task: pin NUMBA_NUM_THREADS=1 in profiling cells + re-measure the DelaunayNN numba Optimal cell

campaigns.yaml

@@ -244,6 +244,19 @@ tasks:
   updated: '2026-10-13'
   next: Unchanged and unissued (PyAutoArray#496 closed 2026-08-27). Select one bounded step when prioritised.
   evidence: https://github.com/PyAutoLabs/PyAutoArray/issues/496
+- id: numba_num_threads_pin_and_delaunay_nn_remeasure
+  campaign: measurement-tools
+  title: Pin NUMBA_NUM_THREADS=1 in profiling cells; re-measure the DelaunayNN numba Optimal cell after PyAutoArray#677 releases
+  status: ready
+  priority: medium
+  path: tasks/numba_num_threads_pin_and_delaunay_nn_remeasure.md
+  updated: '2026-11-05'
+  next: 'Filed 2026-11-05 (human). pin_thread_env omits NUMBA_NUM_THREADS (numba ignores OMP_NUM_THREADS), so the
+    day-zero optimal cell imaging/delaunay_nn/hst numba_cpu_sparse_fp64@hpc_ral_cpu ran its parallel Sibson on up to 8
+    cores (laptop Sibson 14 ms at 8 threads vs 56 ms at 1, about 3-10% of the 1.3 s call). Step 1 (pin, no compute) is
+    ready now; step 2 (re-measure, RAL go) waits for the release carrying PyAutoArray#677; step 3 (day-zero
+    designation) is a human decision.'
+  evidence: https://github.com/PyAutoLabs/PyAutoArray/pull/677
 - id: search_settings_estimation_infrastructure
   campaign: measurement-tools
   title: Search settings-estimation + profiling infrastructure (n_starts / batch_size / n_batch)

dashboard.md

@@ -1,6 +1,6 @@
 # PyAutoPulse — profiling dashboard
 
-<!-- pulse-campaigns:1d875169f93a57e6620c6e3593398f7e518ab029907c025c3c4666ee477ef154 -->
+<!-- pulse-campaigns:0df98084819c074752f8e1899e1f75b939fab981778ea01ac8f75d9f4fdbca13 -->
 ## Profiling Check In
 
 <details><summary>Full check-in prompt</summary>
@@ -130,6 +130,7 @@ Reviewed 2026-10-17. Timing audit #410–#412 and raw HST/MGE pilot producer #41
 - [MGE likelihood_breakdown steps are cumulative and `linear_gaussians` is reported as 0](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/mge_likelihood_breakdown_steps_are_cumulative_an.md) — ready: Unchanged and unissued (no GitHub refs, 2026-10-04). Select one bounded step when measurement-tools is prioritised.
 - [A gradient-cost probe: forward vs `value_and_grad` ms/eval and a strict FD check, on any registry cell](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/gradient_cost_probe.md) — ready: Unissued. New context 2026-10-07: Heart unit-test timing flagged PyAutoLens forward-gradient tests +150%/+96% after PyAutoLens#768 rewrote that test file; not a controlled comparison, and this probe is the right instrument. Select one bounded step when prioritised.
 - [Numba breakdown harness: perturb the instance so the operated-matrix memo cannot hide a step](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/numba_breakdown_harness_memo_blind.md) — ready: Unchanged and unissued (PyAutoArray#496 closed 2026-08-27). Select one bounded step when prioritised.
+- [Pin NUMBA_NUM_THREADS=1 in profiling cells; re-measure the DelaunayNN numba Optimal cell after PyAutoArray#677 releases](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/numba_num_threads_pin_and_delaunay_nn_remeasure.md) — ready: Filed 2026-11-05 (human). pin_thread_env omits NUMBA_NUM_THREADS (numba ignores OMP_NUM_THREADS), so the day-zero optimal cell imaging/delaunay_nn/hst numba_cpu_sparse_fp64@hpc_ral_cpu ran its parallel Sibson on up to 8 cores (laptop Sibson 14 ms at 8 threads vs 56 ms at 1, about 3-10% of the 1.3 s call). Step 1 (pin, no compute) is ready now; step 2 (re-measure, RAL go) waits for the release carrying PyAutoArray#677; step 3 (day-zero designation) is a human decision.
 - [Search settings-estimation + profiling infrastructure (n_starts / batch_size / n_batch)](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/search_settings_estimation_infrastructure.md) — ready: Unchanged and unissued (autolens_profiling#82 closed 2026-08-18). Select one bounded step when prioritised.
 - [jax_compile/probe.py lost its cell builder with the searches tier — give profiling its own](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/jax_compile_probe_needs_own_cell_builder.md) — active: Manual HST/MGE fp64 jit builder #414 and separate raw pilot runner #416 merged. Bounded Mind implementations complete; wider builder coverage/global compile dispatch still unavailable. Agree a next capability or separately authorize a qualified pilot; no production compute authorized. Check-in 2026-10-17: autolens_profiling#418 (pilot qualification design) open since 2026-10-13 with CI green and no review; #417 open; no new measurements.
 - [profile_lens_aggregator.py cannot run from the autolens_workspace_developer root: no config/ directory](https://github.com/PyAutoLabs/PyAutoPulse/blob/main/tasks/profile_lens_aggregator_needs_config_dir.md) — ready: Unchanged: workspace_developer main still has no root config/ (bug presumed live, not re-run). Select one bounded step when prioritised.

