Wire in stateless ONNX runtime evaluation

13 files changed, 335 insertions, 70 deletions

diff --git a/config-model/src/test/cfg/application/ml_serving/models/add_mul.onnx b/config-model/src/test/cfg/application/ml_serving/models/add_mul.onnx
new file mode 100644
index 00000000000..ab054d112e9
--- /dev/null
+++ b/config-model/src/test/cfg/application/ml_serving/models/add_mul.onnx
@@ -0,0 +1,24 @@
+
+add_mul.py:£

+
+input1
+input2output1"Mul
+
+input1
+input2output2"Addadd_mulZ
+input1
+
+

+
Z
+input2
+
+

+
b
+output1
+
+

+
b
+output2
+
+

+
B
\ No newline at end of file
diff --git a/config-model/src/test/cfg/application/ml_serving/models/mnist_softmax.onnx b/config-model/src/test/cfg/application/ml_serving/models/mnist_softmax.onnx
deleted file mode 100644
index a86019bf53a..00000000000
--- a/config-model/src/test/cfg/application/ml_serving/models/mnist_softmax.onnx
+++ /dev/null
diff --git a/config-model/src/test/cfg/application/ml_serving/models/sqrt.onnx b/config-model/src/test/cfg/application/ml_serving/models/sqrt.onnx
new file mode 100644
index 00000000000..04a6420002c
--- /dev/null
+++ b/config-model/src/test/cfg/application/ml_serving/models/sqrt.onnx
@@ -0,0 +1,11 @@
+sqrt.py:V
+
+input
out/layer/1:1"SqrtsqrtZ
+input
+
+

+
b
+
out/layer/1:1
+
+

+
B
\ No newline at end of file
diff --git a/config-model/src/test/cfg/application/ml_serving/models/sqrt.py b/config-model/src/test/cfg/application/ml_serving/models/sqrt.py
new file mode 100644
index 00000000000..b7b99b3850c
--- /dev/null
+++ b/config-model/src/test/cfg/application/ml_serving/models/sqrt.py
@@ -0,0 +1,23 @@
+# Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+import onnx
+from onnx import helper, TensorProto
+
+INPUT = helper.make_tensor_value_info('input', TensorProto.FLOAT, [1])
+OUTPUT = helper.make_tensor_value_info('out/layer/1:1', TensorProto.FLOAT, [1])
+
+nodes = [
+    helper.make_node(
+        'Sqrt',
+        ['input'],
+        ['out/layer/1:1'],
+    ),
+]
+graph_def = helper.make_graph(
+    nodes,
+    'sqrt',
+    [INPUT],
+    [OUTPUT],
+)
+model_def = helper.make_model(graph_def, producer_name='sqrt.py', opset_imports=[onnx.OperatorSetIdProto(version=12)])
+onnx.save(model_def, 'sqrt.onnx')
diff --git a/config-model/src/test/cfg/application/onnx/files/add.onnx b/config-model/src/test/cfg/application/onnx/files/add.onnx
new file mode 100644
index 00000000000..28318dbba4d
--- /dev/null
+++ b/config-model/src/test/cfg/application/onnx/files/add.onnx
@@ -0,0 +1,16 @@
+add.py:f
+
+input1
+input2output"AddaddZ
+input1
+
+

