For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python function

interleaved_block_scales_shape

interleaved_block_scales_shape()

max.nn.moe.interleaved_block_scales_shape(rows, cols)

source

Returns the SM100 interleaved layout’s shape for [rows, cols] scales.

The layout is [rows / 128, cols / 4, 32, 4, 4], the shape block_scales_interleave() produces. Scale row r, column c is stored at [r // 128, c // 4, r % 32, (r % 128) // 32, c % 4], so a range of whole row granules or whole column granules of the interleaved tensor is the interleave of the matching slice of the row-major scales.

Parameters:

• rows (int) – The number of scale rows, one per weight output row.
• cols (int) – The number of scale columns, one per 32-element block.

Returns:

The rank-5 interleaved shape.

Raises:

ValueError – If rows is not a multiple of 128 or cols is not a multiple of 4.

Return type:

list[int]