Gibbs wastefully re-evaluates model once per component sampler per iteration

[penelopeysm]

In these lines, component samplers (usually) return a tuple of (Turing.Inference.Transition, SamplerState).

The construction of Turing.Inference.Transition requires a model re-evaluation (because it runs values_as_in_model, and after #2630, also recalculates logp).

However, in the Gibbs sampler we just throw that away:

Turing.jl/src/mcmc/gibbs.jl

Lines 442 to 451 in d0510b1

	# Take initial step with the current sampler.
	_, new_state = step_function(
	rng,
	conditioned_model,
	sampler;
	# FIXME: This will cause issues if the sampler expects initial params in unconstrained space.
	# This is not the case for any samplers in Turing.jl, but will be for external samplers, etc.
	initial_params=initial_params_local,
	kwargs...,
	)

Turing.jl/src/mcmc/gibbs.jl

Lines 678 to 679 in d0510b1

	# Take a step with the local sampler.
	new_state = last(step_function(rng, conditioned_model, sampler, state; kwargs...))

We should either:

(1) Avoid triggering the Transition re-evaluation unless it's needed (not obvious how to do it).
(2) Investigate whether we can use the Transition return value to avoid the re-evaluation in setparams_varinfo!!.

Note that either of these options may require storing the new varinfo itself in the Turing.Inference.Transition struct.

[mhauru]

Note that either of these options may require storing the new varinfo itself in the Turing.Inference.Transition struct.

This could go nicely with TuringLang/AbstractMCMC.jl#149