tasks/numba_num_threads_pin_and_delaunay_nn_remeasure.md

@@ -0,0 +1,64 @@
+Filed: 2026-11-05 (human request in chat: "yes file the profiling follow-up in Pulse")
+
+# Pin `NUMBA_NUM_THREADS=1` in profiling cells and re-measure the DelaunayNN numba Optimal cell after PyAutoArray#677 is released
+
+Type: feature
+Target: autolens_profiling
+Repos:
+- @autolens_profiling
+- @PyAutoPulse
+Themes:
+- numba-cpu
+- profiling
+Difficulty: small
+Priority: medium
+
+## Finding
+
+`autolens_profiling/_production_config.pin_thread_env` pins `OPENBLAS_NUM_THREADS`, `MKL_NUM_THREADS`,
+`OMP_NUM_THREADS`, `VECLIB_MAXIMUM_THREADS` and `NUMEXPR_NUM_THREADS` to 1, but **not `NUMBA_NUM_THREADS`**.
+numba ignores `OMP_NUM_THREADS`. On the laptop (numba 0.68.1, omp layer), `OMP_NUM_THREADS=1` still gave
+`numba.get_num_threads() == 8` and about 5× CPU per wall-second. numba's default is the node's
+`cpu_count` (252 on RAL); the SLURM cgroup bounds it to the job's CPUs.
+
+Until PyAutoArray#677 (merged 2026-11-05, pending release) the Sibson interpolator was always a numba
+`parallel=True` kernel. So the day-zero cell
+`imaging/delaunay_nn/hst numba_cpu_sparse_fp64@hpc_ral_cpu` (class `optimal`, in
+`snapshots/day_zero/lens.json`; runtime `warm_iid_median_s` 1.295 s, breakdown per call 1.385 s; RAL job
+412955, an 8-CPU array task) ran its Sibson step on up to 8 cores. In production every Nautilus worker is
+one process on one core, and since #677 a worker always runs the serial Sibson build, so the profiled
+number is optimistic.
+
+Size (laptop, Sibson kernel only, 1,280 vertices × 61,600 over-sampled queries, the profiled HST size):
+8 threads 14.1 ms, 4 threads 17.1 ms, 1 thread 56.3 ms. That is about +42 ms per likelihood call for one
+interpolation, roughly 3% of the profiled call. RAL cores are slower and a call may interpolate more than
+once, so the estimate is 3–10%. That is not enough to reorder setups, but the Optimal row should be the
+single-core number.
+
+Not affected:
+- every other numba cell: all other numba kernels are serial by default (`general.yaml numba: parallel: false`);
+- the interferometer `direct_conv` cells: `parallel: false`;
+- the `misc/streaming_scaling` study, which varies threads deliberately;
+- the fixed-light breakdown submits, which already export `NUMBA_NUM_THREADS=1`;
+- the `delaunay-nn-cpu` campaign measurements (PyAutoArray#637), which were 1-thread.
+
+## Steps
+
+1. **Pin (now, no compute):** add `NUMBA_NUM_THREADS` to `_production_config.THREAD_ENV_VARS`, so
+   `configuration.thread_env` records it. It must be set before numba is imported. Cells call
+   `pin_thread_env` above `import numpy`, so check that numba isn't imported earlier.
+2. **Re-measure after the release carrying PyAutoArray#677:** the DelaunayNN numba runtime + breakdown cell
+   on `hpc_ral_cpu` (human go for the RAL job). Mark the 2026.10.25.1 rows `superseded` and keep them.
+3. **Pulse:** decide with the human whether the new row replaces the day-zero designation for that cell (a
+   day-zero re-freeze of one cell) or stands as a post-day-zero current record. Record the decision.
+4. Phase-4 cross-check: the autolens_inference validation leg (RAL 414189, pin + #677) measures wall
+   time with serial Sibson in workers. Its implied per-call can be compared with the step-2 row.
+
+## Evidence
+
+- Failure and fix: PyAutoArray#676 / #677 (numba `prange` kernels serial in multiprocessing workers);
+  Mind record `complete/2026/11/numba-fork-safe-threading.md`.
+- Row provenance: `results/imaging/delaunay_nn/hst/numba_cpu_sparse_fp64/hpc_ral_cpu/breakdown.json`
+  (`configuration.thread_env.vars` lacks `NUMBA_NUM_THREADS`; `device.cpu_count` 252).
+- Human doctrine: numba CPU cost is per process and per core ("one process per core under Nautilus"),
+  and numba parallel threads never run inside multiprocessing workers (2026-11-05).