Issues

Each lumped RLC element adds its own block to the compiled step, so compile time grows ~28 ms per element on CPU (11 s at 400 loads); per-step cost unchanged

bk-squared/rfx#1946 · opened

on Oct 20, 2026

· v2.1 — distributed execution

1

A TF/SF source silently replaces declared CPML or PMC/PEC side faces with periodic ones (warns only through preflight)

bk-squared/rfx#1945 · opened

on Oct 20, 2026

1

Waveguide port with waveform="differentiated_gaussian" re-injects a wrapped precursor at the record end: S21 off by up to 11 dB in an empty matched guide

bk-squared/rfx#1942 · opened

on Oct 20, 2026

2

A graded dz_profile that does not sum to the declared domain z builds silently on every path

bk-squared/rfx#1941 · opened

on Oct 20, 2026

1

subpixel_smoothing='kottke_pec' with conformal PEC silently drops the conformal weights

bk-squared/rfx#1935 · opened

on Oct 20, 2026

· v2.0.0 — release

A Drude/Lorentz medium above its time-step limit runs unchecked on four routes (traced mesh, two-device override, Debye rows, prebuilt coefficients)

bk-squared/rfx#1934 · opened

on Oct 20, 2026

3

Weekly shard 5 aborts on LLVM section memory after RSS climbs over the pole-media weekly files (cause undetermined)

bk-squared/rfx#1932 · opened

on Oct 20, 2026

2

PEC occupancy gradient is half the one-sided derivative at binary designs (clip slope 1/2 at its bounds)

bk-squared/rfx#1926 · opened

on Oct 20, 2026

· v2.0.0 — release

2

weekly scientific-validation lane is red

bk-squared/rfx#1900 · opened

on Oct 19, 2026

Declared vs realized domain length is never compared on PEC/PMC-bounded axes: an empty WR-90 at dx 1.5 mm is solved 5 % wide, TE10 cutoff 4.75 % low, silently

bk-squared/rfx#1887 · opened

on Oct 19, 2026

4

Multi-device assembly is outside declaration_setup, so each host-side piece needs its own trace-time wrapper

bk-squared/rfx#1844 · opened

on Oct 17, 2026

run(until_decay=) with the default pulse never stops on an open plane wave: the remaining energy is near-static electric (cause undetermined: injection or stop rule)

bk-squared/rfx#1839 · opened

on Oct 17, 2026

· v2.1 — distributed execution

19

Complex arccosh has no kernel on the Apple-silicon jax-mps backend: the coax two-port's traced reflection fails there

bk-squared/rfx#1838 · opened

on Oct 17, 2026

A waveguide port gets PEC side walls on the uniform path but CPML on all six faces on the graded path: one declaration, two structures

bk-squared/rfx#1826 · opened

on Oct 17, 2026

A small Debye, Lorentz or Drude inclusion makes every step update dispersion on the whole grid: a 16³ block costs 69 % of the step at 160³

bk-squared/rfx#1825 · opened

on Oct 17, 2026

· v2.1 — distributed execution

2

A Floquet port scanned off normal is refused on every path: no oblique-incidence reflection of a periodic surface (post-2.0)

bk-squared/rfx#1822 · opened

on Oct 17, 2026

A periodic axis or Bloch phase is refused on both multi-device runners: a periodic cell runs on one device only (post-2.0)

bk-squared/rfx#1821 · opened

on Oct 17, 2026

The model layer imports rfx.nonuniform (and through it rfx.stepping): rfx.measurement -> rfx.model -> rfx.nonuniform -> rfx.stepping

bk-squared/rfx#1818 · opened

on Oct 17, 2026

run(until_decay) on a closed TF/SF box rounds its incident line differently from the fixed-step run of the same request (up to 15.5 float32 ULP of the probe peak)

bk-squared/rfx#1814 · opened

on Oct 16, 2026

grad through a declared shape dimension dies with ConcretizationTypeError at rasterization on run() and forward()

