mlx 0.2.0

Dart bindings for Apple's MLX array framework on Apple Silicon: tensors, neural-network layers, FFT, quantization, and safetensors — via FFI to mlx-c.

Changelog

0.2.0 - 2026-08-06

• Added explicit bundle-first and bundle-exclusive native artifact authority. Release hosts can now validate a self-contained runtime bundle with a pinned version and required SHA-256 checksums for all four native artifacts before FFI initialization, then fail closed instead of consulting developer paths.
• Added setCacheLimit(bytes), wrapping mlx-c's mlx_set_cache_limit. It caps the pool of cached (unused) buffers and returns the previous limit, so a long-lived host process running small, bursty inferences can bound MLX's memory the way Swift's GPU.set(cacheLimit:) does.

0.1.0 - 2026-07-19

Initial release: Dart FFI bindings to mlx-c for Apple's MLX on Apple Silicon.

• Core: the MlxArray tensor type with DType, lazy eval, streams, and memory helpers (cache clearing, peak-memory tracking).
• Ops: elementwise math and operators, matmul, the softmax family, reductions and orderings, array creation, indexing / shape manipulation, and channels-last convolutions.
• FFT: real-input rfft / irfft.
• fast: fused kernels — scaled dot-product attention, layer norm, RMS norm, and RoPE.
• random: keyed sampling — randomKey, randomNormal, categorical, and seeding.
• Quantization: group-wise affine quantize / dequantize / quantizedMatmul.
• nn: an MLXNN-style neural-network layer with a Module base; linear / conv / embedding / norm / RNN layers; activations; a KV-cache; and quantization machinery whose registered parameter names match MLX-converted checkpoints 1:1.
• I/O: safetensors loading.