.5.1)
DEBUG Selecting: torch==2.4.1 [compatible] (torch-2.4.1-cp38-none-macosx_11_0_arm64.whl)

--

[ThomasHezard]

One more experiment.

I compiled the version from #8686 and ran

uv lock --python 3.8 --multi-version lastest

on the pyproject.toml from my initial message.
=> uv still resolves torch 2.5.1 for all Python versions (3.8 to 3.13)
=> uv still fails to sync with Python 3.8 as it tries to install torch2.5.1 (same error as my first message).

So unfortunately, #8686 won't help in that case and the mystery of why my two commands do not behave the same way is still bothering me a bit 🙂

I can provide logs if that can help.

[charliermarsh]

We look at the requires-python, not the available wheels. We look at the wheels to determine whether there's at least one wheel in the space of supported Python versions. The assumption is that other versions will build from source. PyTorch doesn't ship source, so you end up with a version that may not support (e.g.) Python 3.8. Hence the open issue that I linked earlier.

[ThomasHezard]

Thank you, absolutely all clear, now I understand why the two commands end up with different versions 👌

This totally makes sense and is quite clever to get a fast resolution while taking advantage of PyPI API.

I must say that the thing that bugs me the most here is that nothing happened in both torch release process and PyPI backend when faced to what seems to me two ill-formed releases, in the sense that the requires-python is not respected by the either wheels or sdists 😮
But maybe there is something I'm missing here.

It is quite funny how uv helped me spot several eronous dependency specs in packages really quickly (see #9211 (comment) for another example).
For me that is a huge advantage of uv, this is the first time I find such an efficiently rigorous (which I love) environment manager in the Python world 👏
It helps me test my packages on different environment really efficiently and make absolutely reliable releases 🤩

[ThomasHezard]

@charliermarsh

FYI: the problem on PyTorch side is being taken care of: pytorch/pytorch#141781.

No sure when, but this (and #5182 if I understood well) should be fixed in the near future, without having to specify complex version constraints in order to fix in uv what is in fact a bug on PyTorch side.