v0.1.0

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

v0.0.11

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

v0.0.10

Highlights:

• Faster matmul: up to 2.5x faster for certain sizes, benchmarks
• 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

v0.0.9

Highlights:

• Initial (and experimental) GGUF support
• Support Python buffer protocol (easy interoperability with NumPy, Jax, Tensorflow, PyTorch, etc)
• at[] syntax for scatter style operations: x.at[idx].add(y), (min, max, prod, etc)

Core

• Array creation from other mx.array’s (mx.array([x, y]))
• Complete support for Python buffer protocol
• mx.inner, mx.outer
• mx.logical_and, mx.logical_or, and operator overloads
• Array at syntax for scatter ops
• Better support for in-place operations (+=, *=, -=, ...)
• VJP for scatter and scatter add
• Constants (mx.pi, mx.inf, mx.newaxis, …)

NN

• GLU activation
• cosine_similarity loss
• Cache for RoPE and ALiBi

Bugfixes / Misc

• Fix data type with tri
• Fix saving non-contiguous arrays
• Fix graph retention for inlace state, and remove retain_graph
• Multi-output primitives
• Better support for loading devices

v0.0.7

Core

• Support for loading and saving HuggingFace's safetensor format
• Transposed quantization matmul kernels
• mlx.core.linalg sub-package with mx.linalg.norm (Frobenius, infininty, p-norms)
• tensordot and repeat

NN

• Layers
  • Bilinear,Identity, InstanceNorm
  • Dropout2D, Dropout3D
  • more customizable Transformer (pre/post norm, dropout)
  • More activations: SoftSign, Softmax, HardSwish, LogSoftmax
  • Configurable scale in RoPE positional encodings
• Losses: hinge, huber, log_cosh

Misc

• Faster GPU reductions for certain cases
• Change to memory allocation to allow swapping

v0.0.6

Core

• quantize, dequantize, quantized_matmul
• moveaxis, swapaxes, flatten
• stack
• floor, ceil, clip
• tril, triu, tri
• linspace

Optimizers

• RMSProp, Adamax, Adadelta, Lion

NN

• Layers: QuantizedLinear, ALiBi positional encodings
• Losses: Label smoothing, Smooth L1 loss, Triplet loss

Misc

• Bug fixes

v0.0.5

• Core ops remainder, eye, identity
• Additional functionality in mlx.nn
  • Losses: binary cross entropy, kl divergence, mse, l1
  • Activations: PRELU, Mish, and several others
• More optimizers: AdamW, Nesterov momentum, Adagrad
• Bug fixes