maint: vram-first validation — probe override, record-and-flag pins (polish phase 2)

.github/workflows/lint.yml

@@ -2,8 +2,8 @@
 #
 # Cheap, fast (<5 min), CPU-only. Runs on every PR and on push to main.
 # Does NOT produce real result artifacts — the smoke step short-circuits
-# every profile script via AUTOLENS_PROFILING_SMOKE=1 (lands as part of
-# Phase 5; see scripts under likelihood/, simulators/, searches/nautilus/).
+# every profile script via AUTOLENS_PROFILING_SMOKE=1 (see scripts under
+# likelihood_runtime/, likelihood_breakdown/, simulators/, searches/nautilus/).
 #
 # Profile re-runs that actually produce results live in `profile.yml`.
 
@@ -60,12 +60,15 @@ jobs:
 
       - name: Install runtime dependencies
         run: |
-          python -m pip install --upgrade pip
-          pip install numpy scipy matplotlib jax jaxlib
-          pip install nautilus-sampler==1.0.5
-          pip install numba
-          # autoconf / autofit / autoarray / autogalaxy / autolens are imported
-          # from the sibling checkouts via PYTHONPATH below — no install needed.
+          python -m pip install --upgrade pip setuptools wheel
+          # Install each sibling checkout in dependency order (the PyAutoHeart
+          # lib-tests.yml pattern): satisfies the full third-party dep chain
+          # (pyyaml, astropy, jax, numba, ...) without hand-maintaining a
+          # package list. PYTHONPATH below still resolves the PyAuto* imports
+          # from source, so the checkouts stay authoritative.
+          for r in PyAutoConf PyAutoFit PyAutoArray PyAutoGalaxy PyAutoLens; do
+            pip install "./$r[optional]"
+          done
 
       - name: Install lint tools
         run: |
@@ -92,9 +95,11 @@ jobs:
         working-directory: autolens_profiling
         run: |
           # Conservative scope: only README.md files in this repo, exclude
-          # GitHub UI links that lychee tends to false-positive on.
+          # GitHub UI links that lychee tends to false-positive on, and the
+          # Slack invite (302 -> 403 to any non-browser client).
           lychee \
             --exclude '^https://github\.com/.*/(issues|pull|commit|releases)/' \
+            --exclude '^https://join\.slack\.com/' \
             --no-progress \
             --accept '200..=299,429' \
             --max-redirects 5 \
@@ -111,6 +116,7 @@ jobs:
           # Each script reads AUTOLENS_PROFILING_SMOKE at module top and
           # exits 0 immediately after the import + setup section. Catches
           # import-graph breakage without running the full profile.
-          python likelihood/imaging/mge.py
+          python likelihood_runtime/imaging/mge.py
+          python likelihood_breakdown/imaging/mge.py
           python simulators/imaging.py
-          python searches/nautilus/simple.py
+          python searches/nautilus/imaging/mge.py

.github/workflows/profile.yml

@@ -92,7 +92,7 @@ jobs:
           pip install nautilus-sampler==1.0.5
           pip install numba
 
-      - name: Run likelihood/ scripts
+      - name: Run likelihood_runtime/ scripts
         if: ${{ inputs.sections == '' || contains(inputs.sections, 'likelihood') }}
         working-directory: autolens_profiling
         continue-on-error: true
@@ -102,8 +102,8 @@ jobs:
           PYTHONPATH: ${{ github.workspace }}/PyAutoConf:${{ github.workspace }}/PyAutoFit:${{ github.workspace }}/PyAutoArray:${{ github.workspace }}/PyAutoGalaxy:${{ github.workspace }}/PyAutoLens
         run: |
           set +e  # don't abort the whole step on one script's failure
-          for script in likelihood/imaging/*.py likelihood/interferometer/*.py \
-                        likelihood/point_source/*.py likelihood/datacube/delaunay.py; do
+          for script in likelihood_runtime/imaging/*.py likelihood_runtime/interferometer/*.py \
+                        likelihood_runtime/point_source/*.py likelihood_runtime/datacube/delaunay.py; do
             echo "::group::$script"
             python "$script" || echo "::warning::$script failed — see log; dashboard cell will show ERR"
             echo "::endgroup::"
@@ -163,7 +163,7 @@ jobs:
           # gitignored; future hardware-tier work may flip a subset to
           # tracked, in which case this `add -f` becomes redundant but
           # harmless.
-          git add -f results/ likelihood/README.md likelihood/*/README.md \
+          git add -f results/ likelihood_runtime/README.md likelihood_breakdown/README.md \
                      simulators/README.md searches/README.md \
                      searches/nautilus/README.md README.md
           git commit -m "chore: refresh profile artifacts + dashboard tables [skip ci]

