Is this page helpful?

DistributedDecompositionOptions#

class cuquantum.tensornet.experimental.distributed.DistributedDecompositionOptions(
compute_type: int | None = None,
device_id: int | None = None,
handle: int | None = None,
logger: Logger | None = None,
memory_limit: int | str | None = '80%',
blocking: Literal[True, 'auto'] = True,
allocator: BaseCUDAMemoryManager | None = None,
collective_error_agreement: bool = False,
)[source]#

DecompositionOptions, plus options specific to distributed_decompose().

compute_type must be None; decomposition precision follows the operand data type.

collective_error_agreement#

If True, distributed_decompose() ballots across ranks on its Python-side pre-checks (operand, dtype, layout, and output validation, staging, allocation) before any collective native call, so a rank-divergent failure raises consistently on every rank instead of stranding a peer in a collective call it will never reach. When False (the default), a rank-local failure simply raises immediately – the standard SPMD/MPI contract, where it is the caller’s responsibility to construct equivalent, valid arguments on every rank or to recover (e.g. MPI_Abort).

Type:

bool