Alexander Grund

@pytorchbot merge

@pytorchbot rebase

Looks like the vendored packaging module needs to be used. Will the macos build handled in the PR CI to check that really does work?

@etaf Conflicts caused by NVIDIA renaming the Python package from cutlass to cutlass_cppgen Resolved by adding a fallback to cutlass if cutlass_cppgen can't be imported

This fails on the H100 test: https://github.com/pytorch/pytorch/blob/cc8bfd1206f4bff26bd86ce584f6c16b6401ef50/test/inductor/test_cutlass_backend.py#L245-L250 > AssertionError: The length of the sequences mismatch: 8 != 0 I can't see how this change would make a difference for the...

Ok, CI rerun seems to have passed regardless. Adding a proper fix with https://github.com/pytorch/pytorch/pull/167060, @etaf could you check that on your end? Requesting re-review as with the merge I had...

Rebased to make the fallback handling a separate commit and handle the conflict earlier

That doesn't seem to be enough. PyTorch 2.9+ seems to require nvidia-cutlass >= 4 so we might not need the fallback from cutlass_cppgen to cutlass. And `cutlass` is separately imported...

From my experiments it seems that cloning NVIDIA/CUTLASS using the same commit as referenced as a submodule by the used PyTorch version and setting `TORCHINDUCTOR_CUTLASS_DIR` to that works. It doesn't...

> @Flamefire can you please fix lint? (namely run clang-format on your PR) Also, do you mind adding a few more references in the doc to the PR description that...