pub fn dequantize_gradient(
quantized: &[i64],
scale: f64,
modulus: i64,
) -> Vec<f64>Expand description
Inverse of quantize_gradient applied to a vector of canonical
[0, modulus) representatives.
Values strictly greater than modulus / 2 are interpreted as negative
fixed-point numbers (the standard two’s-complement style centring around
zero), then scaled back to f64.