IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
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).

Mojo function

get_scaling_kind

def get_scaling_kind[a_type: DType, scales_dtype: DType, SF_VECTOR_SIZE: Int, b_type: DType]() -> UMMAKind

Selects the UMMA kind for a same-format block-scaled matmul.

block_scaled_umma_kind plus the scale-factor vector-size check. Rejects the mixed W4A8 pair, since the dispatch paths that call this are same-format only. The kernels that accept W4A8 (the grouped 1D-1D kernel and the dense SM100 entry point) take kind::mxf8f6f4 directly.

Parameters:

  • ​a_type (DType): A operand element type (uint8 for MXFP4/NVFP4, float8_e4m3fn for MXFP8).
  • ​scales_dtype (DType): Scale-factor element type.
  • ​SF_VECTOR_SIZE (Int): Number of elements each scale factor covers.
  • ​b_type (DType): B operand element type; must equal a_type.

Returns:

UMMAKind: The UMMAKind matching the provided type combination.

Was this page helpful?