Skip to content

active.md

Latest commit

 

History

History
90 lines (80 loc) · 4.48 KB

active.md

File metadata and controls

90 lines (80 loc) · 4.48 KB

subhalo-redshift-jax-fix

  • issue: PyAutoLabs/PyAutoLens#498

  • session: claude (current CLI)

  • status: fix in progress

  • classification: library

  • suggested-branch: feature/subhalo-redshift-jax-fix

  • worktree: ~/Code/PyAutoLabs-wt/subhalo-redshift-jax-fix

  • repos: PyAutoLens

  • repro-pr: PyAutoLabs/autolens_workspace_test#79 (merged d827d1c)

  • summary: | Reported on Slack by @qiuhan96. Free-parameter subhalo redshift (af.UniformPrior) raises TracerBoolConversionError under JAX. Root cause is Python sorted()/<=/== on traced redshifts inside tracer_util.plane_redshifts_from + grid_2d_at_redshift_from.

    Phase 1 done: reproducer landed in autolens_workspace_test as scripts/jax_likelihood_functions/imaging/subhalo.py (PR #79 merged). Two scenarios — fixed z=0.55 PASS (jit/numpy match within rtol=1e-4), free UniformPrior FAIL with the expected TracerBoolConversionError at tracer_util.py:46 (sorted() trips before the explicit <= at 249).

    Phase 2 (next): implement JAX-friendly reformulation of plane_redshifts_from + grid_2d_at_redshift_from in PyAutoLens. Likely approach: sort the concrete-redshift galaxies once, then splice the traced subhalo redshift in via jnp.searchsorted + jax.lax.switch over candidate insertion positions. Once the fix is up the autolens_workspace_test reproducer flips polarity and becomes the regression test.

smoke-test-optimization

  • issue: PyAutoLabs/PyAutoFit#1183
  • session: claude --resume "profile-smoke-test-runtime"
  • status: profiling-and-optimization
  • location: cli-in-progress
  • branch: main
  • repos: all on main (previous PRs merged)
  • summary: | Done: profiled imaging scripts, achieved 80-96% runtime reductions Next: investigate cosmology distance calc, profile interferometer scripts

Completed optimizations

  • PYAUTO_WORKSPACE_SMALL_DATASETS=1 — caps grids/masks to 15x15, forces over_sample_size=2, skips radial bins (PyAutoArray)
  • PYAUTO_DISABLE_JAX=1 — forces use_jax=False in Analysis.init (PyAutoFit)
  • PYAUTO_FAST_PLOTS=1 — skips tight_layout + savefig + critical curve/caustic overlays (PyAutoArray/Galaxy/Lens)
  • Skip print_vram_use(), model.info, result_info, pre/post-fit I/O in test_mode >= 2 (PyAutoFit)
  • Moved test_mode to autoconf (fixes PyAutoArray CI — no autofit dependency)

Profiled scripts

  • imaging/simulator.py: ~100s → 3.6s (96% reduction)
  • interferometer/simulator.py: ~100s → 4.4s
  • imaging/modeling.py: ~100s → 19.5s (80% reduction)

Remaining work — next session

  • Investigate cosmology distance calc (176 calls for 2-plane lens) in subplot_fit_imaging
  • Investigate repeated ray-tracing in subplot panels
  • Profile interferometer/modeling.py and other scripts
  • Consider caching cosmology distances per redshift pair

psf-oversampling

  • issue: PyAutoLabs/PyAutoArray#299

  • session: claude --resume "psf-oversampling"

  • status: parked

  • parked: 2026-05-06

  • classification: library (then workspace follow-up)

  • suggested-branch: feature/psf-oversampling

  • summary: | Parked — no resources claimed. Task worktree was created during /start_library but removed without edits; local feature/psf-oversampling branches deleted from PyAutoArray and PyAutoGalaxy. Both repos are free for other tasks.

    Affected repos (when resumed): - PyAutoArray (library, primary) - PyAutoGalaxy (library) - autolens_workspace_test (workspace follow-up) - autolens_workspace (workspace follow-up)

    To resume: run /start_library — it will recreate the worktree and the feature/psf-oversampling branches off origin/main. Then start with Phase 1 (over_sample_util helpers) per the agreed phasing below.

    Phasing (smaller tasks, agreed mid-session): 1. over_sample_util: Mask2D upscale-by-N + fine->native sum-reduce helpers + tests 2. Convolver: add convolve_over_sample_size kwarg (default 1, no behaviour change) + test 3. Convolver: bin-down branch in all four conv paths, gated > 1 + brute-force test 4. Imaging dataset: kwargs + 2 construction-time guards (adaptive over-sample, sparse) 5. GridsDataset: expose oversampled grids when > 1 6. OperateImage + FitImaging caller threading (PyAutoGalaxy) 7. Inversion mapping audit + assertion (mapping.py / abstract.py) 8. End-to-end library integration test (workspace) extend convolution.py + new convolution_oversampled.py + simulator.py