fastmath, boundscheck and other options

These flags unlock SIMD vectorization and remove checks, but each changes semantics in specific ways.

The JIT decorators accept options that change what the compiler may assume.

fastmath=True lets LLVM ignore strict IEEE 754 rules: it may reorder additions (enabling SIMD vectorization of reductions), assume no NaNs or infinities, and use approximate reciprocals. Loops summing floats often get several times faster. The price: results can differ in the last bits, and code that relies on NaN checks (x != x, np.isnan) may be optimised away. You can pass a set of flags such as {"nsz", "arcp", "contract", "afn", "reassoc"} to enable only some relaxations, leaving NaN and infinity handling intact.
boundscheck=True turns on index bounds checking. By default Numba does not check bounds, so a[10] on a 5-element array reads or writes arbitrary memory. Enable checks while debugging, or globally with NUMBA_BOUNDSCHECK=1.
error_model="numpy" makes division by zero return inf or nan like NumPy instead of raising ZeroDivisionError as Python does. Removing the check makes division-heavy loops faster.
inline="always" inlines a small @njit function into callers at the Numba IR level; LLVM usually inlines small functions anyway.
nogil=True and cache=True are covered elsewhere.

Treat these as measured optimisations: benchmark with and without, and test numerical results against the strict version.

Example

import numpy as np
from numba import njit

@njit(fastmath=True)
def fast_sum(a):
    s = 0.0
    for x in a:
        s += x            # may be reordered and vectorized
    return s

@njit(error_model="numpy")
def safe_ratio(a, b):
    return a / b

@njit(boundscheck=True)
def get(a, i):
    return a[i]

a = np.random.default_rng(0).random(10_000_000)
print(np.isclose(fast_sum(a), a.sum()))   # True (last bits may differ)
print(safe_ratio(1.0, 0.0))               # inf  (Python error model would raise)
try:
    get(np.zeros(5), 10)
except IndexError as e:
    print("IndexError:", e)               # IndexError: index is out of bounds

Remember

fastmath allows reordering and SIMD; results may change slightly and NaN checks may vanish
Bounds are not checked by default; enable boundscheck while debugging
error_model='numpy' returns inf/nan for division by zero

Watch out for

Using fastmath=True in code that tests for NaN, which may then never detect it.

Check yourself

By default, what happens in @njit code when you read a[10] from a 5-element array?

Official docs: numba.readthedocs.io/en/stable/reference/jit-compilation.html

Example

Remember

Watch out for

Check yourself

Builds on

Unlocks