PyTorch nightly index makes uv download all torch versions in a loop

[ViRb3]

I copied the full PyTorch example from https://docs.astral.sh/uv/guides/integration/pytorch/#using-a-pytorch-index:

[project]
name = "project"
version = "0.1.0"
requires-python = ">=3.12.0"
dependencies = [
  "torch>=2.5.1",
  "torchvision>=0.20.1",
]

[tool.uv.sources]
torch = [
    { index = "pytorch-cpu", marker = "platform_system != 'Darwin'" },
]
torchvision = [
    { index = "pytorch-cpu", marker = "platform_system != 'Darwin'" },
]

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

Only difference is I swapped the index url to https://download.pytorch.org/whl/nightly/cpu

When I run uv sync, it downloads all versions of torch in reverse order. It doesn't really stop until I Ctrl + C it. Without nightly, it works fine.

$ uv version
uv 0.5.6 (Homebrew 2024-12-03)

[zanieb]

We're backtracking on the package because its dependencies can't be solved. Because of the limitations of the torch index, we can't inspect its metadata without downloading the full wheel.

With some verbose logs..

TRACE Selecting candidate for pytorch-triton with range ==3.2.0+git35c6c7c6 with 1 remote versions
TRACE Exhausted all candidates for package pytorch-triton with range ==3.2.0+git35c6c7c6 after 0 steps

There are pytorch-triton wheels available with a matching version, but your index is set to explicit. You'll want to use the index for it, e.g.

[project]
name = "project"
version = "0.1.0"
requires-python = ">=3.12.0"
dependencies = [
  "torch>=2.5.1", "pytorch-triton",
]

[tool.uv.sources]
torch = [
    { index = "pytorch-cpu"},
]
torchvision = [
    { index = "pytorch-cpu"},
]
pytorch-triton = [
    { index = "pytorch-cpu"},
]

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/nightly/cpu"
explicit = true

[charliermarsh]

A big part of the problem here (I assume) is that the PyTorch registry has no fast-path for metadata. So you literally have to download every version (in full) to understand the dependencies at each stage.

[ViRb3]

@zanieb Thanks for the suggestion! I tried that, but it fails on macOS with the following error:

error: Distribution pytorch-triton==3.2.0+git35c6c7c6 @ registry+https://download.pytorch.org/whl/nightly/cpu can't be installed because it doesn't have a source distribution or wheel for the current platform

Removing pytorch-triton from dependencies sends it into the torch download loop again...

[zanieb]

I removed the != Darwin markers to debug, you should restore those (and include it for triton).

[ViRb3]

That worked, thanks! Although, why did it work? As far as I understand, we are specifying to use the nightly index ONLY IF platform_system != 'Darwin'. However, I am on Darwin, so it should fall back to the default pip index, and the nightly index is left unused?

[zanieb]

During lock, we'll solve for Darwin / not Darwin so there will be two versions

❯ uv lock
Resolved 27 packages in 516ms
Updated pytorch-triton v3.2.0+git35c6c7c6 -> v0.0.1, v3.2.0+git35c6c7c6
Updated torch v2.6.0.dev20241204+cpu -> v2.5.1, v2.6.0.dev20241204+cpu

And yeah, the Darwin version will come from PyPI instead of the Pytorch CPU index (which only supports Linux).

During sync, we install a locked version appropriate for the platform.

❯ uv sync
...
 + pytorch-triton==0.0.1
 + torch==2.5.1
...

[zanieb]

You can remove the extraneous pytorch-triton install (that's just a dummy package) by declaring the dependency as "pytorch-triton; platform_system != 'Darwin'",

[ViRb3]

Interesting. I was following Apple's guide on installing PyTorch nightly (https://developer.apple.com/metal/pytorch/) and they specify the following command:

pip3 install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cpu

If I run this using uv pip, it installs the following:

torch==2.6.0.dev20241126
torchaudio==2.5.0.dev20241126
torchvision==0.20.0.dev20241126

This definitely looks like nightly, so it surely should exist on the Pytorch CPU index? I was able to reproduce using the following:

dependencies = [
    "torch",
    "torchaudio",
    "torchvision",
]

[tool.uv]
environments = [
    "platform_system == 'Darwin'",
]

[tool.uv.sources]
torch = { url = "https://download.pytorch.org/whl/nightly/cpu/torch-2.6.0.dev20241126-cp312-none-macosx_11_0_arm64.whl" }
torchaudio = { url = "https://download.pytorch.org/whl/nightly/cpu/torchaudio-2.5.0.dev20241126-cp312-cp312-macosx_11_0_arm64.whl" }
torchvision = { url = "https://download.pytorch.org/whl/nightly/cpu/torchvision-0.20.0.dev20241126-cp312-cp312-macosx_11_0_arm64.whl" }

I'm honestly not sure what the environments part does, I saw it from another GitHub issue since without it I get this error:

  × No solution found when resolving dependencies for split (platform_system != 'Darwin' and python_full_version >= '3.12' and python_full_version < '4'):
  ╰─▶ Because there is no version of pytorch-triton{python_full_version < '3.13' and platform_machine == 'x86_64' and platform_system == 'Linux'}==3.2.0+git35c6c7c6 and torch==2.6.0.dev20241126 depends on pytorch-triton{python_full_version < '3.13' and platform_machine
      == 'x86_64' and platform_system == 'Linux'}==3.2.0+git35c6c7c6, we can conclude that torch==2.6.0.dev20241126 cannot be used.
      And because only torch==2.6.0.dev20241126 is available and your project depends on torch, we can conclude that your project's requirements are unsatisfiable.

Specifying a dependency pytorch-triton; platform_system != 'Darwin' does not help in this case.

Does this error throw because I am hardcoding macOS sources, while uv is trying to solve for all OSes so their dependencies can be included in the lock file?

[zanieb]

environments = [
    "platform_system == 'Darwin'",
]

means "only lock for Darwin". https://docs.astral.sh/uv/concepts/projects/config/#limited-resolution-environments

With uv pip, we only solve for the current platform. https://docs.astral.sh/uv/concepts/resolution/#platform-specific-resolution

I believe the +cpu suffixes on the versions are messing up our resolution. The nightly index seems to have different distributions. I'm not sure what the workaround is for that, @charliermarsh would probably know. I'd try to designate a version without +cpu for Darwin and with +cpu elsewhere.

[zanieb]

Okay trial and error, the proper invocation is

[project]
name = "project"
version = "0.1.0"
requires-python = ">=3.12.0"
dependencies = [
  "torch===2.6.0.dev20241204; platform_system == 'Darwin'",
  "torchvision===0.20.0.dev20241205; platform_system == 'Darwin'",
  "torch==2.6.0.dev20241204+cpu; platform_system != 'Darwin'",
  "torchvision==0.20.0.dev20241205+cpu; platform_system != 'Darwin'",
  "pytorch-triton; platform_system != 'Darwin'",
]

[tool.uv.sources]
torch = [
    { index = "pytorch-cpu" },
]
torchvision = [
    { index = "pytorch-cpu" },
]
pytorch-triton = [
    { index = "pytorch-cpu" },
]

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/nightly"
explicit = true

Sorry it's such a pain, the torch indexes are very inconsistent in their versioning and distribution availability per platform.

[ViRb3]

Got it, makes sense. Thanks for the help!

[charliermarsh]

This should be way better in the next release.