diff --git a/python/sglang/srt/configs/model_config.py b/python/sglang/srt/configs/model_config.py
index db705eae9..15dd56361 100644
--- a/python/sglang/srt/configs/model_config.py
+++ b/python/sglang/srt/configs/model_config.py
@@ -1679,6 +1679,11 @@ def compute_mla_mscale_scaling(rope_scaling: dict, base_scaling: float) -> float
     Used by DeepSeek, BailingMoe, SarvamMLA and similar MLA models.
     Warns if 'factor' is missing from rope_scaling (common in v5 configs).
     """
+    if not rope_scaling.get("apply_yarn_scaling", True) or not rope_scaling.get(
+        "apply_scale", True
+    ):
+        return base_scaling
+
     mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
     if "factor" not in rope_scaling:
         logger.warning(
diff --git a/test/registered/unit/configs/test_model_config_scaling.py b/test/registered/unit/configs/test_model_config_scaling.py
new file mode 100644
index 000000000..f04e1681f
--- /dev/null
+++ b/test/registered/unit/configs/test_model_config_scaling.py
@@ -0,0 +1,52 @@
+import math
+import unittest
+
+from sglang.srt.configs.model_config import compute_mla_mscale_scaling
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=1, suite="base-a-test-cpu")
+
+
+class TestMlaMscaleScaling(CustomTestCase):
+    def test_respects_disabled_yarn_scaling(self):
+        base_scaling = 1 / math.sqrt(128)
+        rope_scaling = {
+            "rope_type": "deepseek_yarn",
+            "factor": 128,
+            "mscale_all_dim": 1,
+            "apply_yarn_scaling": False,
+        }
+
+        self.assertEqual(
+            compute_mla_mscale_scaling(rope_scaling, base_scaling), base_scaling
+        )
+
+    def test_applies_yarn_scaling_by_default(self):
+        base_scaling = 1 / math.sqrt(128)
+        rope_scaling = {
+            "rope_type": "deepseek_yarn",
+            "factor": 128,
+            "mscale_all_dim": 1,
+        }
+
+        self.assertGreater(
+            compute_mla_mscale_scaling(rope_scaling, base_scaling), base_scaling
+        )
+
+    def test_respects_disabled_native_apply_scale(self):
+        base_scaling = 1 / math.sqrt(128)
+        rope_scaling = {
+            "rope_type": "mistral",
+            "factor": 128,
+            "mscale_all_dim": 1,
+            "apply_scale": False,
+        }
+
+        self.assertEqual(
+            compute_mla_mscale_scaling(rope_scaling, base_scaling), base_scaling
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()