Pytorch not compatible with React-native library

I’ve been using react native for one of my module.Recently we started using Pytorch . Pytorch along with react native is crashing.
Below is the error

java.lang.UnsatisfiedLinkError: couldn’t find DSO to load: caused by: dlopen failed: cannot locate symbol “_ZN8facebook3jni10JByteOrder11nativeOrderEv” referenced by “/data/app/*****-4ZGDZJcLItKdhIsz3hj2rQ==/lib/arm64/”.

Please help me.I am in great urgency.

Hello, @Arun_Kumar1

Could you please share your build setup: gradle files, CMakeLists.txt ( if you have it)?

Both react native and pytorch android use fbjni library, while “_ZN8facebook3jni10JByteOrder11nativeOrderEv” was added for pytorch android, it looks like system loads first react native that does not have JByteOrder::nativeOrder method which is needed for

Hi @IvanKobzarev,

Please check the below build.gradle file content.I don’t have CMakeLists.txt.

apply plugin: “”



  • The react.gradle file registers a task for each build variant (e.g. bundleDebugJsAndAssets
  • and bundleReleaseJsAndAssets).
  • These basically call react-native bundle with the correct arguments during the Android build
  • cycle. By default, bundleDebugJsAndAssets is skipped, as in debug/dev mode we prefer to load the
  • bundle directly from the development server. Below you can see all the possible configurations
  • and their defaults. If you decide to add a configuration block, make sure to add it before the
  • apply from: "../../node_modules/react-native/react.gradle" line.
  • project.ext.react = [
  • // the name of the generated asset file containing your JS bundle
  • bundleAssetName: “”,
  • // the entry file for bundle generation
  • entryFile: “”,
  • // whether to bundle JS and assets in debug mode
  • bundleInDebug: false,
  • // whether to bundle JS and assets in release mode
  • bundleInRelease: true,
  • // whether to bundle JS and assets in another build variant (if configured).
  • // See
  • // The configuration property can be in the following formats
  • // ‘bundleIn${productFlavor}${buildType}’
  • // ‘bundleIn${buildType}’
  • // bundleInFreeDebug: true,
  • // bundleInPaidRelease: true,
  • // bundleInBeta: true,
  • // whether to disable dev mode in custom build variants (by default only disabled in release)
  • // for example: to disable dev mode in the staging build type (if configured)
  • devDisabledInStaging: true,
  • // The configuration property can be in the following formats
  • // ‘devDisabledIn${productFlavor}${buildType}’
  • // ‘devDisabledIn${buildType}’
  • // the root of your project, i.e. where “package.json” lives
  • root: “…/…/”,
  • // where to put the JS bundle asset in debug mode
  • jsBundleDirDebug: “$buildDir/intermediates/assets/debug”,
  • // where to put the JS bundle asset in release mode
  • jsBundleDirRelease: “$buildDir/intermediates/assets/release”,
  • // where to put drawable resources / React Native assets, e.g. the ones you use via
  • // require(’./image.png’)), in debug mode
  • resourcesDirDebug: “$buildDir/intermediates/res/merged/debug”,
  • // where to put drawable resources / React Native assets, e.g. the ones you use via
  • // require(’./image.png’)), in release mode
  • resourcesDirRelease: “$buildDir/intermediates/res/merged/release”,
  • // by default the gradle tasks are skipped if none of the JS files or assets change; this means
  • // that we don’t look at files in android/ or ios/ to determine whether the tasks are up to
  • // date; if you have any other folders that you want to ignore for performance reasons (gradle
  • // indexes the entire tree), add them here. Alternatively, if you have JS files in android/
  • // for example, you might want to remove it from here.
  • inputExcludes: [“android/", "ios/”],
  • // override which node gets called and with what additional arguments
  • nodeExecutableAndArgs: [“node”],
  • // supply additional arguments to the packager
  • extraPackagerArgs: []
  • ]

project.ext.react = [
entryFile: “index.js”,
enableHermes: true, // HERE!

apply from: “…/…/node_modules/react-native/react.gradle”


  • Set this to true to create two separate APKs instead of one:
    • An APK that only works on ARM devices
    • An APK that only works on x86 devices
  • The advantage is the size of the APK is reduced by about 4MB.
  • Upload all the APKs to the Play Store and people will download
  • the correct one based on the CPU architecture of their device.
    def enableSeparateBuildPerCPUArchitecture = false


  • Run Proguard to shrink the Java bytecode in release builds.
    def enableProguardInReleaseBuilds = false

android {
compileSdkVersion rootProject.ext.compileSdkVersion
buildToolsVersion ‘28.0.3’

defaultConfig {
    applicationId "com.progressreport"
    minSdkVersion rootProject.ext.minSdkVersion
    targetSdkVersion rootProject.ext.targetSdkVersion
    versionCode 1
    versionName "1.0"
    ndk {
        abiFilters 'x86','armeabi-v7a','arm64-v8a','x86_64'

packagingOptions {
    pickFirst "**/"