.gitignore

@@ -46,12 +46,10 @@ dataset/datacube/sim_*/
 # that's purely an auto-cache.
 dataset/**/lensed_source.fits
 
-# Profiling result artifacts — written by scripts under likelihood/, simulators/, searches/.
-# Each run produces fresh versioned JSON+PNG; Phase 4's dashboard will explicitly select
-# which artifacts to commit. Until then, results stay local.
+# Retired legacy results section (pre likelihood_runtime/breakdown split);
+# nothing writes here since the phase-2 output retarget. Kept ignored so old
+# local checkouts don't surface stale artifacts as untracked.
 results/likelihood/
-results/simulators/
-results/searches/
 
-# Latent profiling sweep outputs — regenerated by latent/sweep.py, not committed.
+# Retired legacy latent output home (latent now writes results/latent/, in-repo).
 latent/results/

_profile_cli.py

@@ -257,3 +257,50 @@ def auto_simulate_if_missing(
         ],
         check=True,
     )
+
+
+def check_pinned(got, expected, *, label: str, rtol: float = 1e-4):
+    """Compare a computed likelihood/evidence against its pinned baseline.
+
+    Profiling runs **record and flag** drift; they never adjudicate library
+    correctness (that is autolens_workspace_test's remit — see
+    ``results/notes/design_lock_in.md``). Returns ``None`` when ``got`` is
+    within ``rtol`` of ``expected``; otherwise prints a loud warning and
+    returns a drift record for the result JSON, which PyAutoHeart's vitals
+    scan picks up. Never raises — a changed computation must not kill a
+    profiling job (the timing numbers are still data; the flag marks them
+    non-comparable to the pinned baseline).
+    """
+    got_f = float(got)
+    rel = abs(got_f - expected) / max(abs(expected), 1e-300)
+    if rel <= rtol:
+        return None
+    print(
+        f"  WARNING: PINNED-VALUE DRIFT [{label}] — got {got_f!r}, "
+        f"pinned {expected!r} (rel diff {rel:.3e} > rtol {rtol:g}). "
+        f"Timings from this run are NOT comparable to the pinned baseline; "
+        f"file a bug / check autolens_workspace_test before trusting trends."
+    )
+    return {
+        "label": label,
+        "expected": expected,
+        "got": got_f,
+        "rel_diff": rel,
+        "rtol": rtol,
+    }
+
+
+def record_pinned_check(json_path, expected, drift_records) -> None:
+    """Merge the pinned-value check outcome into an already-written result JSON.
+
+    Adds ``pinned_expected`` (the baseline value, or ``None`` when the
+    instrument has no pin) and ``pinned_drift`` (list of drift records —
+    empty means every compared value matched). The guard blocks run after
+    the summary JSON is written, so this rewrites the file in place.
+    """
+    import json
+
+    data = json.loads(json_path.read_text())
+    data["pinned_expected"] = expected
+    data["pinned_drift"] = drift_records
+    json_path.write_text(json.dumps(data, indent=2))

dataset/imaging/euclid/positions.json

@@ -0,0 +1,28 @@
+{
+    "type": "instance",
+    "class_path": "autoarray.structures.grids.irregular_2d.Grid2DIrregular",
+    "arguments": {
+        "values": {
+            "type": "ndarray",
+            "array": [
+                [
+                    0.27109374999999997,
+                    1.6729626159565016
+                ],
+                [
+                    -1.49921875,
+                    0.1953067707493031
+                ],
+                [
+                    1.49921875,
+                    -0.1953067707493031
+                ],
+                [
+                    -0.27109374999999997,
+                    -1.6729626159565019
+                ]
+            ],
+            "dtype": "float64"
+        }
+    }
+}

