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

apple_batched_matmul

def apple_batched_matmul[rank: Int, *, transpose_b: Bool = False, elementwise_epilogue_fn: Optional[def[c_type: DType, width: SIMDLength, rank: Int, *, alignment: Int = Int(1)](IndexList[rank], SIMD[c_type, width]) capturing thin -> None] = None](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], c_shape_idx: IndexList[rank])

Performs a batched single-precision matmul via the Apple Accelerate library.

Iterates over all batch dimensions (all axes except the last two), calling apple_matmul for each 2D slice. Requires at least rank-3 tensors and all operands to be DType.float32.

Parameters:

rank (Int): The rank of the output tensor (must be >= 3).
transpose_b (Bool): Whether to treat B as transposed.
elementwise_epilogue_fn (Optional[def[c_type: DType, width: SIMDLength, rank: Int, *, alignment: Int = Int(1)](IndexList[rank], SIMD[c_type, width]) capturing thin -> None]): Optional rank-aware epilogue applied to each output slice.

Args:

c (TileTensor[Storage=c.Storage, address_space=c.address_space, linear_idx_type=c.linear_idx_type]): Output batched tensor tile.
a (TileTensor[Storage=a.Storage, address_space=a.address_space, linear_idx_type=a.linear_idx_type]): Input A batched tensor tile.
b (TileTensor[Storage=b.Storage, address_space=b.address_space, linear_idx_type=b.linear_idx_type]): Input B batched tensor tile.
c_shape_idx (IndexList[rank]): Shape of the output tensor used to compute batch strides.