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

build_sm100_matmul_configs

def build_sm100_matmul_configs[a_type: DType, b_type: DType, c_type: DType, N: Int, K: Int, transpose_b: Bool = True, has_bias: Bool = False]() -> Set[FusedSwiGLUMatmulConfig[a_type, b_type, c_type, transpose_b]]

Build FusedSwiGLUMatmulConfigs for SM100 fused GEMM+SwiGLU.

For (N, K) shapes covered by the SwiGLU tuning table each entry's register_swiglu flag selects the epilogue path:

True (register→GMEM): no extra SMEM overhead.
False (SMEM→TMA): pipeline stages reduced for the double-buffered half-output tiles.

For untuned (N, K) shapes returns an empty set; dispatch falls back to its safety-net config.

Parameters:

a_type (DType): DType of the A operand elements.
b_type (DType): DType of the B operand elements.
c_type (DType): DType of the C output elements.
N (Int): Number of columns in the output C matrix.
K (Int): Contraction dimension of the GEMM (shared inner size of A and B).
transpose_b (Bool): Whether the B operand is stored transposed (defaults to True).
has_bias (Bool): Whether a bias vector is applied in the epilogue (defaults to False).

Returns:

Set[FusedSwiGLUMatmulConfig[a_type, b_type, c_type, transpose_b]]