`predict` on vector of varinfo is incorrect if the varinfo passed in is linked

[penelopeysm]

julia> @model f() = x ~ Beta(2, 2)
f (generic function with 2 methods)

julia> m = f(); v = link!!(VarInfo(m), m)
VarInfo{@NamedTuple{x::DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}}, Float64}((x = DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}(Dict(x => 1), [x], UnitRange{Int64}[1:1], [1.3950616230805561], Beta{Float64}[Beta{Float64}(α=2.0, β=2.0)], [0], Dict{String, BitVector}("del" => [0], "trans" => [1])),), Base.RefValue{Float64}(-1.8839487262608983), Base.RefValue{Int64}(0))

julia> v[@varname(x)] # Should always be in (0, 1), even though it's been linked
0.8013990731398695

julia> pdns = predict(m, [v])
1-element Vector{VarInfo{@NamedTuple{x::DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}}, Float64}}:
 VarInfo{@NamedTuple{x::DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}}, Float64}((x = DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}(Dict(x => 1), [x], UnitRange{Int64}[1:1], [1.3950616230805561], Beta{Float64}[Beta{Float64}(α=2.0, β=2.0)], [0], Dict{String, BitVector}("del" => [0], "trans" => [0])),), Base.RefValue{Float64}(-Inf), Base.RefValue{Int64}(0))

julia> pdns[1][@varname(x)]
1.3950616230805561

julia> getlogp(pdns[1])
-Inf

[penelopeysm]

The removal of setval_and_resample! should fix this, just opening it as a separate issue to track it.

[penelopeysm]

Using #967:

julia> using DynamicPPL, Distributions
Precompiling DynamicPPL...
  2 dependencies successfully precompiled in 6 seconds. 122 already precompiled.

julia> @model f() = x ~ Beta(2, 2)
f (generic function with 2 methods)

julia> m = f(); v = link!!(VarInfo(m), m)
VarInfo{@NamedTuple{x::DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}}, DynamicPPL.AccumulatorTuple{3, @NamedTuple{LogPrior::LogPriorAccumulator{Float64}, LogLikelihood::LogLikelihoodAccumulator{Float64}, NumProduce::NumProduceAccumulator{Int64}}}}((x = DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}(Dict(x => 1), [x], UnitRange{Int64}[1:1], [-1.2446528135384165], Beta{Float64}[Beta{Float64}(α=2.0, β=2.0)], [0], Dict{String, BitVector}("del" => [0], "trans" => [1])),), DynamicPPL.AccumulatorTuple{3, @NamedTuple{LogPrior::LogPriorAccumulator{Float64}, LogLikelihood::LogLikelihoodAccumulator{Float64}, NumProduce::NumProduceAccumulator{Int64}}}((LogPrior = LogPriorAccumulator(-1.7100356686475868), LogLikelihood = LogLikelihoodAccumulator(0.0), NumProduce = NumProduceAccumulator(0))))

julia> v[@varname(x)]
0.223627134715461

julia> pdns = predict(m, [v])
1-element Vector{VarInfo{@NamedTuple{x::DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}}, DynamicPPL.AccumulatorTuple{3, @NamedTuple{LogPrior::LogPriorAccumulator{Float64}, LogLikelihood::LogLikelihoodAccumulator{Float64}, NumProduce::NumProduceAccumulator{Int64}}}}}:
 VarInfo{@NamedTuple{x::DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}}, DynamicPPL.AccumulatorTuple{3, @NamedTuple{LogPrior::LogPriorAccumulator{Float64}, LogLikelihood::LogLikelihoodAccumulator{Float64}, NumProduce::NumProduceAccumulator{Int64}}}}((x = DynamicPPL.Metadata{Dict{VarName{:x, typeof(identity)}, Int64}, Vector{Beta{Float64}}, Vector{VarName{:x, typeof(identity)}}, Vector{Float64}}(Dict(x => 1), [x], UnitRange{Int64}[1:1], [0.223627134715461], Beta{Float64}[Beta{Float64}(α=2.0, β=2.0)], [0], Dict{String, BitVector}("del" => [0], "trans" => [0])),), DynamicPPL.AccumulatorTuple{3, @NamedTuple{LogPrior::LogPriorAccumulator{Float64}, LogLikelihood::LogLikelihoodAccumulator{Float64}, NumProduce::NumProduceAccumulator{Int64}}}((LogPrior = LogPriorAccumulator(0.0408619002902344), LogLikelihood = LogLikelihoodAccumulator(0.0), NumProduce = NumProduceAccumulator(0))))

julia> pdns[1][@varname(x)]
0.223627134715461

julia> getlogp(pdns[1])
(logprior = 0.0408619002902344, loglikelihood = 0.0)

[penelopeysm]

Actually found that this is harder than expected. Grr.

DynamicPPL.jl/src/model.jl

Lines 1201 to 1222 in 509f50e

	return map(chain) do params_varinfo
	vi = deepcopy(varinfo)
	# TODO(penelopeysm): BEWARE - Because `ParamsInit` expects unlinked
	# values, this is bugged in the case where `params_varinfo` is linked.
	# This could be solved by using `values_as_in_model`. However, we run
	# into another problem:
	# - The `model` passed into this function will have the target
	# prediction variables (say `y`) set to `missing`.
	# - Calling `values_as_in_model` on this new model will lead to
	# DynamicPPL attempting to extract the values of `y` from the
	# `params_varinfo`, which will fail since `y` will not have been
	# present.
	# I think that the solution is to pass in the ORIGINAL model, but ALSO
	# pass in a set / iterable of VarNames that are to be predicted against.
	# That way, we can:
	# - Safely use `values_as_in_model` to extract a dict of values.
	# - Drop the values of the prediction variables from the dict.
	# - Then use `ParamsInit` to generate the predictions.
	values_dict = values_as(params_varinfo, Dict{VarName,Any})
	_, vi = DynamicPPL.init!!(rng, model, vi, ParamsInit(values_dict, PriorInit()))
	return vi
	end

[penelopeysm]

We agreed at this week's meeting (2025-07-21) to remove the offending method on AbstractVector{<:AbstractVarInfo} and leave predict to work only on MCMCChains.Chains. #984 implements this.

[penelopeysm]

Closed by #984