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_f4e2m1x2_to_fp16x2

def cast_f4e2m1x2_to_fp16x2(x: UInt8) -> SIMD[DType.float16, SIMDLength(2)]

Converts two FP4 E2M1 nibbles packed in one byte to two float16 values using SM100 PTX.

Issues the cvt.rn.f16x2.e2m1x2 PTX instruction to decode both nibbles in a single hardware operation.

Constraints:

Requires NVIDIA GPU with SM100 or newer (B200 and above).

Args:

x (UInt8): A uint8 containing two packed FP4 E2M1 nibbles (low nibble = element 0).

Returns:

SIMD[DType.float16, SIMDLength(2)]: A two-element float16 SIMD vector with the decoded values.