Summary

When using [tool.uv.sources] with conditional index overrides (via extra), packages resolved from the default index (PyPI) for the non-extra case lose their platform_machine-gated
transitive dependencies.

Reproduction

Minimal pyproject.toml that works:

[project]
name = "test-torch"
version = "0.1.0"
requires-python = ">=3.13"
dependencies = ["torch>=2.0"]

uv lock && grep "nvidia-cudnn" uv.lock
# Output: name = "nvidia-cudnn-cu12"

Adding conditional sources breaks it:

[project]
name = "test-torch"
version = "0.1.0"
requires-python = ">=3.13"
dependencies = ["torch>=2.0"]

[project.optional-dependencies]
cu13 = ["torch>=2.0"]
cpu  = ["torch>=2.0"]

[tool.uv]
conflicts = [
  [{ extra = "cu13" }, { extra = "cpu" }],
]

[[tool.uv.index]]
name = "pytorch-cu130"
url = "https://download.pytorch.org/whl/cu130"
explicit = true

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

[tool.uv.sources]
torch = [
  { index = "pytorch-cu130", extra = "cu13" },
  { index = "pytorch-cpu", extra = "cpu" },
]

uv lock && grep "nvidia-cudnn-cu12" uv.lock
# Output:

Expected behavior

When no extra is specified, torch should resolve from PyPI (the default), and its transitive dependencies with platform_machine == 'x86_64' and sys_platform == 'linux' markers (like
nvidia-cudnn-cu12) should be included in the lockfile.

Actual behavior

The lockfile only contains nvidia-cudnn-cu13 (for the cu13 extra). The default PyPI torch resolution is missing all nvidia-*-cu12 dependencies that have platform_machine markers.

Impact

Running uv sync (without extras) installs torch from PyPI, but the CUDA runtime dependencies are missing:

ImportError: libcudnn.so.9: cannot open shared object file: No such file or directory

Platform

Ubuntu 22.04 x86_64

Version

uv 0.9.27

Python version

Python 3.13.11