Repository navigation
research: RAL route-gate constants and A100 FFT size levers (interferometer) #429
Description
Overview
The Aris L2 witness (PyAutoArray#636, autolens_profiling#425) found the laptop-calibrated route constants (c_fft 7.62e-10, c_dc 1.77e-09, PyAutoArray#632) off on RAL: FFT / full direct_conv measured 0.68 at the real-Aris 1,459-vertex cell against 0.89 predicted. This task measures both curvature arms on a RAL CPU node to fit per-host constants, and runs the A100 FFT size-lever witness from PyAutoPulse task interferometer_w_tilde_fft_size_levers (pruned padded FFT; real-space pixel scale at jvla).
Campaign: PyAutoLabs/PyAutoPulse#62 (tasks interferometer_large_mask_aris and interferometer_w_tilde_fft_size_levers). Human authorised RAL A100 + CPU compute for this phase on 2026-10-17.
Plan
A. Cost-model constants on RAL (CPU)
- Time the FFT route, the full
direct_convkernel and the symmetric kernel (curvature step only, compile excluded) on real Aris data, 512×512 mask (steps 12/8/6/4) and 125×125 mask (steps 3/2/1),ralpartition,--exclusive, serial numba, on a private 2026.10.17.1 checkout plus the PyAutoArray#636 patch. - Fit per-host
c_fft,c_dc(κ kept unless contradicted); table predicted vs measured for every cell, in-sample vs out-of-sample marked. - Recommend: keep laptop constants, replace with RAL constants, or expose them as a settings/config override with documented per-host values. A library change (a stacked PyAutoArray PR on #636) only if the data clearly supports one option inside the existing cost-model design.
B. A100 FFT size levers (measurement only)
- Harness-only pruned padded FFT apply (separable rfft-x on y rows, fft-y, Khat, ifft-y keep y rows, irfft-x keep x) vs library F at alma / alma_high / jvla: per-call time, compile separate, max rel diff, figure-of-merit shift.
- Real-space pixel scale row at jvla (0.01 / 0.015 / 0.02 / 0.025″, mask radius 3.5″): F time, full call, log-evidence shift vs 0.01″.
- fp64; host, GPU, driver, JAX versions, dataset and SHAs recorded. Witness JSON
results/breakdown/interferometer/jvla/delaunay_hpc_a100_fp64_fft_prune.json.
Record: results note, campaign journal entry, catalogue regenerated.
Detailed plan
- New scripts (autolens_profiling):
scripts/interferometer/delaunay/route_constants_ral.py(A, real-Aris arms; data path argument, Aris helpers imported from a temporary copy as in the L2 scripts) andscripts/interferometer/delaunay/fft_size_levers.py(B, reusingscripts/misc/likelihood_breakdown/interferometer_pixelized.pyhelpers:load_dataset,build_delaunay_mesh,build_model,timing.jit_profile). - Submit scripts under
hpc/batch_cpu/andhpc/batch_gpu/. - Results:
results/breakdown/interferometer/aris/route_constants_ral_cpu_fp64.json, the jvla witness JSON above, and a noteresults/notes/interferometer_route_constants_and_fft_levers_2026_10.md. - Avoids
scripts/misc/jax_compile/*(docs: design independent HST MGE pilot qualification #418) and the docs: symmetric direct_conv kernel decision rule and Aris witness (PyAutoArray#635) #425 note. - Fit: least squares on log seconds, per arm,
log t = log c + log W; report the ratioc_fft/c_dc(only the ratio routes), compare against laptop 0.430. - Verdict per Pulse task §3: pruned ≥ 10% of F at jvla with ≤ 1e-12 match → file a PyAutoArray follow-up prompt (not implemented here); pixel scale holds 0.5 nats → file a workspace docs prompt.
Tier: judge — merge mode: human /prm
Activity
results: A100 W~ FFT size levers witness (job 412515) and campaign jo…
Jammy2211 commentedon Oct 17, 2026
Scope change 2026-10-17 (the human, via the architect session): the interferometer CPU lane is closed, so part A (RAL route-gate constants) is dropped and no CPU jobs were submitted. The evidence already in hand is recorded as an open note in the campaign journal: FFT / full direct_conv measured 0.68 on RAL against 0.89 predicted, and the 1,459-vertex cell runs ~1.02x over FFT. The note says the constants refit is deferred. Part B (A100 FFT size levers) ran as RAL job 412515. Results are in PR #433.
Jammy2211 commentedon Oct 17, 2026
Shipped
PR #433 merged 2026-10-17. A100 W~ FFT size-levers witness (RAL job 412515, euclid-ral-gpu-1, exclusive, fp64, private 2026.10.17.1 stack): pruned padded FFT saves 19.4/22.1/29.7 % of F at alma/alma_high/jvla (whole-likelihood 1.08/1.18/1.41×), F rel diff ≤1.7e-15, FoM unchanged; coarser jvla pixel scales fail the 0.5-nat bar (+1,310 to +2,757 nats). Part A (CPU route constants) dropped: human decision 2026-10-17 closed the interferometer CPU lane; evidence in hand recorded as a deferred note. Human decision 2026-10-17: GPU levers filed as wiki research, no source changes (the JAX FFT pruning prompt is not filed).