Crate mlx_c_rust 

Safe ownership boundary around the official MLX C API.

This crate owns exactly one responsibility: the MLX-C C ABI translated into idiomatic Rust. It holds the bindgen-generated raw declarations, the captured-error machinery, the owned handle types (MlxArray, MlxDtype, MlxArrayVector, MlxStream), the per-worker bindings context (MlxBindingsContext: GPU stream and linked version), and every operation wrapper family: creation, shape, padding, elementwise math, activation, normalization, convolution, random, rope, quantized operations, quantization construction, NVFP4, attention, the compiled-graph family, the Metal kernels and capture, and the general operation set. It contains no embedding policy: memory limits, the metallib path selection and verification, safetensors vtable policies, and process I/O accounting belong to the embedding application, which links this crate against an MLX-C library and converts MlxCError values into its own typed runtime errors at its boundary.

API-surface completeness

The bridge is complete, not selective: every public function and type the vendored MLX-C headers declare is carried here. The build script translates the vendored tree in full and emits a bridged-symbol inventory alongside the declarations, so any upstream pin bump that adds, removes, or renames surface shows up as an inventory change the coverage contract can fail on. The recorded state for the current pin lives in this crate’s COVERAGE_INVENTORY.md.

Source-reference audits must never drive removals here: a wrapper with no caller yet is the expected steady state of a complete bridge — the crate exists precisely so that every MLX-C capability is reachable without waiting for a wrapper to be written on demand.

Bindings-only mode

Consumers who provide their own native MLX build — for example a pinned, patched fork produced by their own native orchestration — can take this crate as the bridge alone: depend with default-features = false and the build script emits no link directives at all. The vendored-header declarations, the captured-error machinery, the owned handle types, and every operation wrapper family are identical; the consumer’s own build supplies and links the native libraries. The default link feature carries the whole discovery ladder for everyone else.

Test placement

Tests that never touch MLX-C symbols (the error-channel machinery) live in this crate’s test tree. Anything exercising the bindings themselves needs an MLX-C library at link time, discovered through the MLX_C_RUST_MLX_C_PREFIX environment variable or pkg-config mlx-c (see the build script and README); without one, library compilation and documentation builds still succeed and only test binaries fail to link.

Re-exports

pub use error::MlxCError;

pub use error::check_status;

pub use error::clear_captured_mlx_error;

pub use error::install_non_terminating_error_handler;

pub use error::take_captured_mlx_error;

Modules

error

Captured MLX-C error translation.

raw

Bindgen-generated declarations for the complete official MLX C boundary.

Structs

GraphOutputFile

A rendered-graph destination backed by a C FILE* the caller owns for the duration of one rendering call.

MlxArray

Owned MLX array handle released exactly once through the official C API.

MlxArrayVector

Temporary MLX vector that retains array handles for one aggregate operation.

MlxBindingsContext

Process-global official MLX C bindings context configured for one worker.

MlxClosure

Owned MLX closure handle released exactly once through the C API.

MlxClosureCustom

Owned MLX custom-transform (VJP) closure, released exactly once.

MlxClosureCustomJvp

Owned MLX custom JVP closure, released exactly once.

MlxClosureCustomVmap

Owned MLX custom VMAP closure, released exactly once.

MlxClosureKwargs

Owned MLX closure with keyword arguments, released exactly once.

MlxClosureValueAndGrad

Owned MLX value-and-gradient closure, released exactly once.

MlxCompileCache

Owned MLX compile cache handle released exactly once.

MlxCompiledDualHeadNorm

Owns the compiled dual-output query/key head-normalization graph.

MlxCompiledElementwiseGraphs

Retained shapeless compilations for elementwise model graph composites.

MlxCompiledGatedDeltaRateActivation

Owns the compiled dual-output rate-activation graph.

MlxCompiledMultiOutputGraph

Owns one compiled MLX graph that produces an ordered vector of outputs.

MlxCompiledSwiGlu

Owned shapeless MLX compilation of silu(gate) * input.

MlxCudaKernel

Owned MLX-C CUDA kernel handle.

MlxCudaKernelConfig

Owned MLX-C CUDA kernel configuration handle.

MlxDeviceHandle

Owned MLX device handle released exactly once through the official C API.

MlxDeviceInfo

Owned MLX device information table released exactly once.

MlxDistributedGroup

Owned MLX distributed group handle released exactly once.

MlxFunctionExporter

Owned MLX function exporter released exactly once.

MlxImportedFunction

Owned MLX imported function released exactly once.

MlxIoGguf

Owned MLX GGUF container handle released exactly once.

MlxIoReader

Owned MLX reader transport released exactly once.

MlxIoWriter

Owned MLX writer transport released exactly once.

