litert-runtime

This skill provides step-by-step guidance for building an Android app that runs a

.tflite

model with the LiteRT CompiledModel API (

com.google.ai.edge.litert:litert

2.x; source repository) on the CPU or the GPU, with the runtime bundled in the APK or from Google Play services. Don't use the deprecated

org.tensorflow.lite.Interpreter

API. For open text language models using

.litertlm

files, suggest the

litert-lm

skill.

Prerequisites

A Kotlin Android project (
```
minSdk
```
must be
```
24
```
or higher, as declared by the LiteRT 2.3.0 AARs).
The runtime that Step 1 picks: the Google Play services runtime (
```
com.google.android.gms:play-services-litert
```
, the default; see LiteRT in Google Play services), or the bundled runtime (
```
implementation("com.google.ai.edge.litert:litert:2.3.0")
```
and
```
implementation("com.google.ai.edge.litert:litert-gpu:2.3.0")
```
from Google Maven). For the bundled runtime, from 2.3.0 the GPU accelerator is the second artifact: with
```
litert
```
alone the app builds, and
```
CompiledModel.create
```
with
```
CompiledModel.Options(Accelerator.GPU)
```
throws
```
LiteRtException
```
because the APK has no GPU accelerator library. With
```
CompiledModel.Options(Accelerator.GPU, Accelerator.CPU)
```
it doesn't throw: the model runs on the CPU while the app reports the GPU, and only Logcat shows it (
```
GPU accelerator could not be loaded and registered
```
). No manifest entry is needed:
```
litert-api
```
, which
```
litert
```
depends on, declares the OpenCL driver libraries (
```
uses-native-library
```
) in its own manifest.
A verified
```
.tflite
```
model and its input contract (input dimensions, mean, standard deviation, channel order, and tensor layout). For models with a verified recipe, inspect the LiteRT model recipes.

Critical mandates

Single-thread confinement : One single-thread executor or confined coroutine dispatcher must own the
```
CompiledModel
```
instance. Creation, inference, and closing must execute on that same background thread, never on the main thread.
Buffer lifecycle : Allocate
```
TensorBuffer
```
instances once using
```
createInputBuffers
```
and
```
createOutputBuffers
```
, reuse them across inferences, and close all input and output buffers before closing the
```
CompiledModel
```
.
Synchronous timing :
```
model.run
```
enqueues work on the GPU and returns before GPU execution completes. Always measure
```
model.run
```
and output buffer readback (
```
outputs[0].readFloat
```
) together.
Warm-up and exception safety : Execute one warm-up inference inside the helper constructor's
```
init
```
block so first-run GPU shader compilation doesn't block the first user interaction, and call
```
close
```
in
```
catch
```
before rethrowing so neither
```
CompiledModel
```
nor
```
TensorBuffer
```
handles leak if buffer allocation or warm-up throws.
Runtime choice : Before Step 1, you must read LiteRT in Google Play services and choose the runtime. Use the Google Play services runtime unless the app needs the NPU, custom operators, or a Support library class that this skill can't replace (such as
```
MetadataExtractor
```
), or also targets devices without Google Play services. Never put both runtimes in the same build variant.
Migration : When the app already uses
```
org.tensorflow.lite
```
code, you must read the "Migrating from legacy TFLite Interpreter" section in full before you edit any file. Migrated code must not import any
```
org.tensorflow.lite
```
class (including
```
DataType
```
), every legacy GPU choice (
```
GpuDelegate
```
, or the Support library's
```
Model.Device.GPU
```
) becomes
```
CompiledModel.Options(Accelerator.GPU, Accelerator.CPU)
```
, and an NNAPI choice maps to that same GPU option.

