v0.14.0

Highlights

• Small-size build that JIT compiles kernels and omits the CPU backend which results in a binary <4MB
  • Series of PRs 1, 2, 3, 4, 5
• mx.gather_qmm quantized equivalent for mx.gather_mm which speeds up MoE inference by ~2x
  • Some numbers
• Grouped 2D convolutions
  • Some numbers

Core

• mx.conjugate
• mx.conv3d and nn.Conv3d
• List based indexing
• Started mx.distributed which uses MPI (if installed) for communication across machines
  • mx.distributed.init
  • mx.distributed.all_gather
  • mx.distributed.all_reduce_sum
• Support conversion to and from dlpack
• mx.linalg.cholesky on CPU
• mx.quantized_matmul sped up for vector-matrix products
• mx.trace
• mx.block_masked_mm now supports floating point masks!

Fixes

• Error messaging in eval
• Add some missing docs
• Scatter index bug
• The extensions example now compiles and runs
• CPU copy bug with many dimensions

v0.13.1

🚀

v0.13.0

Highlights

• Block sparse matrix multiply speeds up MoEs by >2x
  • some numbers
• Improved quantization algorithm should work well for all networks
  • see evaluations
• Improved gpu command submission speeds up training and inference
  • some numbers

Core

• Bitwise ops added:
  • mx.bitwise_[or|and|xor], mx.[left|right]_shift, operator overloads
• Groups added to Conv1d
• Added mx.metal.device_info to get better informed memory limits
• Added resettable memory stats
• mlx.optimizers.clip_grad_norm and mlx.utils.tree_reduce added
• Add mx.arctan2
• Unary ops now accept array-like inputs ie one can do mx.sqrt(2)

Bugfixes

• Fixed shape for slice update
• Bugfix in quantize that used slightly wrong scales/biases
• Fixed memory leak for multi-output primitives encountered with gradient checkpointing
• Fixed conversion from other frameworks for all datatypes
• Fixed index overflow for matmul with large batch size
• Fixed initialization ordering that occasionally caused segfaults

v0.12.2

Patch bump (#1067)

* version

* use 0.12.2

v0.12.0

Highlights

• Faster quantized matmul
  • Up to 40% faster QLoRA or prompt processing, some numbers

Core

• mx.synchronize to wait for computation dispatched with mx.async_eval
• mx.radians and mx.degrees
• mx.metal.clear_cache to return to the OS the memory held by MLX as a cache for future allocations
• Change quantization to always represent 0 exactly (relevant issue)

Bugfixes

• Fixed quantization of a block with all 0s that produced NaNs
• Fixed the len field in the buffer protocol implementation

v0.11.0

Core

• mx.block_masked_mm for block-level sparse matrix multiplication
• Shared events for synchronization and asynchronous evaluation

NN

• nn.QuantizedEmbedding layer
• nn.quantize for quantizing modules
• gelu_approx uses tanh for consistency with PyTorch

v0.10.0

Highlights

• Improvements for LLM generation
  • Reshapeless quant matmul/matvec
  • mx.async_eval
  • Async command encoding

Core

• Slightly faster reshapeless quantized gemms
• Option for precise softmax
• mx.metal.start_capture and mx.metal.stop_capture for GPU debug/profile
• mx.expm1
• mx.std
• mx.meshgrid
• CPU only mx.random.multivariate_normal
• mx.cumsum (and other scans) for bfloat
• Async command encoder with explicit barriers / dependency management

NN

• nn.upsample support bicubic interpolation

Misc

• Updated MLX Extension to work with nanobind

Bugfixes

• Fix buffer donation in softmax and fast ops
• Bug in layer norm vjp
• Bug initializing from lists with scalar
• Bug in indexing
• CPU compilation bug
• Multi-output compilation bug
• Fix stack overflow issues in eval and array destruction

v0.9.0

Highlights:

• Fast partial RoPE (used by Phi-2)
• Fast gradients for RoPE, RMSNorm, and LayerNorm
  • Up to 7x faster, benchmarks

Core

• More overhead reductions
• Partial fast RoPE (fast Phi-2)
• Better buffer donation for copy
• Type hierarchy and issubdtype
• Fast VJPs for RoPE, RMSNorm, and LayerNorm

NN

• Module.set_dtype
• Chaining in nn.Module (model.freeze().update(…))

Bugfixes

• Fix set item bugs
• Fix scatter vjp
• Check shape integer overlow on array construction
• Fix bug with module attributes
• Fix two bugs for odd shaped QMV
• Fix GPU sort for large sizes
• Fix bug in negative padding for convolutions
• Fix bug in multi-stream race condition for graph evaluation
• Fix random normal generation for half precision

v0.8.0

Highlights

• More perf!
• mx.fast.rms_norm and mx.fast.layer_norm
• Switch to Nanobind substantially reduces overhead
• Up to 4x faster __setitem__ (e.g. a[...] = b)

Core

• mx.inverse, CPU only
• vmap over mx.matmul and mx.addmm
• Switch to nanobind from pybind11
• Faster setitem indexing
  • Benchmarks
• mx.fast.rms_norm, token generation benchmark
• mx.fast.layer_norm, token generation benchmark
• vmap for inverse and svd
• Faster non-overlapping pooling

Optimizers

• Set minimum value in cosine decay scheduler

Bugfixes

• Fix bug in multi-dimensional reduction

v0.7.0

Highlights

• Perf improvements for attention ops:
  • No copy broadcast matmul (benchmarks)
  • Fewer copies in reshape

Core

• Faster broadcast + gemm
  • benchmarks
• mx.linalg.svd (CPU only)
• Fewer copies in reshape
• Faster small reductions
  • benchmarks

NN

• nn.RNN, nn.LSTM, nn.GRU

Bugfixes

• Fix bug in depth traversal ordering
• Fix two edge case bugs in compilation
• Fix bug with modules with dictionaries of weights
• Fix bug with scatter which broke MOE training
• Fix bug with compilation kernel collision