Issues

• BudgetAccumulator.record() keeps every closed context's raw op log for the life of the process, so memory grows with each scored MLP

  AIcrowd/flopscope#321 · Filter by author logxio opened on Oct 19, 2026

• polyfit covariance output performs unbilled matrix work

  AIcrowd/flopscope#309 · Filter by author dipamc opened on Oct 15, 2026

• Vector quantiles attach surviving symmetry to the wrong output axes

  AIcrowd/flopscope#307 · Filter by author dipamc opened on Oct 15, 2026

• Length-one FFT transforms write buffers for zero FLOPs

  AIcrowd/flopscope#306 · Filter by author dipamc opened on Oct 15, 2026

• Unary ufunc signature and explicit-dtype destination underbilling

  AIcrowd/flopscope#304 · Filter by author dipamc opened on Oct 15, 2026

• einsum(..., symmetry=G) runs the symmetry check without billing it, so a tag costs 0 there and k * (7n - 1) through as_symmetric

  AIcrowd/flopscope#298 · Filter by author AndreasHad04 opened on Oct 11, 2026

• histogram raises TypeError for an integer bins that is not a Python int, and histogram2d / histogramdd bill such a bins with the n fallback

  AIcrowd/flopscope#297 · Filter by author AndreasHad04 opened on Oct 11, 2026

• Ufuncs and matmul with an out= larger than a symmetric-tagged operand raise ValueError or SymmetryError on calls numpy accepts

  AIcrowd/flopscope#296 · Filter by author AndreasHad04 opened on Oct 11, 2026

• histogram2d and histogramdd bill only the binning pass: the edges, the float conversion of the whole count grid and the density=True arithmetic are free

  AIcrowd/flopscope#295 · Filter by author AndreasHad04 opened on Oct 11, 2026

• Elementwise ufuncs with a non-symmetric where= mask keep the input's symmetry tag: SymmetryError with out=, a wrongly tagged result (and symmetric-rate billing) without it

  AIcrowd/flopscope#285 · Filter by author AndreasHad04 opened on Oct 5, 2026

• matvec(axes=), matmul(axes=), matmul(dtype=) and dot(a, b, out) fail on call forms numpy accepts

  AIcrowd/flopscope#283 · Filter by author AndreasHad04 opened on Oct 5, 2026

• Tracked arrays as distribution parameters or as obj= / fill_value= / period= / rcond= / k= raise "FlopscopeArray reached numpy..." (same class as #278)

  AIcrowd/flopscope#282 · Filter by author AndreasHad04 opened on Oct 5, 2026

• multivariate_normal(method=...) and hermitian=True are billed as a full SVD although numpy runs a cheaper factorization

  AIcrowd/flopscope#280 · Filter by author AndreasHad04 opened on Oct 5, 2026

• ufuncs with an out= (or where=) larger than the inputs bill the input size, not the broadcast write

  AIcrowd/flopscope#275 · Filter by author AndreasHad04 opened on Oct 2, 2026

• Complete QR on tall matrices seems to undercount Q formation

  AIcrowd/flopscope#269 · Filter by author NathanZane opened on Sep 30, 2026

• einsum silently ignores explicit dtype, producing a different numerical result

  AIcrowd/flopscope#268 · Filter by author NathanZane opened on Sep 30, 2026

• .shape assignment is free in-process, keeps a stale symmetry tag, and raises on flopscope-client

  AIcrowd/flopscope#267 · Filter by author AndreasHad04 opened on Sep 29, 2026

• random.permutation and choice(replace=False) bill the pool length while returning a full copy, a 1e6-element copy costs 4 FLOPs

  AIcrowd/flopscope#266 · Filter by author thibaudlepan77-svg opened on Sep 28, 2026

• choose with a positional out writes into a symmetric-tagged buffer and the tag survives

  AIcrowd/flopscope#265 · Filter by author thibaudlepan77-svg opened on Sep 28, 2026

• Module-level random.shuffle leaves a stale symmetry tag, contractions on the shuffled array bill at the orbit rate

  AIcrowd/flopscope#264 · Filter by author thibaudlepan77-svg opened on Sep 28, 2026

• Masked sum retains invalid input symmetry and rejects valid NumPy results

  AIcrowd/flopscope#263 · Filter by author srdey2002 opened on Sep 17, 2026

• linalg.pinv(..., rtol=None) silently ignores NumPy's Array API tolerance

  AIcrowd/flopscope#262 · Filter by author srdey2002 opened on Sep 17, 2026

• stats.truncnorm returns infinities or raises for valid finite tail intervals

  AIcrowd/flopscope#259 · Filter by author AminMohamed-3 opened on Sep 8, 2026

• Symmetry-inferred einsum rejects an algebraically symmetric, all-finite float32 result with max deviation = inf; the check is np.allclose(atol=1e-6) on near-zero off-diagonals and the inf is a hard-coded sentinel

  AIcrowd/flopscope#257 · Filter by author jigonyoo opened on Sep 6, 2026

• fnp.einsum("ia,ij,ib->jab", ...) raises a broadcast ValueError for a valid explicit-output equation that NumPy accepts

  AIcrowd/flopscope#256 · Filter by author jigonyoo opened on Sep 6, 2026