diff --git a/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
index 595ad85633..e28605dca2 100644
--- a/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
@@ -42,10 +42,11 @@ using implementation::PreparedModelCallback;
 
 Model createModel(const TestModel& testModel) {
     // Model operands.
-    hidl_vec<Operand> operands(testModel.operands.size());
+    CHECK_EQ(testModel.referenced.size(), 0u);  // Not supported in 1.0.
+    hidl_vec<Operand> operands(testModel.main.operands.size());
     size_t constCopySize = 0, constRefSize = 0;
-    for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-        const auto& op = testModel.operands[i];
+    for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+        const auto& op = testModel.main.operands[i];
 
         DataLocation loc = {};
         if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
@@ -70,9 +71,9 @@ Model createModel(const TestModel& testModel) {
     }
 
     // Model operations.
-    hidl_vec<Operation> operations(testModel.operations.size());
-    std::transform(testModel.operations.begin(), testModel.operations.end(), operations.begin(),
-                   [](const TestOperation& op) -> Operation {
+    hidl_vec<Operation> operations(testModel.main.operations.size());
+    std::transform(testModel.main.operations.begin(), testModel.main.operations.end(),
+                   operations.begin(), [](const TestOperation& op) -> Operation {
                        return {.type = static_cast<OperationType>(op.type),
                                .inputs = op.inputs,
                                .outputs = op.outputs};
@@ -80,8 +81,8 @@ Model createModel(const TestModel& testModel) {
 
     // Constant copies.
     hidl_vec<uint8_t> operandValues(constCopySize);
-    for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-        const auto& op = testModel.operands[i];
+    for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+        const auto& op = testModel.main.operands[i];
         if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
             const uint8_t* begin = op.data.get<uint8_t>();
             const uint8_t* end = begin + op.data.size();
@@ -102,8 +103,8 @@ Model createModel(const TestModel& testModel) {
                 reinterpret_cast<uint8_t*>(static_cast<void*>(mappedMemory->getPointer()));
         CHECK(mappedPtr != nullptr);
 
-        for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-            const auto& op = testModel.operands[i];
+        for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+            const auto& op = testModel.main.operands[i];
             if (op.lifetime == TestOperandLifeTime::CONSTANT_REFERENCE) {
                 const uint8_t* begin = op.data.get<uint8_t>();
                 const uint8_t* end = begin + op.data.size();
@@ -114,8 +115,8 @@ Model createModel(const TestModel& testModel) {
 
     return {.operands = std::move(operands),
             .operations = std::move(operations),
-            .inputIndexes = testModel.inputIndexes,
-            .outputIndexes = testModel.outputIndexes,
+            .inputIndexes = testModel.main.inputIndexes,
+            .outputIndexes = testModel.main.outputIndexes,
             .operandValues = std::move(operandValues),
             .pools = std::move(pools)};
 }
diff --git a/neuralnetworks/1.0/vts/functional/Utils.cpp b/neuralnetworks/1.0/vts/functional/Utils.cpp
index 5b630fd7a6..0dba85acd9 100644
--- a/neuralnetworks/1.0/vts/functional/Utils.cpp
+++ b/neuralnetworks/1.0/vts/functional/Utils.cpp
@@ -42,10 +42,10 @@ constexpr uint32_t kOutputPoolIndex = 1;
 
 Request createRequest(const TestModel& testModel) {
     // Model inputs.
-    hidl_vec<RequestArgument> inputs(testModel.inputIndexes.size());
+    hidl_vec<RequestArgument> inputs(testModel.main.inputIndexes.size());
     size_t inputSize = 0;
-    for (uint32_t i = 0; i < testModel.inputIndexes.size(); i++) {
-        const auto& op = testModel.operands[testModel.inputIndexes[i]];
+    for (uint32_t i = 0; i < testModel.main.inputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.inputIndexes[i]];
         if (op.data.size() == 0) {
             // Omitted input.
             inputs[i] = {.hasNoValue = true};
@@ -59,10 +59,10 @@ Request createRequest(const TestModel& testModel) {
     }
 