+
Z
+input2
+
+

+
b
+output
+
+

+
B
\ No newline at end of file
diff --git a/config-model/src/test/cfg/application/onnx/files/add.py b/config-model/src/test/cfg/application/onnx/files/add.py
new file mode 100755
index 00000000000..63b7dc87796
--- /dev/null
+++ b/config-model/src/test/cfg/application/onnx/files/add.py
@@ -0,0 +1,26 @@
+# Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+import onnx
+from onnx import helper, TensorProto
+
+INPUT_1 = helper.make_tensor_value_info('input1', TensorProto.FLOAT, [1])
+INPUT_2 = helper.make_tensor_value_info('input2', TensorProto.FLOAT, [1])
+OUTPUT = helper.make_tensor_value_info('output', TensorProto.FLOAT, [1])
+
+nodes = [
+    helper.make_node(
+        'Add',
+        ['input1', 'input2'],
+        ['output'],
+    ),
+]
+graph_def = helper.make_graph(
+    nodes,
+    'add',
+    [
+        INPUT_1,
+        INPUT_2
+    ],
+    [OUTPUT],
+)
+model_def = helper.make_model(graph_def, producer_name='add.py', opset_imports=[onnx.OperatorSetIdProto(version=12)])
+onnx.save(model_def, 'add.onnx')
diff --git a/config-model/src/test/cfg/application/onnx/models/mul.onnx b/config-model/src/test/cfg/application/onnx/models/mul.onnx
new file mode 100644
index 00000000000..087e2c3427f
--- /dev/null
+++ b/config-model/src/test/cfg/application/onnx/models/mul.onnx
@@ -0,0 +1,16 @@
+mul.py:f
+
+input1
+input2output"MulmulZ
+input1
+
+

+
Z
+input2
+
+

+
b
+output
+
+

+
B
\ No newline at end of file
diff --git a/config-model/src/test/cfg/application/onnx/models/mul.py b/config-model/src/test/cfg/application/onnx/models/mul.py
new file mode 100755
index 00000000000..db01561c355
--- /dev/null
+++ b/config-model/src/test/cfg/application/onnx/models/mul.py
@@ -0,0 +1,26 @@
+# Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+import onnx
+from onnx import helper, TensorProto
+
+INPUT_1 = helper.make_tensor_value_info('input1', TensorProto.FLOAT, [1])
+INPUT_2 = helper.make_tensor_value_info('input2', TensorProto.FLOAT, [1])
+OUTPUT = helper.make_tensor_value_info('output', TensorProto.FLOAT, [1])
+
+nodes = [
+    helper.make_node(
+        'Mul',
+        ['input1', 'input2'],
+        ['output'],
+    ),
+]
+graph_def = helper.make_graph(
+    nodes,
+    'mul',
+    [
+        INPUT_1,
+        INPUT_2
+    ],
+    [OUTPUT],
+)
+model_def = helper.make_model(graph_def, producer_name='mul.py', opset_imports=[onnx.OperatorSetIdProto(version=12)])
+onnx.save(model_def, 'mul.onnx')
diff --git a/config-model/src/test/cfg/application/onnx/searchdefinitions/test.sd b/config-model/src/test/cfg/application/onnx/searchdefinitions/test.sd
new file mode 100644
index 00000000000..d49782ddf39
--- /dev/null
+++ b/config-model/src/test/cfg/application/onnx/searchdefinitions/test.sd
@@ -0,0 +1,27 @@
+# Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+search test {
+
+    document test {
+        field document_value type tensor<float>(d0[1]) {
+            indexing: attribute
+        }
+    }
+
+    onnx-model my_add {
+        file:  files/add.onnx
+        input  input1: attribute(document_value)
+        input  input2: my_input_func
+        output output: out
+    }
+
+    rank-profile test {
+        function my_function() {
+            expression: tensor<float>(d0[1])(1)
+        }
+        first-phase {
+            expression: onnx(my_add).out{d0:1}
+        }
+    }
+
+}
diff --git a/config-model/src/test/cfg/application/onnx/services.xml b/config-model/src/test/cfg/application/onnx/services.xml
new file mode 100644
index 00000000000..8731558c6f7
--- /dev/null
+++ b/config-model/src/test/cfg/application/onnx/services.xml
@@ -0,0 +1,22 @@
+<?xml version="1.0" encoding="utf-8" ?>
+<!-- Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root. -->
+<services version="1.0">
+
+  <container version="1.0">
+    <model-evaluation/>
+    <nodes>
+      <node hostalias="node1" />
+    </nodes>
+  </container>
+
+  <content id="test" version="1.0">
+    <redundancy>1</redundancy>
+    <documents>
+      <document mode="index" type="test"/>
+    </documents>
+    <nodes>
+      <node distribution-key="0" hostalias="node1" />
+    </nodes>
+  </content>
+
+</services>
diff --git a/config-model/src/test/java/com/yahoo/searchdefinition/processing/RankingExpressionWithOnnxTestCase.java b/config-model/src/test/java/com/yahoo/searchdefinition/processing/RankingExpressionWithOnnxTestCase.java
index 40bf970a313..a64b36b327d 100644
--- a/config-model/src/test/java/com/yahoo/searchdefinition/processing/RankingExpressionWithOnnxTestCase.java
+++ b/config-model/src/test/java/com/yahoo/searchdefinition/processing/RankingExpressionWithOnnxTestCase.java
@@ -44,30 +44,6 @@ public class RankingExpressionWithOnnxTestCase {
     }
 
