GridGain ML Model Deployment and Running Predictions

List<String> urls = Arrays.asList(
    "https://huggingface.co/model-repo/resolve/main/model.onnx",
    "https://huggingface.co/model-repo/resolve/main/tokenizer.json"
);

Path tempModelDir = ModelUtils.downloadAndDeployModelAndJar(
    urls,                // URLs to download model files from
    "model-id",         // Model identifier
    "1.0.0",            // Model version
    sourceDir           // Path to custom classes for JAR packaging
);

String modelUrl = "djl://ai.djl.huggingface.pytorch/deepset/minilm-uncased-squad2";

Path tempModelDir = ModelUtils.downloadAndDeployDJLModelAndJar(
    modelUrl,                           // DJL model URL
    PytorchQAInput.class,              // Input type class
    String.class,                       // Output type class
    "pytorch-qa-model",                // Model identifier
    "1.0.0",                           // Model version
    sourceDir,                          // Path to custom classes
    PytorchQATranslatorFactory.class.getName() // Translator factory
);

List<String> urls = Arrays.asList(
    "https://huggingface.co/model-repo/resolve/main/saved_model.pb",
    "https://huggingface.co/model-repo/resolve/main/variables/variables.data-00000-of-00001",
    "https://huggingface.co/model-repo/resolve/main/variables/variables.index"
);

Path tempModelDir = ModelUtils.downloadAndDeployModelAndJarWithFolderStructure(
    urls,                // URLs to download model files from
    "tf-model-id",      // Model identifier
    "1.0.0",            // Model version
    sourceDir           // Path to custom classes for JAR packaging
);

String parameterizedQuery = "SELECT review_text FROM product_reviews WHERE product_category = ? AND sentiment IS NULL LIMIT ?";
Serializable[] params = {"Electronics", 10};

MlSqlJobParameters jobParams = MlSqlJobParameters.builder()
    // ... other parameters ...
    .sqlQuery(parameterizedQuery)
    .sqlParams(params)
    .limit(10) // Additional limit
    .build();

List<Classifications> results = ml.predictFromSql(jobParams);

start(MODE.EMBEDDED);
mlApi = server.api().ml(); // Direct access to ML API

start(MODE.CLIENT_ONLY);
mlApi = client.ml(); // ML API accessed via client connection

public class CustomTranslator implements NoBatchifyTranslator<CustomInput, CustomOutput> {

    private CustomTranslator(HuggingFaceTokenizer tokenizer, boolean int32) {
        this.tokenizer = tokenizer;
        this.int32 = int32;
    }

    @Override
    public void prepare(TranslatorContext ctx) throws IOException, ModelException {
        Model model = ctx.getModel();
        this.predictor = model.newPredictor(new NoopTranslator((Batchifier)null));
        ctx.getPredictorManager().attachInternal(UUID.randomUUID().toString(),
            new AutoCloseable[]{this.predictor});
    }

    @Override
    public NDList processInput(TranslatorContext ctx, CustomInput input) {
        //process input to create and return an NDList
    }

    @Override
    public CustomOutput processOutput(TranslatorContext ctx, NDList list)
            throws TranslateException {
        CustomInput input = (CustomInput)ctx.getAttachment("input");
        //process the output (NDList list) and return the custom output
        return new CustomOutput(input.getText(), labels, scores);
    }
}

public class CustomTranslatorFactory implements TranslatorFactory, Serializable {
    @Override
    public <I, O> Translator<I, O> newInstance(Class<I> input, Class<O> output,
            Model model, Map<String, ?> arguments) {
        // Custom translator implementation
        return new CustomTranslator();
    }
}

MlSimpleJobParameters<CustomInput> jobParams =
    MlSimpleJobParameters.<CustomInput>builder()
        .id(MODEL_ID)
        .version(MODEL_VERSION)
        .type(ModelType.ONNX)
        .name("model_quantized")
        .config(ModelConfig.builder().build())
        .inputClass("org.apache.ignite.example.ml.multiinput.CustomInput")
        .outputClass("org.apache.ignite.example.ml.multiinput.CustomOutput")
        .translatorFactory(
            "org.apache.ignite.example.ml.multiinput.CustomTranslatorFactory")
        .input(input)
        .build();

CustomOutput result = mlApi.predict(jobParams);
System.out.println("Classification Results:");
System.out.println("Text: " + prompt);
System.out.println("Predictions: " + result.getLabels()[0]);
System.out.println("Scores: " + result.getScores()[0]);

public class CustomComputeJob<I extends MlSimpleJobParameters, O>
        extends MlComputeJob<I, O> {

    @Override
    public CompletableFuture<O> predictAsync(JobExecutionContext context, I arg) {
        if (arg == null) {
            return CompletableFuture.completedFuture(null);
        }
        return context.ignite().ml().predictAsync(arg);
    }

    @Override
    public Marshaller<I, byte[]> inputMarshaller() {
        return new CustomInputMarshaller<>();
    }

    @Override
    public Marshaller<O, byte[]> resultMarshaller() {
        return new CustomOutputMarshaller<>();
    }
}

MlSimpleJobParameters<CustomInput> jobParams =
    MlSimpleJobParameters.<CustomInput>builder()
        // ... standard parameters ...
        .customJobClass(CustomComputeJob.class)
        .customInputMarshaller(new CustomInputMarshaller<>())
        .customOutputMarshaller(new CustomOutputMarshaller<>())
        .build();