     // Model outputs.
-    hidl_vec<RequestArgument> outputs(testModel.outputIndexes.size());
+    hidl_vec<RequestArgument> outputs(testModel.main.outputIndexes.size());
     size_t outputSize = 0;
-    for (uint32_t i = 0; i < testModel.outputIndexes.size(); i++) {
-        const auto& op = testModel.operands[testModel.outputIndexes[i]];
+    for (uint32_t i = 0; i < testModel.main.outputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.outputIndexes[i]];
 
         // In the case of zero-sized output, we should at least provide a one-byte buffer.
         // This is because zero-sized tensors are only supported internally to the driver, or
@@ -90,8 +90,8 @@ Request createRequest(const TestModel& testModel) {
     CHECK(inputPtr != nullptr);
 
     // Copy input data to the memory pool.
-    for (uint32_t i = 0; i < testModel.inputIndexes.size(); i++) {
-        const auto& op = testModel.operands[testModel.inputIndexes[i]];
+    for (uint32_t i = 0; i < testModel.main.inputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.inputIndexes[i]];
         if (op.data.size() > 0) {
             const uint8_t* begin = op.data.get<uint8_t>();
             const uint8_t* end = begin + op.data.size();
diff --git a/neuralnetworks/1.1/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.1/vts/functional/GeneratedTestHarness.cpp
index 7a929d6063..cee15a35a1 100644
--- a/neuralnetworks/1.1/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.1/vts/functional/GeneratedTestHarness.cpp
@@ -49,10 +49,11 @@ using V1_0::implementation::PreparedModelCallback;
 
 Model createModel(const TestModel& testModel) {
     // Model operands.
-    hidl_vec<Operand> operands(testModel.operands.size());
+    CHECK_EQ(testModel.referenced.size(), 0u);  // Not supported in 1.1.
+    hidl_vec<Operand> operands(testModel.main.operands.size());
     size_t constCopySize = 0, constRefSize = 0;
-    for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-        const auto& op = testModel.operands[i];
+    for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+        const auto& op = testModel.main.operands[i];
 
         DataLocation loc = {};
         if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
@@ -77,9 +78,9 @@ Model createModel(const TestModel& testModel) {
     }
 
     // Model operations.
-    hidl_vec<Operation> operations(testModel.operations.size());
-    std::transform(testModel.operations.begin(), testModel.operations.end(), operations.begin(),
-                   [](const TestOperation& op) -> Operation {
+    hidl_vec<Operation> operations(testModel.main.operations.size());
+    std::transform(testModel.main.operations.begin(), testModel.main.operations.end(),
+                   operations.begin(), [](const TestOperation& op) -> Operation {
                        return {.type = static_cast<OperationType>(op.type),
                                .inputs = op.inputs,
                                .outputs = op.outputs};
@@ -87,8 +88,8 @@ Model createModel(const TestModel& testModel) {
 
     // Constant copies.
     hidl_vec<uint8_t> operandValues(constCopySize);
-    for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-        const auto& op = testModel.operands[i];
+    for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+        const auto& op = testModel.main.operands[i];
         if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
             const uint8_t* begin = op.data.get<uint8_t>();
             const uint8_t* end = begin + op.data.size();
@@ -109,8 +110,8 @@ Model createModel(const TestModel& testModel) {
                 reinterpret_cast<uint8_t*>(static_cast<void*>(mappedMemory->getPointer()));
         CHECK(mappedPtr != nullptr);
 
