diff --git a/‎examples/vllm_serve/vllm_serve_fakequant.py‎
Lines changed: 18 additions & 4 deletions b/‎examples/vllm_serve/vllm_serve_fakequant.py‎
Lines changed: 18 additions & 4 deletions
diff --git a/‎tests/gpu_vllm/torch/quantization/test_vllm_dynamic_modules.py‎
Lines changed: 4 additions & 1 deletion b/‎tests/gpu_vllm/torch/quantization/test_vllm_dynamic_modules.py‎
Lines changed: 4 additions & 1 deletion
@@ -263,12 +263,29 @@ def _find_serve_model(rest_argv: list) -> str | None:
     return getattr(args, "model_tag", None) or getattr(args, "model", None)
 
 
+def _run_vllm_cli(argv: list[str]) -> None:
+    """Delegate arguments to the stock vLLM CLI."""
+    sys.argv = ["vllm", *argv]
+    from vllm.entrypoints.cli.main import main as vllm_main
+
+    vllm_main()
+
+
 def main():
+    argv = sys.argv[1:]
+    # Non-serve commands do not consume ModelOpt settings, including env defaults.
+    if argv and argv[0] in _VLLM_SUBCOMMANDS and argv[0] != "serve":
+        _run_vllm_cli(argv)
+        return
+
     modelopt_parser = FlexibleArgumentParser(add_help=False)
     _add_fakequant_args(modelopt_parser)
     add_mlflow_args(modelopt_parser)
     modelopt_args, rest_argv = modelopt_parser.parse_known_args(sys.argv[1:])
     rest_argv = _default_to_serve(rest_argv)
+    if rest_argv and rest_argv[0] != "serve":
+        _run_vllm_cli(argv)
+        return
 
     if (modelopt_args.modelopt_quant_cfg or modelopt_args.modelopt_kv_quant_cfg) and (
         modelopt_args.modelopt_recipe_path
@@ -323,10 +340,7 @@ def main():
             if not _has_flag(rest_argv, "--moe-backend", "--moe_backend"):
                 rest_argv = [*rest_argv, "--moe_backend", "triton"]
 
-    sys.argv = ["vllm", *rest_argv]
-    from vllm.entrypoints.cli.main import main as vllm_main
-
-    vllm_main()
+    _run_vllm_cli(rest_argv)
 
 
 if __name__ == "__main__":
 
@@ -210,7 +210,8 @@ def test_vllm_serve_parser_dependency_error_propagates(monkeypatch):
     assert raised.value is dependency_error
 
 
-def test_fakequant_launcher_passes_through_vllm_launch(monkeypatch):
+@pytest.mark.parametrize("quant_cfg", [None, "FP8_DEFAULT_CFG"])
+def test_fakequant_launcher_passes_through_vllm_launch(monkeypatch, quant_cfg):
     for key in (
         "QUANT_CFG",
         "KV_QUANT_CFG",
@@ -219,6 +220,8 @@ def test_fakequant_launcher_passes_through_vllm_launch(monkeypatch):
         "QUANT_FILE_PATH",
     ):
         monkeypatch.delenv(key, raising=False)
+    if quant_cfg:
+        monkeypatch.setenv("QUANT_CFG", quant_cfg)
     launcher = _load_fakequant_launcher(monkeypatch)
     monkeypatch.setattr(launcher, "resolve_mlflow_args", Mock())
     vllm_main = Mock()