bk-squared/rfx#1806 · opened

on Oct 16, 2026

2

jacobian_fwd value differs from the plain call by 3 float32 ULP on Apple arm64 CPU only (bit-identity tests red on the Mac, green in CI)

bk-squared/rfx#1802 · opened

on Oct 16, 2026

1

A gradient through eps_override on the oblique plane wave dies with a raw TracerArrayConversionError instead of a gradient or a refusal

bk-squared/rfx#1795 · opened

on Oct 16, 2026

Fast-suite shards take up to 54 min on main; balance them and return the job timeout from the 75-min stopgap to 60

bk-squared/rfx#1780 · opened

on Oct 15, 2026

Subpixel smoothing averages a metal (PEC) body as a dielectric with its eps_r, so dielectric edges touching metal read a lower permittivity

bk-squared/rfx#1778 · opened

on Oct 15, 2026

Fused GPU fast step taken automatically without CPML runs at half the speed of the ordinary step

bk-squared/rfx#1773 · opened

on Oct 15, 2026

Navigation Menu

Repository navigation

Issues

Each lumped RLC element adds its own block to the compiled step, so compile time grows ~28 ms per element on CPU (11 s at 400 loads); per-step cost unchanged

A TF/SF source silently replaces declared CPML or PMC/PEC side faces with periodic ones (warns only through preflight)

Waveguide port with waveform="differentiated_gaussian" re-injects a wrapped precursor at the record end: S21 off by up to 11 dB in an empty matched guide

A graded dz_profile that does not sum to the declared domain z builds silently on every path

subpixel_smoothing='kottke_pec' with conformal PEC silently drops the conformal weights

A Drude/Lorentz medium above its time-step limit runs unchecked on four routes (traced mesh, two-device override, Debye rows, prebuilt coefficients)

Weekly shard 5 aborts on LLVM section memory after RSS climbs over the pole-media weekly files (cause undetermined)

PEC occupancy gradient is half the one-sided derivative at binary designs (clip slope 1/2 at its bounds)

weekly scientific-validation lane is red

Declared vs realized domain length is never compared on PEC/PMC-bounded axes: an empty WR-90 at dx 1.5 mm is solved 5 % wide, TE10 cutoff 4.75 % low, silently

Multi-device assembly is outside declaration_setup, so each host-side piece needs its own trace-time wrapper

run(until_decay=) with the default pulse never stops on an open plane wave: the remaining energy is near-static electric (cause undetermined: injection or stop rule)

Complex arccosh has no kernel on the Apple-silicon jax-mps backend: the coax two-port's traced reflection fails there

A waveguide port gets PEC side walls on the uniform path but CPML on all six faces on the graded path: one declaration, two structures

A small Debye, Lorentz or Drude inclusion makes every step update dispersion on the whole grid: a 16³ block costs 69 % of the step at 160³

A Floquet port scanned off normal is refused on every path: no oblique-incidence reflection of a periodic surface (post-2.0)

A periodic axis or Bloch phase is refused on both multi-device runners: a periodic cell runs on one device only (post-2.0)

The model layer imports rfx.nonuniform (and through it rfx.stepping): rfx.measurement -> rfx.model -> rfx.nonuniform -> rfx.stepping

run(until_decay) on a closed TF/SF box rounds its incident line differently from the fixed-step run of the same request (up to 15.5 float32 ULP of the probe peak)

grad through a declared shape dimension dies with ConcretizationTypeError at rasterization on run() and forward()

jacobian_fwd value differs from the plain call by 3 float32 ULP on Apple arm64 CPU only (bit-identity tests red on the Mac, green in CI)

A gradient through eps_override on the oblique plane wave dies with a raw TracerArrayConversionError instead of a gradient or a refusal

Fast-suite shards take up to 54 min on main; balance them and return the job timeout from the 75-min stopgap to 60

Subpixel smoothing averages a metal (PEC) body as a dielectric with its eps_r, so dielectric edges touching metal read a lower permittivity

Fused GPU fast step taken automatically without CPML runs at half the speed of the ordinary step