test(internvl): cover runtime config boundary

src/runtime/models/internvl/MODEL.toml

@@ -6,5 +6,6 @@ runtime_library = "libtrtmc_model_internvl.so"
 runtime_plugins = ["plugin.cpp|register_internvl_plugin"]
 runtime_strategies = ["internvl_vision_language"]
 runtime_tests = [
-  "test_internvl_vl_pipeline|test_internvl_vl_pipeline.cpp|trtmc_model_internvl,trtmc_backend_trt|_|REQUIRES_TRT",
+  "test_internvl_runtime_config_contract|test_internvl_runtime_config_contract.cpp|_|_|_",
+  "test_internvl_vl_pipeline|test_internvl_vl_pipeline.cpp|trtmc_model_internvl,trtmc_backend_trt|_|REQUIRES_TRT,REQUIRES_GPU",
 ]

tests/cpp/models/internvl/test_internvl_runtime_config_contract.cpp

@@ -0,0 +1,54 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+// CPU-only consumer contract for InternVL's serialized runtime config.
+
+#include "trtmc/runtime/pipeline_plugin.h"
+
+#include <iostream>
+#include <string>
+
+int main() {
+    const std::string config = R"({
+        "model_type": "internvl",
+        "text_config": {
+            "vocab_size": 151674,
+            "hidden_size": 1536,
+            "num_hidden_layers": 28,
+            "num_attention_heads": 12,
+            "num_key_value_heads": 2,
+            "head_dim": 128,
+            "bos_token_id": 151643
+        },
+        "vocab_size": 151674,
+        "hidden_size": 1536,
+        "num_hidden_layers": 28,
+        "num_attention_heads": 12,
+        "num_key_value_heads": 2,
+        "head_dim": 128,
+        "bos_token_id": 151643,
+        "runtime_strategy": "internvl_vision_language"
+    })";
+
+    const auto parsed = trtmc::parse_base_config(config, 256);
+    bool ok = true;
+    const auto check = [&](bool condition, const char* name) {
+        if (!condition) {
+            std::cerr << "FAIL: " << name << '\n';
+            ok = false;
+        }
+    };
+
+    check(parsed.runtime_strategy == "internvl_vision_language", "runtime strategy");
+    check(parsed.vocab_size == 151674, "vocabulary size");
+    check(parsed.hidden_size == 1536, "hidden size");
+    check(parsed.num_layers == 28, "layer count");
+    check(parsed.num_heads == 12, "attention head count");
+    check(parsed.num_kv_heads == 2, "KV head count");
+    check(parsed.head_dim == 128, "head dimension");
+    check(parsed.id_bos == 151643, "BOS token");
+    check(parsed.attention_size == 1536, "attention width");
+    return ok ? 0 : 1;
+}

tests/e2e/models/internvl/test_internvl_family_plugin_weights.py

@@ -9,7 +9,8 @@
 
 from __future__ import annotations
 
-
+import json
+from unittest.mock import patch
 
 from tests.builder.family_plugin_test_support import (
     ModelConfig,
@@ -265,6 +266,94 @@ def test_bundle_config_overrides_flatten_text_decoder(self, tmp_path):
             "bos_token_id": 151643,
         }
 
+    def test_mock_bundle_serializes_decoder_geometry_at_top_level(self, tmp_path):
+        """Mock engines while exercising the real InternVL bundle boundary."""
+        from tensorrt_model_connect.engine_builder import build_bundle
+        from tensorrt_model_connect.families.internvl.plugin import (
+            plugin as production_plugin,
+        )
+
+        source_config = {
+            "model_type": "internvl",
+            "text_config": {
+                "vocab_size": 151674,
+                "hidden_size": 1536,
+                "num_hidden_layers": 28,
+                "num_attention_heads": 12,
+                "num_key_value_heads": 2,
+                "head_dim": 128,
+                "bos_token_id": 151643,
+            },
+            "vision_config": {
+                "hidden_size": 1024,
+                "num_hidden_layers": 24,
+                "patch_size": 14,
+            },
+        }
+        _write_config(tmp_path, source_config)
+
+        class MockInternVLPlugin:
+            name = "internvl"
+            runtime_strategy = "internvl_vision_language"
+            requires_tokenizer = False
+
+            @staticmethod
+            def load_weights(_model_dir, _config):
+                return {}
+
+            @staticmethod
+            def build_engine(_config, _weights, _max_cache_length, **_kwargs):
+                return b"MOCK_DECODER_PLAN"
+
+            @staticmethod
+            def build_vision_engine(_model_dir, _config, _weights, **_kwargs):
+                return b"MOCK_VISION_PLAN"
+
+            @staticmethod
+            def get_vl_config(config):
+                return production_plugin.get_vl_config(config)
+
+            @staticmethod
+            def get_bundle_config_overrides(config):
+                return production_plugin.get_bundle_config_overrides(config)
+
+        with (
+            patch(
+                "tensorrt_model_connect.engine_builder.find_plugin",
+                return_value=MockInternVLPlugin(),
+            ),
+            patch(
+                "tensorrt_model_connect.engine_builder._get_trt_version",
+                return_value="11.1.0",
+            ),
+            patch(
+                "tensorrt_model_connect.engine_builder._get_gpu_name",
+                return_value="CPU unit mock",
+            ),
+            patch("tensorrt_model_connect.engine_builder.write_bundle") as write_bundle,
+        ):
+            build_bundle(
+                str(tmp_path),
+                str(tmp_path / "internvl.bundle"),
+                max_cache_length=256,
+            )
+
+        sections = {section.name: section.data for section in write_bundle.call_args.args[2]}
+        runtime_config = json.loads(sections["config.json"])
+        assert runtime_config["text_config"] == source_config["text_config"]
+        assert {
+            key: runtime_config[key]
+            for key in (
+                "vocab_size",
+                "hidden_size",
+                "num_hidden_layers",
+                "num_attention_heads",
+                "num_key_value_heads",
+                "head_dim",
+                "bos_token_id",
+            )
+        } == source_config["text_config"]
+
     def test_no_vl_config_without_vision(self, tmp_path):
         """get_vl_config returns None when no vision_config present."""
         from tensorrt_model_connect.families.internvl import plugin