Revert "Bjorncs/huggingface tokenizer"

6 files changed, 210 insertions, 29 deletions

diff --git a/model-integration/pom.xml b/model-integration/pom.xml
index 681003fdc89..cc5eccff2ac 100644
--- a/model-integration/pom.xml
+++ b/model-integration/pom.xml
@@ -91,6 +91,26 @@
       <artifactId>protobuf-java</artifactId>
     </dependency>
 
+     <dependency>
+      <groupId>ai.djl.huggingface</groupId>
+      <artifactId>tokenizers</artifactId>
+      <version>0.22.1</version>
+       <exclusions>
+         <exclusion>
+           <groupId>com.google.code.gson</groupId>
+           <artifactId>gson</artifactId>
+         </exclusion>
+         <exclusion>
+           <groupId>net.java.dev.jna</groupId>
+           <artifactId>jna</artifactId>
+         </exclusion>
+         <exclusion>
+           <groupId>org.slf4j</groupId>
+           <artifactId>slf4j-api</artifactId>
+         </exclusion>
+       </exclusions>
+    </dependency>
+
     <dependency>
       <groupId>org.lz4</groupId>
       <artifactId>lz4-java</artifactId>
diff --git a/model-integration/src/main/java/ai/vespa/embedding/huggingface/Encoding.java b/model-integration/src/main/java/ai/vespa/embedding/huggingface/Encoding.java
new file mode 100644
index 00000000000..274c29a57b2
--- /dev/null
+++ b/model-integration/src/main/java/ai/vespa/embedding/huggingface/Encoding.java
@@ -0,0 +1,54 @@
+// Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+package ai.vespa.embedding.huggingface;
+
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.List;
+
+/**
+ * @author bjorncs
+ */
+public record Encoding(
+        List<Long> ids, List<Long> typeIds, List<String> tokens, List<Long> wordIds, List<Long> attentionMask,
+        List<Long> specialTokenMask, List<CharSpan> charTokenSpans, List<Encoding> overflowing) {
+
+    public record CharSpan(int start, int end) {
+        public static final CharSpan NONE = new CharSpan(-1, -1);
+        static CharSpan from(ai.djl.huggingface.tokenizers.jni.CharSpan s) {
+            if (s == null) return NONE;
+            return new CharSpan(s.getStart(), s.getEnd());
+        }
+        public boolean isNone() { return this.equals(NONE); }
+    }
+
+    public Encoding {
+        ids = List.copyOf(ids);
+        typeIds = List.copyOf(typeIds);
+        tokens = List.copyOf(tokens);
+        wordIds = List.copyOf(wordIds);
+        attentionMask = List.copyOf(attentionMask);
+        specialTokenMask = List.copyOf(specialTokenMask);
+        charTokenSpans = List.copyOf(charTokenSpans);
+        overflowing = List.copyOf(overflowing);
+    }
+
+    static Encoding from(ai.djl.huggingface.tokenizers.Encoding e) {
+        return new Encoding(
+                toList(e.getIds()),
+                toList(e.getTypeIds()),
+                List.of(e.getTokens()),
+                toList(e.getWordIds()),
+                toList(e.getAttentionMask()),
+                toList(e.getSpecialTokenMask()),
+                Arrays.stream(e.getCharTokenSpans()).map(CharSpan::from).toList(),
+                Arrays.stream(e.getOverflowing()).map(Encoding::from).toList());
+    }
+
+    private static List<Long> toList(long[] array) {
+        if (array == null) return List.of();
+        var list = new ArrayList<Long>(array.length);
+        for (long e : array) list.add(e);
+        return list;
+    }
+}
diff --git a/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceEmbedder.java b/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceEmbedder.java
index 0c1cc80544e..365a50f47b5 100644
--- a/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceEmbedder.java
+++ b/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceEmbedder.java
@@ -3,11 +3,9 @@ package ai.vespa.embedding.huggingface;
 import ai.vespa.modelintegration.evaluator.OnnxEvaluator;
 import ai.vespa.modelintegration.evaluator.OnnxEvaluatorOptions;
 import ai.vespa.modelintegration.evaluator.OnnxRuntime;
-import com.yahoo.api.annotations.Beta;
 import com.yahoo.component.AbstractComponent;
 import com.yahoo.component.annotation.Inject;
 import com.yahoo.embedding.huggingface.HuggingFaceEmbedderConfig;
-import com.yahoo.language.huggingface.HuggingFaceTokenizer;
 import com.yahoo.language.process.Embedder;
 import com.yahoo.tensor.IndexedTensor;
 import com.yahoo.tensor.Tensor;
@@ -16,18 +14,17 @@ import com.yahoo.tensor.TensorType;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
+import java.io.IOException;
 import java.nio.file.Paths;
 import java.util.List;
 import java.util.Map;
 