dataset/imaging/euclid/tracer.json

@@ -0,0 +1,110 @@
+{
+    "type": "instance",
+    "class_path": "autolens.lens.tracer.Tracer",
+    "arguments": {
+        "galaxies": {
+            "type": "list",
+            "values": [
+                {
+                    "type": "instance",
+                    "class_path": "autogalaxy.galaxy.galaxy.Galaxy",
+                    "arguments": {
+                        "redshift": 0.5,
+                        "label": null,
+                        "bulge": {
+                            "type": "instance",
+                            "class_path": "autogalaxy.profiles.light.standard.sersic.Sersic",
+                            "arguments": {
+                                "centre": {
+                                    "type": "tuple",
+                                    "values": [
+                                        0.0,
+                                        0.0
+                                    ]
+                                },
+                                "sersic_index": 3.0,
+                                "intensity": 2.0,
+                                "ell_comps": {
+                                    "type": "tuple",
+                                    "values": [
+                                        0.05263157894736841,
+                                        3.2227547345982974e-18
+                                    ]
+                                },
+                                "effective_radius": 0.6
+                            }
+                        },
+                        "mass": {
+                            "type": "instance",
+                            "class_path": "autogalaxy.profiles.mass.total.isothermal.Isothermal",
+                            "arguments": {
+                                "einstein_radius": 1.6,
+                                "ell_comps": {
+                                    "type": "tuple",
+                                    "values": [
+                                        0.05263157894736841,
+                                        3.2227547345982974e-18
+                                    ]
+                                },
+                                "centre": {
+                                    "type": "tuple",
+                                    "values": [
+                                        0.0,
+                                        0.0
+                                    ]
+                                }
+                            }
+                        },
+                        "shear": {
+                            "type": "instance",
+                            "class_path": "autogalaxy.profiles.mass.sheets.external_shear.ExternalShear",
+                            "arguments": {
+                                "gamma_1": 0.05,
+                                "gamma_2": 0.05
+                            }
+                        }
+                    }
+                },
+                {
+                    "type": "instance",
+                    "class_path": "autogalaxy.galaxy.galaxy.Galaxy",
+                    "arguments": {
+                        "redshift": 1.0,
+                        "label": null,
+                        "bulge": {
+                            "type": "instance",
+                            "class_path": "autogalaxy.profiles.light.standard.sersic_core.SersicCore",
+                            "arguments": {
+                                "gamma": 0.25,
+                                "centre": {
+                                    "type": "tuple",
+                                    "values": [
+                                        0.0,
+                                        0.0
+                                    ]
+                                },
+                                "radius_break": 0.025,
+                                "sersic_index": 1.0,
+                                "intensity": 4.0,
+                                "alpha": 3.0,
+                                "ell_comps": {
+                                    "type": "tuple",
+                                    "values": [
+                                        0.0962250448649376,
+                                        -0.05555555555555551
+                                    ]
+                                },
+                                "effective_radius": 0.1
+                            }
+                        }
+                    }
+                }
+            ]
+        },
+        "cosmology": {
+            "type": "instance",
+            "class_path": "autogalaxy.cosmology.model.Planck15",
+            "arguments": {}
+        }
+    }
+}

hpc/batch_gpu/submit_probe_delaunay_a100

@@ -0,0 +1,55 @@
+#!/bin/bash -l
+#
+# Probe-only sweep — DELAUNAY cells (10-30 min per XLA compile; scipy
+# Delaunay via pure_callback bloats the graph, so probes are single-point
+# per the vram/README.md methodology).
+#
+# Re-validates the vram/config.py batch-size table for the
+# PreOptimizationTimes campaign (autolens_profiling#54). Each probe writes
+#   results/runtime/<class>/delaunay/vmap_probe_delaunay[_sparse].json
+# and exits before any timing loop. See submit_probe_fast_a100 for the
+# fast cells and vram/README.md "Probe-only submits" for the ingest step.
+
+#SBATCH -J probe_delaunay_a100
+#SBATCH --partition=gpu
+#SBATCH --gres=gpu:1
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=4
+#SBATCH --mem=64gb
+#SBATCH --time=8:00:00
+#SBATCH -o output/output.%A.out
+#SBATCH -e error/error.%A.err
+#SBATCH --mail-type=END,FAIL
+#SBATCH --mail-user=james.w.nightingale@durham.ac.uk
+
+export AP_ROOT=/mnt/ral/jnightin/autolens_profiling
+source $AP_ROOT/activate.sh
+
+export JAX_PLATFORM_NAME=cuda
+export JAX_PLATFORMS=cuda,cpu
+export XLA_PYTHON_CLIENT_PREALLOCATE=false
+export JAX_ENABLE_X64=True
+export NUMBA_CACHE_DIR=/tmp/numba_cache
+export MPLCONFIGDIR=/tmp/matplotlib
+
+nvidia-smi
+cd $AP_ROOT
+
+run_probe () {
+    echo "=== probe: $1 --instrument $2 $3 ==="
+    date
+    python3 "$1" --instrument "$2" --vmap-probe $3 || echo "!!! probe FAILED: $1 $2 $3"
+}
+
+# Imaging delaunay — campaign instruments, dense + sparse.
+for inst in ao jwst hst; do
+    run_probe likelihood_runtime/imaging/delaunay.py "$inst" ""
+    run_probe likelihood_runtime/imaging/delaunay.py "$inst" "--sparse"
+done
+
+# Interferometer delaunay — campaign anchor (alma_high) + the cheap sanity cell.
+run_probe likelihood_runtime/interferometer/delaunay.py sma ""
+run_probe likelihood_runtime/interferometer/delaunay.py alma_high ""
+
+echo "Finished."
+date