run_executable

→std::vector<TensorImplPtr>

int run_executable(CompiledExecutable * exe, const int & input_feeds)

source edit

Run a compiled executable.

Parameters

exeCompiledExecutable*

Executable returned by compile_trace; must be non-null.

input_feedsconst std::vector<TensorImplPtr>&

Inputs in the executable's input_ids order. Each feed must live on the same device as the executable; the GPU path materialises the underlying MTLBuffer via the MLX bridge and binds it as an MPSGraphTensorData.

Returns

std::vector<TensorImplPtr>

Freshly-allocated output tensors in output_ids order. Each output's GpuStorage wraps a new MTLBuffer allocated by the executable run — no copy back to MLX.

Notes

Implementation lives in CompiledExecutable.mm (Objective-C++) so all MPSGraph types stay in one translation unit.