     @Test
-    public void testGlobalOnnxModel() throws IOException {
-        ImportedModelTester tester = new ImportedModelTester(name, applicationDir);
-        VespaModel model = tester.createVespaModel();
-        tester.assertLargeConstant(name + "_layer_Variable_1", model, Optional.of(10L));
-        tester.assertLargeConstant(name + "_layer_Variable", model, Optional.of(7840L));
-
-        // At this point the expression is stored - copy application to another location which do not have a models dir
-        Path storedAppDir = applicationDir.append("copy");
-        try {
-            storedAppDir.toFile().mkdirs();
-            IOUtils.copy(applicationDir.append("services.xml").toString(), storedAppDir.append("services.xml").toString());
-            IOUtils.copyDirectory(applicationDir.append(ApplicationPackage.MODELS_GENERATED_DIR).toFile(),
-                                  storedAppDir.append(ApplicationPackage.MODELS_GENERATED_DIR).toFile());
-            ImportedModelTester storedTester = new ImportedModelTester(name, storedAppDir);
-            VespaModel storedModel = storedTester.createVespaModel();
-            tester.assertLargeConstant(name + "_layer_Variable_1", storedModel, Optional.of(10L));
-            tester.assertLargeConstant(name + "_layer_Variable", storedModel, Optional.of(7840L));
-        }
-        finally {
-            IOUtils.recursiveDeleteDir(storedAppDir.toFile());
-        }
-    }
-
-    @Test
     public void testOnnxReferenceWithConstantFeature() {
         RankProfileSearchFixture search = fixtureWith("constant(mytensor)",
                 "onnx_vespa('mnist_softmax.onnx')",
diff --git a/config-model/src/test/java/com/yahoo/vespa/model/ml/ModelEvaluationTest.java b/config-model/src/test/java/com/yahoo/vespa/model/ml/ModelEvaluationTest.java
index d0196ace766..bf35a002e3a 100644
--- a/config-model/src/test/java/com/yahoo/vespa/model/ml/ModelEvaluationTest.java
+++ b/config-model/src/test/java/com/yahoo/vespa/model/ml/ModelEvaluationTest.java
@@ -3,16 +3,13 @@ package com.yahoo.vespa.model.ml;
 
 import ai.vespa.models.evaluation.Model;
 import ai.vespa.models.evaluation.ModelsEvaluator;
-import ai.vespa.models.evaluation.RankProfilesConfigImporter;
 import ai.vespa.models.handler.ModelsEvaluationHandler;
 import com.yahoo.component.ComponentId;
-import com.yahoo.config.FileReference;
 import com.yahoo.config.application.api.ApplicationPackage;
 import com.yahoo.filedistribution.fileacquirer.FileAcquirer;
 import com.yahoo.filedistribution.fileacquirer.MockFileAcquirer;
 import com.yahoo.io.IOUtils;
 import com.yahoo.path.Path;
-import com.yahoo.tensor.Tensor;
 import com.yahoo.tensor.TensorType;
 import com.yahoo.vespa.config.search.RankProfilesConfig;
 import com.yahoo.vespa.config.search.core.OnnxModelsConfig;
@@ -21,7 +18,10 @@ import com.yahoo.vespa.model.VespaModel;
 import com.yahoo.vespa.model.container.ApplicationContainerCluster;
 import org.junit.Test;
 
