I'd love to se a side by side by side comparison of implementing as triton, cuda/c++, just using torch.compile, etc etc with a few example ops. I have broken out triton a lot for things but found that it is very hit or miss how much I will gain over just using torch.compile. Probably a lot of that is my skills and a lot is how much torch.compile can take together and optimize if raw pytorch is made available to it.

Aren't there other options for custom PyTorch operators before going so low-level as CUDA C++, like using Jax or CuPy?

neat -- how to add pt2 support and how to allow fusion of your custom ops with normal ops by inductor?

hmm, dynamo did not support torchbind last time I looked. when did support for torchbind get added? I need to know what version of torch to upgrade to.