
MoE auxiliary load-balancing loss (Kamath Ch 2)
Source:R/kamath_llm2_native.R
morie_kamath_moe_load_balance_loss.Rdalpha * N * sum f_i P_i; equals alpha under perfect balance.
Examples
morie_kamath_moe_load_balance_loss(c(0.7, 0.3), c(0.6, 0.4),
2, 0.01)
#> $estimate
#> [1] 0.0108
#>
#> $loss
#> [1] 0.0108
#>
#> $per_expert
#> [1] 0.42 0.12
#>
#> $balanced_floor
#> [1] 0.01
#>
#> $imbalance_ratio
#> [1] 1.08
#>
#> $alpha
#> [1] 0.01
#>
#> $n
#> [1] 2
#>
#> $method
#> [1] "MoE auxiliary load-balancing loss alpha*N*sum f_i P_i"
#>