+import java.io.File;
 import java.io.IOException;
+import java.util.HashMap;
+import java.util.Map;
 import java.util.Set;
 import java.util.stream.Collectors;
 
@@ -64,7 +64,7 @@ public class ModelEvaluationTest {
         Path storedAppDir = appDir.append("copy");
         try {
             ImportedModelTester tester = new ImportedModelTester("ml_serving", appDir);
-            assertHasMlModels(tester.createVespaModel());
+            assertHasMlModels(tester.createVespaModel(), appDir);
 
             // At this point the expression is stored - copy application to another location which do not have a models dir
             storedAppDir.toFile().mkdirs();
@@ -72,7 +72,7 @@ public class ModelEvaluationTest {
             IOUtils.copyDirectory(appDir.append(ApplicationPackage.MODELS_GENERATED_DIR).toFile(),
                                   storedAppDir.append(ApplicationPackage.MODELS_GENERATED_DIR).toFile());
             ImportedModelTester storedTester = new ImportedModelTester("ml_serving", storedAppDir);
-            assertHasMlModels(storedTester.createVespaModel());
+            assertHasMlModels(storedTester.createVespaModel(), appDir);
         }
         finally {
             IOUtils.recursiveDeleteDir(appDir.append(ApplicationPackage.MODELS_GENERATED_DIR).toFile());
@@ -80,7 +80,7 @@ public class ModelEvaluationTest {
         }
     }
 
-    private void assertHasMlModels(VespaModel model) {
+    private void assertHasMlModels(VespaModel model, Path appDir) {
         ApplicationContainerCluster cluster = model.getContainerClusters().get("container");
         assertNotNull(cluster.getComponentsMap().get(new ComponentId(ModelsEvaluator.class.getName())));
 
@@ -100,12 +100,13 @@ public class ModelEvaluationTest {
         cluster.getConfig(ob);
         OnnxModelsConfig onnxModelsConfig = new OnnxModelsConfig(ob);
 
-        assertEquals(4, config.rankprofile().size());
+        assertEquals(5, config.rankprofile().size());
         Set<String> modelNames = config.rankprofile().stream().map(v -> v.name()).collect(Collectors.toSet());
         assertTrue(modelNames.contains("xgboost_2_2"));
         assertTrue(modelNames.contains("lightgbm_regression"));
-        assertTrue(modelNames.contains("mnist_softmax"));
+        assertTrue(modelNames.contains("add_mul"));
         assertTrue(modelNames.contains("small_constants_and_functions"));
+        assertTrue(modelNames.contains("sqrt"));
 
         // Compare profile content in a denser format than config:
         StringBuilder sb = new StringBuilder();
@@ -113,10 +114,14 @@ public class ModelEvaluationTest {
             sb.append(p.name()).append(": ").append(p.value()).append("\n");
         assertEquals(profile, sb.toString());
 
-        ModelsEvaluator evaluator = new ModelsEvaluator(new ToleratingMissingConstantFilesRankProfilesConfigImporter(MockFileAcquirer.returnFile(null))
-                                                                .importFrom(config, constantsConfig, onnxModelsConfig));
+        Map<String, File> fileMap = new HashMap<>();
+        for (OnnxModelsConfig.Model onnxModel : onnxModelsConfig.model()) {
+            fileMap.put(onnxModel.fileref().value(), appDir.append(onnxModel.fileref().value()).toFile());
+        }
+        FileAcquirer fileAcquirer = MockFileAcquirer.returnFiles(fileMap);
+        ModelsEvaluator evaluator = new ModelsEvaluator(config, constantsConfig, onnxModelsConfig, fileAcquirer);
 
