Error when building PyTorch from source

[cora-codes]

Question

I'm using the latest version of UV.

uv add --editable external/pytorch/

Gives the following error:

  × Failed to build `torch @ file:///home/ubuntu/pytorch`
  ├─▶ The build backend returned an error
  ╰─▶ Call to `setuptools.build_meta:__legacy__.prepare_metadata_for_build_editable` failed (exit
      status: 1)

      [stdout]
      -- Configuring incomplete, errors occurred!
      Building wheel torch-2.8.0a0+git20ba8fe
      -- Building version 2.8.0a0+git20ba8fe
      -- Checkout nccl release tag: v2.26.2-1
      cmake -GNinja -DBUILD_PYTHON=True -DBUILD_TEST=True -DCMAKE_BUILD_TYPE=Release
      -DCMAKE_INSTALL_PREFIX=/home/ubuntu/pytorch/torch
      -DCMAKE_PREFIX_PATH=/home/ubuntu/.cache/uv/builds-v0/.tmpiKso57/lib/python3.10/site-packages
      -DPython_EXECUTABLE=/home/ubuntu/.cache/uv/builds-v0/.tmpiKso57/bin/python
      -DTORCH_BUILD_VERSION=2.8.0a0+git20ba8fe -DUSE_NUMPY=True
      /home/ubuntu/pytorch

      [stderr]
      CMake Error at CMakeLists.txt:26 (project):
        Running

         '/home/ubuntu/.cache/uv/builds-v0/.tmpVYbIyZ/bin/ninja' '--version'

        failed with:

         no such file or directory



      hint: This usually indicates a problem with the package or the build environment.
  help: If you want to add the package regardless of the failed resolution, provide the `--frozen`
        flag to skip locking and syncing.

Could be related to build isolation or something? I've tried cleaning the uv cache and I've ensured that ninja and cmake are installed and up to date.

Platform

No response

Version

No response

[cora-codes]

Just documenting other issues that have came up:

1. I needed to manually install typing extentions, pyyaml and pybind11 (uv add typing_extensions pyyaml pybind11). I also installed ninja and cmake, but I'm unsure if this is best practice or if apt should be used to install them instead.
2. make clean inside the pytorch directory fixed the error above
3. CUDA::cuFile can't be found as a target (I believe this is an issue with poetry too). USE_CUFILE can be turned off, but there isn't a way to have uv automatically set this env variable (yet)

I'm leaving the issue open since it would be nice to make this workflow less painful (by which I mean installing pytorch from source). It's very possible I am missing something and that's why I'm hitting these issues, but I can also believe most people are happy with just using the wheels and so this hasn't been done much (or uv pip is used instead)

[konstin]

PyTorch is a complex, hard to build project with gotchas around different build systems and CUDA versions. You'll likely have a better time following https://github.com/pytorch/pytorch/blob/main/CONTRIBUTING.md than approaching this through uv, which can only use PEP 517.

[cora-codes]

Yes, I am very familiar with the complexity of building PyTorch. I guess I optimistically see a future where uv can build it from source (perhaps with some patches to PyTorch which I'd be happy to submit)...

Also, flash attention works which has a similar level of complexity?