0.33.0

MLX

• new mlx release: ml-explore/mlx@v0.32.2...v0.32.4

API Change

Source Breaking

• compile(inputs:outputs:shapeless:_:) does not return @Sendable closure

  • compile(shapeless:_:) requires a @Sendable closure and returns a @Sendable closure
  • see #469 and ml-explore/mlx-swift-lm#589

• MLXCustomFunction closures are now @Sendable

  • see #481

Note: the compile() change should only be source breaking if it was being used in an incorrect way
(which wasn't clear in the past). In particular:

• functions and closures that capture Module state should be using inputs: and are not Sendable
• functions and closures that are not Sendable cannot be made Sendable through this function

If you encounter a build failure:

• see if you should be using the inputs: variant
• see if you need Sendable on the output
• try to convert the function/closure to a pure function

Potentially Source Breaking

• MLXArray is not final, see MaterializedArray below

  • see #418

• Linear, QuantizedLinear, QuantizedEmbedding changed parameters from bare MLXArray to @ParameterInfo

• Logging (MLXLog) no longer uses cLogSupport library -- much simpler now

  • should be transparent as the library was internal
  • see #494

Behavior Change

• eval() no longer holds evalLock while waiting for the result
  • see #474
• tripletLoss() now uses correct p-norm
  • see #505
• .shardedToAll with index segments now builds correct predicate for QuantizedLinear
  • see #503
• JACCL distributed backend enabled (#502).

New API

MLX

• MaterializedArray (#418): an evaluated, immutable array that can be shared across concurrency domains.
  • It has its own +=, -=, *= and /= operators.
  • MLXArray.materialized() -> MaterializedArray is new.
  • mlx-swift-lm adoption will be in ml-explore/mlx-swift-lm#335 or a follow-on
• materialize(_:): takes an MLXArray or [MLXArray] and returns MaterializedArrays
  • It traps if called inside a transform.
• vmapSendable(_:inAxes:outAxes:) -- transforms @Sendable closures
  @sendable versions for the [MLXArray], 1-argument and 2-argument forms.
• 3-parameter VJP: VJP { primals, cotangents, outputs in … }
• Safetensors loading:
  • Concurrent multi-file loading (#462): loadArrays(urls:stream:) and loadArraysAndMetadata(urls:stream:)
  • Byte-progress reporting (#427):
    • a new LoadProgress struct with url, completedUnitCount, totalUnitCount and fractionCompleted
• MLXFast.gatedDeltaUpdate - chunked gated delta network forward pass
• Memory.bufferSize(_:) - returns the size in bytes of the buffers behind the arrays, counting distinct storage only once
• matrixTransposed(_:stream:)
• uniqued() -- returns fixed sized unique values

MLXNN

• MaterializedModule<LayerType: Module>: an open class that is @unchecked Sendable
  • It wraps a module and seals it.
  • After wrapping, mutating APIs on the original module (update, apply, freeze, train, …) trap.
• Module.logicalParameterCount (open var): counts parameters, accounting for quantized packing.

What's Changed

• Fix attention mask for float16 by @louen in #496
• be able to trigger workflows on main. notes about upcoming releases by @davidkoski in #489
• fix xcframework builds by @davidkoski in #495
• fix lock inversion between evalLock and dispatch_once on globalStreams by @davidkoski in #499
• idea for doing synchronous evals but not holding evalLock by @davidkoski in #474
• Fix leak when _updateInternal replaces an array with siblings by @ShankyShako in #501
• Enable JACCL by @ronaldmannak in #502
• Fix asArray crash on zero-size arrays by @tk1475 in #486
• compile and vmap can produce Sendable closures from Sendable inputs by @davidkoski in #469
• Fix iOS 26.2 launch crash from Logger.isEnabled by @pmanot in #494
• Fix shardInPlace wrong integer split positions on quantized layers by @ronaldmannak in #503
• Add Sendable closures and 3-parameter VJP support to MLXCustomFunction by @shubhransh-gupta in #481
• GPU.deviceInfo(): select the device like MLX, without the window server by @davidtai in #498
• MaterializedArray is a Sendable MLXArray by @davidkoski in #418
• Fix tripletLoss for non-default p-norm by @Silenterc in #505
• Add concurrent synchronous safetensors loading across files by @aleroot in #462
• Add byte-progress reporting for safetensors file loading by @aleroot in #427
• turn down warnings from the C++ code by @davidkoski in #506
• Fix crossEntropy and nllLoss for a class axis other than the last by @HamedShams in #509
• [CUDA] Allow selecting nvcc's host compiler, and invalidate on change by @GoodOlClint in #443
• [CUDA] Make the CUDA include roots configurable by @GoodOlClint in #444
• Add optimizer state save and restore by @sbkoth in #487
• mlx v0.32.4 by @davidkoski in #508
• fix compatibility issue in update(parameters:) by @davidkoski in #510
• debug CI failure by @davidkoski in #514
• Support for 0.32.4 distributed functions by @ronaldmannak in #512
• Fix a potential data race in ErrorBox.check() by @Silenterc in #513

New Contributors

• @ShankyShako made their first contribution in #501
• @tk1475 made their first contribution in #486
• @pmanot made their first contribution in #494
• @davidtai made their first contribution in #498
• @HamedShams made their first contribution in #509
• @sbkoth made their first contribution in #487

Full Changelog: 0.32.3...0.33.0