Slow custom loss function

[Moelf]

using LossFunctions

my_loss_d(ŷ, y) = LossFunctions.deriv(CrossEntropyLoss(), ŷ, y)
my_loss_dd(ŷ, y) = LossFunctions.deriv2(CrossEntropyLoss(), ŷ, y)


xgboost(train_data, my_loss_d, my_loss_dd; tree_method="gpu_hist", objective="binary:logistic", num_round=10)

This is much slower than if not providing my own loss -- I suspect this is because these functions are scalar version?

[Moelf]

XGBoost.jl/src/booster.jl

Lines 387 to 388 in 361022a

	g = convert(Vector{Cfloat}, g)
	h = convert(Vector{Cfloat}, h)

eh, is it because now it's training on CPU?