-@Beta
 public class HuggingFaceEmbedder extends AbstractComponent implements Embedder {
 
     private static final Logger LOG = LoggerFactory.getLogger(HuggingFaceEmbedder.class.getName());
 
     private final String inputIdsName;
     private final String attentionMaskName;
-    private final String tokenTypeIdsName;
     private final String outputName;
     private final int maxTokens;
     private final boolean normalize;
@@ -35,17 +32,13 @@ public class HuggingFaceEmbedder extends AbstractComponent implements Embedder {
     private final OnnxEvaluator evaluator;
 
     @Inject
-    public HuggingFaceEmbedder(OnnxRuntime onnx, HuggingFaceEmbedderConfig config) {
+    public HuggingFaceEmbedder(OnnxRuntime onnx, HuggingFaceEmbedderConfig config) throws IOException {
         maxTokens = config.transformerMaxTokens();
         inputIdsName = config.transformerInputIds();
         attentionMaskName = config.transformerAttentionMask();
-        tokenTypeIdsName = config.transformerTokenTypeIds();
         outputName = config.transformerOutput();
         normalize = config.normalize();
-        tokenizer = new HuggingFaceTokenizer.Builder()
-                .addSpecialTokens(true)
-                .addDefaultModel(Paths.get(config.tokenizerPath().toString()))
-                .build();
+        tokenizer = new HuggingFaceTokenizer(Paths.get(config.tokenizerPath().toString()));
         var onnxOpts = new OnnxEvaluatorOptions();
         if (config.transformerGpuDevice() >= 0)
             onnxOpts.setGpuDevice(config.transformerGpuDevice());
@@ -59,7 +52,6 @@ public class HuggingFaceEmbedder extends AbstractComponent implements Embedder {
         Map<String, TensorType> inputs = evaluator.getInputInfo();
         validateName(inputs, inputIdsName, "input");
         validateName(inputs, attentionMaskName, "input");
-        if (!tokenTypeIdsName.isEmpty()) validateName(inputs, tokenTypeIdsName, "input");
 
         Map<String, TensorType> outputs = evaluator.getOutputInfo();
         validateName(outputs, outputName, "output");
@@ -74,7 +66,8 @@ public class HuggingFaceEmbedder extends AbstractComponent implements Embedder {
 
     @Override
     public List<Integer> embed(String s, Context context) {
-        var tokenIds = tokenizer.embed(s, context);
+        Encoding encoding = tokenizer.encode(s);
+        List<Integer> tokenIds = encoding.ids().stream().map(Long::intValue).toList();
 
         int tokensSize = tokenIds.size();
 
@@ -94,21 +87,18 @@ public class HuggingFaceEmbedder extends AbstractComponent implements Embedder {
 
