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_float_to_fp4e2m1_amd

def cast_float_to_fp4e2m1_amd[dtype: DType, width: SIMDLength, //](input: SIMD[dtype, width], scale: Float32) -> UInt32

Converts up to eight floating-point values to packed FP4 E2M1 using AMD CDNA4 intrinsics.

Applies the provided scale factor before quantization using the llvm.amdgcn.cvt.scalef32.pk.fp4.* intrinsics, packing pairs of values per call.

Constraints:

Requires AMD CDNA4 or newer (MI355X and above).

Parameters:

dtype (DType): Input element type (bfloat16 or float32).
width (SIMDLength): Number of input elements; must be even and at most 8.

Args:

input (SIMD[dtype, width]): Input floating-point vector to convert.
scale (Float32): Scale factor applied before FP4 quantization.

Returns:

UInt32: A UInt32 with the converted FP4 E2M1 nibbles packed in order.