v0.21.0

Highlights

• Support 3 and 6 bit quantization: benchmarks
• Much faster memory efficient attention for headdim 64, 80: benchmarks
• Much faster sdpa inference kernel for longer sequences: benchmarks

Core

• contiguous op (C++ only) + primitive
• Bfs width limit to reduce memory consumption during eval
• Fast CPU quantization
• Faster indexing math in several kernels:
  • unary, binary, ternary, copy, compiled, reduce
• Improve dispatch threads for a few kernels:
  • conv, gemm splitk, custom kernels
• More buffer donation with no-ops to reduce memory use
• Use CMAKE_OSX_DEPLOYMENT_TARGET to pick Metal version
• Dispatch Metal bf16 type at runtime when using the JIT

NN

• nn.AvgPool3d and nn.MaxPool3d
• Support groups in nn.Conv2d

Bug fixes

• Fix per-example mask + docs in sdpa
• Fix FFT synchronization bug (use dispatch method everywhere)
• Throw for invalid *fft{2,n} cases
• Fix OOB access in qmv
• Fix donation in sdpa to reduce memory use
• Allocate safetensors header on the heap to avoid stack overflow
• Fix sibling memory leak
• Fix view segfault for scalars input
• Fix concatenate vmap

v0.20.0

Highlights

• Even faster GEMMs
  • Peaking at 23.89 TFlops on M2 Ultra benchmarks
• BFS graph optimizations
  • Over 120tks with Mistral 7B!
• Fast batched QMV/QVM for KV quantized attention benchmarks

Core

• New Features
  • mx.linalg.eigh and mx.linalg.eigvalsh
  • mx.nn.init.sparse
  • 64bit type support for mx.cumprod, mx.cumsum
• Performance
  • Faster long column reductions
  • Wired buffer support for large models
  • Better Winograd dispatch condition for convs
  • Faster scatter/gather
  • Faster mx.random.uniform and mx.random.bernoulli
  • Better threadgroup sizes for large arrays
• Misc
  • Added Python 3.13 to CI
  • C++20 compatibility

Bugfixes

• Fix command encoder synchronization
• Fix mx.vmap with gather and constant outputs
• Fix fused sdpa with differing key and value strides
• Support mx.array.__format__ with spec
• Fix multi output array leak
• Fix RMSNorm weight mismatch error

v0.19.3

🚀

v0.19.2

🚀🚀

v0.19.1

🚀

v0.19.0

Highlights

• Speed improvements
  • Up to 6x faster CPU indexing benchmarks
  • Faster Metal compiled kernels for strided inputs benchmarks
  • Faster generation with fused-attention kernel benchmarks
• Gradient for grouped convolutions
• Due to Python 3.8's end-of-life we no longer test with it on CI

Core

• New features
  • Gradient for grouped convolutions
  • mx.roll
  • mx.random.permutation
  • mx.real and mx.imag
• Performance
  • Up to 6x faster CPU indexing benchmarks
  • Faster CPU sort benchmarks
  • Faster Metal compiled kernels for strided inputs benchmarks
  • Faster generation with fused-attention kernel benchmarks
  • Bulk eval in safetensors to avoid unnecessary serialization of work
• Misc
  • Bump to nanobind 2.2
  • Move testing to python 3.9 due to 3.8's end-of-life
  • Make the GPU device more thread safe
  • Fix the submodule stubs for better IDE support
  • CI generated docs that will never be stale

NN

• Add support for grouped 1D convolutions to the nn API
• Add some missing type annotations

Bugfixes

• Fix and speedup row-reduce with few rows
• Fix normalization primitive segfault with unexpected inputs
• Fix complex power on the GPU
• Fix freeing deep unevaluated graphs details
• Fix race with array::is_available
• Consistently handle softmax with all -inf inputs
• Fix streams in affine quantize
• Fix CPU compile preamble for some linux machines
• Stream safety in CPU compilation
• Fix CPU compile segfault at program shutdown

v0.18.1

🚀

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

🐛