pub static FP8_E4M3_LUT: [f32; 256]Expand description
2026-09-25: FP8 E4M3 -> f32 lookup table, one entry per byte value.
OCP FP8 E4M3FN: sign(1) | exponent(4) | mantissa(3), bias 7. There are no
infinities; 0x7F / 0xFF are NaN and max finite is +/-448.0 (exp 15,
mant 6).
NaN entries decode to signed zero: one bad byte then cannot turn a whole dequantised tensor into NaN.
Built at compile time, so a decode is one indexed load with no branches.