v0.5.4

• We now register casts from int2 and int4 to all of the custom float types,
  except float6_e2m3fn and float8_e8m0fnu.
• Custom floats may now be constructed from Python integers (#317)
• Fixed bug in byte-swap operation for custom floats (#311)
• Wheels now support Python 3.14 free threading on Windows.

v0.5.3

• NPY_NEEDS_PYAPI was removed from the dtype flags. This should improve the speed of array operations, but it does mean that values pickled using previous versions of ml_dtypes are incompatible with the current release and should be regenerated with the current release.
• Wheels now support Python 3.14.
• Wheels now support Windows 11 ARM.

v0.5.2 release

This is a minor change over v0.5.1. This release reverts the addition of Power wheels, which caused problems with our wheel release process.

Due to an oversight during the release process, the wheels from this release report themselves as v0.5.1. The only difference between the two is the lack of Power wheels.

v0.5.1 release

• Fixed sign bit handling for float4 and float6 types.
• Wheels now support Python 3.13 free-threading.
• Wheels now support the Power architecture.

v0.5.0 release

• Added new 8-bit float types following IEEE 754 convention:
  ml_dtypes.float8_e4m3, ml_dtypes.float8_e3m4
• Added the 8-bit floating point type ml_dtypes.float8_e8m0fnu, which is the
  OpenCompute MX scale format.
• Added new 4-bit and 6-bit float types:
  ml_dtypes.float4_e2m1fn, ml_dtypes.float6_e2m3fn and ml_dtypes.float6_e3m2fn.
• Fix outputs of float divmod and floor_divide when denominator is zero.

v0.4.1 release

• Updates build requirements to use NumPy 2.0 release

v0.4.0 release

• Updates ml_dtypes for compatibility with the upcoming NumPy 2.0 release.
• Wheels are built against NumPy 2.0.0rc1.

v0.4.0b1 pre-release

• Updates ml_dtypes for compatibility with future NumPy 2.0 release.
• Wheels for the release candidate are built against NumPy 2.0.0b1.

Version 0.3.2 release

• Fixed spurious invalid value warnings when casting between floating point
  types on Mac ARM.
• Remove pybind11 build requirement
• Update C++ sources for compatibility with NumPy 2.0

v0.3.1 release

• Added support for int4 casting to wider integers such as int8
• Added support to cast np.float32 and np.float64 into int4