Update source_plane regression to Richardson-converged truth#24
Merged
Conversation
EXPECTED_LOG_LIKELIHOOD_SOURCE_PLANE updated to -4491.83220547254 (np and jnp now agree). JIT regression rtol tightened from 2e-3 to 1e-4. Removed annotations referencing the pre-fix np/jnp divergence. Regenerated result JSON/PNG artifacts. Upstream PR: PyAutoLabs/PyAutoGalaxy#358
Merged
3 tasks
Merged
4 tasks
Jammy2211
pushed a commit
that referenced
this pull request
May 11, 2026
Lens light: single Sersic → MGE-60 (60 linear Gaussians) in both
canonical refs (jax_profiling/jit/imaging/{pixelization,delaunay}.py)
and corresponding result artifacts. The MGE columns enter the
inversion mapping matrix, growing F+H from NxN to (N+60)x(N+60).
EXPECTED_LOG_EVIDENCE_HST recomputed:
- Pixelization (1225 + MGE-60): 24746.105672366088 (was 25918.026 single-Sersic)
- Delaunay (1231 + MGE-60): 26288.321397232066 (was 27433.903)
Configs shipped:
- Rectangular: 6 (CPU+GPU x fp64+mp, A100 fp64+mp) — full coverage
- Delaunay: 4 (GPU fp64+mp, A100 fp64+mp) — local CPU still hangs (Task #24)
Companion z_projects/profiling commit on local main updates
_setup_pixelization.py + _setup_delaunay.py to MGE-60, fixes
pixelization_profile.py to use eager-extracted H from inversion (the
old source-only constant_regularization_matrix_from gave a (N,N) shape
that mismatched F's (N+60,N+60) when lens is linear-MGE), and adds
delaunay_vmap_probe.py + rectangular_vmap_probe.py for per-step
vmap/single decomposition.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Jammy2211
added a commit
that referenced
this pull request
May 11, 2026
…231 src + MGE-60 lens) (#60) * profiling: re-run rectangular + Delaunay at ~1250 source pixel science fiducial Bumps mesh sizes: - Rectangular: 28x28 = 784 → 35x35 = 1225 source pixels - Delaunay: 26x26 Overlay = 570 vertices → 39x39 Overlay = 1231 vertices EXPECTED_LOG_EVIDENCE_HST recomputed: - Rectangular: 25918.02569499014 (was 26232.07) - Delaunay: 27433.90296505439 (was 29179.95) Both canonical references in jax_profiling/jit/imaging/{pixelization, delaunay}.py and the result artifacts under jax_profiling/results/jit/ imaging/ have been updated together so canonical and per-config stay aligned at the new fiducial. Configs shipped: - Rectangular: 6 (CPU+GPU x fp64+mp, A100 fp64+mp) — full coverage - Delaunay: 4 (GPU fp64+mp, A100 fp64+mp) — local CPU configs hang local_cpu_fp64 + local_cpu_mp for Delaunay at 1231 vertices both hang indefinitely at full_pipeline_first_call after compile succeeds. Identical futex_wait_queue_me signature for both precisions — extends Task #24 (was thought to be mp-specific; size-related). Rectangular CPU configs at 1225 work fine. PR ships Delaunay without CPU rows. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * profiling: bump rect + Delaunay sweeps to MGE-60 lens fiducial Lens light: single Sersic → MGE-60 (60 linear Gaussians) in both canonical refs (jax_profiling/jit/imaging/{pixelization,delaunay}.py) and corresponding result artifacts. The MGE columns enter the inversion mapping matrix, growing F+H from NxN to (N+60)x(N+60). EXPECTED_LOG_EVIDENCE_HST recomputed: - Pixelization (1225 + MGE-60): 24746.105672366088 (was 25918.026 single-Sersic) - Delaunay (1231 + MGE-60): 26288.321397232066 (was 27433.903) Configs shipped: - Rectangular: 6 (CPU+GPU x fp64+mp, A100 fp64+mp) — full coverage - Delaunay: 4 (GPU fp64+mp, A100 fp64+mp) — local CPU still hangs (Task #24) Companion z_projects/profiling commit on local main updates _setup_pixelization.py + _setup_delaunay.py to MGE-60, fixes pixelization_profile.py to use eager-extracted H from inversion (the old source-only constant_regularization_matrix_from gave a (N,N) shape that mismatched F's (N+60,N+60) when lens is linear-MGE), and adds delaunay_vmap_probe.py + rectangular_vmap_probe.py for per-step vmap/single decomposition. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Jammy2211 <JNightingale2211@gmail.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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
1 participant
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.
Summary
The library-side Richardson extrapolation in
LensCalc.hessian_from(PyAutoGalaxy#358) closes the prior ~0.1% np/jnp log-likelihood divergence in the source-plane point-source pipeline. This PR updates the regression assertion to the new converged truth and tightens the tolerance.Scripts Changed
jax_profiling/point_source/source_plane.py— updatedEXPECTED_LOG_LIKELIHOOD_SOURCE_PLANEto-4491.83220547254(Richardson-converged), tightened JIT regressionrtolfrom2e-3to1e-4, removed annotations referencing the pre-fix np/jnp divergencejax_profiling/point_source/results/{source,image}_plane_summary_v2026.4.13.6.{json,png}— regenerated artifactsUpstream PR
PyAutoLabs/PyAutoGalaxy#358
Test Plan
log_likelihood matches -4491.832205rtol=1e-4:log_likelihood matches -4491.832205🤖 Generated with Claude Code