-        assertEquals(4, evaluator.models().size());
+        assertEquals(5, evaluator.models().size());
 
         Model xgboost = evaluator.models().get("xgboost_2_2");
         assertNotNull(xgboost);
@@ -128,31 +133,29 @@ public class ModelEvaluationTest {
         assertNotNull(lightgbm.evaluatorOf());
         assertNotNull(lightgbm.evaluatorOf("lightgbm_regression"));
 
-        Model onnx_mnist_softmax = evaluator.models().get("mnist_softmax");
-        assertNotNull(onnx_mnist_softmax);
-        assertEquals(1, onnx_mnist_softmax.functions().size());
-        assertNotNull(onnx_mnist_softmax.evaluatorOf());
-        assertNotNull(onnx_mnist_softmax.evaluatorOf("default"));
-        assertNotNull(onnx_mnist_softmax.evaluatorOf("default", "add"));
-        assertNotNull(onnx_mnist_softmax.evaluatorOf("default.add"));
-        assertNotNull(onnx_mnist_softmax.evaluatorOf("add"));
-        assertNotNull(onnx_mnist_softmax.evaluatorOf("serving_default"));
-        assertNotNull(onnx_mnist_softmax.evaluatorOf("serving_default", "add"));
-        assertNotNull(onnx_mnist_softmax.evaluatorOf("serving_default.add"));
-        assertNotNull(evaluator.evaluatorOf("mnist_softmax", "default.add"));
-        assertNotNull(evaluator.evaluatorOf("mnist_softmax", "default", "add"));
-        assertNotNull(evaluator.evaluatorOf("mnist_softmax", "add"));
-        assertNotNull(evaluator.evaluatorOf("mnist_softmax", "serving_default.add"));
-        assertNotNull(evaluator.evaluatorOf("mnist_softmax", "serving_default", "add"));
-        assertEquals(TensorType.fromSpec("tensor<float>(d0[],d1[784])"), onnx_mnist_softmax.functions().get(0).argumentTypes().get("Placeholder"));
+        Model add_mul = evaluator.models().get("add_mul");
+        assertNotNull(add_mul);
+        assertEquals(2, add_mul.functions().size());
+        assertNotNull(add_mul.evaluatorOf("output1"));
+        assertNotNull(add_mul.evaluatorOf("output2"));
+        assertNotNull(evaluator.evaluatorOf("add_mul", "output1"));
+        assertNotNull(evaluator.evaluatorOf("add_mul", "output2"));
+        assertEquals(TensorType.fromSpec("tensor<float>(d0[1])"), add_mul.functions().get(0).argumentTypes().get("input1"));
+        assertEquals(TensorType.fromSpec("tensor<float>(d0[1])"), add_mul.functions().get(0).argumentTypes().get("input2"));
+
+        Model sqrt = evaluator.models().get("sqrt");
+        assertNotNull(sqrt);
+        assertEquals(1, sqrt.functions().size());
+        assertNotNull(sqrt.evaluatorOf());
+        assertNotNull(sqrt.evaluatorOf("out_layer_1_1"));  // converted from "out/layer/1:1"
+        assertNotNull(evaluator.evaluatorOf("sqrt"));
+        assertNotNull(evaluator.evaluatorOf("sqrt", "out_layer_1_1"));
+        assertEquals(TensorType.fromSpec("tensor<float>(d0[1])"), sqrt.functions().get(0).argumentTypes().get("input"));
     }
 
