What's New: Intel® Distribution for Python* 2025.3

[StefR_Intel]

Data Parallel Control Library* includes the following improvements:

• Implemented tensor.isin membership test
• Allow NumPy arrays for advanced indexing
• Reduced binary size by making constexpr variables in headers inline and others static, and removing all anonymous namespaces from headers
• Fixed DLL search path when initializing dpctl, fixing failure to see devices when importing in venv on Windows
• Fixed failure to copy arrays of double dtype from device with double precision to a device without double precision

Data Parallel Extension for Numpy* includes the following improvements:

• Added implementation of dpnp.piecewise, dpnp.special.erf, dpnp.special.erfc, dpnp.linalg.lu_solve and dpnp.linalg.lu_factor functions
• Implemented new dpnp.ndarray.view, dpnp.ndarray.__contains__ methods and dpnp.ndarray.data, dpnp.ndarray.data.ptr attributes to improve CuPy compatibility
• Improved performance of dpnp.isclose, dpnp.linalg.det and dpnp.linalg.slogdet functions
• Improved performance for a specialized matrix multiplication where the result is a symmetric matrix
• The order of individual FFTs over axes for dpnp.fft.irfftn is changed to be in forward order
• The intermediate steps of ND FFT are updated to perform in-place FFT
• dpnp.size accepts a tuple of integers for axes keyword
• dpnp.pad accepts a dictionary for pad_width keyword
• The license handling is aligned with PEP-639
• The new dpnp release is compatible with NumPy 2.3.3
• Including fixes in:
  • dpnp.ndarray constructor passed with dpnp.ndarray.data value for buffer keyword
  • dpnp.random functions to allow any value of size keyword where each element is castable to Py_ssize_t type
  • dpnp.linalg.cond to always return a real data type
  • dpnp.unique with 1D input array and axis=0, equal_nan=True passed where the produced result didn't collapse the NaNs
  • dpnp.resize with unsigned integer size

Intel® Distribution for Python's NumPy-based Python interface to Intel® oneAPI Math Kernel Library (oneMKL) Fourier Transform Functions includes following improvements:

• Permit building with Python 3.13
• Removed scipy.fftpack interfaced (deprecated in scipy)
• Resolved a bug in N-D FFTs when given out and s keyword arguments

Intel® Distribution for Python's mkl_umath package implementing NumPy's UFuncs based on SVML and MKL VML:

• Changed to dynamic linking and to use LP64, permitting arrays larger than 2**31
• Added new MKL function implementations
• Removed MKL implementations for functions which showed no performance improvement compared to stock NumPy pending further evaluation

Other improvements:

• Intel® Distribution for Python's package for run-time control of Intel® oneAPI Math Kernel Library (oneMKL) now permits building with Python 3.14