-        for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-            const auto& op = testModel.operands[i];
+        for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+            const auto& op = testModel.main.operands[i];
             if (op.lifetime == TestOperandLifeTime::CONSTANT_REFERENCE) {
                 const uint8_t* begin = op.data.get<uint8_t>();
                 const uint8_t* end = begin + op.data.size();
@@ -121,8 +122,8 @@ Model createModel(const TestModel& testModel) {
 
     return {.operands = std::move(operands),
             .operations = std::move(operations),
-            .inputIndexes = testModel.inputIndexes,
-            .outputIndexes = testModel.outputIndexes,
+            .inputIndexes = testModel.main.inputIndexes,
+            .outputIndexes = testModel.main.outputIndexes,
             .operandValues = std::move(operandValues),
             .pools = std::move(pools),
             .relaxComputationFloat32toFloat16 = testModel.isRelaxed};
diff --git a/neuralnetworks/1.2/vts/functional/CompilationCachingTests.cpp b/neuralnetworks/1.2/vts/functional/CompilationCachingTests.cpp
index 2130a76b75..10dec791cf 100644
--- a/neuralnetworks/1.2/vts/functional/CompilationCachingTests.cpp
+++ b/neuralnetworks/1.2/vts/functional/CompilationCachingTests.cpp
@@ -207,10 +207,10 @@ TestModel createLargeTestModelImpl(TestOperationType op, uint32_t len) {
     };
 
     return {
-            .operands = std::move(operands),
-            .operations = std::move(operations),
-            .inputIndexes = {1},
-            .outputIndexes = {len * 2 + 1},
+            .main = {.operands = std::move(operands),
+                     .operations = std::move(operations),
+                     .inputIndexes = {1},
+                     .outputIndexes = {len * 2 + 1}},
             .isRelaxed = false,
     };
 }
diff --git a/neuralnetworks/1.2/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.2/vts/functional/GeneratedTestHarness.cpp
index 599fd1d9be..4c8fede8b2 100644
--- a/neuralnetworks/1.2/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.2/vts/functional/GeneratedTestHarness.cpp
@@ -75,10 +75,11 @@ struct TestConfig {
 
 Model createModel(const TestModel& testModel) {
     // Model operands.
-    hidl_vec<Operand> operands(testModel.operands.size());
+    CHECK_EQ(testModel.referenced.size(), 0u);  // Not supported in 1.1.
+    hidl_vec<Operand> operands(testModel.main.operands.size());
     size_t constCopySize = 0, constRefSize = 0;
-    for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-        const auto& op = testModel.operands[i];
+    for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+        const auto& op = testModel.main.operands[i];
 
         DataLocation loc = {};
         if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
@@ -110,9 +111,9 @@ Model createModel(const TestModel& testModel) {
     }
 
     // Model operations.
-    hidl_vec<Operation> operations(testModel.operations.size());
-    std::transform(testModel.operations.begin(), testModel.operations.end(), operations.begin(),
-                   [](const TestOperation& op) -> Operation {
+    hidl_vec<Operation> operations(testModel.main.operations.size());
+    std::transform(testModel.main.operations.begin(), testModel.main.operations.end(),
+                   operations.begin(), [](const TestOperation& op) -> Operation {
                        return {.type = static_cast<OperationType>(op.type),
                                .inputs = op.inputs,
                                .outputs = op.outputs};
@@ -120,8 +121,8 @@ Model createModel(const TestModel& testModel) {
 
     // Constant copies.
     hidl_vec<uint8_t> operandValues(constCopySize);
-    for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-        const auto& op = testModel.operands[i];
+    for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+        const auto& op = testModel.main.operands[i];
         if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
             const uint8_t* begin = op.data.get<uint8_t>();
             const uint8_t* end = begin + op.data.size();
@@ -142,8 +143,8 @@ Model createModel(const TestModel& testModel) {
                 reinterpret_cast<uint8_t*>(static_cast<void*>(mappedMemory->getPointer()));
         CHECK(mappedPtr != nullptr);
 
