For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python class

JitExecutor

JitExecutor

class max.experimental.executor.JitExecutor(*, interpreter, sync_on_interpreter_fallback=True)

source

Bases: object

Executes via the interpreter while compiling in the background.

The first call for a given graph (keyed by structure) submits the graph to a background compile process (ProcessCompilePool) and caches the resulting future for the life of this executor. The future resolves to an initialized model: session.init runs as soon as the compile lands, off the calling thread, so a call that finds a done future executes immediately. While the compile is pending, calls execute via the interpreter. When the interpreter refuses a graph, the call waits for that graph’s compile.

Failed compilations are assumed to not be transient and are cached.

Initializes the executor.

Parameters:

• interpreter (Executor) – Executor serving graphs whose compile is pending.
• sync_on_interpreter_fallback (bool) – Whether a graph the interpreter refuses waits for its compile. When False such a graph raises UnsupportedGraphError instead.

cache

cache: dict[tuple[str, tuple[tuple[str, str], ...]], Future[Model]]

source

compile()

compile(graph)

source

Submits the graph to the compiler on first sight.

Returns:

A future resolving to the initialized model.

Parameters:

graph (Graph)

Return type:

Future[Model]

execute()

execute(graph, inputs)

source

Executes graph, compiling in the background on first call.

Raises:

Exception – The interpreter’s own error, when it fails and sync_on_interpreter_fallback is False.

Parameters:

• graph (Graph)
• inputs (Sequence[Buffer])

Return type:

Sequence[Buffer | None]

lock

lock: lock

source