diff --git a/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py b/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
index fcb31631c8f9..35cd7c1a0f1f 100644
--- a/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
+++ b/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
@@ -547,9 +547,7 @@ def do_mem_analysis(self, output_batch: OutputBatch):
 
         residency_hint = (
             f" Components that can remain on GPU: {can_stay_resident}. "
-            "Make it explicit with --component-residency <name>=resident; "
-            "--performance-mode auto with server warmup applies safe "
-            "adjustments automatically."
+            "Make it explicit with --component-residency <name>=resident."
             if can_stay_resident
             else ""
         )
diff --git a/python/sglang/multimodal_gen/runtime/server_args/auto_tune.py b/python/sglang/multimodal_gen/runtime/server_args/auto_tune.py
index 12bfc720118e..cf012ceee519 100644
--- a/python/sglang/multimodal_gen/runtime/server_args/auto_tune.py
+++ b/python/sglang/multimodal_gen/runtime/server_args/auto_tune.py
@@ -93,7 +93,8 @@ def auto_residency_args_skip_reason(server_args: ServerArgs) -> str | None:
         return "dynamic batching enabled"
     if not current_platform.is_cuda():
         return "requires CUDA"
-    return None
+    # nothing applies calibrated placements yet, so warmup skips the probe and step floor
+    return "residency planner not available"
 
 
 PERFORMANCE_MODES = ("manual", "auto", "speed", "memory")
diff --git a/python/sglang/multimodal_gen/test/unit/test_auto_residency.py b/python/sglang/multimodal_gen/test/unit/test_auto_residency.py
index 31299282eb35..42f6b141c942 100644
--- a/python/sglang/multimodal_gen/test/unit/test_auto_residency.py
+++ b/python/sglang/multimodal_gen/test/unit/test_auto_residency.py
@@ -7,18 +7,24 @@
 
 from sglang.multimodal_gen.configs.pipeline_configs.base import ModelTaskType
 from sglang.multimodal_gen.configs.pipeline_configs.longlive2 import LongLive2T2VConfig
+from sglang.multimodal_gen.configs.sample.sampling_params import SamplingParams
 from sglang.multimodal_gen.runtime.managers.memory_managers.auto_residency import (
     ACTIVATION_EXTRAPOLATION_MARGIN,
     GIB_BYTES,
     WarmupMemoryRecord,
     estimate_default_workload_peak_bytes,
 )
+from sglang.multimodal_gen.runtime.platforms import current_platform
 from sglang.multimodal_gen.runtime.warmup_request_builder import (
     SERVER_WARMUP_MAX_VIDEO_FRAMES,
+    SERVER_WARMUP_VIDEO_STEPS,
     _resolve_auto_residency_warmup_shape,
     _resolve_warmup_num_frames,
+    build_warmup_reqs,
 )
 
+PLANNER_UNAVAILABLE = "residency planner not available"
+
 
 def _record(
     *,
@@ -455,8 +461,9 @@ def test_fixed_loading_paths_still_calibrate_other_components(
     ):
         monkeypatch.delenv("SGLANG_DIFFUSION_DISABLE_AUTO_RESIDENCY", raising=False)
         monkeypatch.delenv("SGLANG_CACHE_DIT_ENABLED", raising=False)
+        monkeypatch.setattr(current_platform, "is_cuda", lambda: True)
         reason = self._skip_reason(self._base_args(**overrides))
-        assert reason is None or reason == "requires CUDA"
+        assert reason == PLANNER_UNAVAILABLE
 
     def test_cache_dit_excluded(self, monkeypatch):
         monkeypatch.delenv("SGLANG_DIFFUSION_DISABLE_AUTO_RESIDENCY", raising=False)
@@ -472,15 +479,53 @@ def test_cache_dit_excluded(self, monkeypatch):
             {"use_fsdp_inference": True},
         ],
     )
-    def test_parallel_paths_reach_platform_gate(self, monkeypatch, overrides):
+    def test_parallel_paths_reach_planner_gate(self, monkeypatch, overrides):
         monkeypatch.delenv("SGLANG_DIFFUSION_DISABLE_AUTO_RESIDENCY", raising=False)
         monkeypatch.delenv("SGLANG_CACHE_DIT_ENABLED", raising=False)
+        monkeypatch.setattr(current_platform, "is_cuda", lambda: True)
         reason = self._skip_reason(self._base_args(**overrides))
-        assert reason is None or reason == "requires CUDA"
+        assert reason == PLANNER_UNAVAILABLE
 
-    def test_eligible_path_reaches_platform_gate(self, monkeypatch):
+    def test_eligible_path_reaches_planner_gate(self, monkeypatch):
         monkeypatch.delenv("SGLANG_DIFFUSION_DISABLE_AUTO_RESIDENCY", raising=False)
         monkeypatch.delenv("SGLANG_CACHE_DIT_ENABLED", raising=False)
+        monkeypatch.setattr(current_platform, "is_cuda", lambda: True)
         reason = self._skip_reason(self._base_args())
-        # on a CUDA host everything passes; CPU CI stops at the platform gate
-        assert reason is None or reason == "requires CUDA"
+        assert reason == PLANNER_UNAVAILABLE
+
+    def test_eligible_server_warmup_runs_no_probe_or_step_floor(self, monkeypatch):
+        # no calibration consumer yet: one bounded warmup, not probe + 4-step floor
+        monkeypatch.delenv("SGLANG_DIFFUSION_DISABLE_AUTO_RESIDENCY", raising=False)
+        monkeypatch.delenv("SGLANG_CACHE_DIT_ENABLED", raising=False)
+        monkeypatch.setattr(current_platform, "is_cuda", lambda: True)
+        args = self._base_args(
+            warmup_steps=1,
+            enable_cfg_parallel=False,
+            num_gpus=1,
+            pipeline_config=SimpleNamespace(
+                task_type=ModelTaskType.T2V,
+                supports_auto_residency=True,
+                adjust_num_frames=lambda value: value,
+                vae_stride=None,
+                vae_scale_factor=None,
+                vae_config=SimpleNamespace(arch_config=None),
+            ),
+            is_arg_explicitly_set=lambda _name: False,
+        )
+        monkeypatch.setattr(
+            "sglang.multimodal_gen.runtime.warmup_request_builder.get_model_sampling_defaults",
+            lambda _args: SamplingParams(
+                width=1280,
+                height=720,
+                num_frames=81,
+                num_inference_steps=35,
+                adjust_frames=False,
+                supported_resolutions=[(1280, 720), (832, 480)],
+            ),
+        )
+        reqs = build_warmup_reqs(
+            args, warmup_resolutions=None, server_based_warmup=True
+        )
+        assert len(reqs) == 1
+        assert not reqs[0].extra.get("auto_residency_full_shape_probe")
+        assert reqs[0].num_inference_steps == SERVER_WARMUP_VIDEO_STEPS