-        for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-            const auto& op = testModel.operands[i];
+        for (uint32_t i = 0; i < testModel.main.operands.size(); i++) {
+            const auto& op = testModel.main.operands[i];
             if (op.lifetime == TestOperandLifeTime::CONSTANT_REFERENCE) {
                 const uint8_t* begin = op.data.get<uint8_t>();
                 const uint8_t* end = begin + op.data.size();
@@ -154,15 +155,15 @@ Model createModel(const TestModel& testModel) {
 
     return {.operands = std::move(operands),
             .operations = std::move(operations),
-            .inputIndexes = testModel.inputIndexes,
-            .outputIndexes = testModel.outputIndexes,
+            .inputIndexes = testModel.main.inputIndexes,
+            .outputIndexes = testModel.main.outputIndexes,
             .operandValues = std::move(operandValues),
             .pools = std::move(pools),
             .relaxComputationFloat32toFloat16 = testModel.isRelaxed};
 }
 
 static bool isOutputSizeGreaterThanOne(const TestModel& testModel, uint32_t index) {
-    const auto byteSize = testModel.operands[testModel.outputIndexes[index]].data.size();
+    const auto byteSize = testModel.main.operands[testModel.main.outputIndexes[index]].data.size();
     return byteSize > 1u;
 }
 
@@ -302,17 +303,17 @@ void EvaluatePreparedModel(const sp<IPreparedModel>& preparedModel, const TestMo
             // either empty, or have the same number of elements as the number of outputs.
             ASSERT_EQ(ErrorStatus::NONE, executionStatus);
             ASSERT_TRUE(outputShapes.size() == 0 ||
-                        outputShapes.size() == testModel.outputIndexes.size());
+                        outputShapes.size() == testModel.main.outputIndexes.size());
             break;
         case OutputType::UNSPECIFIED:
             // If the model output operands are not fully specified, outputShapes must have
             // the same number of elements as the number of outputs.
             ASSERT_EQ(ErrorStatus::NONE, executionStatus);
-            ASSERT_EQ(outputShapes.size(), testModel.outputIndexes.size());
+            ASSERT_EQ(outputShapes.size(), testModel.main.outputIndexes.size());
             break;
         case OutputType::INSUFFICIENT:
             ASSERT_EQ(ErrorStatus::OUTPUT_INSUFFICIENT_SIZE, executionStatus);
-            ASSERT_EQ(outputShapes.size(), testModel.outputIndexes.size());
+            ASSERT_EQ(outputShapes.size(), testModel.main.outputIndexes.size());
             ASSERT_FALSE(outputShapes[0].isSufficient);
             return;
     }
@@ -320,7 +321,7 @@ void EvaluatePreparedModel(const sp<IPreparedModel>& preparedModel, const TestMo
     // Go through all outputs, check returned output shapes.
     for (uint32_t i = 0; i < outputShapes.size(); i++) {
         EXPECT_TRUE(outputShapes[i].isSufficient);
-        const auto& expect = testModel.operands[testModel.outputIndexes[i]].dimensions;
+        const auto& expect = testModel.main.operands[testModel.main.outputIndexes[i]].dimensions;
         const std::vector<uint32_t> actual = outputShapes[i].dimensions;
         EXPECT_EQ(expect, actual);
     }
diff --git a/neuralnetworks/1.3/vts/functional/CompilationCachingTests.cpp b/neuralnetworks/1.3/vts/functional/CompilationCachingTests.cpp
index 0bd24daec0..ac18c8ffcc 100644
--- a/neuralnetworks/1.3/vts/functional/CompilationCachingTests.cpp
+++ b/neuralnetworks/1.3/vts/functional/CompilationCachingTests.cpp
@@ -209,10 +209,10 @@ TestModel createLargeTestModelImpl(TestOperationType op, uint32_t len) {
     };
 