Detailed steps

Step 1: Configure Gradle and place the model file

Search the project for

libs.versions.toml

gradle.properties

app/build.gradle.kts

, and any existing

org.tensorflow.lite

imports before editing build files. Then pick the runtime:

Google Play services runtime (default) : Use it unless one of the bundled runtime cases applies, also when the request doesn't name a runtime. You must read LiteRT in Google Play services now: its dependencies replace the ones in this step, and its
```
Environment
```
replaces the
```
environment
```
property of Step 2. Use version
```
17.0.0
```
exactly, and run
```
./gradlew assembleDebug
```
(not
```
--dry-run
```
) after you add it. Switch to the bundled runtime only when that build reports
```
Could not find com.google.android.gms:play-services-litert:17.0.0
```
, and tell the user, also when the user asked for Google Play services. Never leave the build broken.
Bundled runtime : Use it only when the app needs the NPU, custom operators, or a Support library class that this skill can't replace (such as
```
MetadataExtractor
```
), when the app also targets devices without Google Play services, or when the user asks for it.
```
FileUtil
```
,
```
ImageProcessor
```
,
```
TensorImage
```
, and
```
TensorLabel
```
don't count: replace them as LiteRT in Google Play services describes. For the bundled runtime, configure
```
minSdk
```
and add the LiteRT runtime and GPU accelerator dependencies in
```
app/build.gradle.kts
```
:
android { defaultConfig { minSdk = 24 } }

dependencies { implementation("com.google.ai.edge.litert:litert:2.3.0") implementation("com.google.ai.edge.litert:litert-gpu:2.3.0") }

Place bundled models at

app/src/main/assets/model.tflite

(AGP stores

.tflite

files uncompressed by default). Deliver a model that exceeds the base APK size limit through Play for On-Device AI as an AI pack, or download it into

context.filesDir

(

bundletool --local-testing

installs AI packs locally without Google Play). Read an install-time AI pack through

AssetManager

. Load a fast-follow or on-demand AI pack or a downloaded file from its on-disk

filePath

using

CompiledModel.create(filePath, CompiledModel.Options(accelerator), environment)

in place of the asset call in Step 2. Obtain the pack's directory from the AI Delivery library (

com.google.android.play:ai-delivery

) using

AiPackManagerFactory.getInstance(context).getPackLocation(name)?.assetsPath()

, which returns

null

until the pack has finished downloading.

With 2.3.0, nothing is added to

gradle.properties

litert

2.3.0 declares the namespace

com.google.ai.edge.litert.impl

and its dependency

litert-api

2.3.0 declares

com.google.ai.edge.litert

, so AGP 9.x builds the project. With 2.2.0 or 2.1.6, both libraries declare

com.google.ai.edge.litert

, and AGP 9.x stops at

processDebugMainManifest

(AGP 8.x reports it as a warning and builds); for those two versions, add

android.uniquePackageNames=false

gradle.properties

, and remove it when moving to 2.3.0 because it also hides the same clash between any other two libraries. Excluding

litert-api

instead doesn't work:

CompiledModel

and

Accelerator

are in

litert-api

, and the build then fails at

compileDebugKotlin

Step 2: Implement the inference helper class

Isolate

CompiledModel

and its

TensorBuffer

instances in a dedicated

AutoCloseable

helper class free of Android UI types:

import android.content.Context
import com.google.ai.edge.litert.Accelerator
import com.google.ai.edge.litert.CompiledModel
import com.google.ai.edge.litert.Environment
import com.google.ai.edge.litert.TensorBuffer

const val INPUT_SIZE = 224 * 224 * 3
private val environment by lazy { Environment.create() }

class Classifier(context: Context, accelerator: Accelerator) : AutoCloseable {
    private val model = CompiledModel.create(
        context.assets,
        "model.tflite",
        CompiledModel.Options(accelerator),
        environment
    )
    private val inputs = mutableListOf<TensorBuffer>()
    private val outputs = mutableListOf<TensorBuffer>()

    init {
        try {
            inputs += model.createInputBuffers()
            outputs += model.createOutputBuffers()
            infer(FloatArray(INPUT_SIZE))
        } catch (e: Exception) {
            close()
            throw e
        }
    }

    fun infer(input: FloatArray): FloatArray {
        inputs[0].writeFloat(input)
        model.run(inputs, outputs)
        return outputs[0].readFloat()
    }

    override fun close() {
        (inputs + outputs).forEach { it.close() }
        model.close()
    }
}

Accelerator.CPU

requires nothing from the device.

