[shravanngoswamii]

Each latent built a fresh memo for every candidate value, so recovering one draw summed out the rest of the model once per latent.
A memo key already holds the values of the earlier latents the rest of the model depends on, which is why one log density evaluation can share a memo across all its branches.
Recovering a draw drops from about 3 s to 7 ms on Eye Tracking and from about 110 s to 0.05 s on Cervix, with the same draws for a given seed.

[github-actions]

JuliaBUGS.jl documentation for PR #573 is available at:
https://TuringLang.github.io/JuliaBUGS.jl/previews/PR573/

[github-actions]

Benchmark Results

Julia 1.13.4 on AMD EPYC 9V74 80-Core Processor (Linux)

Ratio = JuliaBUGS/Stan (lower is better for JuliaBUGS)

# Params shows Stan / JuliaBUGS when parameter counts differ.

Log Density uses generated evaluation. Gradient columns measure log density + gradient.

McRvs = Mooncake reverse mode with generated evaluation; FwdDiff = ForwardDiff with graph evaluation.

Model	# Params	Log Density Ratio	McRvs Ratio	FwdDiff Ratio
rats	65	0.44x	1.97x	100.97x
pumps	12	1.66x	2.38x	19.41x
bones	13	0.62x	3.96x	0.28x
oxford	244	1.81x	5.40x	699.83x
epil	303	0.50x	1.91x	766.24x
lsat	1006	0.79x	2.55x	2158.94x
schools	133	0.34x	1.71x	126.63x
beetles	2	2.55x	7.00x	10.45x
air	5	1.86x	1.98x	12.57x

Note: Stan benchmarks use hand-optimized Stan models, not direct BUGS translations. Comparison is illustrative only.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.13%. Comparing base (5fd788c) to head (cabddd4).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #573      +/-   ##
==========================================
- Coverage   88.49%   88.13%   -0.37%     
==========================================
  Files          30       30              
  Lines        3851     3851              
==========================================
- Hits         3408     3394      -14     
- Misses        443      457      +14     

Flag	Coverage Δ
callbacks	28.50% <0.00%> (ø)
compilation_model	50.00% <0.00%> (ø)
compilation_source_gen	44.85% <0.00%> (ø)
distributed_sampling	38.55% <0.00%> (ø)
elementary	27.89% <0.00%> (ø)
frontend	47.02% <0.00%> (ø)
gibbs	39.63% <100.00%> (ø)
graphs	21.14% <0.00%> (ø)
inference_chains	37.07% <100.00%> (ø)
inference_flexichains	?
inference_hmc	29.90% <0.00%> (ø)
inference_marginalization	31.10% <100.00%> (ø)
inference_mh	35.46% <100.00%> (ø)
inference_slice_sampling	26.27% <0.00%> (ø)
log_density	59.15% <100.00%> (ø)
model_operations	43.01% <0.00%> (ø)
parallel_sampling	?
params_with_stats	38.18% <100.00%> (ø)
serialization	34.28% <0.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.