ucc icon indicating copy to clipboard operation
ucc copied to clipboard

Add some CPU collectives to the NCCL TL

Open pavanbalaji opened this issue 2 years ago • 1 comments

What

Adds support for CPU collectives to NCCL TL (only covers a subset for now).

Why ?

Currently UCC requires two TLs (NCCL and UCP) to fully support NVIDIA GPU platforms. This patch allows for the NCCL TL to support CPU collectives too, so we do not need to rely on two different TLs for the full coverage.

How ?

Stages CPU data through the GPU and makes NCCL calls on the GPU-resident data.

pavanbalaji avatar Jul 20 '22 01:07 pavanbalaji

Can one of the admins verify this patch?

swx-jenkins3 avatar Jul 20 '22 01:07 swx-jenkins3