For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo struct
Communicators
struct Communicators
Holds NCCL/RCCL communicator handles for a fixed set of GPUs.
Stores one ncclComm_t handle per GPU (up to MAX_GPUS). Instances are
initialized lazily by _get_global_comms and cached process-wide; call
init_comms() from a single thread before using multi-threaded collectives
to avoid the check-then-create race in _get_global_comms.
Fields
- ngpus (
Int): The number of GPUs participating in the communicator group. - comms (
Array[Optional[Pointer[NoneType, MutUntrackedOrigin, _safe=False]], Int(8)]): Per-GPU communicator handles, valid for indices0..ngpus-1.
Implemented traits
AnyType,
Copyable,
ImplicitlyCopyable,
ImplicitlyDeletable,
Movable
Methods
__init__
def __init__(out self, *, copy: Self)
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!