FLOPs per Token: the 6N Rule

For a transformer with N non-embedding parameters, the approximate FLOPs to do one forward + backward pass on one token is 6 * N (Kaplan et al. 2020).

Signature: def flops_per_token(n_params: int) -> int

Return 6 * n_params.