Repository navigation
fix: avoid redundant Nautilus completion pass #1651
Description
Overview
Nautilus factor searches that converge in their first call still trigger an intermediate result update and a second no-op sampler call. Remove this overhead, following the completed Dynesty EP optimisation in #1642.
Plan
- Recognize convergence immediately while preserving finite-budget updates.
- Cover convergence, chunked execution and likelihood limits with regression tests.
- Compare sampler/update calls, seeded results and timing, then run the full PyAutoFit suite serially.
Detailed implementation plan
- Repository: PyAutoFit; clean main, no competing claim. Existing hook-maintenance branch does not overlap.
- Branch:
feature/ep-nautilus-single-pass. - Inspect the installed Nautilus run completion contract and likelihood counter before selecting the completion condition.
- Update
autofit/non_linear/search/nest/nautilus/search.py(Nautilus.call_search, and budget handling if required). Do not infer likelihood budget exhaustion from posterior sample counts. - Extend
test_autofit/non_linear/search/nest/test_nautilus.pywith converged and finite-cadence cases, global limits and a seeded real-search comparison. No public API change intended. - Validate targeted tests and full serial
test_autofit; measure overhead without changing sampler settings or EP mathematics.
Original user request
We have begun a task speeding up the PyAutoFit source code to make EP run faster without PyautoFit overheads, continue this task
The human approved this bounded follow-up after explanation: “ok, do it.”
Existing prompt
PyAutoMind/draft/bug/autofit/nautilus_converged_run_double_pass.md records the redundant completion pass and the requirement to preserve finite-cadence sampling.
Activity
Jammy2211 commentedon Sep 29, 2026
Implementation is complete in
feature/ep-nautilus-single-pass(base69bb11d54, source edits not yet committed). Nautilus uses its own convergence return, advances cumulative budgets usingn_like, and stops at a global cap even if a batch overshoots it.Validation:
Fresh Heart tick (2026-09-29T19:22:53Z) is RED:
These refer to canonical checkouts; the task branch is based on current PyAutoFit origin/main. Awaiting a human development-only override or cleared Heart before commit/push/PR. No merge or release requested. Detailed local evidence paths and resume instructions are in the Mind prompt's handoff.
Jammy2211 commentedon Sep 29, 2026
The live human authorized the named #1651 development override and same-turn merge: "yeah, and then do a $prm".
Exact Heart RED reasons previously surfaced:
PyAutoFit: 1 commit(s) behind originPyAutoLens: 1 commit(s) behind originPassed branch gates: 2934 tests passed / 2 skipped serially; 16 Nautilus tests passed; 8 autofit workspace scripts + 2 notebooks passed; seeded posterior samples/likelihoods/weights bit-identical. The broader downstream smoke run has resumed. No release or CI/protection bypass is authorized; merge remains conditional on every required CI run and matrix leg passing.
Jammy2211 commentedon Sep 29, 2026
Downstream smoke validation
All eight configured workspace groups passed against this task's PyAutoFit source, using
pyauto-heart smokeand isolated disposable workspace checkouts.CTI environments used the documented no-root GSL extraction into task scratch; no system packages changed. Full log:
.worktrees/ep-nautilus-single-pass/smoke-resume.log; workspace JSON reports:.smoke-root/*/test-results/within that task.Source validation remains 2934 tests passed / 2 skipped; all 8 new regression cases fail against the original methods. The human's named Heart RED development override and same-turn
/prmauthorization remain recorded above; no release is authorized.Jammy2211 commentedon Sep 30, 2026
Shipped. Merged in #1652 (merge 5cf687d, 2026-09-30): a converged single-chunk Nautilus fit no longer runs a second no-op
run()pass with a during_analysis update. Pending the next PyAutoFit release.