Affected Repositories

• autolens_profiling (primary, workspace tier)

Branch Survey

Claim guard: worktree_check_conflict reports autolens_profiling claimed by hst-mge-pilot-qualification (PR #418 open, awaiting merge). This task is registered in Mind planned.md as blocked by it; the A100 measurement runs in RAL scratch (no edit to this repo) and the commit/PR waits for the claim to clear.

Suggested branch: feature/point-source-image-gpu-admission

Implementation Steps

1. RAL: git clone --depth 1 --branch 2026.10.17.1 PyAutoNerves/Fit/Array/Galaxy/Lens into /mnt/ral/jnightin/profiling/PyAuto_wt/ps-image-admission-2026.10.17.1/; clone autolens_profiling main into /mnt/ral/jnightin/profiling/autolens_profiling_wt/ps-image-admission/.
2. New submit hpc/batch_gpu/submit_breakdown_point_source_image_admission_a100_fp64 (derived from the phase-0+1 submit): --exclusive, private PYTHONPATH prepended after activate.sh, refuse unless every library imports from the private base at the tag SHAs, assert GPU backend; legs (a) likelihood_breakdown.py --config-name hpc_ral_a100_fp64_r2026_10_17_1, (b) gpu_bottleneck_map.py fp64 with job 366916's arguments, same config label. No fp32 what-if (not mixed precision).
3. hpc/sync pull; copy JSON/PNG into results/breakdown/point_source_image/, log into results/logs/point_source_image/; build_readme.py, build_catalogue.py, build_dashboard.py from a clean worktree.
4. Journal entry + ledger section; WALL-BASIS block from the measured wall.

Key Files

• scripts/point_source_image/image_plane/likelihood_breakdown.py — shared harness (unchanged)
• scripts/point_source_image/image_plane/gpu_bottleneck_map.py — phase-0+1 cell (unchanged)
• hpc/batch_gpu/submit_breakdown_point_source_image_gpu_bottleneck_map_a100_fp64 — template
• wiki/campaigns/point_source_gpu_breakdown.md, results/notes/point_source_gpu_breakdown_2026_09.md

Point-source image-plane A100 admission measurement on released 2026.10.17.1

Type: research
Target: autolens_profiling
Repos:

• autolens_profiling
  Themes:
• point-source
• profiling
• jax-gpu
  Difficulty: medium
  Autonomy: supervised
  Priority: high
  Status: formalised
  Consequence: judge
  Filed: 2026-10-17
  Campaign: PyAutoPulse point-source-image-gpu / task tasks/point_source_image_plane_gpu_breakdown.md

Request (verbatim, human 2026-10-17)

"Do all that, I authorise all" — A100 compute on RAL for the image-plane admission
measurement of the point-source A100 campaign on the released 2026.10.17.1 stack, and the
development flow to record it. Merge stays human.

Scope (one bounded phase)

Measure whether the released 2026.10.17.1 image-plane point-source likelihood (after the
correctness fix PyAutoLens#768, which closed PyAutoLens#767) is admissible as the campaign's
A100 performance baseline, using the shared breakdown harness
(scripts/point_source_image/image_plane/likelihood_breakdown.py) plus the phase-0+1
bottleneck-map cell for a like-for-like pairing with job 366916:

• full fused likelihood s/eval and the 8-step solver prefix breakdown; compile separate from execution;
• fp64 only (the cell defines no mixed-precision switch);
• one RAL A100, --gres=gpu:1, one task, --exclusive;
• private library checkouts at tag 2026.10.17.1 under /mnt/ral/jnightin/profiling/ (never the shared mirror, never HPCPullPyAuto);
• record host, GPU, driver, JAX/jaxlib, dataset, model configuration and SHAs;
• pair against job 366916 (last qualified pre-#768 A100 fp64 row) only if hardware/precision/dataset match, else "no compatible comparison".

No optimization. Admissibility is stated as evidence only; acceptance of a baseline is the human's call.