Accelerator.GPU

compiles the graph for the GPU; when the GPU can't execute the model, the constructor throws

LiteRtException

(from

create

for an operator the GPU doesn't support, and on the Android emulator from

create

or from buffer allocation). A missing asset or a file that isn't a valid model also throws

LiteRtException

from

create

The buffers are created once with the model, reused for every inference, and closed before the model. One

Environment

serves every model in the process and stays open, so LiteRT loads the GPU accelerator library once rather than on every

create

call. The constructor ends with one warm-up inference (the first GPU run includes shader compilation) and closes the model and its buffers if any step in

init

throws.

INPUT_SIZE

is the model's input element count:

writeFloat

throws on a longer array and writes a shorter one without an error, so a wrong size produces an incorrect result. For

int8

uint8

tensors, use

writeInt8

and

readInt8

TensorBuffer

instead of

writeFloat

and

readFloat

When memory matters, use the following options:

Load from a path :
```
create(context.assets, ...)
```
keeps a copy of the asset's bytes in the native heap while the model is open. Copy the asset once into
```
context.noBackupFilesDir
```
(a downloaded or AI pack file is already on disk) and call
```
create(filePath, ...)
```
so the file is mapped and reclaimable. Version the staged filename, bump the version when the bundled model changes, and delete the previous version's file. Copy to a temporary file and rename it so a copy interrupted by process death is never loaded:
import android.content.Context import com.google.ai.edge.litert.Accelerator import com.google.ai.edge.litert.CompiledModel import java.io.File

fun stageModelAsset( context: Context, assetName: String = "model.tflite", version: String = "v1" ): File = File(context.noBackupFilesDir, "$version-$assetName").also { file -> if (!file.isFile) { context.noBackupFilesDir.listFiles()?.forEach { if (it.name.endsWith("-$assetName")) it.delete() } val tmp = File("${file.path}.tmp") context.assets.open(assetName).use { input -> tmp.outputStream().use { output -> input.copyTo(output) } } check(tmp.renameTo(file)) } }

fun memoryOptions( context: Context, accelerator: Accelerator = Accelerator.CPU, assetName: String = "model.tflite", version: String = "v1" ): CompiledModel.Options = CompiledModel.Options(accelerator).apply { val prefix = "$assetName-" val cache = File(context.codeCacheDir, "$prefix$version.xnnpack_cache") context.codeCacheDir.listFiles()?.forEach { if (it.name.startsWith(prefix) && it.name.endsWith(".xnnpack_cache") && it != cache) { it.delete() } } cpuOptions = CompiledModel.CpuOptions( numThreads = 4, xnnPackFlags = 0x23, xnnPackWeightCachePath = cache.path ) }
Weight cache (
xnnPackWeightCachePath
): Written on the first load (~1x an FP32 or ~2x an FP16 model's file size) and mapped on later loads, moving packed FP32 weights off the heap. Version the cache filename and delete outdated caches.
Transient indirection buffer (
xnnPackFlags = 0x23
) : Default
```
0x3
```
plus flag
```
0x20
```
, replacing per-operator convolution pointer tables with one shared buffer (byte-identical outputs, ~1% latency overhead at 4 threads, ~3% single-threaded). Because the value replaces the default,
```
0x20
```
alone leaves
```
int8
```
and
```
uint8
```
operators to fallback kernels.
```
cpuOptions
```
is ignored on the GPU.
Close after a task: Closing frees the runtime workspace; recreating with a warm cache takes ~20 ms, whereas the first inference after any load pays a full warm-up run, so close when the next use isn't imminent.