     return {
-            .operands = std::move(operands),
-            .operations = std::move(operations),
-            .inputIndexes = {1},
-            .outputIndexes = {len * 2 + 1},
+            .main = {.operands = std::move(operands),
+                     .operations = std::move(operations),
+                     .inputIndexes = {1},
+                     .outputIndexes = {len * 2 + 1}},
             .isRelaxed = false,
     };
 }
diff --git a/neuralnetworks/1.3/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.3/vts/functional/GeneratedTestHarness.cpp
index 82f34ff779..404c2a1e03 100644
--- a/neuralnetworks/1.3/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.3/vts/functional/GeneratedTestHarness.cpp
@@ -169,7 +169,8 @@ class DeviceMemoryAllocator {
         if constexpr (ioType == IOType::INPUT) {
             if (buffer != nullptr) {
                 // TestBuffer -> Shared memory.
-                const auto& testBuffer = kTestModel.operands[kTestModel.inputIndexes[index]].data;
+                const auto& testBuffer =
+                        kTestModel.main.operands[kTestModel.main.inputIndexes[index]].data;
                 ASSERT_GT(testBuffer.size(), 0);
                 hidl_memory tmp = nn::allocateSharedMemory(testBuffer.size());
                 sp<IMemory> inputMemory = mapMemory(tmp);
@@ -195,26 +196,42 @@ class DeviceMemoryAllocator {
     const TestModel& kTestModel;
 };
 
-}  // namespace
+Subgraph createSubgraph(const TestSubgraph& testSubgraph, uint32_t* constCopySize,
+                        std::vector<const TestBuffer*>* constCopies, uint32_t* constRefSize,
+                        std::vector<const TestBuffer*>* constReferences) {
+    CHECK(constCopySize != nullptr);
+    CHECK(constCopies != nullptr);
+    CHECK(constRefSize != nullptr);
+    CHECK(constReferences != nullptr);
 
-Model createModel(const TestModel& testModel) {
-    // Model operands.
-    hidl_vec<Operand> operands(testModel.operands.size());
-    size_t constCopySize = 0, constRefSize = 0;
-    for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-        const auto& op = testModel.operands[i];
+    // Operands.
+    hidl_vec<Operand> operands(testSubgraph.operands.size());
+    for (uint32_t i = 0; i < testSubgraph.operands.size(); i++) {
+        const auto& op = testSubgraph.operands[i];
 
         DataLocation loc = {};
         if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
-            loc = {.poolIndex = 0,
-                   .offset = static_cast<uint32_t>(constCopySize),
-                   .length = static_cast<uint32_t>(op.data.size())};
-            constCopySize += op.data.alignedSize();
+            loc = {
+                    .poolIndex = 0,
+                    .offset = *constCopySize,
+                    .length = static_cast<uint32_t>(op.data.size()),
+            };
+            constCopies->push_back(&op.data);
+            *constCopySize += op.data.alignedSize();
         } else if (op.lifetime == TestOperandLifeTime::CONSTANT_REFERENCE) {
-            loc = {.poolIndex = 0,
-                   .offset = static_cast<uint32_t>(constRefSize),
-                   .length = static_cast<uint32_t>(op.data.size())};
-            constRefSize += op.data.alignedSize();
+            loc = {
+                    .poolIndex = 0,
+                    .offset = *constRefSize,
+                    .length = static_cast<uint32_t>(op.data.size()),
+            };
+            constReferences->push_back(&op.data);
+            *constRefSize += op.data.alignedSize();
+        } else if (op.lifetime == TestOperandLifeTime::SUBGRAPH) {
+            loc = {
+                    .poolIndex = 0,
+                    .offset = *op.data.get<uint32_t>(),
+                    .length = 0,
+            };
         }
 
         V1_2::Operand::ExtraParams extraParams;
@@ -233,25 +250,52 @@ Model createModel(const TestModel& testModel) {
                        .extraParams = std::move(extraParams)};
     }
 
-    // Model operations.
-    hidl_vec<Operation> operations(testModel.operations.size());
-    std::transform(testModel.operations.begin(), testModel.operations.end(), operations.begin(),
-                   [](const TestOperation& op) -> Operation {
+    // Operations.
+    hidl_vec<Operation> operations(testSubgraph.operations.size());
+    std::transform(testSubgraph.operations.begin(), testSubgraph.operations.end(),
+                   operations.begin(), [](const TestOperation& op) -> Operation {
                        return {.type = static_cast<OperationType>(op.type),
                                .inputs = op.inputs,
                                .outputs = op.outputs};
                    });
 