     @Override
     public Tensor embed(String s, Context context, TensorType tensorType) {
-        var encoding = tokenizer.encode(s, context.getLanguage());
-        Tensor inputSequence = createTensorRepresentation(encoding.ids(), "d1");
-        Tensor attentionMask = createTensorRepresentation(encoding.attentionMask(), "d1");
-        Tensor tokenTypeIds = createTensorRepresentation(encoding.typeIds(), "d1");
-
-
-        Map<String, Tensor> inputs;
-        if (tokenTypeIds.isEmpty()) {
-            inputs = Map.of(inputIdsName, inputSequence.expand("d0"),
-                            attentionMaskName, attentionMask.expand("d0"));
-        } else {
-            inputs = Map.of(inputIdsName, inputSequence.expand("d0"),
-                            attentionMaskName, attentionMask.expand("d0"),
-                            tokenTypeIdsName, tokenTypeIds.expand("d0"));
-        }
+        List<Integer> tokenIds = embed(s.toLowerCase(), context);
+        return embedTokens(tokenIds, tensorType);
+    }
+
+    Tensor embedTokens(List<Integer> tokenIds, TensorType tensorType) {
+        Tensor inputSequence = createTensorRepresentation(tokenIds, "d1");
+        Tensor attentionMask = createAttentionMask(inputSequence);
+
+        Map<String, Tensor> inputs = Map.of(
+                inputIdsName, inputSequence.expand("d0"),
+                attentionMaskName, attentionMask.expand("d0")
+        );
 
         Map<String, Tensor> outputs = evaluator.evaluate(inputs);
         Tensor tokenEmbeddings = outputs.get(outputName);
@@ -146,7 +136,7 @@ public class HuggingFaceEmbedder extends AbstractComponent implements Embedder {
         return builder.build();
     }
 