Step 3: Wire the ViewModel and Compose screen

Drive the helper from a single-thread executor inside an

AndroidViewModel

, and fall back to

Accelerator.CPU

when GPU compilation fails:

import android.app.Application
import androidx.lifecycle.AndroidViewModel
import com.google.ai.edge.litert.Accelerator
import com.google.ai.edge.litert.LiteRtException
import java.util.concurrent.Executors
import kotlinx.coroutines.CoroutineScope
import kotlinx.coroutines.SupervisorJob
import kotlinx.coroutines.asCoroutineDispatcher
import kotlinx.coroutines.cancel
import kotlinx.coroutines.flow.MutableStateFlow
import kotlinx.coroutines.flow.StateFlow
import kotlinx.coroutines.flow.asStateFlow
import kotlinx.coroutines.flow.update
import kotlinx.coroutines.launch

data class UiState(
    val ready: Boolean = false,
    val accelerator: Accelerator = Accelerator.CPU,
    val result: List<Float>? = null,
    val error: String? = null
)

class MainViewModel(app: Application) : AndroidViewModel(app) {
    private val executor = Executors.newSingleThreadExecutor()
    private val scope = CoroutineScope(
        SupervisorJob() + executor.asCoroutineDispatcher()
    )
    private var classifier: Classifier? = null
    private val _uiState = MutableStateFlow(UiState())
    val uiState: StateFlow<UiState> = _uiState.asStateFlow()

    fun load(preferred: Accelerator = Accelerator.GPU) {
        scope.launch {
            if (classifier != null) return@launch
            val appContext = getApplication<Application>()
            val selected = try {
                classifier = Classifier(appContext, preferred)
                preferred
            } catch (e: LiteRtException) {
                if (preferred != Accelerator.CPU) {
                    try {
                        classifier = Classifier(appContext, Accelerator.CPU)
                        Accelerator.CPU
                    } catch (cpuError: LiteRtException) {
                        _uiState.value = UiState(error = cpuError.message)
                        return@launch
                    }
                } else {
                    _uiState.value = UiState(error = e.message)
                    return@launch
                }
            }
            _uiState.value = UiState(ready = true, accelerator = selected)
        }
    }

    fun classify(input: FloatArray) {
        scope.launch {
            try {
                classifier?.infer(input)?.let { result ->
                    _uiState.update {
                        it.copy(result = result.toList(), error = null)
                    }
                }
            } catch (e: LiteRtException) {
                _uiState.update { it.copy(error = e.message) }
            }
        }
    }

    override fun onCleared() {
        scope.launch { classifier?.close() }.invokeOnCompletion {
            scope.cancel()
        }
        executor.shutdown()
    }
}

One single-thread executor owns the model: creation, inference, and closing execute only on it, never on the main thread. The

ViewModel

keeps its own scope because

viewModelScope

is canceled before

onCleared

runs, so a close launched there wouldn't run;

onCleared

cancels

scope

once

close

completes.

load

is a no-op when a model is already loaded across configuration changes; a failed

load

leaves

classifier

null

classify

reports any

LiteRtException

(in package

com.google.ai.edge.litert

) in

error

and keeps the model.

The following Compose screen takes the picture (

bitmap

) from the photo picker or CameraX and runs

preprocess

in the click handler (a few milliseconds for 224x224; move

preprocess

onto the executor for images exceeding 224x224).

viewModel

and

collectAsStateWithLifecycle

come from

androidx.lifecycle:lifecycle-viewmodel-compose

and

androidx.lifecycle:lifecycle-runtime-compose

import android.graphics.Bitmap
import androidx.compose.foundation.layout.Column
import androidx.compose.material3.Button
import androidx.compose.material3.Text
import androidx.compose.runtime.Composable
import androidx.compose.runtime.LaunchedEffect
import androidx.compose.runtime.getValue
import androidx.lifecycle.compose.collectAsStateWithLifecycle
import androidx.lifecycle.viewmodel.compose.viewModel
import com.google.ai.edge.litert.Accelerator

@Composable
fun MainScreen(
    bitmap: Bitmap,
    viewModel: MainViewModel = viewModel()
) {
    val state by viewModel.uiState.collectAsStateWithLifecycle()
    LaunchedEffect(Unit) { viewModel.load(Accelerator.GPU) }
    Column {
        state.error?.let { Text(it) }
        Button(
            onClick = { viewModel.classify(preprocess(bitmap)) },
            enabled = state.ready
        ) {
            Text("Run (${state.accelerator})")
        }
        state.result?.let { scores ->
            Text("Top class: ${scores.indices.maxBy { i -> scores[i] }}")
        }
    }
}

