Crate mlx_core 

Safe, ergonomic Rust API for MLX tensors and operations.

mlx-core provides the foundational types (Tensor, Device, DType, Shape) and a backend-agnostic interface for lazy tensor computation.

Architecture

Operations on tensors build a lazy computation graph. Calling eval() (or to_vec_f32()) triggers a topological walk that dispatches each node to the active Backend. A default CPU reference backend is provided out of the box.

Backends

• Built-in CPU reference: simple, safe Rust — always available
• ffi feature: delegates to the MLX C++ runtime via mlx-sys
• mlx-cpu crate: optimized pure-Rust CPU backend (future)
• mlx-metal crate: native Apple Silicon acceleration (future)

Re-exports

pub use graph::NodeId;

pub use tensor::Device;

pub use tensor::Tensor;

pub use types::DType;

pub use types::Shape;

Modules

backend

Backend trait and Stream — pluggable compute engine for tensor evaluation.

cpu_kernels

Built-in CPU reference backend — correctness oracle.

graph

Lazy computation graph IR.

tensor

Tensor type — a lazy handle to a node in the computation graph.

types

Core type definitions: DType, Shape.

Enums

MlxError

Type Aliases

Result