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
pack_transposed_b_ndbuffer
def pack_transposed_b_ndbuffer[b_type: DType, //, a_type: DType, c_type: DType](b_input: TileTensor[b_type, Storage=b_input.Storage, linear_idx_type=b_input.linear_idx_type], output_buffer: TileTensor[b_type, Storage=output_buffer.Storage, linear_idx_type=output_buffer.linear_idx_type], kernel_type_m: Int = Int(0))
TileTensor primary implementation of pack_transposed_b_ndbuffer.
Takes kernel_type_m directly instead of extracting it from a_shape
static shape params (0 = dynamic M).
Parameters:
- b_type (
DType): Element type of the B operand being packed (inferred). - a_type (
DType): Element type of the A operand of the matmul. - c_type (
DType): Element type of the C output of the matmul.
Args:
- b_input (
TileTensor[b_type, Storage=b_input.Storage, linear_idx_type=b_input.linear_idx_type]): Read-only rank-2TileTensorcontaining the original B matrix to pack, stored as[N, K]. - output_buffer (
TileTensor[b_type, Storage=output_buffer.Storage, linear_idx_type=output_buffer.linear_idx_type]): Pre-allocated mutable rank-2TileTensorthat receives the packed B matrix. - kernel_type_m (
Int): M dimension used to select the matmul kernel variant (defaults to 0, meaning dynamic M).
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!