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

decode_e2m1_to_f32

def decode_e2m1_to_f32[width: SIMDLength, //](nibble: SIMD[DType.uint16, width]) -> SIMD[DType.float32, width]

Decodes E2M1 nibbles to float32 with branch-free bit arithmetic.

The float32-native twin of decode_e2m1_to_bf16: it builds the float32 bit pattern directly instead of constructing bf16 and widening. All 16 E2M1 values are exactly representable in float32, so the result is bit-identical to E2M1_TO_FLOAT32[nibble] (and therefore to decode_e2m1_to_bf16(nibble).cast[float32]()). Use it on the dequant path where the next step is a float32 scale multiply -- it removes the bf16->f32 widen per element while keeping the bit-exact-vs-table contract.

Construction (float32 layout s | 8-bit exp (bias 127) | 23-bit mantissa): let E = bits2:1 and m = bit0.

Normal (E >= 1): exponent field E + 126, mantissa MSB m (mantissa field m << 22).
Subnormal-class (E == 0): 0.5 * m, i.e. float32 0x3F000000 (+0.5) when m == 1, else 0x00000000. Selected branch-free.
Sign bit (s, nibble bit 3) shifted into float32 bit 31 (s << 28).

Parameters:

width (SIMDLength): SIMD width (lane count) of the nibble vector.

Args:

nibble (SIMD[DType.uint16, width]): One E2M1 nibble per lane in the low 4 bits (0..15).

Returns:

SIMD[DType.float32, width]: The decoded values as SIMD[DType.float32, width], bit-identical to indexing E2M1_TO_FLOAT32[nibble].