Repository navigation
Description
Tracks SEP 49: The TDSE response on native, and numpy_ref as a slim test reference (on its draft branch until the draft pull request merges; then docs/dev/seps/sep-0049-tdse-native-slim-reference.md on main).
It proposes two phases. First, the TDSE response runs on the native CPU build: native's Adams march evaluates the source through a host source hook, a Python callable that the native session binds and the compiled march calls holding the GIL at each source evaluation. That is an ADR exception to "no Python call in a trial step", limited to this hook; device builds refuse it. Second, numpy_ref becomes a slim test and development reference: the contract arithmetic (linear, source, Runge–Kutta and Adams trials, statistics) and the decision specification (adams_core, rk_core), with no step loop, sessions, groups, batch or registry row, and no longer selectable. Full-run parity becomes native's run replayed on the reference along its own steps, plus decision traces. Estimated cost: about −11.5k source lines, −12k test lines, +0.3k C++ and +0.4k native adapter lines.
Where it stands
- Status: Draft. Review: Not reviewed. Implementation: not started.
- Requires: ADR-077 (two backends, binary64 only; Lean rules: ADR-077 (two backends, binary64 only), withdraw SEP 48, private-docstring rule, size report #763) and the change it calls for (single precision, CF4, the comparison methods and
PIControlleroverrides removed; native importing nothing fromnumpy_ref); SEPs 12, 20, 26, 39, 45, 47 and 25.
What comes next
- Merge the draft so it can be discussed in the developer documentation.
- Human and AI reviews of Part I; answer the open questions Q1–Q10 (§12).
- Acceptance by the maintainers.
To resolve first
- ADR-077 merged, and the change it calls for merged, before phase 1.
- The Adams controller's decision traces cover only the exponential path #270 (the Adams decision traces cover only the exponential path) before phase 2, since the traces then become the only check of the fused formulation's decisions.
Related
- Answered or closed: Write a SEP for the step loops and the kernel seam: one march protocol, the batch axis, the integrator and its tolerance #110 (its step-loop question), The Python Adams march: per-step allocations and casts #104, the compiled-coupling item of The cost of a TDSE run: provider reductions, a compiled coupling and the stepper on the actual source #247.
- Changed: A leading batch axis outlives SEP 47 in the collar check, the split crossing, the boundary monitor and
Backend.max_batch#618, One diverging point fails every member of itsBatchgroup #171, Group construction consumes starts and closes owners before successful admission #623, No test holdsnumpy_refand native together outside the active band under the split collar on a general guide #230,on_axis_spectrum:nativeand the Python backends differ at a field grid's Nyquist bin when a hand-built grid keeps it active #155, One host statistics module fed by a backend hook #107, Splitdriver/simulation.pyandbackends/multistep.py#109, CI's Linux native wheel needs glibc 2.35 and a system FFTW, so a development build can't go on a cluster #570,solprop-nativecannot be built from a source archive or offline #312. - Unchanged, for the record: Qualifying the TDSE response: the window, the calibration, the numerics and a full-length run #246, Publish Pyroraptor so that the TDSE response can be installed from a released package #248, The TDSE steady-current certificate can refuse a resolved taper on a band far above its frequencies #249, The TDSE consistency cache uses a Linux-style directory on macOS and Windows #583.
Done when
The SEP reaches Final, Rejected, Withdrawn or Superseded, and this issue closes with it. A Deferred SEP keeps it open.
Activity
SEP 49: name its tracking issue, #764
jtravs commentedon Nov 12, 2026
Draft pull request: #765.
[-]SEP 49: The TDSE response on native, and numpy_ref as a slim test reference[/-][+]SEP 49: The TDSE response on native[/+]jtravs commentedon Nov 13, 2026
Done in #771 (merged as 22f886e). SEP 49 is Final and archived: the TDSE response runs on native through the host-source hook, with Pyroraptor always in a worker process (ADR-078). A/B against main: geometric mean 0.986 on 8 cases (limit 1.015); non-TDSE runs bitwise identical on 7 cases with estimate plans. Follow-ups: #777 (strict-FMA checker bound), #785 (GPU TDSE comparison).