HuggingFaceTokenizer android load

[soonguKwon]

Description

I tried to load HuggingFaceTokenizer on android but failed.

HuggingFaceTokenizer tokenizer = HuggingFaceTokenizer.newInstance(assetFilePath(this, "tokenizer.json"), options);

Error Message

ai.djl.engine.EngineException: Failed to load Huggingface native library.

[frankfliu]

Currently Huggingface Tokenizer extension is not available for Android. We only added PyTorch and OnnxRuntime to Android.

[soonguKwon]

Do you have any support plans?

Currently Huggingface Tokenizer extension is not available for Android. We only added PyTorch and OnnxRuntime to Android.

[frankfliu]

We don't have bandwidth currently. It shouldn't be hard to add Android support. No java code change is needed. The most of work is setup CI to build android native library. We can work with you if you want to contribute.

[BigSquirrel2000]

@frankfliu do you know if someone is already working on this?

[lanking520]

https://github.com/mozilla/rust-android-gradle

[BigSquirrel2000]

currently I managed to recompile rust

but during System.loadLibrary("tokenizers") in android emulator, following error occurs:
java.lang.UnsatisfiedLinkError: dlopen failed: library "libc++_shared.so" not found: needed by /lib/x86_64/libtokenizers.so in namespace classloader-namespace

any clues what could be a solution?

[BigSquirrel2000]

update: it seems like gradle issue gradle/gradle#23023
for that moment, quickfix described in above issue - seems to work.

[lanking520]

Thanks for your discovery @BigSquirrel2000. Are you willing to contribute your working solution to DJL repo? For Android NDK build

[naveen521kk]

Hey, I would like to contribute to this issue. I came across this library while looking for a way to run the tokenizer in my Android app project.

I've figured out how to build the rust library using cargo-ndk. However, I wasn't able to figure out how to build OpenSSL, so I've disabled the functions which will use Hugging Face API - that is only local paths would work on Android. I've also created a build native library for Android like pytorch-native to distribute the rust library.

See https://github.com/deepjavalibrary/djl/compare/master...naveen521kk:djl:tokenizer-android-support?expand=1 for the changes I made.

I've tested the resultant build output in my Android app and it works - made a small wrapper for it in react native too.

I could make a PR out of that branch if someone could help me configure the CI.

Thanks!

cc @frankfliu

[frankfliu]

@naveen521kk

Really appreciate for your contribution. Your changes looks great, and I tried locally, it works. Please raise a PR, once it merged, we can add CI to publish it.

[naveen521kk]

@naveen521kk

Really appreciate for your contribution. Your changes looks great, and I tried locally, it works. Please raise a PR, once it merged, we can add CI to publish it.

Hey, I've created #3531, have a look at it.