[CI][Test] Tier diffusion offloader CI and reduce gpu_4_queue merge usage and fix #5182 (#5185)

Signed-off-by: wangyu <410167048@qq.com>

tests/entrypoints/openai_api/test_qwen3_tts_websocket.py

@@ -102,7 +102,7 @@ async def _run_ws_session(host: str, port: int, model: str) -> dict:
 class TestQwen3TTSWebSocket:
     @pytest.mark.advanced_model
     @pytest.mark.tts
-    @hardware_test(res={"cuda": "L4"}, num_cards=4)
+    @hardware_test(res={"cuda": "L4"}, num_cards=1)
     @pytest.mark.parametrize("omni_server", tts_ws_server_params, indirect=True)
     def test_streaming_pcm_output(self, omni_server) -> None:
         result = asyncio.run(_run_ws_session(omni_server.host, omni_server.port, omni_server.model))

tests/entrypoints/openai_api/test_serving_speech_batch.py

@@ -88,7 +88,7 @@ class TestSpeechBatchE2E:
 
     @pytest.mark.advanced_model
     @pytest.mark.tts
-    @hardware_test(res={"cuda": "H100"}, num_cards=1)
+    @hardware_test(res={"cuda": "L4"}, num_cards=1)
     @pytest.mark.parametrize("omni_server", default_server_params, indirect=True)
     def test_batch_basic_two_items(self, omni_server) -> None:
         """Batch with two items returns two successful base64-encoded results."""