NumPy

Warnings

• breaking ABI break in NumPy 2.0. Packages built against NumPy 1.x will not import with NumPy 2.x — raises ImportError about binary incompatibility. This affected TensorFlow, PyTorch, scipy, and many others at release.
  Fix: Upgrade all binary dependencies to versions built against NumPy 2.x. Check https://github.com/numpy/numpy/issues/24300 for ecosystem compatibility status.
• breaking ~100 np namespace members removed in 2.0 including: np.float_, np.int_, np.complex_, np.object_, np.bool_, np.str_, np.long, np.unicode_, np.cfloat, np.Inf, np.Infinity, np.NAN, np.PINF, np.NINF. AttributeError on first use.
  Fix: Replace with explicit dtypes: np.float_ → np.float64, np.int_ → np.intp, np.complex_ → np.complex128, np.bool_ → np.bool_, np.Inf → np.inf, np.NAN → np.nan. Use ruff --select NPY201 to auto-fix.
• breaking Type promotion rules changed (NEP 50) in NumPy 2.0. Scalar precision is now preserved: np.float32(3) + 3. now returns float32 instead of float64. Code with mixed scalar/array operations may silently produce lower-precision results.
  Fix: Audit mixed-precision arithmetic. Explicitly cast where float64 precision is required: np.float64(3) + arr or arr.astype(np.float64).
• breaking np.core namespace renamed to np._core (private) in 2.0. Code importing from np.core (common in older library code) will raise ImportError.
  Fix: Access all public members from the main np namespace. Never import from np.core or np._core directly.
• deprecated Legacy random API (np.random.seed, np.random.rand, np.random.randn, np.random.randint etc.) is not formally deprecated yet but the new Generator API is strongly preferred. Legacy API has known issues with reproducibility in multiprocessing.
  Fix: Use np.random.default_rng(seed) to create a Generator. Replace np.random.randn(n) with rng.standard_normal(n), np.random.randint(a, b) with rng.integers(a, b).
• gotcha np.float (Python built-in alias) was removed in NumPy 1.24, not 2.0. Already gone for 2+ years but still extremely common in LLM-generated code.
  Fix: Use Python float or np.float64 explicitly. Same for np.int → int or np.int64, np.complex → complex or np.complex128.
• gotcha np.string_ is still available but is bytes-based (fixed-width). For variable-length string arrays use np.dtypes.StringDType() (new in 2.0) or Python object arrays.
  Fix: For variable-length strings: arr = np.array(['hello', 'world'], dtype=np.dtypes.StringDType()). np.string_ creates fixed-width byte arrays.

Install

• pip install numpy Core

Imports

• numpy

  import numpy as np
  
  arr = np.array([1.0, 2.0], dtype=np.float64)
  scalar = np.float64(3.14)

  np.float_, np.int_, np.complex_, np.bool_, np.object_, np.str_ aliases all removed in NumPy 2.0. Use np.float64, np.int64, np.complex128, np.bool_ → bool, etc.
• random

  rng = np.random.default_rng(seed=42)
  arr = rng.standard_normal(100)

  np.random.seed() + np.random.randn() is the legacy RandomState API. The new Generator API (np.random.default_rng()) is preferred and provides better statistical properties.

Quickstart

Basic array creation and operations. Use new random Generator API.

import numpy as np

# Array creation
arr = np.array([1, 2, 3, 4], dtype=np.float64)
matrix = np.zeros((3, 4), dtype=np.float32)

# Operations
result = arr * 2 + 1
dot = np.dot(matrix.T, matrix)

# Random (new API)
rng = np.random.default_rng(seed=42)
samples = rng.standard_normal((100, 10))

# Type check
print(arr.dtype)   # float64
print(arr.shape)   # (4,)