Triton JIT support for device APIs

[Gui-Yom]

There is already support for Numba CUDA JIT but Triton JIT interacts much better with frameworks like PyTorch.

[ZzEeKkAa]

Hello @Gui-Yom ! Thank you for interesting in the product! Could you provide more details what feature are you looking for? Is it FFT or matrix multiplication?

[Gui-Yom]

It would be nice to be able to use a blocked FFT.execute() from inside a blocked Triton JIT kernel. I think Triton has support for external functions already though I'm unsure how to bridge this with nvmath-python own codegen. The FFT operation is rather low level and can't be expressed with the triton language directly.

[samaid]

This is very useful input, Gui-Yom. We are exploring this.