Mario Lezcano Casado

I am not the right person to review this PR. Perhaps @jbschlosser or @albanD...

@pytorchbot merge

Note that the general pattern here would be `a*x - F(a*x)` = `a*(x-F(x))` for `F` linear, as the same thing may happen for any function `F` that's linear and that...

yeah, the more general pattern just asks for F to be [1-homogeneous](https://en.wikipedia.org/wiki/Homogeneous_function), but we don't have a good way of marking which ops are 1-homogeneous. (I realised of this after...

Why did the last commit fixes a perf regression?

I put up a fix, but I was not able to test whether it works (my triton version is acting up with the repro). Mind checking if it fixes the...

I figured a way would be to try to cudagraph the relevant code and see that it was able to do so

I'm a bit worried that the user does not have access to these. I wonder whether the cure is worse than the disease...

What I wonder is whether it makes more sense to have FMA on by default and then have a `torch.nofma` op for when it may be an issue. That is,...

@pytorchbot merge