+    return {.operands = std::move(operands),
+            .operations = std::move(operations),
+            .inputIndexes = testSubgraph.inputIndexes,
+            .outputIndexes = testSubgraph.outputIndexes};
+}
+
+void copyTestBuffers(const std::vector<const TestBuffer*>& buffers, uint8_t* output) {
+    uint32_t offset = 0;
+    for (const TestBuffer* buffer : buffers) {
+        const uint8_t* begin = buffer->get<uint8_t>();
+        const uint8_t* end = begin + buffer->size();
+        std::copy(begin, end, output + offset);
+        offset += buffer->alignedSize();
+    }
+}
+
+}  // namespace
+
+Model createModel(const TestModel& testModel) {
+    uint32_t constCopySize = 0;
+    uint32_t constRefSize = 0;
+    std::vector<const TestBuffer*> constCopies;
+    std::vector<const TestBuffer*> constReferences;
+
+    Subgraph mainSubgraph = createSubgraph(testModel.main, &constCopySize, &constCopies,
+                                           &constRefSize, &constReferences);
+    hidl_vec<Subgraph> refSubgraphs(testModel.referenced.size());
+    std::transform(testModel.referenced.begin(), testModel.referenced.end(), refSubgraphs.begin(),
+                   [&constCopySize, &constCopies, &constRefSize,
+                    &constReferences](const TestSubgraph& testSubgraph) {
+                       return createSubgraph(testSubgraph, &constCopySize, &constCopies,
+                                             &constRefSize, &constReferences);
+                   });
+
     // Constant copies.
     hidl_vec<uint8_t> operandValues(constCopySize);
-    for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-        const auto& op = testModel.operands[i];
-        if (op.lifetime == TestOperandLifeTime::CONSTANT_COPY) {
-            const uint8_t* begin = op.data.get<uint8_t>();
-            const uint8_t* end = begin + op.data.size();
-            std::copy(begin, end, operandValues.data() + operands[i].location.offset);
-        }
-    }
+    copyTestBuffers(constCopies, operandValues.data());
 
     // Shared memory.
     hidl_vec<hidl_memory> pools = {};
@@ -266,27 +310,18 @@ Model createModel(const TestModel& testModel) {
                 reinterpret_cast<uint8_t*>(static_cast<void*>(mappedMemory->getPointer()));
         CHECK(mappedPtr != nullptr);
 
-        for (uint32_t i = 0; i < testModel.operands.size(); i++) {
-            const auto& op = testModel.operands[i];
-            if (op.lifetime == TestOperandLifeTime::CONSTANT_REFERENCE) {
-                const uint8_t* begin = op.data.get<uint8_t>();
-                const uint8_t* end = begin + op.data.size();
-                std::copy(begin, end, mappedPtr + operands[i].location.offset);
-            }
-        }
+        copyTestBuffers(constReferences, mappedPtr);
     }
 
-    return {.main = {.operands = std::move(operands),
-                     .operations = std::move(operations),
-                     .inputIndexes = testModel.inputIndexes,
-                     .outputIndexes = testModel.outputIndexes},
+    return {.main = std::move(mainSubgraph),
+            .referenced = std::move(refSubgraphs),
             .operandValues = std::move(operandValues),
             .pools = std::move(pools),
             .relaxComputationFloat32toFloat16 = testModel.isRelaxed};
 }
 
 static bool isOutputSizeGreaterThanOne(const TestModel& testModel, uint32_t index) {
-    const auto byteSize = testModel.operands[testModel.outputIndexes[index]].data.size();
+    const auto byteSize = testModel.main.operands[testModel.main.outputIndexes[index]].data.size();
     return byteSize > 1u;
 }
 
@@ -320,10 +355,10 @@ static std::pair<Request, std::vector<sp<IBuffer>>> createRequest(
     std::vector<uint32_t> tokens;
 
