Skip to content

CHANGES.md

Latest commit

 

History

History
230 lines (169 loc) · 8.94 KB

CHANGES.md

File metadata and controls

230 lines (169 loc) · 8.94 KB
Raw

3.8.0 (under development)

  • Run faster on Linux, where the new /proc/self/maps mechanism in 3.7.0 added quite a bit of overhead. #250

  • Added official support for Python 3.15, and declared support for Python 3.14 in package metadata. #251

3.7.0 (2026-09-15)

  • Fixed an intermittent OSError on Windows when DLLs are loaded or unloaded concurrently during library discovery (for example when importing conda-forge OpenCV). On Python 3.14+, discovery uses ctypes.util.dllist when available. If dllist raises OSError, threadpoolctl emits a RuntimeWarning instead of crashing so the failure can be reported upstream with a minimal reproducer. Older Pythons use a Toolhelp snapshot enumerator, with graceful per-module fallbacks. #219

  • Added the ability to check whether a limiting API affects just the current thread or the whole process. Mainly aimed at debugging and diagnostics, and somewhat unreliable, it is therefore enabled by default only for command-line usage. #213

  • Only warn about simultaneous libomp and libiomp usage on Linux, where the incompatibility is known to cause crashes. #222

  • Fixed a deadlock triggered by getting or setting MKL's number of threads from parallel threads when using MKL with libiomp (Intel threading) on Linux. #228

  • Going forward, setting the number of threads will only have a thread-local impact if feasible (for example, at minimum the underlying library must support this option, and many don't.) #228

  • For MKL, setting the number of threads is now thread-local, i.e. limiting the number of threads won't impact MKL's thread pool size when using MKL in other Python threads. #228

  • For OpenBLAS compiled with OpenMP on Linux and macOS, setting the number of threads is now thread-local, i.e. won't impact OpenBLAS thread pool size in other Python threads. On Windows behavior is likely process-wide, but this may depend on how OpenBLAS was compiled with OpenMP. #228

  • Fix OpenBLAS detection for conda package on Windows #240

  • Fixed a deadlock on Linux when using threadpoolctl from multiple threads. #243

    • Start using Python 3.14's built-in support for listing shared libraries.

    • On Linux, start using /proc/self/maps for listing shared libraries.

  • Avoid importing ctypes.util on Linux (and load libc with ctypes.CDLL(None)) so threadpool_info() does not create libffi closures that can abort after os.fork() on some libffi builds. #242

  • Dropped official support for Python 3.9. #255

3.6.0 (2025-03-13)

  • Added support for libraries with a path longer than 260 on Windows. The supported path length is now 10 times higher but not unlimited for security reasons. #189

  • Dropped official support for Python 3.8. #186 #191

3.5.0 (2024-04-29)

3.4.0 (2024-03-20)

  • Added support for Python interpreters statically linked against libc or linked against alternative implementations of libc like musl (on Alpine Linux for instance). #171

  • Added support for Pyodide #169

3.3.0 (2024-02-14)

  • Extended FlexiBLAS support to be able to switch backend at runtime. #163

  • Added support for FlexiBLAS #156

  • Fixed a bug where an unsupported library would be detected because it shares a common prefix with one of the supported libraries. Now the symbols are also checked to identify the supported libraries. #151

3.2.0 (2023-07-13)

  • Dropped support for Python 3.6 and 3.7.

  • Added support for custom library controllers. Custom controllers must inherit from the threadpoolctl.LibController class and be registered to threadpoolctl using the threadpoolctl.register function. #138

  • A warning is raised on macOS when threadpoolctl finds both Intel OpenMP and LLVM OpenMP runtimes loaded simultaneously by the same Python program. See details and workarounds at https://github.com/joblib/threadpoolctl/blob/master/multiple_openmp.md. #142

3.1.0 (2022-01-31)

  • Fixed a detection issue of the BLAS libraires packaged by conda-forge on Windows. #112

  • threadpool_limits and ThreadpoolController.limit now accept the string "sequential_blas_under_openmp" for the limits parameter. It should only be used for the specific case when one wants to have sequential BLAS calls within an OpenMP parallel region. It takes into account the unexpected behavior of OpenBLAS with the OpenMP threading layer. #114

3.0.0 (2021-10-01)

  • New object threadpooctl.ThreadpoolController which holds controllers for all the supported native libraries. The states of these libraries is accessible through the info method (equivalent to threadpoolctl.threadpool_info()) and their number of threads can be limited with the limit method which can be used as a context manager (equivalent to threadpoolctl.threadpool_limits()). This is especially useful to avoid searching through all loaded shared libraries each time. #95

  • Added support for OpenBLAS built for 64bit integers in Fortran. #101

  • Added the possibility to use threadpoolctl.threadpool_limits and threadpooctl.ThreadpoolController as decorators through their wrap method. #102

  • Fixed an attribute error when using old versions of OpenBLAS or BLIS that are missing version query functions. #88 #91

  • Fixed an attribute error when python is run with -OO. #87

2.2.0 (2021-07-09)

  • threadpoolctl.threadpool_info() now reports the architecture of the CPU cores detected by OpenBLAS (via openblas_get_corename) and BLIS (via bli_arch_query_id and bli_arch_string).

  • Fixed a bug when the version of MKL was not found. The "version" field is now set to None in that case. #82

2.1.0 (2020-05-29)

  • New commandline interface:

    python -m threadpoolctl -i numpy

    will try to import the numpy package and then return the output of threadpoolctl.threadpool_info() on STDOUT formatted using the JSON syntax. This makes it easier to quickly introspect a Python environment.

2.0.0 (2019-12-05)

  • Expose MKL, BLIS and OpenBLAS threading layer in information displayed by threadpool_info. This information is referenced in the threading_layer field. #48 #60

  • When threadpoolctl finds libomp (LLVM OpenMP) and libiomp (Intel OpenMP) both loaded, a warning is raised to recall that using threadpoolctl with this mix of OpenMP libraries may cause crashes or deadlocks. #49

1.1.0 (2019-09-12)

  • Detect libraries referenced by symlinks (e.g. BLAS libraries from conda-forge). #34

  • Add support for BLIS. #23

  • Breaking change: method get_original_num_threads on the threadpool_limits context manager to cheaply access the initial state of the runtime:

    • drop the user_api parameter;
    • instead return a dict {user_api: num_threads};
    • fixed a bug when the limit parameter of threadpool_limits was set to None.

    #32

1.0.0 (2019-06-03)

Initial release.