test(phi4-multimodal): cover the runtime config boundary

src/runtime/models/phi4_multimodal/MODEL.toml

@@ -6,5 +6,6 @@ runtime_library = "libtrtmc_model_phi4_multimodal.so"
 runtime_plugins = ["plugin.cpp|register_phi4_multimodal_plugin"]
 runtime_strategies = ["phi4_multimodal_vision_language"]
 runtime_tests = [
+  "test_phi4_multimodal_runtime_config_contract|test_phi4_multimodal_runtime_config_contract.cpp|_|image_preprocessor.cpp|EXTRA_INCLUDES,third_party/stb",
   "test_phi4_multimodal_vl_pipeline|test_phi4_multimodal_vl_pipeline.cpp|trtmc_model_phi4_multimodal,trtmc_backend_trt|_|REQUIRES_TRT,REQUIRES_GPU",
 ]

tests/cpp/models/phi4_multimodal/test_phi4_multimodal_runtime_config_contract.cpp

@@ -0,0 +1,99 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+// CPU-only consumer contract for Phi-4-multimodal's serialized runtime config.
+//
+// Unlike the families that nest their decoder under "text_config", Phi-4
+// multimodal is already flat at the top level, so it needs no
+// get_bundle_config_overrides(). What it does carry is a nested
+// "img_processor" block holding the image tower's own hidden size, head count
+// and layer count. Those keys survive into the bundle's config.json, so pin
+// that the decoder contract is read from the top level and the vision tower's
+// dimensions never leak into it.
+
+#include "runtime/models/phi4_multimodal/image_preprocessor.h"
+#include "trtmc/runtime/pipeline_plugin.h"
+
+#include <iostream>
+#include <string>
+
+int main() {
+    // Shape follows the checkpoint config: flat decoder geometry at the top
+    // level, image tower geometry nested under "img_processor". The nested
+    // values are deliberately small so a consumer that picked them up instead
+    // would fail these checks.
+    const std::string config = R"({
+        "model_type": "phi4mm",
+        "img_processor": {
+            "image_size": 448,
+            "patch_size": 14,
+            "hidden_size": 64,
+            "num_attention_heads": 4,
+            "num_hidden_layers": 2,
+            "intermediate_size": 128,
+            "image_token_id": 999,
+            "vision_output_dim": 777,
+            "num_image_pad_tokens": 888
+        },
+        "vocab_size": 200064,
+        "hidden_size": 3072,
+        "num_hidden_layers": 32,
+        "num_attention_heads": 24,
+        "num_key_value_heads": 8,
+        "head_dim": 128,
+        "bos_token_id": 199999,
+        "eos_token_id": 199999,
+        "max_position_embeddings": 131072,
+        "image_token_id": 200010,
+        "fixed_image_size": 448,
+        "patch_size": 14,
+        "num_image_pad_tokens": 721,
+        "vision_output_dim": 3072,
+        "preprocessor_type": "phi4_hd_chw",
+        "interpolation": "bilinear",
+        "image_token_str": "<|endoftext10|>",
+        "runtime_strategy": "phi4_multimodal_vision_language"
+    })";
+
+    const auto parsed = trtmc::parse_base_config(config, 768);
+    bool ok = true;
+    const auto check = [&](bool condition, const char* name) {
+        if (!condition) {
+            std::cerr << "FAIL: " << name << '\n';
+            ok = false;
+        }
+    };
+
+    // Decoder geometry, read from the top level.
+    check(parsed.runtime_strategy == "phi4_multimodal_vision_language", "runtime strategy");
+    check(parsed.vocab_size == 200064, "vocabulary size");
+    check(parsed.num_layers == 32, "layer count");
+    check(parsed.num_kv_heads == 8, "KV head count");
+    check(parsed.head_dim == 128, "head dimension");
+    check(parsed.id_bos == 199999, "BOS token");
+    check(parsed.id_eos == 199999, "EOS token");
+    check(parsed.max_cache_length == 768, "max cache length override");
+
+    // The image tower's own dimensions must not reach the decoder contract.
+    check(parsed.hidden_size == 3072, "hidden size is the decoder's, not img_processor's");
+    check(parsed.num_heads == 24, "head count is the decoder's, not img_processor's");
+    check(parsed.attention_size == 3072, "attention width");
+
+    // The VL contract, through the family's own consumer: plugin.cpp builds
+    // its Phi4MultimodalPreprocessConfig with exactly this call, so a
+    // regression that stops reading a key, reaches into img_processor
+    // instead, or silently falls back to a default fails here.
+    const auto vl = trtmc::phi4_multimodal_parse_preprocess_config(config, "");
+    check(vl.image_token_id == 200010, "image token id reaches the VL config");
+    check(vl.vision_output_dim == 3072, "vision output dim is the decoder width");
+    check(vl.num_image_pad_tokens == 721, "image pad token count");
+    check(vl.fixed_image_size == 448, "fixed image size");
+    check(vl.patch_size == 14, "patch size");
+    check(vl.preprocessor_type == "phi4_hd_chw", "preprocessor strategy");
+    check(vl.interpolation == "bilinear", "interpolation");
+    check(vl.image_token_str == "<|endoftext10|>", "image placeholder token");
+
+    return ok ? 0 : 1;
+}