diff --git a/benchmarks/kernels/benchmark_rope.py b/benchmarks/kernels/benchmark_rope.py
index 73fc9e9dbf461..784b1cf9844e4 100644
--- a/benchmarks/kernels/benchmark_rope.py
+++ b/benchmarks/kernels/benchmark_rope.py
@@ -31,7 +31,7 @@ def benchmark_rope_kernels_multi_lora(
     # batched RoPE can take multiple scaling factors
     batched_rope = get_rope(head_size, rotary_dim, max_position, base,
                             is_neox_style, {
-                                "type": "linear",
+                                "rope_type": "linear",
                                 "factor": tuple(scaling_factors)
                             })
     # non-batched RoPE takes only one scaling factor, we create multiple
@@ -41,7 +41,7 @@ def benchmark_rope_kernels_multi_lora(
         non_batched_ropes.append(
             get_rope(head_size, rotary_dim, max_position, base, is_neox_style,
                      {
-                         "type": "linear",
+                         "rope_type": "linear",
                          "factor": (scaling_factor, )
                      }))
 
diff --git a/requirements-common.txt b/requirements-common.txt
index 1178143409e2e..ca09f9d35909e 100644
--- a/requirements-common.txt
+++ b/requirements-common.txt
@@ -4,7 +4,7 @@ numpy < 2.0.0
 requests >= 2.26.0
 tqdm
 py-cpuinfo
-transformers >= 4.45.0  # Required for Llama 3.2.
+transformers >= 4.45.2  # Required for Llama 3.2 and Qwen2-VL.
 tokenizers >= 0.19.1  # Required for Llama 3.
 protobuf # Required by LlamaTokenizer.
 fastapi >= 0.107.0, < 0.113.0; python_version < '3.9'
diff --git a/tests/kernels/test_pos_encoding.py b/tests/kernels/test_pos_encoding.py
index ba9d2d4389b21..94da00915d40e 100644
--- a/tests/kernels/test_pos_encoding.py
+++ b/tests/kernels/test_pos_encoding.py
@@ -105,7 +105,7 @@ def test_batched_rotary_embedding(
     if rotary_dim is None:
         rotary_dim = head_size
     rope = get_rope(head_size, rotary_dim, max_position, base, is_neox_style, {
-        "type": "linear",
+        "rope_type": "linear",
         "factor": (1, )
     })
     rope = rope.to(dtype=dtype)
@@ -166,7 +166,7 @@ def test_batched_rotary_embedding_multi_lora(
         rotary_dim = head_size
     scaling_factors: List[int] = [1, 2, 4]
     rope = get_rope(head_size, rotary_dim, max_position, base, is_neox_style, {
-        "type": "linear",
+        "rope_type": "linear",
         "factor": tuple(scaling_factors)
     })
     rope = rope.to(dtype=dtype)
@@ -211,10 +211,10 @@ def test_rope_module_cache():
     MAX_POSITIONS = [123, 1234]
     BASES = [10000, 1000000]
     ROPE_SCALINGS = (None, {
-        "type": "linear",
+        "rope_type": "linear",
         "factor": (1, )
     }, {
-        "type": "dynamic",
+        "rope_type": "dynamic",
         "factor": 1
     })
     settings = (HEAD_SIZES, ROTARY_DIMS, MAX_POSITIONS, BASES, IS_NEOX_STYLE,
diff --git a/tests/lora/test_layers.py b/tests/lora/test_layers.py
index e3233c6b60696..db877219a285c 100644
--- a/tests/lora/test_layers.py
+++ b/tests/lora/test_layers.py
@@ -951,7 +951,7 @@ def test_rotary_embedding_long_context(dist_init, num_loras, device,
     lora_rope.create_lora_weights(max_loras, lora_config)
     linear_rope = get_rope(head_size, rotary_dim, max_position, base,
                            is_neox_style, {
-                               "type": "linear",
+                               "rope_type": "linear",
                                "factor": scaling_factors
                            })
     linear_rope = linear_rope.to(dtype=dtype)
diff --git a/tests/test_config.py b/tests/test_config.py
index 225d71c0bc0ea..b89429005e1d0 100644
--- a/tests/test_config.py
+++ b/tests/test_config.py
@@ -64,9 +64,9 @@ def test_get_sliding_window():
 
 
 def test_rope_customization():
-    TEST_ROPE_SCALING = {"type": "dynamic", "factor": 2.0}
+    TEST_ROPE_SCALING = {"rope_type": "dynamic", "factor": 2.0}
     TEST_ROPE_THETA = 16_000_000.0
-    LONGCHAT_ROPE_SCALING = {"type": "linear", "factor": 8.0}
+    LONGCHAT_ROPE_SCALING = {"rope_type": "linear", "factor": 8.0}
 
