v0.18.0

Highlights

• Speed improvements:
  • Up to 2x faster I/O: benchmarks.
  • Faster transposed copies, unary, and binary ops
    • CPU benchmarks here.
    • GPU benchmarks here and here.
• Transposed convolutions
• Improvements to mx.distributed (send/recv/average_gradients)

Core

• New features:

  • mx.conv_transpose{1,2,3}d
  • Allow mx.take to work with integer index
  • Add std as method on mx.array
  • mx.put_along_axis
  • mx.cross_product
  • int() and float() work on scalar mx.array
  • Add optional headers to mx.fast.metal_kernel
  • mx.distributed.send and mx.distributed.recv
  • mx.linalg.pinv

• Performance

  • Up to 2x faster I/O
  • Much faster CPU convolutions
  • Faster general n-dimensional copies, unary, and binary ops for both CPU and GPU
  • Put reduction ops in default stream with async for faster comms
  • Overhead reductions in mx.fast.metal_kernel
  • Improve donation heuristics to reduce memory use

• Misc

  • Support Xcode 160

NN

• Faster RNN layers
• nn.ConvTranspose{1,2,3}d
• mlx.nn.average_gradients data parallel helper for distributed training

Bug Fixes

• Fix boolean all reduce bug
• Fix extension metal library finding
• Fix ternary for large arrays
• Make eval just wait if all arrays are scheduled
• Fix CPU softmax by removing redundant coefficient in neon_fast_exp
• Fix JIT reductions
• Fix overflow in quantize/dequantize
• Fix compile with byte sized constants
• Fix copy in the sort primitive
• Fix reduce edge case
• Fix slice data size
• Throw for certain cases of non captured inputs in compile
• Fix copying scalars by adding fill_gpu
• Fix bug in module attribute set, reset, set
• Ensure io/comm streams are active before eval
• Fix mx.clip
• Override class function in Repr so mx.array is not confused with array.array
• Avoid using find_library to make install truly portable
• Remove fmt dependencies from MLX install
• Fix for partition VJP
• Avoid command buffer timeout for IO on large arrays

v0.17.3

🚀

v0.17.1

🐛

v0.17.0

Highlights

• mx.einsum: PR
• Big speedups in reductions: benchmarks
• 2x faster model loading: PR
• mx.fast.metal_kernel for custom GPU kernels: docs

Core

• Faster program exits
• Laplace sampling
• mx.nan_to_num
• nn.tanh gelu approximation
• Fused GPU quantization ops
• Faster group norm
• bf16 winograd conv
• vmap support for mx.scatter
• mx.pad "edge" padding
• More numerically stable mx.var
• mx.linalg.cholesky_inv/mx.linalg.tri_inv
• mx.isfinite
• Complex mx.sign now mirrors NumPy 2.0 behaviour
• More flexible mx.fast.rope
• Update to nanobind 2.1

Bug Fixes

• gguf zero initialization
• expm1f overflow handling
• bfloat16 hadamard
• large arrays for various ops
• rope fix
• bf16 array creation
• preserve dtype in nn.Dropout
• nn.TransformerEncoder with norm_first=False
• excess copies from contiguity bug

v0.16.3

🚀

v0.16.2

🚀🚀

0.16.1

🚀

v0.16.0

Highlights

• @mx.custom_function for custom vjp/jvp/vmap transforms
• Up to 2x faster Metal GEMV and fast masked GEMV
  • benchmarks
• Fast hadamard_transform
  • benchmarks

Core

• Metal 3.2 support
• Reduced CPU binary size
• Added quantized GPU ops to JIT
• Faster GPU compilation
• Added grads for bitwise ops + indexing

Bug Fixes

• 1D scatter bug
• Strided sort bug
• Reshape copy bug
• Seg fault in mx.compile
• Donation condition in compilation
• Compilation of accelerate on iOS

v0.15.2

🚀

v0.15.1

🚀