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

block_scales_interleave

def block_scales_interleave[scales_dtype: DType, //, *, SF_VECTOR_SIZE: Int, target: StringSlice[ImmStaticOrigin] = StringSlice("cpu")](output_scales_device: TileTensor[scales_dtype, Storage=output_scales_device.Storage, linear_idx_type=output_scales_device.linear_idx_type], input_scales_device: TileTensor[scales_dtype, Storage=input_scales_device.Storage, linear_idx_type=input_scales_device.linear_idx_type], ctx: DeviceContext)

Reinterleaves rank-2 scale factors into the 5D TCGEN layout on SM100 hardware.

Delegates to block_scales_interleave_fp4 after validating that the output scales are rank-5 and the input scales are rank-2.

Parameters:

scales_dtype (DType): Element type of the block scale-factor tensors (inferred).
SF_VECTOR_SIZE (Int): Number of elements covered by each block scale factor: 16 for NVFP4 or 32 for MXFP4.
target (StringSlice[ImmStaticOrigin]): Trace target string identifying the device in profiling output (defaults to "cpu").