different behavior of uv pip install for x86 and aarch64 archs

[senysenyseny16]

Hello, I encountered the following problem:

x86     | pip install torch==2.1.0+cpu --extra-index-url https://download.pytorch.org/whl/cpu       | ✅
x86     | pip install torch==2.1.0 --extra-index-url https://download.pytorch.org/whl/cpu           | ✅ --+
x86     | uv pip install torch==2.1.0+cpu --extra-index-url https://download.pytorch.org/whl/cpu    | ✅   |
x86     | uv pip install torch==2.1.0 --extra-index-url https://download.pytorch.org/whl/cpu        | ❌   |
                                                                                                     [intersection]
aarch64 | pip install torch==2.1.0+cpu --extra-index-url https://download.pytorch.org/whl/cpu       | ❌   |
aarch64 | pip install torch==2.1.0 --extra-index-url https://download.pytorch.org/whl/cpu           | ✅ --+
aarch64 | uv pip install torch==2.1.0+cpu --extra-index-url https://download.pytorch.org/whl/cpu    | ❌
aarch64 | uv pip install torch==2.1.0 --extra-index-url https://download.pytorch.org/whl/cpu        | ✅

With pip, you can use pip install torch==2.1.0 --extra-index-url https://download.pytorch.org/whl/cpu command and it will work on both architectures, but for uv, you'll have to add a condition based on the architecture, which is very inconvenient, especially in the case of automation.

[charliermarsh]

It's because the PyTorch index itself is inconsistent. Some of the wheels on that index use +cpu, and others don't, but they're all CPU wheels. I don't see this as something we can solve without redoing local variant supports and I don't anticipate doing that in the near future.

[senysenyseny16]

ok, thanks for the answer, I think I need to ask the Torch team why they did it that way.

[charliermarsh]

I'm also interested, I really want to improve the Torch experience. I've seen https://github.com/pytorch/pytorch/issues/136275.