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

cast_uint_to_fp4e2m1

def cast_uint_to_fp4e2m1[in_dtype: DType, in_width: SIMDLength, //, *, out_dtype: DType, out_width: Int](x: SIMD[in_dtype, in_width]) -> SIMD[out_dtype, out_width]

Unpacks FP4 E2M1 nibbles packed inside unsigned integers to floating-point values.

Each input integer (uint8, uint16, or uint32) holds multiple packed 4-bit E2M1 values. The function extracts each nibble, looks up its float32 value in E2M1_TO_FLOAT32, casts to out_dtype, and assembles the result vector.

Parameters:

in_dtype (DType): Unsigned integer type holding packed FP4 nibbles (uint8, uint16, or uint32).
in_width (SIMDLength): Number of input integer elements.
out_dtype (DType): Output floating-point element type.
out_width (Int): Total output elements; must equal in_width * (bit_width(in_dtype) // 4).

Args:

x (SIMD[in_dtype, in_width]): Packed FP4 E2M1 input vector.

Returns:

SIMD[out_dtype, out_width]: Decoded floating-point vector of length out_width.