cuda.core.SMResourceOptions

class cuda.core.SMResourceOptions(

count: int | SequenceABC | None = None,

coscheduled_sm_count: int | SequenceABC | None = None,

preferred_coscheduled_sm_count: int | SequenceABC | None = None,

backfill: bool | SequenceABC = False,

)

Customizable SMResource.split options.

Each field accepts a scalar (for a single group) or a Sequence (for multiple groups). count drives the number of groups; other Sequence fields must match its length.

count

Requested SM count per group. None means discovery mode (auto-detect). (Default to None)

Type:

int or Sequence[int], optional

coscheduled_sm_count

Minimum number of SMs guaranteed to be co-scheduled in each group. (Default to None)

Type:

int or Sequence[int], optional

preferred_coscheduled_sm_count

Preferred co-scheduled SM count; the driver tries to satisfy this but may fall back to coscheduled_sm_count. (Default to None)

Type:

int or Sequence[int], optional

backfill

If True, allow the driver to relax the co-scheduling constraint when assigning SMs. This enables requesting arbitrary aligned SM counts that the driver would otherwise reject due to hardware topology constraints. (Default to False)

Type:

bool or Sequence[bool], optional