#include <mlx_backend.hpp>

Inheritance diagram for continuum::backend::MLXBackend:

Public Member Functions
BackendCapabilities	capabilities () const override

std::string	tensor_backend_type () const override

BackendRunResult	run_with_cache (const ir::Node &node, const std::vector< continuum::Value > &inputs, const std::optional< BackendState > &prefix_state, std::int32_t remaining_tokens) override

Public Member Functions inherited from continuum::backend::Backend
virtual	~Backend ()=default

virtual std::vector< std::uint8_t >	export_state (const BackendState &) const

virtual std::optional< BackendState >	import_state (const std::vector< std::uint8_t > &)

Static Public Member Functions
static std::string	runtime ()
	"mlx <version> (gpu\|cpu)" when built against MLX, else "reference (built without MLX)".

Detailed Description

Tensor backend behind the mlx tensor type (CONTINUUM_TENSOR_BACKEND=mlx).

Built with Apple's MLX (CONTINUUM_USE_MLX, on by default when the mlx package is installed on Apple silicon), every op runs as an mlx::core kernel: on the Metal GPU on Apple silicon (CONTINUUM_MLX_DEVICE=cpu to pin the CPU), on the CPU with Linux wheels. Built without MLX, the same ops fall back to portable C++ reference kernels. runtime() reports which is in use. Tensors are row-major float32 MlxTensorValues. Ops: identity / input / id, relu, softmax (attr 0 = dim, default -1), add (same shape), matmul (1D x 1D dot, or 2D x 2D). Tensor-only: no token path, no cache, no portable state. See the capability matrix in docs/design/abi.md.

Member Function Documentation

◆ capabilities()

BackendCapabilities continuum::backend::MLXBackend::capabilities ( ) const

overridevirtual

Implements continuum::backend::Backend.

◆ run_with_cache()

BackendRunResult continuum::backend::MLXBackend::run_with_cache	(	const ir::Node &	node,
		const std::vector< continuum::Value > &	inputs,
		const std::optional< BackendState > &	prefix_state,
		std::int32_t	remaining_tokens
	)

overridevirtual

Implements continuum::backend::Backend.

◆ runtime()

static std::string continuum::backend::MLXBackend::runtime ( )

static

"mlx <version> (gpu|cpu)" when built against MLX, else "reference (built without MLX)".

◆ tensor_backend_type()

std::string continuum::backend::MLXBackend::tensor_backend_type ( ) const

inlineoverridevirtual

Reimplemented from continuum::backend::Backend.

The documentation for this class was generated from the following file:

include/continuum/backend/mlx_backend.hpp