MlxMapStringToArray

Owned MLX map from strings to arrays, released exactly once.

MlxMapStringToArrayIterator

Owned iterator over an MLX string-to-array map.

MlxMapStringToString

Owned MLX map from strings to strings, released exactly once.

MlxMapStringToStringIterator

Owned iterator over an MLX string-to-string map.

MlxMetalKernel

Owned MLX custom Metal kernel function compiled from a source string.

MlxMetalKernelConfig

MlxMetalKernelOutput

Output shape and dtype requested from one custom Metal kernel launch.

MlxNodeNamer

Owned MLX node namer released exactly once.

MlxStream

Owned MLX stream handle used to preserve runtime thread affinity.

MlxStreamThreadLocal

Per-thread MLX stream slot for one device.

MlxString

Owned MLX string handle released exactly once through the official C API.

MlxVectorInt

Owned MLX integer vector handle released exactly once through the C API.

MlxVectorStream

Owned MLX vector of streams; elements are copied out as owned stream handles.

MlxVectorString

Owned MLX vector of strings; elements are borrowed from the vector for each read and copied into Rust memory.

MlxVectorVectorArray

Owned MLX vector of array vectors.

Enums

MlxDtype

MLX array element types exposed without leaking generated C declarations.

MlxMetalKernelTemplateArgument

Compile-time template argument passed to an MLX custom Metal kernel launch.

Functions

active_memory_bytes

Samples process-local MLX active-memory bytes.

apply_metal_kernel_in_graph

Applies one custom Metal kernel inside a compiled-graph trace.

array_from_vector

Reads one array out of a closure’s input vector inside a graph builder.

cache_memory_bytes

Samples process-local MLX allocator-cache bytes.

clear_allocator_cache

Releases reclaimable MLX allocator-cache allocations.

clear_streams

Frees every stream created in this process except the defaults in use.

cuda_is_available

Whether the CUDA backend reports availability in this process.

disable_compile

Disables MLX’s compilation transform globally.

distributed_is_available

Whether MLX’s distributed backend is initialized in this process.

enable_compile

Re-enables MLX’s compilation transform globally.

export_function

Exports a closure and example arguments to an MLX function file.

export_function_with_keywords

Exports a kwargs closure and example arguments to an MLX function file.

export_graph_to_dot^⚠

Writes the graph behind the outputs to a Graphviz DOT file.

graph_output_array

Runs one raw MLX operation inside a graph builder and owns its lazy output.

initialize_distributed

Initializes MLX’s distributed backend.

jvp_of_closure

Computes the Jacobian-vector product of a closure.

load_array_from_reader

Loads one array from an open reader transport.

load_safetensors

Loads the string-to-array and string-to-string safetensors maps from a file.

load_safetensors_from_reader

Loads the safetensors maps from an open reader transport.

memory_limit_bytes

Reads the allocator limit currently enforced by MLX.

metal_is_available

Whether the MLX-Metal backend is available in this process.

metallib_path

The linked MLX-Metal backend’s metallib path.

peak_memory_bytes

Samples process-local MLX peak active-memory bytes.

print_graph^⚠

Prints the graph behind the outputs as text.

reset_peak_memory

Resets MLX’s process-local peak active-memory counter.

save_array_to_writer

Saves one array through an open writer transport.

save_safetensors

Saves the parameter and metadata maps to a safetensors file.

save_safetensors_to_writer

Saves the parameter and metadata maps through an open writer transport.

set_cache_limit

Applies the allocator-cache retention limit and returns the previous limit in bytes.

set_compile_mode

Selects the global compilation mode.

set_default_stream

Replaces the process’s default stream.

set_graph_output^⚠

Publishes the single output of a single-output graph builder.

set_graph_output_vector^⚠

Appends one lazy output to a multi-output graph’s result vector.

set_memory_limit

Applies the graph-evaluation active-memory guidance and returns the previous limit in bytes.

set_metallib_path

Overrides the metallib MLX loads.

set_wired_limit

Raises the wired-memory ceiling for MLX allocations.

synchronize

Waits for submitted work on the given stream to complete.

synchronize_all_defaults

Synchronizes work across all devices’ default streams.

value_and_grad_of_closure

Builds the value-and-gradient closure for the given closure.

vjp_of_closure

Computes the vector-Jacobian product of a closure.

vmap_replace

Replaces one traced array in a vector-of-arrays graph, the building block of the VMAP transform.

vmap_trace

Traces a closure for the VMAP transform, producing traced inputs and outputs plus the resolved axes.

Type Aliases

MlxGraphBuilder

The MLX C closure ABI every compiled-graph builder follows: it receives the output vector to populate and the input vector to read, and returns zero on success.