-    private IndexedTensor createTensorRepresentation(List<Long> input, String dimension) {
+    private IndexedTensor createTensorRepresentation(List<Integer> input, String dimension) {
         int size = input.size();
         TensorType type = new TensorType.Builder(TensorType.Value.FLOAT).indexed(dimension, size).build();
         IndexedTensor.Builder builder = IndexedTensor.Builder.of(type);
diff --git a/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceTokenizer.java b/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceTokenizer.java
new file mode 100644
index 00000000000..e6765a4cc8a
--- /dev/null
+++ b/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceTokenizer.java
@@ -0,0 +1,47 @@
+// Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+package ai.vespa.embedding.huggingface;
+
+import java.io.IOException;
+import java.nio.file.Path;
+
+/**
+ * Wrapper around Deep Java Library's HuggingFace tokenizer.
+ *
+ * @author bjorncs
+ */
+public class HuggingFaceTokenizer implements AutoCloseable {
+
+    private final ai.djl.huggingface.tokenizers.HuggingFaceTokenizer instance;
+
+    public HuggingFaceTokenizer(Path path) throws IOException { this(path, HuggingFaceTokenizerOptions.defaults()); }
+
+    public HuggingFaceTokenizer(Path path, HuggingFaceTokenizerOptions opts) throws IOException {
+        var original = Thread.currentThread().getContextClassLoader();
+        Thread.currentThread().setContextClassLoader(HuggingFaceTokenizer.class.getClassLoader());
+        try {
+            instance = createInstance(path, opts);
+        } finally {
+            Thread.currentThread().setContextClassLoader(original);
+        }
+    }
+
+    public Encoding encode(String text) { return Encoding.from(instance.encode(text)); }
+
+    @Override public void close() { instance.close(); }
+
+    private static ai.djl.huggingface.tokenizers.HuggingFaceTokenizer createInstance(
+            Path path, HuggingFaceTokenizerOptions opts) throws IOException {
+        var builder = ai.djl.huggingface.tokenizers.HuggingFaceTokenizer.builder().optTokenizerPath(path);
+        opts.addSpecialToken().ifPresent(builder::optAddSpecialTokens);
+        opts.truncation().ifPresent(builder::optTruncation);
+        if (opts.truncateFirstOnly()) builder.optTruncateFirstOnly();
+        if (opts.truncateSecondOnly()) builder.optTruncateSecondOnly();
+        opts.padding().ifPresent(builder::optPadding);
+        if (opts.padToMaxLength()) builder.optPadToMaxLength();
+        opts.maxLength().ifPresent(builder::optMaxLength);
+        opts.padToMultipleOf().ifPresent(builder::optPadToMultipleOf);
+        opts.stride().ifPresent(builder::optStride);
+        return builder.build();
+    }
+}
diff --git a/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceTokenizerOptions.java b/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceTokenizerOptions.java
new file mode 100644
index 00000000000..74f80756603
--- /dev/null
+++ b/model-integration/src/main/java/ai/vespa/embedding/huggingface/HuggingFaceTokenizerOptions.java
@@ -0,0 +1,71 @@
+// Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+package ai.vespa.embedding.huggingface;
+
+import java.util.Optional;
+import java.util.OptionalInt;
+
+/**
+ * @author bjorncs
+ */
+public class HuggingFaceTokenizerOptions {
+
+    private final Boolean addSpecialToken;
+    private final Boolean truncation;
+    private final boolean truncateFirstOnly;
+    private final boolean truncateSecondOnly;
+    private final Boolean padding;
+    private final boolean padToMaxLength;
+    private final Integer maxLength;
+    private final Integer padToMultipleOf;
+    private final Integer stride;
+
+    private HuggingFaceTokenizerOptions(Builder b) {
+        addSpecialToken = b.addSpecialToken;
+        truncation = b.truncation;
+        truncateFirstOnly = b.truncateFirstOnly;
+        truncateSecondOnly = b.truncateSecondOnly;
+        padding = b.padding;
+        padToMaxLength = b.padToMaxLength;
+        maxLength = b.maxLength;
+        padToMultipleOf = b.padToMultipleOf;
+        stride = b.stride;
+    }
+
+    public static Builder custom() { return new Builder(); }
+    public static HuggingFaceTokenizerOptions defaults() { return new Builder().build(); }
+
+    Optional<Boolean> addSpecialToken() { return Optional.ofNullable(addSpecialToken); }
+    Optional<Boolean> truncation() { return Optional.ofNullable(truncation); }
+    boolean truncateFirstOnly() { return truncateFirstOnly; }
+    boolean truncateSecondOnly() { return truncateSecondOnly; }
+    Optional<Boolean> padding() { return Optional.ofNullable(padding); }
+    boolean padToMaxLength() { return padToMaxLength; }
+    OptionalInt maxLength() { return maxLength != null ? OptionalInt.of(maxLength) : OptionalInt.empty(); }
+    OptionalInt padToMultipleOf() { return padToMultipleOf != null ? OptionalInt.of(padToMultipleOf) : OptionalInt.empty(); }
+    OptionalInt stride() { return stride != null ? OptionalInt.of(stride) : OptionalInt.empty(); }
+
+    public static class Builder {
+        private Boolean addSpecialToken;
+        private Boolean truncation;
+        private boolean truncateFirstOnly;
+        private boolean truncateSecondOnly;
+        private Boolean padding;
+        private boolean padToMaxLength;
+        private Integer maxLength;
+        private Integer padToMultipleOf;
+        private Integer stride;
+
+        public Builder addSpecialToken(boolean enabled) { addSpecialToken = enabled; return this; }
+        public Builder truncation(boolean enabled) { truncation = enabled; return this; }
+        public Builder truncateFirstOnly() { truncateFirstOnly = true; return this; }
+        public Builder truncateSecondOnly() { truncateSecondOnly = true; return this; }
+        public Builder padding(boolean enabled) { padding = enabled; return this; }
+        public Builder padToMaxLength() { padToMaxLength = true; return this; }
+        public Builder maxLength(int length) { maxLength = length; return this; }
+        public Builder padToMultipleOf(int num) { padToMultipleOf = num; return this; }
+        public Builder stride(int stride) { this.stride = stride; return this; }
+        public HuggingFaceTokenizerOptions build() { return new HuggingFaceTokenizerOptions(this); }
+    }
+
+}
diff --git a/model-integration/src/main/resources/configdefinitions/hugging-face-embedder.def b/model-integration/src/main/resources/configdefinitions/hugging-face-embedder.def
index 97515818f14..1dccea0ddf6 100644
--- a/model-integration/src/main/resources/configdefinitions/hugging-face-embedder.def
+++ b/model-integration/src/main/resources/configdefinitions/hugging-face-embedder.def
@@ -12,7 +12,6 @@ transformerMaxTokens int default=512
 # Input names
 transformerInputIds      string default=input_ids
 transformerAttentionMask string default=attention_mask
-transformerTokenTypeIds  string default=token_type_ids
 
 # Output name
 transformerOutput string default=last_hidden_state