diff --git a/python/sglang/srt/utils/hf_transformers_patches.py b/python/sglang/srt/utils/hf_transformers_patches.py
index dbf9726a0..17a394190 100644
--- a/python/sglang/srt/utils/hf_transformers_patches.py
+++ b/python/sglang/srt/utils/hf_transformers_patches.py
@@ -136,41 +136,18 @@ def _ensure_gguf_version():
 
 
 def _patch_rope_parameters_validation():
-    """Fix rope_parameters validation for unregistered model types.
+    """Guard ``standardize_rope_params()`` against missing
+    ``max_position_embeddings``.
 
-    For unregistered model types (e.g. ``deepseek_v32``), the generic
-    ``PretrainedConfig`` lacks a ``rope_parameters`` field so the conversion
-    that injects ``rope_theta`` from the top-level config is skipped.
-    Additionally, ``standardize_rope_params()`` accesses
+    For ``PretrainedConfig``, ``standardize_rope_params()`` accesses
     ``self.max_position_embeddings`` during ``__post_init__`` before extra
     kwargs are set as attributes, causing ``AttributeError``.
 
-    Fix: (1) patch ``from_dict`` to inject ``rope_theta`` into
-    ``rope_scaling``, (2) guard ``standardize_rope_params`` against missing
+    Fix: guard ``standardize_rope_params`` against missing
     ``max_position_embeddings``.
-
-    TODO(upstream): remove once unregistered model types handle rope
-    standardization correctly in transformers.
     """
     from transformers import PretrainedConfig
 
-    original = PretrainedConfig.from_dict.__func__
-
-    @classmethod  # type: ignore[misc]
-    def patched(cls, config_dict, **kwargs):
-        rope_scaling = config_dict.get("rope_scaling")
-        rope_theta = config_dict.get("rope_theta")
-        if (
-            isinstance(rope_scaling, dict)
-            and rope_theta is not None
-            and "rope_theta" not in rope_scaling
-        ):
-            config_dict = config_dict.copy()
-            config_dict["rope_scaling"] = {**rope_scaling, "rope_theta": rope_theta}
-        return original(cls, config_dict, **kwargs)
-
-    PretrainedConfig.from_dict = patched
-
     # standardize_rope_params accesses self.max_position_embeddings before
     # __post_init__ sets extra kwargs — skip when the attribute is absent.
     if hasattr(PretrainedConfig, "standardize_rope_params"):
diff --git a/test/registered/unit/utils/test_hf_transformers.py b/test/registered/unit/utils/test_hf_transformers.py
index b29839d0e..81b0122d5 100644
--- a/test/registered/unit/utils/test_hf_transformers.py
+++ b/test/registered/unit/utils/test_hf_transformers.py
@@ -480,6 +480,15 @@ class TestPatchRemovedSymbols(unittest.TestCase):
 
 
 class TestPatchRopeParametersValidation(unittest.TestCase):
+    # -----------------------------------------------------------------------
+    # Test ``rope_theta`` injection into ``rope_scaling``.
+    #
+    # Upstream `transformers.PretrainedConfig` now natively handles this
+    # logic. While the manual injection patch has been removed, these
+    # test cases are retained to ensure regression testing of the
+    # configuration's injection behavior.
+    # -----------------------------------------------------------------------
+
     def test_injects_rope_theta_into_rope_scaling(self):
         config_dict = {
             "model_type": "llama",