Add Llama model support for embeddings and chat completions

Author: Jan-Kazlouski-elastic

server/src/main/java/org/elasticsearch/TransportVersions.java

@@ -324,6 +324,7 @@ static TransportVersion def(int id) {
     public static final TransportVersion ML_INFERENCE_COHERE_API_VERSION = def(9_110_0_00);
     public static final TransportVersion ESQL_PROFILE_INCLUDE_PLAN = def(9_111_0_00);
     public static final TransportVersion MAPPINGS_IN_DATA_STREAMS = def(9_112_0_00);
+    public static final TransportVersion ML_INFERENCE_LLAMA_ADDED = def(9_113_0_00);
 
     /*
      * STOP! READ THIS FIRST! No, really,

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/InferenceNamedWriteablesProvider.java

@@ -103,6 +103,8 @@
 import org.elasticsearch.xpack.inference.services.jinaai.embeddings.JinaAIEmbeddingsTaskSettings;
 import org.elasticsearch.xpack.inference.services.jinaai.rerank.JinaAIRerankServiceSettings;
 import org.elasticsearch.xpack.inference.services.jinaai.rerank.JinaAIRerankTaskSettings;
+import org.elasticsearch.xpack.inference.services.llama.completion.LlamaChatCompletionServiceSettings;
+import org.elasticsearch.xpack.inference.services.llama.embeddings.LlamaEmbeddingsServiceSettings;
 import org.elasticsearch.xpack.inference.services.mistral.completion.MistralChatCompletionServiceSettings;
 import org.elasticsearch.xpack.inference.services.mistral.embeddings.MistralEmbeddingsServiceSettings;
 import org.elasticsearch.xpack.inference.services.openai.completion.OpenAiChatCompletionServiceSettings;
@@ -172,6 +174,7 @@ public static List<NamedWriteableRegistry.Entry> getNamedWriteables() {
         addJinaAINamedWriteables(namedWriteables);
         addVoyageAINamedWriteables(namedWriteables);
         addCustomNamedWriteables(namedWriteables);
+        addLlamaNamedWriteables(namedWriteables);
 
         addUnifiedNamedWriteables(namedWriteables);
 
@@ -271,8 +274,25 @@ private static void addMistralNamedWriteables(List<NamedWriteableRegistry.Entry>
                 MistralChatCompletionServiceSettings::new
             )
         );
+        // no task settings for Mistral
+    }
 
-        // note - no task settings for Mistral embeddings...
+    private static void addLlamaNamedWriteables(List<NamedWriteableRegistry.Entry> namedWriteables) {
+        namedWriteables.add(
+            new NamedWriteableRegistry.Entry(
+                ServiceSettings.class,
+                LlamaEmbeddingsServiceSettings.NAME,
+                LlamaEmbeddingsServiceSettings::new
+            )
+        );
+        namedWriteables.add(
+            new NamedWriteableRegistry.Entry(
+                ServiceSettings.class,
+                LlamaChatCompletionServiceSettings.NAME,
+                LlamaChatCompletionServiceSettings::new
+            )
+        );
+        // no task settings for Llama
     }
 
     private static void addAzureAiStudioNamedWriteables(List<NamedWriteableRegistry.Entry> namedWriteables) {

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/InferencePlugin.java

@@ -132,6 +132,7 @@
 import org.elasticsearch.xpack.inference.services.huggingface.elser.HuggingFaceElserService;
 import org.elasticsearch.xpack.inference.services.ibmwatsonx.IbmWatsonxService;
 import org.elasticsearch.xpack.inference.services.jinaai.JinaAIService;
+import org.elasticsearch.xpack.inference.services.llama.LlamaService;
 import org.elasticsearch.xpack.inference.services.mistral.MistralService;
 import org.elasticsearch.xpack.inference.services.openai.OpenAiService;
 import org.elasticsearch.xpack.inference.services.sagemaker.SageMakerClient;
@@ -399,6 +400,7 @@ public List<InferenceServiceExtension.Factory> getInferenceServiceFactories() {
             context -> new JinaAIService(httpFactory.get(), serviceComponents.get()),
             context -> new VoyageAIService(httpFactory.get(), serviceComponents.get()),
             context -> new DeepSeekService(httpFactory.get(), serviceComponents.get()),
+            context -> new LlamaService(httpFactory.get(), serviceComponents.get()),
             ElasticsearchInternalService::new,
             context -> new CustomService(httpFactory.get(), serviceComponents.get())
         );

x-pack/plugin/inference/src/main/java/org/elasticsearch/xpack/inference/services/llama/LlamaModel.java

@@ -0,0 +1,67 @@
+/*
+ * Copyright Elasticsearch B.V. and/or licensed to Elasticsearch B.V. under one
+ * or more contributor license agreements. Licensed under the Elastic License
+ * 2.0; you may not use this file except in compliance with the Elastic License
+ * 2.0.
+ */
+
+package org.elasticsearch.xpack.inference.services.llama;
+
+import org.elasticsearch.inference.EmptySecretSettings;
+import org.elasticsearch.inference.ModelConfigurations;
+import org.elasticsearch.inference.ModelSecrets;
+import org.elasticsearch.inference.SecretSettings;
+import org.elasticsearch.inference.ServiceSettings;
+import org.elasticsearch.xpack.inference.services.RateLimitGroupingModel;
+import org.elasticsearch.xpack.inference.services.settings.DefaultSecretSettings;
+import org.elasticsearch.xpack.inference.services.settings.RateLimitSettings;
+
+import java.net.URI;
+import java.net.URISyntaxException;
+import java.util.Map;
+import java.util.Objects;
+
+public abstract class LlamaModel extends RateLimitGroupingModel {
+    protected String modelId;
+    protected URI uri;
+    protected RateLimitSettings rateLimitSettings;
+
+    protected LlamaModel(ModelConfigurations configurations, ModelSecrets secrets) {
+        super(configurations, secrets);
+    }
+
+    protected LlamaModel(RateLimitGroupingModel model, ServiceSettings serviceSettings) {
+        super(model, serviceSettings);
+    }
+
+    public String model() {
+        return this.modelId;
+    }
+
+    public URI uri() {
+        return this.uri;
+    }
+
+    @Override
+    public RateLimitSettings rateLimitSettings() {
+        return this.rateLimitSettings;
+    }
+
+    @Override
+    public int rateLimitGroupingHash() {
+        return Objects.hash(modelId, uri, getSecretSettings());
+    }
+
+    // Needed for testing only
+    public void setURI(String newUri) {
+        try {
+            this.uri = new URI(newUri);
+        } catch (URISyntaxException e) {
+            // swallow any error
+        }
+    }
+
+    protected static SecretSettings retrieveSecretSettings(Map<String, Object> secrets) {
+        return (secrets != null && secrets.isEmpty()) ? EmptySecretSettings.INSTANCE : DefaultSecretSettings.fromMap(secrets);
+    }
+}