`uv sync` handling third party dependencies in `setup.py`

[jamesbraza]

I am using uv==0.4.16, and there are two packages that make my life annoying because their setup.py depends on torch:

• https://github.com/facebookresearch/xformers/blob/v0.0.28/setup.py#L24-L31 (PyPI)
• https://github.com/Dao-AILab/flash-attention/blob/v2.6.3/setup.py#L21-L29 (not on PyPI)

How can uv sync work with this scenario? Currently, it successfully makes a uv.lock, but then fails on uv sync:

[[package]]
name = "xformers"
version = "0.0.28"
source = { registry = "https://pypi.org/simple" }
dependencies = [
    { name = "numpy" },
    { name = "torch" },
]
sdist = { url = "https://files.pythonhosted.org/packages/be/8e/1b9bc5317a7a26d33af2db63f5db112c0cfe373befe7d5653e1d639db231/xformers-0.0.28.tar.gz", hash = "sha256:bc74e7033d2675962d4ddc708fc66328b6f6cbe5aedff7491eab98c338623bf2", size = 7757812 }
wheels = [
    { url = "https://files.pythonhosted.org/packages/d1/a0/bb5a668186f5f77a73043a8ff8b86176c7d3cbed10cbdafcca67237ee432/xformers-0.0.28-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:c7c6b1887d6bb71fe26464492b84d6cdbfbfdada5669a4a606c186a384e226f5", size = 16653563 },
    { url = "https://files.pythonhosted.org/packages/9b/e4/efead2c5c89cd5766c7c4f5320448c80a51f3a497bbe560b8000e2ebc930/xformers-0.0.28-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:9b3a5e6f2e02d6b2bb84a38351dcb181cfdb037451ab8cd57f39c161af33d83b", size = 16653138 },
]

error: Failed to prepare distributions
  Caused by: Failed to fetch wheel: xformers==0.0.28
  Caused by: Build backend failed to determine extra requires with `build_wheel()` with exit status: 1
--- stdout:

--- stderr:
Traceback (most recent call last):
  File "<string>", line 14, in <module>
  File "/Users/user/Library/Caches/uv/builds-v0/.tmpCHWnFA/lib/python3.12/site-packages/setuptools/build_meta.py", line 332, in get_requires_for_build_wheel
    return self._get_build_requires(config_settings, requirements=[])
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/user/Library/Caches/uv/builds-v0/.tmpCHWnFA/lib/python3.12/site-packages/setuptools/build_meta.py", line 302, in _get_build_requires
    self.run_setup()
  File "/Users/user/Library/Caches/uv/builds-v0/.tmpCHWnFA/lib/python3.12/site-packages/setuptools/build_meta.py", line 503, in run_setup
    super().run_setup(setup_script=setup_script)
  File "/Users/user/Library/Caches/uv/builds-v0/.tmpCHWnFA/lib/python3.12/site-packages/setuptools/build_meta.py", line 318, in run_setup
    exec(code, locals())
  File "<string>", line 24, in <module>
ModuleNotFoundError: No module named 'torch'
---

[charliermarsh]

There are some tips on that here: https://docs.astral.sh/uv/concepts/projects/#build-isolation. It's still a bit of a pain though.

[jamesbraza]

Ah I see, thank you! Yeah so currently installation is a two-step process:

1. pip install torch
2. pip install xformers

I was sort of hoping uv could somehow convert that to be a one-step process, but it doesn't seem to be the case:

1. uv sync --extra build
   • Note: the extra build= ["torch"]
2. uv sync --extra xformers
   • Note: relies on no-build-isolation-package = ["xformers"] being in tool.uv

Do you think it's a reasonable request to uv to somehow support this use case as a one-step process? In other words, make it less of a special case from an installer's perspective

[charliermarsh]

I would like it to be one-step, yeah. It's a reasonable request. It likely won't happen immediately though.

