(Feature request) DirectML support

[NotNite]

I have an AMD GPU, so my way of getting around not having CUDA is by using DirectML. Unfortunately, this depends on CUDA APIs (and is Linux only), so I can't use it.

[RoARene317]

Just use Linux, make sure your GPU support ROCm. Case closed.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

[yoda5477]

This would be greatly appeciated now that it is needed for flux nf4. not sure how complexe to solve the current issue link to only cuda compatible

[ethrx]

Can we re-open this?

It's a valid use-case.

[matthewdouglas]

@ethrx I agree, especially for consumer AMD GPUs not otherwise supported in ROCm. I'll reopen this, but we won't be able to get to this right away.