prompt: pause next-wave for the night — resume plan (GPU 48-start result, SVGD/SV-CMA-ES on RAL)

Author: Jammy2211

active.md

@@ -3,7 +3,9 @@
 
 ## next-wave-population-optimizers
 - issue: https://github.com/PyAutoLabs/autolens_workspace_developer/issues/97
-- status: workspace-dev, Phase-1 done (committed aacdac2, pushed WIP; NOT PR'd — supervised parks ship). CHECKPOINT for user direction. Result: interacting-population needs BOTH diversity AND gradients. cmaes COLLAPSES (r_E7.999, 0/16 — worse than single start); sv_cmaes (Stein repulsion, gradient-free) PREVENTS collapse (r_E2.605, still improving, 0/8 — diverse but slow); svgd (repulsion+gradient = ideal) COMPILE-PROHIBITIVE on CPU (~16x fused grad graphs >10GB) → GPU/HPC. Winner stays multi-start Adam (independence+gradient). Remaining/optional: SVGD-on-GPU, multi-start ADABelief/Lion, jaxns cameo.
+- status: PAUSED for the night 2026-07-13 (user→bed, resume tomorrow). Phase-1 done+committed (aacdac2/197456a, pushed WIP on feature/next-wave-population-optimizers; NOT PR'd — supervised). Result: interacting-population needs BOTH diversity AND gradients. cmaes COLLAPSES (r_E7.999, 0/16); sv_cmaes (Stein repulsion, gradient-free) PREVENTS collapse (r_E2.605, still improving, 0/8, slow); svgd (repulsion+gradient=ideal) CPU-compile-prohibitive → needs GPU. Winner stays multi-start Adam.
+- RESUME: (1) read output/gpu_multi_start_adam_n48_summary.txt — laptop-GPU GIGA-Lens-scaling run (48 starts) was IN PROGRESS at pause (does p_hit→~1?); if it fit 6GB sweep N∈{24,48,96}, else reduce. (2) SVGD + SV-CMA-ES need RAL A100 (laptop PyAutoGPU has blackjax 0.1.0b1 too old + no evosax). RAL not reachable from local (no SSH host; runs ON the RAL login node /mnt/ral/jnightin, user jnightin) & COSMA8 2FA-gated → HUMAN must submit on RAL. Next: confirm RAL checkout layout, write RAL sbatch for searches_minimal/{svgd,cmaes,sv_cmaes}.py. (3) optional multi-start ADABelief/Lion (CPU). Full state: searches_minimal/next_wave_findings.md "GPU/HPC status".
+- also-pending: PR#96 (task jax-gradient-optimizer-benchmark) awaiting human MERGE; on merge this branch rebases onto main.
 - worktree: ~/Code/PyAutoLabs-wt/next-wave-population-optimizers
 - autonomy: supervised (--auto continued from #95 session via "go...do next task" 2026-07-13; experiment cap; ship parks for human sign-off)
 - note: branch feature/next-wave-population-optimizers STACKED on feature/jax-gradient-optimizer-benchmark (PR#96 unmerged) — reuses _grad_setup.py + multi_start_adam.py; REBASE onto main once #96 merges. Reuses #95 MAP harness. Metric = fraction of starts/particles → truth basin (r_E≈1.6) + wallclock + evals vs multi-start Adam baseline.