Fix Bug in MX Builds

[drisspg]

Stacked PRs:

• ->Fix Bug in MX Builds #2284

---

Fix Bug in MX Builds

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2284

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 919480d with merge base e51ffd9 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[drisspg]

This fixes it but there is another problem that is if no torch_cuda_arch_list is set the cutlass files dont' get build...

[gau-nernst]

Actually as long as we have a sufficiently new CUDA toolkit (12.8 I think?), we can compile for sm100a? Regardless of actual compute capability available at compile time.

The downside is that compile time is longer for non-sm100a users building from source.

[drisspg]

So there was 2 problems:

1. the accidental downgrade of Cutlass fixed in the PR
2. The second is that cuda extensions doesn't let you compile w/ perfile flags, In the past for cutlass utilzing the 'a' variants if you try to compile with anything but that sm+a you would get a compile time error, e.g. Fixes MX formats build for blackwell #2214

I think the problem right now is that I unconditionally remove remove the sources

e.g.
if I set 90a, and 100a I get

INFO:root:copying build/lib.linux-x86_64-cpython-312/torchao/_C.abi3.so -> torchao
INFO:root:copying build/lib.linux-x86_64-cpython-312/torchao/_C_cutlass_90a.abi3.so -> torchao
INFO:root:copying build/lib.linux-x86_64-cpython-312/torchao/_C_cutlass_100a.abi3.so -> torchao

which is right

ahh its just 90a nd 100a are not in the arch_flags w/ no TORCH_CUDA_ARCHLIST set