Group Operations

Free functions and helpers for batching NCCL operations into groups. See Group Calls for usage details.

group

nccl.core.group() → Generator[None, None, None]

Context manager for NCCL group operations.

Automatically calls group_start() on entry and group_end() on exit, ensuring proper cleanup even if an exception occurs.

Simulation mode is not supported here. To simulate, call group_start() and group_end() directly and pass simulate=True to group_end().

group_start

nccl.core.group_start() → None

Starts a group of NCCL operations.

All NCCL operations called after this will be batched together and executed when group_end() is called. This can improve performance by allowing NCCL to optimize the operation sequence.

group_end

nccl.core.group_end(*, simulate: Literal[False] = False) → None

nccl.core.group_end(*, simulate: Literal[True]) → GroupSimInfo

nccl.core.group_end(*, simulate: bool) → GroupSimInfo | None

Ends a group of NCCL operations.

By default, executes all operations queued since the last group_start(). When simulate=True, the queued operations are simulated instead of executed, and the estimated execution time is returned in a GroupSimInfo.

Parameters:: simulate – When True, simulates the group instead of executing it and returns a GroupSimInfo carrying the estimated time. Defaults to False.
Returns:: None when simulate=False; a GroupSimInfo with the simulation result when simulate=True.

GroupSimInfo

class nccl.core.GroupSimInfo(estimated_time: float)

Bases: object

Result of an NCCL group simulation.

Returned by group_end() when called with simulate=True.

estimated_time: float: Estimated execution time for the simulated group operations, in seconds.