[Deploy] Reduce MiniCPM-o single-GPU startup memory (#5447)

Signed-off-by: R2-Y <ruiruyang2@gmail.com>

tests/model_executor/models/minicpmo_4_5/test_pipeline.py

@@ -159,10 +159,11 @@ class TestDeployTopology:
         if filename == "minicpmo_4_5.yaml":
             assert [stage.yaml_engine_args["max_num_seqs"] for stage in stages] == [4, 4, 4]
             assert [stage.yaml_engine_args["gpu_memory_utilization"] for stage in stages] == [
-                0.65,
-                0.15,
-                0.15,
+                0.55,
+                0.22,
+                0.22,
             ]
+            assert stages[0].yaml_engine_args["limit_mm_per_prompt"] == {"video": {"count": 1, "num_frames": 32}}
         elif filename in {"minicpmo_4_5_batching.yaml", "minicpmo_4_5_2gpu.yaml"}:
             assert [stage.yaml_engine_args["gpu_memory_utilization"] for stage in stages] == [
                 0.9,