     llama_model_config = ModelConfig(
         "meta-llama/Meta-Llama-3-8B-Instruct",
diff --git a/vllm/config.py b/vllm/config.py
index 7a3248f4087ae..33005ebbd5219 100644
--- a/vllm/config.py
+++ b/vllm/config.py
@@ -1739,16 +1739,10 @@ def _get_and_verify_max_len(
 
     rope_scaling = getattr(hf_config, "rope_scaling", None)
     if rope_scaling is not None:
-        if "type" in rope_scaling:
-            rope_type = rope_scaling["type"]
-        elif "rope_type" in rope_scaling:
-            rope_type = rope_scaling["rope_type"]
-        else:
-            raise ValueError(
-                "rope_scaling must have a 'type' or 'rope_type' key.")
+        # No need to consider "type" key because of patch_rope_scaling when
+        # loading HF config
+        rope_type = rope_scaling["rope_type"]
 
-        # The correct one should be "longrope", kept "su" here
-        # to be backward compatible
         if rope_type not in ("su", "longrope", "llama3"):
             if disable_sliding_window:
                 # TODO(robertgshaw): Find a model that supports rope_scaling
@@ -1758,11 +1752,10 @@ def _get_and_verify_max_len(
                     "with rope_scaling. Please raise an issue so we can "
                     "investigate.")
 