Step 4: Preprocess inputs to match the model contract

Convert the

Bitmap

into the model's input with

preprocess

. Match the model's exact input dimensions, normalization mean and standard deviation, channel order, and memory layout (this example uses NHWC, RGB, scaled to

-1.0..1.0

). Mismatched

mean

std

values produce corrupted predictions that look like a broken model:

import android.graphics.Bitmap

fun preprocess(
    bitmap: Bitmap,
    size: Int = 224,
    mean: Float = 127.5f,
    std: Float = 127.5f
): FloatArray {
    val scaled = checkNotNull(
        Bitmap.createScaledBitmap(bitmap, size, size, true)
            .copy(Bitmap.Config.ARGB_8888, false)
    )
    val pixels = IntArray(size * size).also {
        scaled.getPixels(it, 0, size, 0, 0, size, size)
    }
    val out = FloatArray(size * size * 3)
    pixels.forEachIndexed { i, p ->
        out[i * 3] = ((p shr 16 and 0xFF) - mean) / std
        out[i * 3 + 1] = ((p shr 8 and 0xFF) - mean) / std
        out[i * 3 + 2] = ((p and 0xFF) - mean) / std
    }
    return out
}

The scaled bitmap is copied to

Bitmap.Config.ARGB_8888

because

ImageDecoder

returns

Bitmap.Config.HARDWARE

bitmaps by default on Android 9 and higher (including from a photo picker

Uri

), scaling preserves that config, and

getPixels

can't read

HARDWARE

bitmaps; the copy is of the 224x224 scaled bitmap rather than the full image. Scaling doesn't preserve the aspect ratio; crop first if the model requires center cropping. An audio or embedding model replaces

preprocess

with its own input encoding and keeps the inference class.

Step 5: Verify output on a physical device

Run the app on a physical device that runs Android from Android Studio. Compare the app's output with the model's reference output for one fixed input on

Accelerator.CPU

first and then on

Accelerator.GPU

; when verifying model outputs on a physical device, you must follow device verification.

Migrating from legacy TFLite Interpreter

An existing app on

org.tensorflow.lite.Interpreter

(from

org.tensorflow:tensorflow-lite*

or from

com.google.ai.edge.litert:litert

1.x) moves to the

CompiledModel

code of Steps 1 to 5 in seven steps, in this order. Before you change any file of an existing app, you must read migration, which gives the commands and code for each step. An app on the Interpreter API in Google Play services (

play-services-tflite-java

) follows LiteRT in Google Play services instead:

Find the runtime: Read the build files and the resolved dependencies, and note the delegate, the thread count, and the Support library classes that the app uses.
Capture a golden output first : While the legacy runtime is still in the project, record the app's output for one fixed input on the CPU. This step needs a device or an emulator. Run the capture test with
```
am instrument
```
, not
```
connectedDebugAndroidTest
```
, which uninstalls the app and deletes the file.
Swap the dependencies : Delete every legacy artifact line (never leave it commented out) and add the dependencies of the runtime that Step 1 picks: by default those of LiteRT in Google Play services, with the same fallback to the bundled runtime when Gradle can't find
```
play-services-litert
```
. On the Play services runtime, replace the Support library classes as LiteRT in Google Play services describes. On the bundled runtime, an app keeps them with
```
com.google.ai.edge.litert:litert-support:2.3.0
```
, with the package
```
org.tensorflow.lite.support
```
renamed to
```
com.google.ai.edge.litert.support
```
(LiteRT Support migration guide).
Replace the delegates :
```
GpuDelegate
```
becomes
```
CompiledModel.Options(Accelerator.GPU, Accelerator.CPU)
```
, so the CPU runs the operators that the GPU doesn't support. An NNAPI choice in the app maps to the same GPU option, not to
```
Accelerator.CPU
```
alone or to an NPU accelerator.
```
setNumThreads(n)
```
becomes new CPU options with
```
cpuOptions = CompiledModel.CpuOptions(numThreads = n)
```
.
Rewrite the inference code : Replace
```
Interpreter
```
and
```
FileUtil.loadMappedFile
```
with the
```
Classifier
```
helper of Step 2, created and run only on the single-thread executor of Step 3, never in a constructor that the main thread calls. When no caller retries on the CPU as in Step 3, create the helper's GPU option with the CPU as a second accelerator (
```
CompiledModel.Options(Accelerator.GPU, Accelerator.CPU)
```
).
Verify in two stages: The build passes and the searches for legacy references print nothing; then the migrated output matches the golden output on a device.
Report what changed and what didn't migrate : When the app used NNAPI, tell the user that it now runs on the GPU with the CPU as a second accelerator, which changes the speed and the numerical results of a quantized model that ran on a DSP or an NPU before; point them to NPU acceleration with LiteRT for the NPU. Tell the user about the TensorFlow Lite Task Library, C++ modules, and Select TensorFlow operators, which this skill doesn't migrate.