     private final String profile =
-            "rankingExpression(imported_ml_function_small_constants_and_functions_exp_output).rankingScript: map(input, f(a)(exp(a)))\n" +
-            "rankingExpression(imported_ml_function_small_constants_and_functions_exp_output).type: tensor<float>(d0[3])\n" +
-            "rankingExpression(default.output).rankingScript: join(rankingExpression(imported_ml_function_small_constants_and_functions_exp_output), reduce(join(join(reduce(rankingExpression(imported_ml_function_small_constants_and_functions_exp_output), sum, d0), tensor<float>(d0[1])(1.0), f(a,b)(a * b)), 9.999999974752427E-7, f(a,b)(a + b)), sum, d0), f(a,b)(a / b))\n" +
-            "rankingExpression(default.output).input.type: tensor<float>(d0[3])\n" +
-            "rankingExpression(default.output).type: tensor<float>(d0[3])\n";
+            "rankingExpression(output).rankingScript: onnxModel(small_constants_and_functions).output\n" +
+            "rankingExpression(output).type: tensor<float>(d0[3])\n";
 
     private RankProfilesConfig.Rankprofile.Fef findProfile(String name, RankProfilesConfig config) {
         for (RankProfilesConfig.Rankprofile profile : config.rankprofile()) {
@@ -162,17 +165,4 @@ public class ModelEvaluationTest {
         throw new IllegalArgumentException("No profile named " + name);
     }
 
