[TRTLLM-17557][fix] Make the configuration-database CLI test runnable

docs/source/_static/config_db.json

@@ -24,6 +24,18 @@
       "model_url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4",
       "scenario": "Max Throughput"
     },
+    {
+      "command": "trtllm-serve nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 --config ${TRTLLM_DIR}/examples/configs/curated/nemotron-super-marlin.yaml",
+      "config_filename": "nemotron-super-marlin.yaml",
+      "config_github_url": "https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/configs/curated/nemotron-super-marlin.yaml",
+      "config_path": "examples/configs/curated/nemotron-super-marlin.yaml",
+      "config_raw_url": "https://raw.githubusercontent.com/NVIDIA/TensorRT-LLM/main/examples/configs/curated/nemotron-super-marlin.yaml",
+      "gpu_compatibility": "H100, H200",
+      "model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4",
+      "model_display_name": "Nemotron v3 Super (NVFP4)",
+      "model_url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4",
+      "scenario": "NVFP4 on Hopper (Marlin)"
+    },
     {
       "command": "trtllm-serve Qwen/Qwen3-Next-80B-A3B-Thinking --config ${TRTLLM_DIR}/examples/configs/curated/qwen3-next.yaml",
       "config_filename": "qwen3-next.yaml",

examples/configs/curated/lookup.yaml

@@ -9,6 +9,11 @@
   config_path: examples/configs/curated/nemotron-3-ultra-throughput.yaml
   scenario: Max Throughput
   gpu_compatibility: "B200, B300, GB200, GB300, H100, H200"
+- model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
+  arch: NemotronHForCausalLM
+  config_path: examples/configs/curated/nemotron-super-marlin.yaml
+  scenario: NVFP4 on Hopper (Marlin)
+  gpu_compatibility: "H100, H200"
 - model: Qwen/Qwen3-Next-80B-A3B-Thinking
   arch: Qwen3NextForCausalLM
   config_path: examples/configs/curated/qwen3-next.yaml
@@ -56,6 +61,18 @@
   config_path: examples/configs/curated/qwen3.5.yaml
   scenario: Max Throughput
   gpu_compatibility: "B200, B300, GB200, GB300"
+- model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  arch: Qwen3_5MoeForCausalLM
+  config_path: examples/configs/curated/qwen3.5-gb300-agentic-prefill.yaml
+  scenario: Agentic Disaggregated Prefill
+  gpu_compatibility: "GB300"
+  disagg: true
+- model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  arch: Qwen3_5MoeForCausalLM
+  config_path: examples/configs/curated/qwen3.5-gb300-agentic-decode.yaml
+  scenario: Agentic Disaggregated Decode
+  gpu_compatibility: "GB300"
+  disagg: true
 - model: Qwen/Qwen3-30B-A3B
   arch: Qwen3MoeForCausalLM
   config_path: examples/configs/curated/qwen3.yaml

tests/unittest/llmapi/test_config_database.py

@@ -45,7 +45,14 @@
 CURATED_DIR = CONFIG_ROOT / "curated"
 DATABASE_DIR = CONFIG_ROOT / "database"
 
-CURATED_CONFIGS = collect_yaml_files(CURATED_DIR, "**/*.yaml", exclude_names={"lookup.yaml"})
+# curated/eplb/ holds MoE expert-placement tables that curated configs reference through
+# moe_config.load_balancer; like lookup.yaml they are not standalone serving configs.
+CURATED_EPLB_DIR = CURATED_DIR / "eplb"
+CURATED_CONFIGS = [
+    path
+    for path in collect_yaml_files(CURATED_DIR, "**/*.yaml", exclude_names={"lookup.yaml"})
+    if CURATED_EPLB_DIR not in path.parents
+]
 DATABASE_CONFIGS = collect_yaml_files(DATABASE_DIR, "**/*.yaml", exclude_names={"lookup.yaml"})
 ALL_CONFIGS = sorted(CURATED_CONFIGS + DATABASE_CONFIGS)
 ALL_LOOKUP_PATHS = sorted((DATABASE_DIR / "lookup.yaml", CURATED_DIR / "lookup.yaml"))
@@ -207,6 +214,13 @@ def _serve_cli_args(config_path: Path, port: int = 17999):
         str(config_path),
         "--port",
         str(port),
+        # With more than one frontend, launch_server spawns attached frontend
+        # processes by re-executing sys.argv (here: the pytest command line);
+        # a single frontend keeps the real launch_server in-process. Passing
+        # the flag skips the default frontend-count resolution, so this test
+        # covers the single-frontend path only.
+        "--num_serve_frontends",
+        "1",
     ]
 
 
@@ -218,6 +232,8 @@ class _MockOpenAIServer:
     def __init__(self, generator, model, **kwargs):
         self.generator = generator
         self.model = model
+        # launch_server registers --middleware objects on the server's app.
+        self.app = FastAPI()
 
     def __call__(self, host, port, sockets=None):
         return _noop_serve(host, port, sockets)