Computes one update step. `params` holds the current parameter values θ_{t-1}. When `weight_decay` is `0.0` the update is standard Adam; any positive value applies the AdamW decoupled decay term `λ * θ_{t-1}` directly to the parameters, independent of the adaptive scaling. # Panics Panics if `gradients` and `params` have different lengths.
(&mut self, gradients: &[f64], params: &[f64])
source not stored for this graph (policy: none)