ADTypes.AutoMooncake(; config=nothing) doesn't work

[penelopeysm]

Initialising ADTypes.AutoMooncake(; config=nothing) doesn't seem to use a config of Mooncake.Config() by default.

The following MWE uses a feature branch of Turing.jl (TuringLang/Turing.jl#2341):

using Turing
import Mooncake
backend = Turing.AutoMooncake(; config=nothing)  # reexported from ADTypes
@model f() = x ~ Normal(0, 1)
sample(f(), HMC(0.1, 3; adtype=backend), 10)

# works with backend = Turing.AutoMooncake(; config=Mooncake.Config())

The error:

ERROR: type Nothing has no field debug_mode
Stacktrace:
  [1] getproperty(x::Nothing, f::Symbol)
    @ Base ./Base.jl:37
  [2] ADgradient(x::AutoMooncake{…}, ℓ::LogDensityFunction{…})
    @ MooncakeLogDensityProblemsADExt ~/.julia/packages/Mooncake/eEGqO/ext/MooncakeLogDensityProblemsADExt.jl:63
  [3] ADgradient(ℓ::LogDensityFunction{DynamicPPL.TypedVarInfo{…}, DynamicPPL.Model{…}, DynamicPPL.SamplingContext{…}})
    @ Turing.Inference ~/ppl/lib/src/mcmc/Inference.jl:154
  [4] initialstep(rng::Random.TaskLocalRNG, model::DynamicPPL.Model{…}, spl::DynamicPPL.Sampler{…}, vi_original::DynamicPPL.TypedVarInfo{…}; initial_params::Nothing, nadapts::Int64, kwargs::@Kwargs{})
    @ Turing.Inference ~/ppl/lib/src/mcmc/hmc.jl:159
[...]

which points to at least this:

Mooncake.jl/ext/MooncakeLogDensityProblemsADExt.jl

Lines 61 to 72 in 942eead

	# Interop with ADTypes.
	function ADgradient(x::ADTypes.AutoMooncake, ℓ)
	debug_mode = x.config.debug_mode
	if debug_mode
	msg = "Running Mooncake in debug mode. This mode is computationally expensive, " *
	"should only be used when debugging a problem with AD, and turned off in " *
	"general use. Do this by using " *
	"AutoMooncake(; config=Mooncake.Config(debug_mode=false))."
	@info msg
	end
	return ADgradient(Val(:Mooncake), ℓ; debug_mode)
	end