Migration mandates that apply to every migrated app:

No
org.tensorflow.lite
imports : Migrated code must not import any
```
org.tensorflow.lite
```
class, including
```
org.tensorflow.lite.DataType
```
; the LiteRT artifacts keep these classes only for backward compatibility. Replace Support calls that need
```
DataType
```
(
```
TensorImage(DataType.FLOAT32)
```
,
```
TensorBuffer.createFixedSize
```
, and
```
TensorLabel
```
) with
```
preprocess
```
of Step 4, the
```
CompiledModel
```
buffers, and labels paired with the output scores. Remove every existing
```
org.tensorflow.lite
```
import line, not only the ones that you add.

Antipatterns

Never create, run, or close
```
CompiledModel
```
on the Android main thread.
Never allocate new
```
TensorBuffer
```
instances on every inference call; failing to close per-frame
```
TensorBuffer
```
objects leaks C++ heap memory.
Never benchmark
```
model.run
```
without including output buffer readback (
```
outputs[0].readFloat
```
).

Never mix

org.tensorflow:tensorflow-lite*

artifacts (such as

org.tensorflow:tensorflow-lite-support

) with

com.google.ai.edge.litert:*

artifacts; both ship the

org.tensorflow.lite.*

classes and cause duplicate class build failures.

Never use
```
Accelerator.GPU
```
on LiteRT 2.3.0 without adding the
```
com.google.ai.edge.litert:litert-gpu:2.3.0
```
dependency alongside
```
com.google.ai.edge.litert:litert:2.3.0
```
; with the CPU as a second accelerator, the model silently runs on the CPU.
Never set
```
xnnPackFlags = 0x20
```
on its own to get the transient indirection buffer; the value replaces the runtime's default
```
0x3
```
, so pass
```
0x23
```
.

Checklist

[ ] Is

minSdk

set to

or higher and are both

com.google.ai.edge.litert:litert:2.3.0

and

com.google.ai.edge.litert:litert-gpu:2.3.0

added in

app/build.gradle.kts

[ ] Are all legacy

org.tensorflow:tensorflow-lite*

dependencies and

org.tensorflow.lite.*

imports removed?

[ ] Do
```
CompiledModel.create
```
,
```
model.run
```
, and
```
close
```
all execute on the same single-thread background dispatcher?
[ ] Are
```
TensorBuffer
```
instances allocated once, reused across calls, and closed before
```
model.close
```
, including when
```
init
```
or warm-up throws?
[ ] Does
```
load
```
execute one warm-up inference and fall back to
```
Accelerator.CPU
```
if
```
Accelerator.GPU
```
throws
```
LiteRtException
```
?
[ ] Does the project compile with
```
./gradlew assembleDebug
```
?
[ ] If memory matters, is the model created from a path with a versioned weight cache and
```
xnnPackFlags = 0x23
```
on the CPU, and closed after a task when the next use isn't imminent?