-            if rope_type == "mrope":
-                scaling_factor = 1
-            else:
-                assert "factor" in rope_scaling
-                scaling_factor = rope_scaling["factor"]
+            # NOTE: rope_type == "default" does not define factor
+            # https://github.com/huggingface/transformers/blob/v4.45.2/src/transformers/modeling_rope_utils.py
+            scaling_factor = rope_scaling.get("factor", 1.0)
+
             if rope_type == "yarn":
                 derived_max_model_len = rope_scaling[
                     "original_max_position_embeddings"]
diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
index 1b132cf76a10d..040b8c1bdd0a2 100644
--- a/vllm/engine/arg_utils.py
+++ b/vllm/engine/arg_utils.py
@@ -454,11 +454,12 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
                             'None, we assume the model weights are not '
                             'quantized and use `dtype` to determine the data '
                             'type of the weights.')
-        parser.add_argument('--rope-scaling',
-                            default=None,
-                            type=json.loads,
-                            help='RoPE scaling configuration in JSON format. '
-                            'For example, {"type":"dynamic","factor":2.0}')
+        parser.add_argument(
+            '--rope-scaling',
+            default=None,
+            type=json.loads,
+            help='RoPE scaling configuration in JSON format. '
+            'For example, {"rope_type":"dynamic","factor":2.0}')
         parser.add_argument('--rope-theta',
                             default=None,
                             type=float,
diff --git a/vllm/model_executor/layers/rotary_embedding.py b/vllm/model_executor/layers/rotary_embedding.py
index d4e9ed87ed54f..2ed44e2093bbe 100644
--- a/vllm/model_executor/layers/rotary_embedding.py
+++ b/vllm/model_executor/layers/rotary_embedding.py
@@ -920,13 +920,10 @@ def get_rope(
         rotary_emb = RotaryEmbedding(head_size, rotary_dim, max_position, base,
                                      is_neox_style, dtype)
     else:
-        scaling_type = rope_scaling[
-            "type"] if "type" in rope_scaling else rope_scaling["rope_type"]
-        # The correct one should be "longrope" but keep "su" here
-        # for backward compatible
-        if scaling_type not in {"su", "longrope"}:
-            scaling_factor = rope_scaling.get("factor", 1.0)
+        scaling_type = rope_scaling["rope_type"]
+
         if scaling_type == "llama3":
+            scaling_factor = rope_scaling["factor"]
             low_freq_factor = rope_scaling["low_freq_factor"]
             high_freq_factor = rope_scaling["high_freq_factor"]
             original_max_position = rope_scaling[
@@ -937,16 +934,39 @@ def get_rope(
                                                scaling_factor, low_freq_factor,
                                                high_freq_factor,
                                                original_max_position)
+        elif scaling_type == "default":
+            if "mrope_section" in rope_scaling:
+                rotary_emb = MRotaryEmbedding(
+                    head_size,
+                    rotary_dim,
+                    max_position,
+                    base,
+                    is_neox_style,
+                    dtype,
+                    mrope_section=rope_scaling["mrope_section"],
+                )
+            else:
+                rotary_emb = RotaryEmbedding(
+                    head_size,
+                    rotary_dim,
+                    max_position,
+                    base,
+                    is_neox_style,
+                    dtype,
+                )
         elif scaling_type == "linear":
+            scaling_factor = rope_scaling["factor"]
             rotary_emb = LinearScalingRotaryEmbedding(head_size, rotary_dim,
                                                       max_position, base,
                                                       is_neox_style,
                                                       scaling_factor, dtype)
         elif scaling_type == "dynamic":
+            scaling_factor = rope_scaling["factor"]
             rotary_emb = DynamicNTKScalingRotaryEmbedding(
                 head_size, rotary_dim, max_position, base, is_neox_style,
                 scaling_factor, dtype)
         elif scaling_type == "yarn":
+            scaling_factor = rope_scaling["factor"]
             original_max_position = rope_scaling[
                 "original_max_position_embeddings"]
             extra_kwargs = {
@@ -961,6 +981,7 @@ def get_rope(
                                                     scaling_factor, dtype,
                                                     **extra_kwargs)
         elif scaling_type == "deepseek_yarn":
+            scaling_factor = rope_scaling["factor"]
             original_max_position = rope_scaling[
                 "original_max_position_embeddings"]
             # assert max_position == original_max_position * scaling_factor
@@ -973,9 +994,7 @@ def get_rope(
             rotary_emb = DeepseekScalingRotaryEmbedding(
                 head_size, rotary_dim, original_max_position, base,
                 is_neox_style, scaling_factor, dtype, **extra_kwargs)
-        # The correct one should be "longrope" but keep "su" here
-        # for backward compatible
-        elif scaling_type == "su" or scaling_type == "longrope":
+        elif scaling_type == "longrope":
             short_factor = rope_scaling["short_factor"]
             long_factor = rope_scaling["long_factor"]
             original_max_position = rope_scaling[
@@ -989,16 +1008,6 @@ def get_rope(
                 head_size, rotary_dim, max_position, original_max_position,
                 base, is_neox_style, dtype, short_factor, long_factor,
                 **extra_kwargs)
-        elif scaling_type == "mrope":
-            rotary_emb = MRotaryEmbedding(
-                head_size,
-                rotary_dim,
-                max_position,
-                base,
-                is_neox_style,
-                dtype,
-                mrope_section=rope_scaling["mrope_section"],
-            )
         else:
             raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
     _ROPE_DICT[key] = rotary_emb
diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py
index 702be7b7f5ed9..38114836bfdbb 100644
--- a/vllm/model_executor/models/deepseek_v2.py
+++ b/vllm/model_executor/models/deepseek_v2.py
@@ -242,7 +242,7 @@ def __init__(
                                         bias=False,
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.o_proj")
-        rope_scaling['type'] = 'deepseek_yarn'
+        rope_scaling["rope_type"] = 'deepseek_yarn'
         self.rotary_emb = get_rope(qk_rope_head_dim,
                                    rotary_dim=qk_rope_head_dim,
                                    max_position=max_position_embeddings,
diff --git a/vllm/model_executor/models/phi3_small.py b/vllm/model_executor/models/phi3_small.py
index 4cfeb3bb3496f..3a7afc606bb9a 100644
--- a/vllm/model_executor/models/phi3_small.py
+++ b/vllm/model_executor/models/phi3_small.py
@@ -179,7 +179,7 @@ def __init__(
                 rope_scaling["factor"] = self.rope_position_scale
         else:
             rope_scaling = {
-                "type": "linear",
+                "rope_type": "linear",
                 "factor": self.rope_position_scale,
             }
 
diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py
index 4a39b3fbe5a41..bdc21df8b6563 100644
--- a/vllm/model_executor/models/qwen2_vl.py
+++ b/vllm/model_executor/models/qwen2_vl.py
@@ -34,6 +34,8 @@
 from transformers.image_utils import (get_image_size,
                                       infer_channel_dimension_format,
                                       to_numpy_array)
+from transformers.models.qwen2_vl.configuration_qwen2_vl import (
+    Qwen2VLConfig, Qwen2VLVisionConfig)
 from transformers.models.qwen2_vl.image_processing_qwen2_vl import (
     make_batched_images, make_batched_videos, smart_resize)
 
@@ -62,8 +64,7 @@
 from vllm.multimodal.image import cached_get_image_processor
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors, SequenceData
-from vllm.transformers_utils.configs.qwen2vl import (Qwen2VLConfig,
-                                                     Qwen2VLVisionConfig)
+from vllm.transformers_utils.config import uses_mrope
 from vllm.transformers_utils.processor import get_processor
 from vllm.utils import is_cpu
 
