chains methods claim to be generic but only work on MCMCChains

[penelopeysm]

DynamicPPL.jl/src/pointwise_logdensities.jl

Lines 227 to 229 in 9bd8f16

	function pointwise_logdensities(
	model::Model, chain, ::Type{KeyType}=VarName, ::Val{whichlogprob}=Val(:both)
	) where {KeyType,whichlogprob}

DynamicPPL.jl/src/varinfo.jl

Lines 1837 to 1839 in 9bd8f16

	function setval!(vi::VarInfo, chains::AbstractChains, sample_idx::Int, chain_idx::Int)
	return setval!(vi, chains.value[sample_idx, :, chain_idx], keys(chains))
	end

These functions claim to work either on chain::Any or chain::AbstractMCMC.AbstractChains. In practice the first one calls the second, so the first should only work on AbstractChains. However there is no guarantee that an AbstractChains type can be indexed into using chains.value[iter, :, chain]. This makes assumptions about the internal structure of chains which are only true of MCMCChains.Chains.

DynamicPPL.jl/src/pointwise_logdensities.jl

Lines 260 to 261 in 9bd8f16

	niters = size(chain, 1)
	nchains = size(chain, 3)

There is also no guarantee that size(chain, 1) or size(chain, 3) return the number of iterations or chains respectively. Again this is only true for MCMCChains.

All these functions' signatures should be restricted to MCMCChains and moved to MCMCChainsExt

[penelopeysm]

• pointwise_logdensities should actually use InitContext. I neglected to change this in my series of PRs earlier. Switching this over should let us get rid of more code such as setval!.

[devmotion]

That problem sounded familiar: #170 😄

[penelopeysm]

Aha, yes, it was done correctly for predict but then we must have forgotten about pointwise_... :)

[penelopeysm]

fixed by #1076