     // Model inputs.
-    hidl_vec<RequestArgument> inputs(testModel.inputIndexes.size());
+    hidl_vec<RequestArgument> inputs(testModel.main.inputIndexes.size());
     size_t inputSize = 0;
-    for (uint32_t i = 0; i < testModel.inputIndexes.size(); i++) {
-        const auto& op = testModel.operands[testModel.inputIndexes[i]];
+    for (uint32_t i = 0; i < testModel.main.inputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.inputIndexes[i]];
         if (op.data.size() == 0) {
             // Omitted input.
             inputs[i] = {.hasNoValue = true};
@@ -350,10 +385,10 @@ static std::pair<Request, std::vector<sp<IBuffer>>> createRequest(
     }
 
     // Model outputs.
-    hidl_vec<RequestArgument> outputs(testModel.outputIndexes.size());
+    hidl_vec<RequestArgument> outputs(testModel.main.outputIndexes.size());
     size_t outputSize = 0;
-    for (uint32_t i = 0; i < testModel.outputIndexes.size(); i++) {
-        const auto& op = testModel.operands[testModel.outputIndexes[i]];
+    for (uint32_t i = 0; i < testModel.main.outputIndexes.size(); i++) {
+        const auto& op = testModel.main.operands[testModel.main.outputIndexes[i]];
         if (preferDeviceMemory) {
             SCOPED_TRACE("Output index = " + std::to_string(i));
             auto [buffer, token] = allocator.allocate<IOType::OUTPUT>(i);
@@ -398,9 +433,9 @@ static std::pair<Request, std::vector<sp<IBuffer>>> createRequest(
     CHECK(inputMemory.get() != nullptr);
     uint8_t* inputPtr = static_cast<uint8_t*>(static_cast<void*>(inputMemory->getPointer()));
     CHECK(inputPtr != nullptr);
-    for (uint32_t i = 0; i < testModel.inputIndexes.size(); i++) {
+    for (uint32_t i = 0; i < testModel.main.inputIndexes.size(); i++) {
         if (!inputs[i].hasNoValue && inputs[i].location.poolIndex == kInputPoolIndex) {
-            const auto& op = testModel.operands[testModel.inputIndexes[i]];
+            const auto& op = testModel.main.operands[testModel.main.inputIndexes[i]];
             const uint8_t* begin = op.data.get<uint8_t>();
             const uint8_t* end = begin + op.data.size();
             std::copy(begin, end, inputPtr + inputs[i].location.offset);
@@ -443,7 +478,7 @@ static std::vector<TestBuffer> getOutputBuffers(const TestModel& testModel, cons
         if (outputLoc.poolIndex == kOutputPoolIndex) {
             outputBuffers.emplace_back(outputLoc.length, outputPtr + outputLoc.offset);
         } else {
-            const auto& op = testModel.operands[testModel.outputIndexes[i]];
+            const auto& op = testModel.main.operands[testModel.main.outputIndexes[i]];
             if (op.data.size() == 0) {
                 outputBuffers.emplace_back();
             } else {
@@ -638,17 +673,17 @@ void EvaluatePreparedModel(const sp<IDevice>& device, const sp<IPreparedModel>&
             // either empty, or have the same number of elements as the number of outputs.
             ASSERT_EQ(ErrorStatus::NONE, executionStatus);
             ASSERT_TRUE(outputShapes.size() == 0 ||
-                        outputShapes.size() == testModel.outputIndexes.size());
+                        outputShapes.size() == testModel.main.outputIndexes.size());
             break;
         case OutputType::UNSPECIFIED:
             // If the model output operands are not fully specified, outputShapes must have
             // the same number of elements as the number of outputs.
             ASSERT_EQ(ErrorStatus::NONE, executionStatus);
-            ASSERT_EQ(outputShapes.size(), testModel.outputIndexes.size());
+            ASSERT_EQ(outputShapes.size(), testModel.main.outputIndexes.size());
             break;
         case OutputType::INSUFFICIENT:
             ASSERT_EQ(ErrorStatus::OUTPUT_INSUFFICIENT_SIZE, executionStatus);
-            ASSERT_EQ(outputShapes.size(), testModel.outputIndexes.size());
+            ASSERT_EQ(outputShapes.size(), testModel.main.outputIndexes.size());
             ASSERT_FALSE(outputShapes[0].isSufficient);
             return;
     }
@@ -656,7 +691,7 @@ void EvaluatePreparedModel(const sp<IDevice>& device, const sp<IPreparedModel>&
     // Go through all outputs, check returned output shapes.
     for (uint32_t i = 0; i < outputShapes.size(); i++) {
         EXPECT_TRUE(outputShapes[i].isSufficient);
-        const auto& expect = testModel.operands[testModel.outputIndexes[i]].dimensions;
+        const auto& expect = testModel.main.operands[testModel.main.outputIndexes[i]].dimensions;
         const std::vector<uint32_t> actual = outputShapes[i].dimensions;
         EXPECT_EQ(expect, actual);
     }
@@ -862,7 +897,7 @@ INSTANTIATE_GENERATED_TEST(FencedComputeTest,
                            [](const TestModel& testModel) { return !testModel.expectFailure; });
 
