public final class InferenceSession

Object
InferenceSession

Reusable, native on-device session for a TensorFlow Lite model.

Opening and execution are asynchronous because model allocation and delegates can be expensive. Metadata and resize operations are synchronous. A session is not usable after close(); applications should retain and reuse one session instead of reopening the model for every input.

Methods

`public static boolean isSupported()`	Tests whether the current port includes a native LiteRT runtime.
`public static` `AsyncResource<InferenceSession> open(ModelSource source, InferenceOptions options)`	Opens and allocates a model session off the EDT.
`public` `TensorInfo[] getInputs()`	Returns the model’s current input metadata.
`public` `TensorInfo[] getOutputs()`	Returns the model’s current output metadata.
`public` `AsyncResource<Tensor[]> run(Tensor[] inputs)`	Copies input tensors to native memory, invokes the model, and returns every output tensor.
`public void resizeInput(String name, int[] shape)`	Resizes an input and reallocates native tensors before the next run.
`public void close()`	Releases the interpreter, delegates, and any temporary staged model.

Inherited methods

From Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait, clone

Method details

isSupported

public static boolean isSupported()

Tests whether the current port includes a native LiteRT runtime.

Returns

true when sessions can be opened on this target

open

public static AsyncResource<InferenceSession> open(ModelSource source, InferenceOptions options)

Opens and allocates a model session off the EDT.

The option values are copied before asynchronous backend work is scheduled. Reusing or changing the supplied InferenceOptions after this method returns therefore cannot alter the pending open. Canceling the returned resource prevents session publication; if the native backend finishes opening afterward, its handle is closed automatically.

Parameters

source ModelSource: bytes, resource, or file containing a .tflite model
options InferenceOptions: execution options; null uses defaults

Returns

an asynchronous session, failed with InferenceException when the model or requested accelerator cannot be used

getInputs

public TensorInfo[] getInputs()

Returns the model’s current input metadata. Shapes reflect the most recent successful resizeInput(String, int[]) call. Metadata cannot be queried while run(Tensor[]) is pending because the native interpreter is mutable and may be updating tensor state.

Returns

a defensive copy of the input metadata array

Throws

IllegalStateException: if the session is closed or a run is pending

getOutputs

public TensorInfo[] getOutputs()

Returns the model’s current output metadata. Backends refresh this information after an invocation so models with dynamically resolved output dimensions report the shape used to decode the returned tensor. Metadata cannot be queried while run(Tensor[]) is pending because the native interpreter is mutable and may be updating tensor state.

Returns

a defensive copy of the output metadata array

Throws

IllegalStateException: if the session is closed or a run is pending

run

public AsyncResource<Tensor[]> run(Tensor[] inputs)

Copies input tensors to native memory, invokes the model, and returns every output tensor. Named tensors are matched by name; unnamed tensors are matched by position. Calling close() while this operation is pending prevents new work immediately but defers native release until the returned resource succeeds or fails. A session accepts one run at a time because the underlying native interpreter is mutable. The array container is defensively copied before it is handed to the asynchronous backend, so replacing an element after this method returns cannot change the pending invocation. Each Tensor is itself immutable. Canceling the returned resource suppresses result publication but does not interrupt an invocation that has already entered LiteRT. The session remains busy, and a pending close() remains deferred, until the native operation actually succeeds or fails.

Parameters

inputs Tensor[]: one tensor for each model input; null is treated as an empty input array

Returns

asynchronous output tensors in model output order

Throws

IllegalStateException: if the session is closed or already running
IllegalArgumentException: if an input name, count, or shape does not match the model’s current input metadata

resizeInput

public void resizeInput(String name, int[] shape)

Resizes an input and reallocates native tensors before the next run.

This method throws while an asynchronous run(Tensor[]) is pending because native runtimes cannot safely reallocate tensors during an invocation.

Parameters

name String: model input name, or null for the first input
shape int[]: new non-negative dimensions

Throws

IllegalStateException: if the session is closed or a run is pending

close

public void close()

Releases the interpreter, delegates, and any temporary staged model. The original file supplied by ModelSource.file(String) is never deleted. If a run is pending, release is deferred until that run settles. Calling this method more than once has no effect.