diff --git a/tests/e2e/offline_inference/test_qwen3_omni.py b/tests/e2e/offline_inference/test_qwen3_omni.py
index b33d693ef..de64c9045 100644
--- a/tests/e2e/offline_inference/test_qwen3_omni.py
+++ b/tests/e2e/offline_inference/test_qwen3_omni.py
@@ -13,6 +13,7 @@ import os
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 
 import pytest
+from vllm import SamplingParams
 
 from tests.helpers.mark import hardware_test
 from tests.helpers.media import generate_synthetic_video
@@ -126,11 +127,19 @@ def test_structured_multistage_config_reaches_runtime(omni_runner, offline_clien
         },
     )
     assert len(engine.default_sampling_params_list) == len(expected_sampling)
-    for runtime_params, expected in zip(engine.default_sampling_params_list, expected_sampling, strict=True):
-        # SamplingParams normalization may add backend defaults (for example
-        # ``detokenize=True``). Verify deploy-owned values without rejecting
-        # those non-lossy normalized fields.
-        assert all(getattr(runtime_params, name) == value for name, value in expected.items())
+    for stage, runtime_params, expected in zip(
+        resolved_stages, engine.default_sampling_params_list, expected_sampling, strict=True
+    ):
+        # The resolver keeps the deploy values verbatim; the pipeline only adds
+        # its own ``detokenize`` / ``stop_token_ids`` constraints on top.
+        assert expected.items() <= (stage.model_config.default_sampling_params or {}).items()
+        # The runtime object is a SamplingParams built from those values, so it
+        # carries vLLM's normalization: greedy sampling (temperature 0) stores
+        # top_k=0, top_p=1.0 and min_p=0.0 even when the YAML spells the
+        # disabled top_k as -1, and backend defaults such as ``detokenize=True``
+        # appear. Compare against the same normalization, not the raw literals.
+        normalized = SamplingParams(**expected)
+        assert all(getattr(runtime_params, name) == getattr(normalized, name) for name in expected)
 
     if current_omni_platform.is_cuda():
         assert thinker_vllm.model_config.enforce_eager is False