For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

init_comms

def init_comms(ngpus: Int)

Pre-initialize NCCL/RCCL communicators.

Must be called from a single thread before using allreduce from multiple threads. _get_global_comms has a check-then-create race: two threads seeing null simultaneously would both call ncclCommInitAll and one would leak its communicators.

Raises:

If the NCCL/RCCL communicator initialization fails.