English

Product LaunchesMLX Swift

MLX Swift v0.32.2 Public — Distributed support and CUDA performance improvements

By the editors ／ Model: gemma-4-26b-a4b-it-qat-6bitAs of: 2026-09-29

MLX Swift v0.32.2 has been released. This update includes significant improvements to distributed support, CUDA build performance on Linux, and several API changes designed to better align the Swift implementation with the original Python MLX library.

New Features and Improvements

  • Added distributed support for MLX.
  • Improved CUDA NVCC spawn performance on Linux.
  • Added fp8 conversion utilities.
  • Added global scale support to quantized layers.
  • Added countNonzero to the Swift API.
  • Integrated more robust integration tests.

Bug Fixes and Behavioral Changes

  • Fixed a deadlock between CompiledFunction.lock and evalLock, as well as a data race in the compiler cache.
  • Fixed memory leaks in MLXArray (rawPointer:_:dtype:finalizer:).
  • Fixed convolve padRight behavior for even kernels in .same mode.
  • Fixed MLXNN.Pool padding width issues.
  • Updated tensordot default axes and nanToNum defaults to match the Python MLX implementation.
  • Updated linspace to handle dtype arguments, ensuring integer arguments default to float32.
  • Adjusted several optimizer and layer behaviors, including ALiBi, GRU, Adafactor, and Lion, to be more consistent with the Python version.

Sources

  1. 0.32.2 （2026-09-28）

Related