Numba: High Performance Python Compiler

Warnings

• gotcha Numba treats global variables as compile-time constants. Modifying a global variable after a `@jit` decorated function has been compiled will not affect the compiled function's behavior unless the function is explicitly recompiled using `.recompile()` or the global is passed as an argument.
  Fix: Either recompile the function after modifying the global variable or, preferably, refactor your code to pass the variable as an argument to the jitted function.
• gotcha Type inference failures are a common reason for Numba compilation errors, especially in `nopython` mode. Numba needs to determine concrete types for all variables; if it cannot, compilation will fail.
  Fix: Ensure that all operations and data types within a `@njit` (nopython mode) function are supported by Numba and that types are consistent. Explicitly casting types (e.g., using `np.int64(0)` instead of `0`) can sometimes help with type propagation.
• gotcha Running Numba-compiled scripts twice in IDEs like Spyder can lead to `TypeError: No matching definition for argument type(s)` due to module reloading issues.
  Fix: Configure Spyder to exclude `numba` from its User Module Reloader (UMR). Go to `Preferences -> Console -> Advanced Settings`, click 'Set UMR excluded modules', and add `numba`. Restart the IPython console after applying the setting.
• deprecated The `target` kwarg for the `numba.jit` decorator family has been deprecated.
  Fix: Avoid using the `target` kwarg. Numba's decorators like `jit` and `njit` infer the target (CPU by default) or specific GPU decorators like `cuda.jit` should be used for explicit targets.
• deprecated Reflection for Python `List` and `Set` types within Numba-compiled code is deprecated. This feature ensured changes to mutable Python containers were visible after the function returned.
  Fix: Numba aims to replace this with a better implementation. Users relying on reflection for these types might need to adapt their code if they observe unexpected behavior or warnings in future versions. Pinning Numba dependency is recommended if relying on this behavior.
• breaking Numba 0.64.0 supports NumPy 2.3 and 2.4. NumPy 2.0 introduces binary incompatible changes and a new type system. Numba has evolved to accommodate this, but users should be prepared to test and verify their codebases, as NumPy 2.0 may also impact non-JIT compiled code output.
  Fix: Update to Numba 0.64.0 or later to ensure compatibility with NumPy 2.x. Review Numba's documentation and NumPy 2.0 migration guides for potential code adjustments, especially concerning type interactions.
• breaking The experimental RVSDG (Region-based Value Stream Dependence Graph) frontend was removed in Numba 0.61.0.
  Fix: Code relying on the experimental RVSDG frontend will no longer function. Users should remove any explicit enabling of this frontend from their Numba configurations.

Install

• pip install numba PyPI
• conda install numba Anaconda/Miniconda

Imports

• jit

  from numba import jit

  `jit` is typically imported directly from the `numba` package as a decorator, not as a submodule.
• njit

  from numba import njit

  `njit` is a shorthand for `@jit(nopython=True)`, ensuring Numba generates code that doesn't fall back to the Python interpreter, which is often desired for performance.
• guvectorize

  from numba import guvectorize

Quickstart

This quickstart demonstrates the core usage of Numba's `@njit` decorator to compile a Python function for numerical computation. The function `sum_array` iterates over a NumPy array, and when decorated with `@njit`, Numba compiles it to highly optimized machine code at runtime, significantly speeding up its execution compared to pure Python.

import numpy as np
from numba import njit

@njit
def sum_array(arr):
    total = 0.0
    for x in arr:
        total += x
    return total

# Example usage
data = np.arange(1000000, dtype=np.float64)
result = sum_array(data)
print(f"Sum of array: {result}")