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

padded_copy_kernel

def padded_copy_kernel[InputLayoutType: TensorLayout, input_origin: ImmOrigin, OutputLayoutType: TensorLayout, output_origin: MutOrigin, dtype: DType, simd_width: Int](input_tensor: TileTensor[dtype, InputLayoutType, input_origin], output_tensor: TileTensor[dtype, OutputLayoutType, output_origin], rows_per_sm: Int, total_rows: Int, row_length: Int)

Copies rows from input_tensor into the unpadded region of output_tensor using a cooperatively vectorized, coalesced row copy.

Each block processes rows_per_sm rows (clamped to total_rows), and threads along the x-dimension cooperate on a single row via _vectorized_copy_row, advancing through rows in steps of block_dim.y. The input is treated as a flat row-major buffer of total_rows rows each of length row_length, while the output offset for each row is derived from output_tensor's layout so the copy lands in the correct position within the padded output.

Parameters:

InputLayoutType (TensorLayout): Layout type of the input TileTensor.
input_origin (ImmOrigin): Origin (mutability) of the input TileTensor.
OutputLayoutType (TensorLayout): Layout type of the output TileTensor.
output_origin (MutOrigin): Origin (mutability) of the output TileTensor.
dtype (DType): Element type of the tensors.
simd_width (Int): SIMD vector width used for the coalesced row copy.

Args:

input_tensor (TileTensor[dtype, InputLayoutType, input_origin]): Source TileTensor of contiguous row-major data.
output_tensor (TileTensor[dtype, OutputLayoutType, output_origin]): Destination TileTensor whose layout maps each row to its padded output position.
rows_per_sm (Int): Maximum rows assigned to a single thread block.
total_rows (Int): Total number of rows to copy.
row_length (Int): Number of elements in each row.