-    // We don't have function file distribution so just return empty tensor constants
-    private static class ToleratingMissingConstantFilesRankProfilesConfigImporter extends RankProfilesConfigImporter {
-
-        public ToleratingMissingConstantFilesRankProfilesConfigImporter(FileAcquirer fileAcquirer) {
-            super(fileAcquirer);
-        }
-
-        protected Tensor readTensorFromFile(String name, TensorType type, FileReference fileReference) {
-            return Tensor.from(type, "{}");
-        }
-
-    }
-
 }
diff --git a/config-model/src/test/java/com/yahoo/vespa/model/ml/StatelessOnnxEvaluationTest.java b/config-model/src/test/java/com/yahoo/vespa/model/ml/StatelessOnnxEvaluationTest.java
new file mode 100644
index 00000000000..5dea4a04229
--- /dev/null
+++ b/config-model/src/test/java/com/yahoo/vespa/model/ml/StatelessOnnxEvaluationTest.java
@@ -0,0 +1,108 @@
+// Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+package com.yahoo.vespa.model.ml;
+
+import ai.vespa.models.evaluation.FunctionEvaluator;
+import ai.vespa.models.evaluation.Model;
+import ai.vespa.models.evaluation.ModelsEvaluator;
+import ai.vespa.models.evaluation.RankProfilesConfigImporter;
+import ai.vespa.models.handler.ModelsEvaluationHandler;
+import com.yahoo.component.ComponentId;
+import com.yahoo.config.FileReference;
+import com.yahoo.config.application.api.ApplicationPackage;
+import com.yahoo.filedistribution.fileacquirer.FileAcquirer;
+import com.yahoo.filedistribution.fileacquirer.MockFileAcquirer;
+import com.yahoo.io.IOUtils;
+import com.yahoo.path.Path;
+import com.yahoo.tensor.Tensor;
+import com.yahoo.tensor.TensorType;
+import com.yahoo.vespa.config.search.RankProfilesConfig;
+import com.yahoo.vespa.config.search.core.OnnxModelsConfig;
+import com.yahoo.vespa.config.search.core.RankingConstantsConfig;
+import com.yahoo.vespa.model.VespaModel;
+import com.yahoo.vespa.model.container.ApplicationContainerCluster;
+import org.junit.Test;
+
+import java.io.File;
+import java.io.IOException;
+import java.util.HashMap;
+import java.util.Map;
+import java.util.Set;
+import java.util.stream.Collectors;
+
+import static org.junit.Assert.assertEquals;
+import static org.junit.Assert.assertNotNull;
+import static org.junit.Assert.assertNull;
+import static org.junit.Assert.assertTrue;
+
+/**
+ * Tests stateless model evaluation (turned on by the "model-evaluation" tag in "container")
+ * for ONNX models.
+ *
+ * @author lesters
+ */
+public class StatelessOnnxEvaluationTest {
+
+    @Test
+    public void testStatelessOnnxModelEvaluation() throws IOException {
+        Path appDir = Path.fromString("src/test/cfg/application/onnx");
+        Path storedAppDir = appDir.append("copy");
+        try {
+            ImportedModelTester tester = new ImportedModelTester("onnx_rt", appDir);
+            assertModelEvaluation(tester.createVespaModel(), appDir);
+
+            // At this point the expression is stored - copy application to another location which does not have a models dir
+            storedAppDir.toFile().mkdirs();
+            IOUtils.copy(appDir.append("services.xml").toString(), storedAppDir.append("services.xml").toString());
+            IOUtils.copyDirectory(appDir.append(ApplicationPackage.MODELS_GENERATED_DIR).toFile(),
+                    storedAppDir.append(ApplicationPackage.MODELS_GENERATED_DIR).toFile());
+            IOUtils.copyDirectory(appDir.append(ApplicationPackage.SEARCH_DEFINITIONS_DIR).toFile(),
+                    storedAppDir.append(ApplicationPackage.SEARCH_DEFINITIONS_DIR).toFile());
+            ImportedModelTester storedTester = new ImportedModelTester("onnx_rt", storedAppDir);
+            assertModelEvaluation(storedTester.createVespaModel(), appDir);
+
+        } finally {
+            IOUtils.recursiveDeleteDir(appDir.append(ApplicationPackage.MODELS_GENERATED_DIR).toFile());
+            IOUtils.recursiveDeleteDir(storedAppDir.toFile());
+        }
+    }
+
+    private void assertModelEvaluation(VespaModel model, Path appDir) {
+        ApplicationContainerCluster cluster = model.getContainerClusters().get("container");
+        assertNotNull(cluster.getComponentsMap().get(new ComponentId(ModelsEvaluator.class.getName())));
+
+        RankProfilesConfig.Builder b = new RankProfilesConfig.Builder();
+        cluster.getConfig(b);
+        RankProfilesConfig config = new RankProfilesConfig(b);
+
+        RankingConstantsConfig.Builder cb = new RankingConstantsConfig.Builder();
+        cluster.getConfig(cb);
+        RankingConstantsConfig constantsConfig = new RankingConstantsConfig(cb);
+
+        OnnxModelsConfig.Builder ob = new OnnxModelsConfig.Builder();
+        cluster.getConfig(ob);
+        OnnxModelsConfig onnxModelsConfig = new OnnxModelsConfig(ob);
+
+        assertEquals(1, config.rankprofile().size());
+        Set<String> modelNames = config.rankprofile().stream().map(v -> v.name()).collect(Collectors.toSet());
+        assertTrue(modelNames.contains("mul"));
+
+        // This is actually how ModelsEvaluator is injected
+        Map<String, File> fileMap = new HashMap<>();
+        for (OnnxModelsConfig.Model onnxModel : onnxModelsConfig.model()) {
+            fileMap.put(onnxModel.fileref().value(), appDir.append(onnxModel.fileref().value()).toFile());
+        }
+        FileAcquirer fileAcquirer = MockFileAcquirer.returnFiles(fileMap);
+        ModelsEvaluator modelsEvaluator = new ModelsEvaluator(config, constantsConfig, onnxModelsConfig, fileAcquirer);
+        assertEquals(1, modelsEvaluator.models().size());
+
+        Model mul = modelsEvaluator.models().get("mul");
+        FunctionEvaluator evaluator = mul.evaluatorOf();  // or "default.output" - or actually use name of model output
+
+        Tensor input1 = Tensor.from("tensor<float>(d0[1]):[2]");
+        Tensor input2 = Tensor.from("tensor<float>(d0[1]):[3]");
+        Tensor output = evaluator.bind("input1", input1).bind("input2", input2).evaluate();
+        assertEquals(6.0, output.sum().asDouble(), 1e-9);
+
+    }
+
+}