Struct GraphExecutor 

pub struct GraphExecutor { /* private fields */ }

Batched Metal dispatch — encodes multiple ops into a single CommandEncoder.

Create one per model (or per forward-pass loop). Call begin at the start of each forward pass to get a GraphSession that holds the shared encoder.

Implementations

impl GraphExecutor

pub fn new(device: MlxDevice) -> Self

Create a new graph executor backed by the given device.

pub fn begin(&self) -> Result<GraphSession<'_>>

Begin a new forward pass (direct-dispatch mode).

Returns a GraphSession that holds a fresh CommandEncoder. All ops encoded through the session share this single encoder. Call GraphSession::finish to commit and wait.

pub fn begin_recorded(&self) -> Result<GraphSession<'_>>

Begin a new forward pass in capture (record) mode.

All op calls are recorded into a ComputeGraph instead of being dispatched to Metal. When GraphSession::finish is called, the recorded graph is replayed into a fresh encoder via ComputeGraph::encode_sequential().

The API is identical to begin() — callers do not need to change any op call code. The only behavioral difference: GPU work happens at finish() time rather than at each op call.

pub fn device(&self) -> &MlxDevice

Borrow the underlying device.