Mlx

Highlights:

- Partial shapeless compilation
- Default shapeless compilation for all activations
- Can be more than 5x faster than uncompiled versions
- CPU kernel fusion
- Some functions can be up to [10x faster](https://github.com/ml-explore/mlx/pull/691#issuecomment-1947708207)

Core

- CPU compilation
- Shapeless compilation for some cases
- `mx.compile(function, shapeless=True)`
- Up to 10x faster scatter: [benchmarks](https://github.com/ml-explore/mlx/pull/709#issuecomment-1951733610)
- `mx.atleast_1d`, `mx.atleast_2d`, `mx.atleast_3d`

Bugfixes

- Bug with `tolist` with `bfloat16` and `float16`
- Bug with `argmax` on M3

Highlights:

- `mx.fast` subpackage
- Custom `mx.fast.rope` up to [20x faster](https://github.com/ml-explore/mlx/pull/676#issuecomment-1940672929)

Core

- Support metadata with `safetensors`
- Up to 5x faster scatter and 30% faster gather
- 40% faster `bfloat16` quantizated matrix-vector multiplies
- `mx.fast` subpackage with a fast RoPE
- Context manager `mx.stream` to set the default device

NN

- Average and Max pooling layers for 1D and 2D inputs

Optimizers

- Support schedulers for e.g. learning rates
- A few basic schedulers:
- `optimizers.step_decay`
- `optimizers.cosine_decay`
- `opimtizers.exponential_decay`

Bugfixes

- Fix bug in remainder with negative numerators and integers
- Fix bug with slicing into softmax
- Fix quantized matmuls with non 32 multiples

Highlights:

- `mx.compile` **makes stuff go fast**
- Some functions are up to 10x faster ([benchmarks](https://github.com/ml-explore/mlx/pull/614#issuecomment-1929056286))
- Training models anywhere from 10% to twice as fast ([benchmarks](https://github.com/ml-explore/mlx-examples/pull/420#issuecomment-1932422605))
- Simple syntax for compiling full training steps

Core

- `mx.compile` function transformation
- Find devices properly for iOS
- Up to 10x faster GPU gather
- `__abs__` overload for `abs` on arrays
- `loc` and `scale` in parameter for `mx.random.normal`

NN

- Margin ranking loss
- BCE loss with weights

Bugfixes

- Fix for broken eval during function transformations
- Fix `mx.var` to give `inf` with `doff >= nelem`
- Fix loading empty modules in `nn.Sequential`

Highlights

- Memory use improvements:
- Gradient checkpointing for training with `mx.checkpoint`
- Better graph execution order
- Buffer donation

Core

- Gradient checkpointing with `mx.checkpoint`
- CPU only QR factorization `mx.linalg.qr`
- Release Python GIL during `mx.eval`
- Depth-based graph execution order
- Lazy loading arrays from files
- Buffer donation for reduced memory use
- `mx.diag`, `mx.diagonal`
- Breaking: `array.shape` is a Python tuple
- GPU support for `int64` and `uint64` reductions
- vmap over reductions and arg reduction:
- `sum`, `prod`, `max`, `min`, `all`, `any`
- `argmax`, `argmin`

NN

- Softshrink activation

Bugfixes

- Comparisons with `inf` work, and fix `mx.isinf`
- Bug fix with RoPE cache
- Handle empty Matmul on the CPU
- Negative shape checking for `mx.full`
- Correctly propagate `NaN` in some binary ops
- `mx.logaddexp`, `mx.maximum`, `mx.minimum`
- Fix > 4D non-contiguous binary ops
- Fix `mx.log1p` with `inf` input
- Fix SGD to apply weight decay even with 0 momentum

Highlights:
- GGUF improvements:
- Native quantizations `Q4_0`, `Q4_1`, and `Q8_0`
- Metadata

Core
- Support for reading and writing GGUF metadata
- Native GGUF quantization (`Q4_0`, `Q4_1`, and `Q8_0`)
- Quantize with group size of 32 (2x32, 4x32, and 8x32)

NN
- `Module.save_weights` supports safetensors
- `nn.init` package with several commonly used neural network initializers
- Binary cross entropy and cross entropy losses can take probabilities as targets
- `Adafactor` in `nn.optimizers`

Bugfixes

- Fix `isinf` and friends for integer types
- Fix array creation from list Python ints to `int64`, `uint`, and `float32`
- Fix power VJP for `0` inputs
- Fix out of bounds `inf` reads in `gemv`
- `mx.arange` crashes on NaN inputs

Highlights:

- Faster matmul: up to 2.5x faster for certain sizes, [benchmarks](https://github.com/ml-explore/mlx/pull/424#issuecomment-1898815724)
- Fused matmul + addition (for faster linear layers)

Core

- Quantization supports sizes other than multiples of 32
- Faster GEMM (matmul)
- ADMM primitive (fused addition and matmul)
- `mx.isnan`, `mx.isinf`, `isposinf`, `isneginf`
- `mx.tile`
- VJPs for `scatter_min` and `scatter_max`
- Multi output split primitive

NN
- Losses: Gaussian negative log-likelihood

Misc

- Performance enhancements for graph evaluation with lots of outputs
- Default PRNG seed is based on current time instead of 0
- Primitive VJP takes output as input. Reduces redundant work without need for simplification
- PRNGs default seed based on system time rather than fixed to 0
- Format boolean printing in Python style when in Python

Bugfixes

- Scatter < 32 bit precision and integer overflow fix
- Overflow with `mx.eye`
- Report Metal out of memory issues instead of silent failure
- Change `mx.round` to follow NumPy which rounds to even