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

small_MN_gemms

def small_MN_gemms[config: TuningConfigSmallMNGemms, elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None, pdl_level: PDLLevel = PDLLevel()](c: TileTensor[Storage=c.Storage, address_space=c.address_space, linear_idx_type=c.linear_idx_type], a: TileTensor[Storage=a.Storage, address_space=a.address_space, linear_idx_type=a.linear_idx_type], b: TileTensor[Storage=b.Storage, address_space=b.address_space, linear_idx_type=b.linear_idx_type], ctx: DeviceContext)

Launches a small-MN GEMM via the configured split-K GEMV or MMA-CPasync kernel.

Selects between gemm_mma_cpasync (for GEMM_MMA_CPASYNC kernel kind) and gemv_split_k (otherwise) based on config.kernel_kind, then enqueues the chosen kernel with the runtime M, N, K derived from the input tiles.

Parameters:

config (TuningConfigSmallMNGemms): Tuning config selecting the kernel kind (GEMM_MMA_CPASYNC or split-K GEMV) and the tile shapes, thread count, and unroll factor for the launched kernel.
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional epilogue applied to each output element (defaults to None).
pdl_level (PDLLevel): Programmatic dependent launch level for the dispatched kernel (defaults to PDLLevel()).