Repository navigation
InnerLayer.GGJ - BUGFIX - Accept profile-only options in the :ray solve, warn on very large ray grids, and use a relative ggj_reference floor - #595
Merged
Conversation
… solve and warn on very large ray grids
A GGJModel{:ray} built with npc or certify_rtol threw a MethodError in solve_inner, which
forwarded them to solve_ray. They are now accepted and ignored there, as real_profiles already
was. solve_ray warns when the Hermite grid exceeds warn_cells (default 6000), and its docstring
states that the grid is fixed rather than refined against a residual, with the expected accuracy
and how to get an error bar.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…es and use it for ggj_reference A quantity may set relative_threshold, a fraction of the first ref's magnitude below which a change is noise, alongside the absolute noise_threshold. It is stored with each run (an older database gains the column on open). ggj_reference's Δ values are up to ~2e4, where the absolute 1e-10 floor was ~5e-15 relative, below cross-machine rounding; they now use 1e-10 relative. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Collaborator
|
@d-burg what problems are you getting large grids on? The inward march was designed to solve needing huge galerkin grids to get representative frobenius series at the edge cells, IME I have only had huge BVP grids needed when Q > ~100, which should be pretty huge for something like DIII-D no? Just curious what region of parameter space this is in |
Collaborator
Author
|
@matt-pharr Nothing in practice — this is the "warn on a large cell count" item from your #536 reply. I measured it on the Glasser–Wang and q4 benchmark surfaces: 1,024–2,700 cells for |Q| from 0.01 to 5000i, and only Glasser–Wang at Q = 5000i reaches 7,418. So it agrees with your experience, and the 6000 threshold only fires deep in that regime. |
… a large ray grid and leave profile-only options untyped Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Collaborator
|
@d-burg — completed the requested full SDCC fanout against develop Result: no changes in comparable quantities; harness exit code 0, with no run failures.
Environment: Julia 1.12.6, x86_64 Linux, CI manifest ./regress_fanout.sh --refs 7817f0e7a,2dc99e1c8 --harness-ref 2dc99e1c8 --name pr595_2dc99e1c8_vs_7817f0e7a_20261028SDCC report: Full regression report (21 cases)Third-party human review and approval are required before merge. |
Collaborator
Author
|
@matt-pharr on the relative floor for |
d-burg
marked this pull request as ready for review
October 29, 2026 01:26
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Release note
Follow-ups to #536. A
GGJModel(; solver=:ray)built with the profile optionsnpcorcertify_rtolthrew aMethodErrorwhen solved, becausesolve_innerforwarded them tosolve_ray. They are now ignored there, asreal_profilesalready was.solve_raynow warns once when its Hermite grid exceedswarn_cells(default 6000), and its docstring states that the grid is fixed rather than refined against a residual, with the expected accuracy and how to get an error bar.Changes
solve_inner(::GGJModel{:ray})accepts and ignoresreal_profiles,npcandcertify_rtol, the profile-only options ofsolve_inner_profile, as untyped keywords so they do not look like they act. Test: a model built with all three gives the bare model's Δ exactly.solve_raytakeswarn_cells(default 6000) and warns once perwarn_cellsvalue when the grid is larger, namingcfl,min_cellsand the local rates that set it; the solve itself is unchanged. Test:warn_cells = 100warns, and Δ is unchanged.solve_raydocstring: the mesh is fixed by the local rates (cfl,min_cells), no run-time error estimate is returned, and Δ is accurate to about 1e-8–2e-7 relative at small |Q|. It points todelta_convergencefor an error bar. I didn't add a residual field toRaySolveResult, becausedelta_convergencealready gives that error bar.Regression report
Latest SDCC fanout rerun: develop
7817f0e7avs2dc99e1c8— 21 cases; no changed comparable quantities; harness exit 0. Full report and limitations.gal_resistive_pehas all eight quantities missing on both refs and provides no numerical validation. Earlier reports are retained below.runtests_innerlayer.jlpasses on feynman (GGJ 10, :ray backend 10, :ray internals 15, profile interface 24). The report below is the earlier run at98e9167b9.regress --cases ggj_reference,ggj_ray_q500i,gal_resistive_diiid_galinner,diiid_slayer_n1 --refs bca7ce751,98e9167b9 --forceon feynman (SLURM, 8 threads, manifest pinned): every case is unchanged. The tests inruntests_innerlayer.jlandruntests_slayer_runner.jlpass (GGJ 10, :ray backend 10, :ray internals 15, profile interface 24, runner 88).ggj_referenceggj_ray_q500igal_resistive_diiid_galinnerdiiid_slayer_n1Notes for reviewers
No relative noise floor. InnerLayer.GGJ - REFACTOR! - Solve the :ray BVP with the Galerkin assembly and use one parity convention #536 asked for
ggj_reference's floor to be relative. Measured first: feynman reproduces all four Δ values to the last bit across runs, and the harness always compares its two refs on the same machine, so develop's absolute 1e-10 floor raises no false alarms and stays as it is. Across machines the values do differ (Mac arm64 / Julia 1.11.6 vs feynman x86 / 1.11.9, same pinned manifest):The noise follows |Δ₊| ≈ 2.07e4 (4.8e-9 relative) rather than each part, so a relative floor of 1e-10 on each part would still flag Re Δ₊. The measurement is recorded in Regression - Read noise floors from the case file at compare time #607 for when cross-machine comparison is wanted.
Develop merged (
526141d4a), no conflicts.🤖 Generated with Claude Code