Hey yall I just found out about this project a few days ago, very cool stuff. Currently evaluating this for use in my team at Anduril. One of the main blockers tho is the lack of a easy way to directly pass in device buffers to kernels (I'm sure theres some not so straightforward ways) . There are several reasons to want to do this:

• You are using cubecl in conjunction with some other library, tool, langauge, etc that produces arrays/tensors that are already on the GPU. You'll have no choice but to copy it back to the host, and then have the runtime copy it back in order to call a kernel on it.
• You want to use unified / pinned memory, GPU RDMA, etc for performance reasons.

I've sketched out an implementation here and I'm looking for feedback before I get too deep. I've basically just added an ArrayArg variant that takes the Resource of the associated ComputeStorage. In the end you'll be able to mix it with normal cubecl arrays and it'll look something like this:

let device_buffer = CudaResource::from_device_pointer(...);

some_kernel::launch_unchecked::<...>(
    &client,
    CubeCount::Static(...),
    CubeDim::new(...),
    ArrayArg::from_raw_resource(device_buffer),
    ArrayArg::from_raw_parts(&normal_handle, ...),
);