3.1 KiB
Experimental uniform DFG pilot
Numerical source remains926ca27407c26ca11283282a6e53249925725376. This test-only followup uses a separate dfg_source_manifest.json, pins the official reference-cycle file, and records actual execution HEAD independently. The historical bodyless manifest is unchanged and correctly rejects this checkout because the shared runner gained optional hooks. Use historical939c10c for that historical entrypoint; do not repin old evidence.
The new run_dfg.py is uniform-only, with the original D80 cylinder, SRT tau.524, shiftedFP32 double buffer, original initialization, noLES/force, pressure-NEQ outlet all-nine damping.5. Only inlet scheme is experimental_regularized_full9. It reuses run_refined_dfg.run_case and the existing pressure, force, physical mass/flux and periodic/reference helpers. Legacy run_case callers retain the original inlet and dense pilot cadence.
This pilot advances40,000steps to t*=5 with sparse observations every4000steps (t*=.5), matching the standard pre240 cadence. Every100-step batch contributes its full momentum impulse; force time is the interval midpoint, pressure time its endpoint. First100steps provide warmup; the next four100-step batches measure warmed step+force cost. All500steps are included in the physical trajectory and observation impulse; no reset or discarded evolution. Calibration durations and total step/force/startup/wall timings are recorded. Batch force download synchronizes GPU work through the existing API.
A first unhealthy sample identifies only the first observed failure, not its onset. Receipts separate executed_steps, last_observed_step and the existing steps field; timeout retains available progress, history and calibration. No per-step field download or new health threshold is introduced. Runtime admission uses180seconds internally, with an external240second watchdog. Later full-run admission must use measured batched timings, not the earlier per-step snapshot diagnostic rate. Same-size historical MRT full campaigns took roughly354–373seconds, but this pilot measures the present SRT/full9 case.
CPU checks from this checkout:
python tests/validation/experimental_full9/test_dfg.py
python tests/validation/experimental_full9/test_full9.py
python -m py_compile tests/validation/experimental_full9/run_dfg.py tests/validation/run_refined_dfg.py
python tests/validation/experimental_full9/run_dfg.py --output /tmp/new-full9-dfg-preflight
After independent review and explicit GPU admission only:
CUDA_VISIBLE_DEVICES=<admitted-device> timeout --signal=TERM --kill-after=30s 240s python tests/validation/experimental_full9/run_dfg.py --execute --gpu-admitted --wall-seconds 180 --output output/new-full9-dfg-pilot
This entrypoint intentionally has no full-run execution option. Existing official-reference assessment is reused and accuracy_pass requires health, periodicity and reference agreement for future qualification; pilot accuracy_qualified is alwaysfalse. A healthy pilot qualifies bounded admission/timing only. No grid convergence, canonical outlet, refined/MRT/other precision, long-time accuracy or general stability claim follows.