What's the curent state of rust gpu support?

I am currently learning c++ because omp is the gold standard for running code on gpu at least from the background of the ppl teaching me hpc.

Rust seems cool tho and i am considering learning it as an hpc languge but I kinda need to know how it works on gpu and what's the compiler support like to things like cuda.

Like I seen burn as an option to replace pytorch but if u r using the torch backend i feel like u r losing a lot. A lot of the fast c++ extensions use the raw cuda pointer and then go into their own computations.

I feel like rusts concurrency model could be really good for something that continuously fed the gpu. Especially because most modern ml is memory bound.

Moving ownership representing a physical gpu move seems reasonable and i think u may be able to get nice stuff with reusable memory with a good enough compiler