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 struct
TuningConfigSM90
struct TuningConfigSM90
Tuning record for a single (M, N, K) SM90 matmul configuration.
Stores the full set of kernel hyperparameters: MMA shape, block tile shape, pipeline stages, cluster shape, consumer count, multicast flag, optional grid shape, scheduling strategy, split-K count, and raster order, corresponding to one entry in the pre-tuned BF16 lookup table.
Fields
- M (
Int): - N (
Int): - K (
Int): - mma_shape (
IndexList[Int(3)]): - block_tile_shape (
IndexList[Int(3)]): - num_pipeline_stages (
Int): - cluster_shape (
IndexList[Int(3)]): - num_consumer (
Int): - partitioned_multicast (
Bool): - grid_shape (
OptionalReg[IndexList[Int(2)]]): - schedule (
MatmulSchedule): - splits (
OptionalReg[Int]): - raster_order (
OptionalReg[RasterOrder]): - dispatch_group (
TuningGroup):
Implemented traits
AnyType,
Copyable,
Deinitable,
ImplicitlyCopyable,
Movable,
RegisterPassable,
TrivialRegisterPassable,
TuningConfig,
Writable
Methods
__init__
def __init__(M: Int, N: Int, K: Int, mma_shape: IndexList[Int(3)], block_tile_shape: IndexList[Int(3)], num_pipeline_stages: Int, cluster_shape: IndexList[Int(3)], num_consumer: Int, partitioned_multicast: Bool, grid_shape: OptionalReg[IndexList[Int(2)]] = None, schedule: MatmulSchedule = MatmulSchedule.NONE, splits: OptionalReg[Int] = None, raster_order: OptionalReg[RasterOrder] = None, dispatch_group: TuningGroup = TuningGroup.CORE) -> Self
write_to
def write_to(self, mut writer: T)
Writes the tuning config as a string.
Args:
- writer (
T): The writer to write to.