Repository navigation
Numba AdvancedSubtensor and AdvancedIncSubtensor don't bounds-check integer indices #2442
Description
Activity
Context: a Cast{int64} of a NaN float gives INT64_MIN on x86. After negative-index wraparound it stays negative, and the scatter writes one element past the end of its buffer. glibc later reports that as corrupted size vs. prev_size or double free or corruption, far from the Op that caused it.
Don't?
The C backend raises IndexError for the same graph.
Guess which backend is historically slow on indexing operations.
Bound checks are expensive (or can be), and break vectorization. You can enable them with numba checkbounds flag or should be able to.
import os os.environ["NUMBA_BOUNDSCHECK"] = "1" # Set before importing Numba/PyTensor import numpy as np import pytensor import pytensor.tensor as pt x, idx = pt.vector("x"), pt.lvector("idx") f = pytensor.function( [pytensor.In(x, mutable=True), idx], pt.inc_subtensor(x[idx], 7.0), mode="NUMBA", ) backing = np.zeros(12) f(backing[:11], np.array([np.iinfo(np.int64).min])) # Raises IndexError
- added a commit that references this issue
on Oct 11, 2026
Under the numba backend,
x[idx]andinc_subtensor(x[idx], v)with an out-of-range integer index read and write outside the array instead of raising. The C backend raisesIndexErrorfor the same graph.Indices 20 and -20 also return without error, while the C backend raises
IndexErrorfor all three.Context: a
Cast{int64}of a NaN float givesINT64_MINon x86. After negative-index wraparound it stays negative, and the scatter writes one element past the end of its buffer. glibc later reports that ascorrupted size vs. prev_sizeordouble free or corruption, far from the Op that caused it.