Skip to content

numba-utils

Battle-tested building blocks for production Numba workloads.

numba-utils is a standard library for numerical software built with Numba. Every piece is callable from your own @njit code, benchmarked against the standard alternative, and honest about when not to use it.

from numba import njit
from numba_utils import topk

@njit
def winners(scores):
    return topk(scores, 10)     # O(n), no full sort — runs in nopython mode

What's inside

  • Kernels — search, selection, sorts, transforms, and rolling windows that run inside compiled code.
  • Containers — PriorityQueue, SparseSet, RingBuffer, BitSet and more, as jitclasses you can construct inside @njit.
  • Parallel — complete parallel operations (not prange wrappers) with serial fallbacks and bit-exact results where it matters.
  • Tooling — honest benchmarking (compare), diagnostics for compiled functions (diagnostics.check), and a testing module built around reference validation.

Why it exists

Production Numba behaves differently than tutorial Numba: caches that crash across processes, prange barrier costs, silent out-of-bounds corruption. numba-utils ships that hard-won knowledge as code, diagnostics, and documentation — not just functions. Read the philosophy for the identity behind every decision.

Next steps