CUDA.math_mode!(...; precision) option not working

[carstenbauer]

The precision option of CUDA.math_mode! isn't working.

julia> CUDA.math_mode!(CUDA.FAST_MATH; precision=:Float16)
ERROR: MethodError: Cannot `convert` an object of type typeof(CUDA.math_precision) to an object of type Symbol
Closest candidates are:
  convert(::Type{T}, ::T) where T at /scratch/pc2-mitarbeiter/bauerc/.julia/juliaup/julia-1.7.2+0~x64/share/julia/base/essentials.jl:218
  Symbol(::Any...) at /scratch/pc2-mitarbeiter/bauerc/.julia/juliaup/julia-1.7.2+0~x64/share/julia/base/strings/basic.jl:229
Stacktrace:
 [1] setproperty!(x::CUDA.TaskLocalState, f::Symbol, v::Function)
   @ Base ./Base.jl:43
 [2] math_mode!(mode::CUDA.MathMode; precision::Symbol)
   @ CUDA /scratch/pc2-mitarbeiter/bauerc/.julia/packages/CUDA/t62lT/lib/cudadrv/state.jl:326
 [3] top-level scope
   @ REPL[2]:1
 [4] top-level scope
   @ /scratch/pc2-mitarbeiter/bauerc/.julia/packages/CUDA/t62lT/src/initialization.jl:52

This is CUDA.jl version v3.8.1. At least, this line is broken. I guess it is a function nowadays (for task-localness), so parentheses seem to be missing. However, note that even if that was fixed, the following line only sets default_math_mode and doesn't influence any tasks that already exist (or am I misunderstanding something?). So even with parentheses added we have

julia> CUDA.math_mode!(CUDA.FAST_MATH; precision=:Float16)

julia> CUDA.math_precision()
:TensorFloat32