[bugfix] Fix CLI flags that modify model_config when using --deploy-config

tests/config/test_omni_config.py

@@ -179,6 +179,42 @@ def test_nested_stage_override_deep_merges_structured_model_config() -> None:
     assert model_config["policy_server_config"]["action_space"] == "joint_position"
 
 
+def test_global_model_config_override_deep_merges_structured_model_config() -> None:
+    pipeline = _resolve_pipeline_or_skip("cosmos3_omni_deploy")
+    # cosmos3_omni.yaml already ships ``guardrails: false``; flip it back to True
+    # (plus a nested key) so the CLI override is observable.
+    deploy = load_deploy_config(_DEPLOY_DIR / "cosmos3_omni.yaml")
+    deploy.stages[0].engine_extras["model_config"] = {
+        "guardrails": True,
+        "inference_overrides": {"history_mode": "full"},
+    }
+
+    baseline = VllmOmniConfig.from_pipeline_config(pipeline, user_deploy_config=deploy)
+    assert baseline.stage_by_id(0).diffusion_config.model_config["guardrails"] is True
+
+    # --no-guardrails emits a global (not stage-scoped) model_config override.
+    config = VllmOmniConfig.from_pipeline_config(
+        pipeline,
+        user_deploy_config=deploy,
+        cli_overrides={"model_config": {"guardrails": False}},
+    )
+
+    model_config = config.stage_by_id(0).diffusion_config.model_config
+    assert model_config["guardrails"] is False
+    assert model_config["inference_overrides"]["history_mode"] == "full"
+
+
+def test_global_model_config_override_skips_llm_stages() -> None:
+    overrides = {"model_config": {"guardrails": False}}
+
+    assert "model_config" not in omni_config_module._stage_cli_overrides(
+        0, overrides, execution_type=StageExecutionType.LLM_AR
+    )
+    assert omni_config_module._stage_cli_overrides(0, overrides, execution_type=StageExecutionType.DIFFUSION)[
+        "model_config"
+    ] == {"guardrails": False}
+
+
 def _build_single_diffusion_config(
     *,
     cli_overrides: dict | None = None,

vllm_omni/config/omni_config.py

@@ -103,6 +103,10 @@
 _QuantizationConfigType: TypeAlias = QuantizationConfig | str | Mapping[str, Any] | None
 
 _DIFFUSION_SHARED_ONLY_ENGINE_FIELDS = frozenset({"kv_cache_dtype", "seed"})
+# Global CLI inputs that are not EngineArgs/deploy fields but still target the
+# stages that own them (``--no-guardrails`` rides on ``model_config``). Stages
+# that do not own them skip them silently instead of rejecting the flag.
+_OWNER_ONLY_GLOBAL_STAGE_CLI_FIELDS = frozenset({"model_config"})
 
 
 class _QuantizationEngineOverrides(TypedDict, total=False):
@@ -367,7 +371,7 @@ def _stage_cli_overrides(
             **stage_scoped,
         }
     runtime_overrides = build_stage_runtime_overrides(stage_id, dict(cli_overrides))
-    global_stage_fields = _global_stage_cli_fields()
+    global_stage_fields = _global_stage_cli_fields() | _OWNER_ONLY_GLOBAL_STAGE_CLI_FIELDS
     owned_fields = None if execution_type is None else _STAGE_ENGINE_FIELDS_BY_EXECUTION_TYPE[execution_type]
     result: dict[str, Any] = {}
     for key, value in runtime_overrides.items():