Oh open an issue for that. Seems like the way we automatically generate the functions doesn’t play nicely with CUDA.jl. So then you’d have to do the function generation completely manually, i.e.:
sys_f = eval(generate_function(sys)[2])
sys_jac = eval(generate_jacobian(sys)[2])
sys_tgrad = eval(generate_tgrad(sys)[2])
func = ODEFunction(sys_f ,jac=sys_jac,tgrad=sys_tgrad)