Inductor can not fuse cat with a pointwise

[shunting314]

🐛 Describe the bug

Check this example

    def test_concat_and_downcast(self):
        M = 30522
        N = 768
        PAD = 6

        @torch.compile
        def f(x):
            z = torch.cat([x, torch.zeros([PAD, N])], dim=0)
            return z, x.to(torch.float16)

        x = torch.randn(M, N)
        f(x)

The cat and the downcast can not be fused right now. But in principal we should be able to fuse them and save one whole load of tensor 'x'.

cc @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @jansel @Chillee @eellison as a FYI.

Error logs

No response

Minified repro

No response

Versions

.

cc @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang

[Chillee]

We'd fuse it with the pointwise cat codegen?

[shunting314]

We'd fuse it with the pointwise cat codegen?

That requires the awareness of the other pointwise node that can be fused when we lower cat as a piontwise? I think it's a bit hard since the other pointwise node that can be fused can happen somewhere else.

Or an alternative is to do this in scheduler since the scheduler does global search and can try to fuse the pointwise node with the cat node. But the tricky thing here is these 2 nodes has different numels.