 INSTANTIATE_GENERATED_TEST(QuantizationCouplingTest, [](const TestModel& testModel) {
-    return testModel.hasQuant8CoupledOperands() && testModel.operations.size() == 1;
+    return testModel.hasQuant8CoupledOperands() && testModel.main.operations.size() == 1;
 });
 
 }  // namespace android::hardware::neuralnetworks::V1_3::vts::functional
diff --git a/neuralnetworks/1.3/vts/functional/QualityOfServiceTests.cpp b/neuralnetworks/1.3/vts/functional/QualityOfServiceTests.cpp
index 2f1e05c5c6..e07eb4750e 100644
--- a/neuralnetworks/1.3/vts/functional/QualityOfServiceTests.cpp
+++ b/neuralnetworks/1.3/vts/functional/QualityOfServiceTests.cpp
@@ -237,12 +237,13 @@ void runExecutionTest(const sp<IPreparedModel>& preparedModel, const TestModel&
 
     // If the model output operands are fully specified, outputShapes must be either
     // either empty, or have the same number of elements as the number of outputs.
-    ASSERT_TRUE(outputShapes.size() == 0 || outputShapes.size() == testModel.outputIndexes.size());
+    ASSERT_TRUE(outputShapes.size() == 0 ||
+                outputShapes.size() == testModel.main.outputIndexes.size());
 
     // Go through all outputs, check returned output shapes.
     for (uint32_t i = 0; i < outputShapes.size(); i++) {
         EXPECT_TRUE(outputShapes[i].isSufficient);
-        const auto& expect = testModel.operands[testModel.outputIndexes[i]].dimensions;
+        const auto& expect = testModel.main.operands[testModel.main.outputIndexes[i]].dimensions;
         const std::vector<uint32_t> actual = outputShapes[i].dimensions;
         EXPECT_EQ(expect, actual);
     }
diff --git a/neuralnetworks/1.3/vts/functional/ValidateModel.cpp b/neuralnetworks/1.3/vts/functional/ValidateModel.cpp
index b9ea430694..09e9922a23 100644
--- a/neuralnetworks/1.3/vts/functional/ValidateModel.cpp
+++ b/neuralnetworks/1.3/vts/functional/ValidateModel.cpp
@@ -182,6 +182,7 @@ static float getInvalidScale(OperandType type) {
         case OperandType::TENSOR_FLOAT16:
         case OperandType::TENSOR_FLOAT32:
         case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
+        case OperandType::SUBGRAPH:
             return 1.0f;
         case OperandType::TENSOR_INT32:
             return -1.0f;
@@ -220,6 +221,7 @@ static std::vector<int32_t> getInvalidZeroPoints(OperandType type) {
         case OperandType::TENSOR_FLOAT32:
         case OperandType::TENSOR_INT32:
         case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
+        case OperandType::SUBGRAPH:
             return {1};
         case OperandType::TENSOR_QUANT8_ASYMM:
             return {-1, 256};