add comment

Author: sunxd3

src/abstractmcmc.jl

@@ -31,18 +31,17 @@ getmetric(state::HMCState) = state.metric
 getintegrator(state::HMCState) = state.κ.τ.integrator
 
 function AbstractMCMC.getparams(state::HMCState)
-    # TODO(sunxd): should we return a copy?
     return state.transition.z.θ
 end
 
+# Using @set to update state.transition.z.θ can lead to inconsistencies:
+# - It retains cached log-joint and gradient computations that become invalid
+# - This can cause incorrect evaluations in subsequent steps (e.g. MH)
+# 
+# TODO: adopt https://github.com/TuringLang/MCMCTempering.jl/blob/deb96684496f3fbd011b9f70f28c49a161def23f/ext/MCMCTemperingAdvancedHMCExt.jl#L10-L17
+# if in the future the interface provides access to the log density function
 function AbstractMCMC.setparams!!(state::HMCState, params)
-    hamiltonian = AdvancedHMC.Hamiltonian(state.metric, model)
-    return Setfield.@set state.transition.z = AdvancedHMC.phasepoint(
-        hamiltonian,
-        params,
-        state.transition.z.r;
-        ℓκ = state.transition.z.ℓκ,
-    )
+    return @set state.transition.z.θ = θ
 end
 
 """
@@ -429,4 +428,4 @@ end
 
 function make_kernel(spl::HMCSampler, integrator::AbstractIntegrator)
     return spl.κ
-end
+end