[vwxyzjn]

Hi @charliermarsh just to confirm, the current best approach still requires two commands, right? I needed to run

uv sync
uv sync --compile

with

[project]
name = "open-instruct"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = ">=3.10"
dependencies = [
    "numpy<2",
    "packaging>=24.2",
    "setuptools>=75.6.0",
    "torch>=2.5.1",
    # "flash-attn>=2.7.0.post2",
]

# pytroch related setups
[tool.uv.sources]
torch = [
  { index = "pytorch-cu121", marker = "platform_system != 'Darwin'"},
]
[[tool.uv.index]]
name = "pytorch-cu121"
url = "https://download.pytorch.org/whl/cu121"
explicit = true
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

# flash-attn related setups
[project.optional-dependencies]
compile = ["flash-attn>=2.7.0.post2"]
[tool.uv]
no-build-isolation-package = ["flash-attn"]
[[tool.uv.dependency-metadata]]
name = "flash-attn"
version = "2.7.0.post2"
requires-dist = ["torch", "setuptools"]

I am slightly puzzled as I thought the tool.uv.dependency-metadata would make it a one command process, because it already tells uv to wait for torch and setuptools 👀

[AyhamSaffar]

I am struggling with the same issue.

xformers is needed to run any hugging face models in PyTorch, so i would imagine alot of people are getting stuck on this issue.

I am sure xformers' environment is really poorly configured given it doesn't automatically install when you install the hugging face libraries that depend on it, but supporting these kinds of libraries without any hard-to-find tricks would be amazing.

[AyhamSaffar]

Also @charliermarsh couldn't get xformers to install using your linked page.

Setup:

[project]
name = "demo"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = "==3.12.*"
dependencies = []

[project.optional-dependencies]
build = ["torch", "setuptools"]
compile = ["xformers"]

[tool.uv]
no-build-isolation-package = ["xformers"]

[[tool.uv.dependency-metadata]]
name = "xformers"
version = "0.0.29.post1"
requires-dist = ["torch", "setuptools"]

Then running this:

uv sync --extra build
uv sync --extra build --extra compile

The following keeps getting raised:

building 'xformers._C' extension ... UserWarning: Attempted to use ninja as the BuildExtension backend but we could not find ninja.. Falling back to using the slow distutils backend. ... UserWarning: Error checking compiler version for cl: [WinError 2] The system cannot find the file specified error: Microsoft Visual C++ 14.0 or greater is required. Get it with ...

Every time I add more to the pyproject.toml dependencies, a different error comes up with the same final line.

Is this a UV thing or an xformers thing?

[charliermarsh]

That would be an xformers thing... Unfortunately they only ship Linux wheels, so you have to build from source.=, and your machine doesn't seem to have the dependencies it needs.

[AyhamSaffar]

thanks! everything works fine on wsl. You do not even need the whole no-build-isolation trick.

I guess the library setup.py is only run when there isn't a dist and wheel / whatever binaries for your platform on PyPI?

[zanieb]

You'll probably want the MS C++ toolchain for other packages as well, https://visualstudio.microsoft.com/visual-cpp-build-tools/

[charliermarsh]

Yeah that's right. You can see here that they only ship Linux-compatible "wheels":

So if you're not on Linux, we have to download the source distribution and try to build it.

[AyhamSaffar]

Cheers @zanieb and @charliermarsh. I appreciate the help.

[muhammad-fiaz]

is there any good solution for this?? or any example working reference, for me the main issue is with xformers it always conflict with cuda enabled pytorch when installing with uv pyproject.toml

[zanieb]

We made this work in one-step now, see https://docs.astral.sh/uv/concepts/projects/config/#augmenting-build-dependencies

@muhammad-fiaz if you have questions after reading that please open a new issue with details. I don't think this one is specific enough for us to help you.

[muhammad-fiaz]

Hello @zanieb Thanks for replying, i also saw that docs but it does automate some features specific to pytorch and flash attention but for me the main issue is with xformer usage with uv i am not sure it always conflict with pytorch and installing incorrectly when ever there is some new cuda so i needed to manually install using pip for each dependencies based dependencies so i have created an #16522 feature request support for multiple automated version for same dependencies

BTW! i am using uv version (v0.9.6)