splits {

    // Configures multiple APKs based on screen density.
    splits {
        abi {
            enable true
            include 'armeabi-v7a'
            universalApk false
buildTypes {
    release {
        minifyEnabled enableProguardInReleaseBuilds
        proguardFiles getDefaultProguardFile("proguard-android.txt"), ""
compileOptions {
    sourceCompatibility 1.8
    targetCompatibility 1.8
packagingOptions {
    pickFirst '/armeabi-v7a/'
    pickFirst '/x86/'
    pickFirst '/arm64-v8a/'
    pickFirst '/x86_64/'
    pickFirst '/x86/'
    pickFirst '/armeabi-v7a/'
    pickFirst '/arm64-v8a/'
// applicationVariants are e.g. debug, release
applicationVariants.all { variant ->
    variant.outputs.each { output ->
        // For each separate APK per architecture, set a unique version code as described here:
        def versionCodes = [ "x86":1,"x86_64":2,"armeabi-v7a":3,"arm64-v8a":2]
        def abi = output.getFilter(OutputFile.ABI)
        if (abi != null) {  // null for the universal-debug, universal-release variants
            output.versionCodeOverride =
                    versionCodes.get(abi) * 1048576 + defaultConfig.versionCode


dependencies {
implementation project(’:react-native-svg’)
implementation project(’:lottie-react-native’)
implementation project(’:react-native-localization’)
implementation project(’:react-native-linear-gradient’)
implementation project(’:react-native-vector-icons’)
implementation project(’:react-native-gesture-handler’){}
implementation fileTree(dir: “libs”, include: ["*.jar"])

implementation "${rootProject.ext.supportLibVersion}"
implementation ("com.facebook.react:react-native:+" ){
    exclude  module: 'fbjni-java-only'
} // From node_modules
def jscFlavor = 'org.webkit:android-jsc:+'
def enableHermes = project.ext.react.get("enableHermes", false);
if (enableHermes) {
    // for RN 0.61+
    def hermesPath = "$rootDir/../node_modules/hermes-engine/android/";
    debugImplementation files(hermesPath + "hermes-debug.aar")
    releaseImplementation files(hermesPath + "hermes-release.aar")
} else {
    implementation jscFlavor
def useIntlJsc = false
if (useIntlJsc) {
    implementation ('org.webkit:android-jsc-intl:+')
} else {
    implementation ('org.webkit:android-jsc:+')
                exclude group:'com.facebook.fbjni',module:  "fbjni-java-only-0.0.3"
            } // From node_modules/ From node_modules

implementation(name:'pytorch_android-1.5.0-SNAPSHOT', ext:'aar')
implementation(name:'pytorch_android_fbjni-1.5.0-SNAPSHOT', ext:'aar')
implementation 'com.facebook.soloader:nativeloader:0.8.0'

implementation "androidx.multidex:multidex:2.0.1"


// Run this once to be able to run the application with BUCK
// puts all compile dependencies into folder libs for BUCK to use
task copyDownloadableDepsToLibs(type: Copy) {
from configurations.compile
into ‘libs’

1 Like

Hi @Arun_Kumar1,
I have reproduced your problem. We have contacted react-native team to align our’s. It might take some time and updating to react-native master.

If you desperately need to unblock yourself you could try hacky way to use from pytorch-android in react-native.

  1. Downloading nightly build of pytorch android that contains inside,
wget -O pa.aar

# We will need absolute path to downloaded file
  1. Edit node_modules/react-native/ReactAndroid/build.gradle ( changing task extractJNIFiles which is at the moment is used only for fbjni
copy {
    from zipTree(file)
    into "$projectDir/src/main/jni/first-party/$packageName/"
    include "jni/**/*"


copy {
    from zipTree(new File("${PA_AAR_ABSOLUTE_PATH}"))
    into "$projectDir/src/main/jni/first-party/$packageName/"
    include "jni/**/"

where new File(…) argument is absolute path to pa.aar that was downloaded on step 1.
It worked for me.
(The main idea of this hack - to use from pytorch-android)

I will reply here when react-native and pytorch-android libfbjni will be aligned.
Sorry for inconvenience.

Thanks @IvanKobzarev for quick followup.
Will follow the same steps as mentioned to unblock ourselves for now.

1 Like

Hi, is there an update on the .so alignment between RN & PyTorch? Or an issue to follow?
I wanted to use PyTorch from RN but I’m blocked and don’t really want to force a fbjni that may cause issue in production for RN.

After some more investigation, using the release 1.5.0 android, I was able to get the app to compile but not run the PyTorch. So I got almost the same issue but with a slightly different error message:

E/SoLoader: couldn't find DSO to load: caused by: dlopen failed: 
cannot locate symbol "_ZN8facebook3jni11JByteBuffer5orderENS0_9alias_refINS0_10JByteOrderEEE" 
referenced by "/data/app/com.offline-EWIa3xVhdQYckXATRsmgrw==/lib/arm64/"... result: 0

The symbol of _ZN8facebook3jni11JByteBuffer5orderENS0_9alias_refINS0_10JByteOrderEEE differ from _ZN8facebook3jni10JByteOrder11nativeOrderEv of the original question.

When trying the proposed solution (which is to get the fbjni from PyTorch rather than RN one, it produced the same error

@HugoGresse Please upgrade RN to latest version.If you are using custom build Pytorch library.
Please do the following.

exclude module: ‘fbjni-java-only’

I make an upgrading to RN 0.62.2, but I still have the same error as @HugoGresse mentioned.
Do you have another way to fix this bug ?
Sorry for my bad English!

Guys, importing pytorch like that seems to work for me:

    implementation ('org.pytorch:pytorch_android:1.5.0'){
     exclude module: 'fbjni-java-only'
    implementation ('org.pytorch:pytorch_android_torchvision:1.5.0'){
     exclude module: 'fbjni-java-only'
    implementation ('org.pytorch:pytorch_android_fbjni:1.5.0'){
     exclude module: 'fbjni-java-only'
     exclude module: 'fbjni-java-only'

I also have

android {
        packagingOptions {
	pickFirst '**/*.so'

so that it won’t go crazy over those multiple I’m getting for some reason. Hope that helps!