[FEA] Calling NVSHMEM put/get inside Cute-DSL kernel

[tridao]

Which component requires the feature?

CuTe DSL

Feature Request

We're planning a project that involves calling NVSHMEM put/get inside the kernel. Is there plan for cute-dsl to support this (maybe through FFI?)
We could write it in CUDA but Cute-DSL is quite a bit nicer to work with :D

[benhg]

Hi @tridao, perfect timing :) We have just pushed this feature to NVSHMEM repo as Github Preview. We are working on finishing up the QA process and we'll have the official release out very soon.

You can look here for the implementation: https://github.com/NVIDIA/nvshmem/tree/devel/nvshmem4py/nvshmem/core/device/cute

[tridao]

Perfect, thank you very much! Looking forward to GA!

[benhg]

We did the release last week. You can see the docs here: https://docs.nvidia.com/nvshmem/api/api/language_bindings/python/device/cute/index.html