Question: about CTran Send Recv

Hi, Thank you for your nice work! I have read about your paper Collective Communication for 100k+ GPUs especially Chapter 5.1 PP: Zero-copy and SM-free Send/Receive.

I wander how to use CTran to achieve SM-free and Zero Copy send recv in async way, without using NCCL copy-based send/recv or RDMA which would rely on pre-allocate buffer, or maybe register user tensor as RDMA MR every time when we launch send/recv? Is there a best practice?
In the evaluation chapter of your paper Collective Communication for 100k+ GPUs, you have mention SM-Free and Zero Copy send recv. So I really really want to try it :)

Also I notice send recv in ncclx backend still using nccl, which is not SM Free and Zero Copy, I wander why we don't use Ctran to implement a better version that is SM Free and Zero Copy?

The issue names CTran and the ncclx backend but does not identify a source file or test. Read the CTran and ncclx send/recv entry points and Chapter 5.1 of the linked paper first; define the supported async, SM-free, zero-copy behavior and a reproducible validation path before proposing implementation.