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 function

default_config_sm90

def default_config_sm90[a_type: DType, b_type: DType, c_type: DType, transpose_b: Bool, wgmma_shape: IndexList[Int(3)]]() -> MatmulConfig[a_type, b_type, c_type, transpose_b]

Returns the default SM90 matmul config for the given WGMMA shape.

Sets BM=128, BN from the WGMMA N dimension, BK to the maximum TMA-aligned value, 4 pipeline stages, 2 consumer warp groups, and no multicast.

Parameters:

a_type (DType): A-matrix (activations) element type; drives BK via TMA alignment.
b_type (DType): B-matrix (expert weights) element type.
c_type (DType): Output element type.
transpose_b (Bool): Whether B is stored transposed.
wgmma_shape (IndexList[Int(3)]): WGMMA instruction shape (M, N, K); N sets the block tile BN.

Returns:

MatmulConfig[a_type, b_type, c_type, transpose_b]: A MatmulConfig suitable for use as the default grouped matmul config.