0.32.2

IMPORTANT

Use release 0.32.3 -- important patch for #491

Highlights

• New mlx release, https://github.com/ml-explore/mlx/compare/v0.31.1...v0.32.2
• Distributed support by @ronaldmannak in #482
• replaced and expanded integration tests - #477
• add thin logging API - #484

API Changes Changes

• Device.deviceType is no longer optional
• Device equality now compares index
• Device.gpu and Device.cpu now return the current device per the Stream, e.g. on a multi-gpu CUDA system it might return something other than device 0
• Stream.gpu and Stream.cpu now return the current GPU or CPU stream (task aware)
• Stream.defaultStream (new) returns the current stream
• MLXNN.Lion.betas default values changed (now correct -- matching python)

Deprecations

• StreamOrDevice.device(_:) → prefer .stream(Stream)
• Stream.init() and Stream.init(_ device:) → use Stream.withNewDefaultStream
• Stream.defaultStream(_ device: Device) → use defaultStream(DeviceType) / withNewDefaultStream(Device)
• Device.setDefaultDevice message tightened: only valid before any mlx resources are created

Behavior Changes

• tensordot(_:_:axes:) default axes: 1 → 2 (matches Python).
• nanToNum(posInf:negInf:) defaults 0 → nil (now replaces with dtype max/min, matching Python).
• linspace now takes dtype:; integer args default to float32 (previously T.dtype), Double args resolve to float32 instead of float64.
• convolve .same mode: padRight fix for even kernels (padLeft/2 - 1 → padLeft - 1).
• MLXNN.Pool padding widths fix (was emitting an extra leading/trailing pair).
• ALiBi.alibiSlope rewritten to match Python for non-power-of-2 head counts; alibiMatrix x2 axis fix (real distance matrix now).
• GRU: hidden bias bhn now applied when there is no incoming hidden state.
• Adafactor: uses broadcast instead of matmul (works for rank > 2), and lazily creates state instead of force-unwrapping.
• Lion: default beta2 0.999 → 0.99.
• Muon: Newton–Schulz now uses norm/addMM — small numeric drift vs. before.
• Device/Stream semantics: streams are now pooled and default device/stream are task-local pairs; Device.cpu/.gpu reflect the enclosing withDefaultDevice/withNewDefaultStream scope rather than fixed index-0 devices.

What's Changed

• Add fp8 conversion utilities by @lucasnewman in #435
• build iOS in CI by @davidkoski in #440
• Add global scale support to quantized layers by @aleroot in #426
• Fix FinalizerCaptureState leak in MLXArray(rawPointer:_:dtype:finalizer:) by @agerjura in #448
• Derive the jit-source lists in update-mlx.sh instead of hand-listing them by @GoodOlClint in #445
• Improve CUDA NVCC spawn performance (on Linux) by @Joannis in #451
• Condition CUDA build plugin on Linux hosts by @ThorFuchs in #447
• Nested initializers by @louen in #456
• [CUDA] Build the CUDA argument lists in steps by @GoodOlClint in #442
• Fix a deadlock between CompiledFunction.lock and evalLock, and a data race on the compiler cache by @DanielMercado60660 in #461
• Fix StreamOrDevice.stream(_:) ignoring its argument by @aleroot in #463
• test Module parameters and compile by @davidkoski in #465
• update for mlx v0.32.2 by @CharlieTLe in #450
• Fix wired-memory ticket lifecycle during cancellation by @aleroot in #471
• Fix typo condiiton -> condition by @Silenterc in #473
• fix(ops): fix convolve padRight for even kernels in same mode by @shubhransh-gupta in #467
• Add countNonzero to the Swift API by @Silenterc in #479
• Distributed support by @ronaldmannak in #482
• add thin logging shim by @davidkoski in #484
• patch for Cuda builds by @davidkoski in #478
• pool Streams, fix Device inheritance by @davidkoski in #472
• replace and improve integration tests by @davidkoski in #477
• pick up fix-custom-io from mlx-c by @davidkoski in #476
• package: exclude mlx/build cmake artifacts from Cmlx target by @Nicolas-nwb in #402

New Contributors

• @agerjura made their first contribution in #448
• @ThorFuchs made their first contribution in #447
• @DanielMercado60660 made their first contribution in #461
• @Silenterc made their first contribution in #473
• @shubhransh-gupta made their first contribution in #467

Full Changelog: 0.31.6...0.32.2