opt-in for scatter operation in nvfuser

[jjsjann123]

No description provided.

[t-vi]

Cool to see NVFuser expand the ops!

What would be the circumstances in which one would opt in and not opt in? I can see why linear is a thing that is more tricky in terms of perf, but this seems to be a good option to always activate.

I still think all options are more or less evil, but after #2057 , I'm more optimistic that this one could be temporary.

[jjsjann123]

I still think all options are more or less evil, but after #2057 , I'm more optimistic that this one could be temporary.

Agree that these options are not good user experience. These are going to be a temporary thing.

The reason I'm adding a flag is that we have concurrent work going on.

• nvfuser is adding codegen support for these operations, so eventually we should be able to take them as-is by nvfuser.
• Meanwhile, our multi-gpu folks are working on models that needed these operations inside nvfuser (for parallelization strategy), so we want to have them, even though they are running as a fallback.

We'll see a couple similar ops added in the coming weeks (topk groupedgemm e.t.c.). We are not expecting any user to rely on these flags, since fallbacks don't necessarily provide any kernel perf gain.

[jjsjann123]

Looks like scatter related tests run fine. I'm not sure what the failure on stable torch 2.7 is about.

I'll clean up the PR