@@ -1061,8 +1062,7 @@ def forward(
             if image_input is None and video_input is None:
                 inputs_embeds = None
             else:
-                rope_scaling = getattr(self.config, "rope_scaling", {})
-                if rope_scaling.get("type", None) == "mrope":
+                if uses_mrope(self.config):
                     assert positions.ndim == 2 and positions.size(0) == 3, (
                         "multimodal section rotary embedding requires "
                         f"(3, seq_len) positions, but got {positions.size()}")
diff --git a/vllm/transformers_utils/config.py b/vllm/transformers_utils/config.py
index b33449c42ecf5..46405f3529215 100644
--- a/vllm/transformers_utils/config.py
+++ b/vllm/transformers_utils/config.py
@@ -23,8 +23,8 @@
                                              MedusaConfig, MllamaConfig,
                                              MLPSpeculatorConfig, MPTConfig,
                                              NemotronConfig, NVLM_D_Config,
-                                             Qwen2VLConfig, RWConfig,
-                                             SolarConfig, UltravoxConfig)
+                                             RWConfig, SolarConfig,
+                                             UltravoxConfig)
 # yapf: enable
 from vllm.transformers_utils.utils import check_gguf_file
 
@@ -57,7 +57,6 @@
     "NVLM_D": NVLM_D_Config,
     "solar": SolarConfig,
     "ultravox": UltravoxConfig,
-    "qwen2_vl": Qwen2VLConfig,
     **_CONFIG_REGISTRY_OVERRIDE_HF
 }
 
@@ -91,6 +90,43 @@ def file_or_path_exists(model: Union[str, Path], config_name, revision,
         return False
 
 
+def patch_rope_scaling(config: PretrainedConfig) -> None:
+    """Provide backwards compatibility for RoPE."""
+    text_config = getattr(config, "text_config", None)
+    if text_config is not None:
+        patch_rope_scaling(text_config)
+
+    rope_scaling = getattr(config, "rope_scaling", None)
+    if rope_scaling is not None:
+        patch_rope_scaling_dict(rope_scaling)
+
+
+def patch_rope_scaling_dict(rope_scaling: Dict[str, Any]) -> None:
+    if "rope_type" not in rope_scaling and "type" in rope_scaling:
+        rope_scaling["rope_type"] = rope_scaling["type"]
+        logger.info("Replacing legacy 'type' key with 'rope_type'")
+
+    if "rope_type" not in rope_scaling:
+        raise ValueError("rope_scaling should have a 'rope_type' key")
+
+    if rope_scaling["rope_type"] == "su":
+        rope_scaling["rope_type"] = "longrope"
+        logger.warning("Replacing legacy rope_type 'su' with 'longrope'")
+    elif rope_scaling["rope_type"] == "mrope":
+        assert "mrope_section" in rope_scaling
+        rope_scaling["rope_type"] = "default"
+        logger.warning("Replacing legacy rope_type 'mrope' with 'default'")
+
+
+def uses_mrope(config: PretrainedConfig) -> bool:
+    """Detect if the model with this config uses M-ROPE."""
+    rope_scaling = getattr(config, "rope_scaling", None)
+    if rope_scaling is None:
+        return False
+
+    return "mrope_section" in rope_scaling
+
+
 def get_config(
     model: Union[str, Path],
     trust_remote_code: bool,
@@ -191,6 +227,8 @@ def get_config(
             )
             config.update({key: value})
 
+    patch_rope_scaling(config)
+
     return config
 
 
diff --git a/vllm/transformers_utils/configs/__init__.py b/vllm/transformers_utils/configs/__init__.py
index 8d6385d42d002..f0d79197a82c5 100644
--- a/vllm/transformers_utils/configs/__init__.py
+++ b/vllm/transformers_utils/configs/__init__.py
@@ -14,8 +14,6 @@
 from vllm.transformers_utils.configs.mpt import MPTConfig
 from vllm.transformers_utils.configs.nemotron import NemotronConfig
 from vllm.transformers_utils.configs.nvlm_d import NVLM_D_Config
-from vllm.transformers_utils.configs.qwen2vl import (Qwen2VLConfig,
-                                                     Qwen2VLVisionConfig)
 from vllm.transformers_utils.configs.solar import SolarConfig
 from vllm.transformers_utils.configs.ultravox import UltravoxConfig
 
@@ -35,6 +33,4 @@
     "NVLM_D_Config",
     "SolarConfig",
     "UltravoxConfig",
-    "Qwen2VLConfig",
-    "Qwen2VLVisionConfig",
 ]
diff --git a/vllm/transformers_utils/configs/qwen2vl.py b/vllm/transformers_utils/configs/qwen2vl.py
deleted file mode 100644
index 92dd962790bc8..0000000000000
--- a/vllm/transformers_utils/configs/qwen2vl.py
+++ /dev/null
@@ -1,131 +0,0 @@
-# coding=utf-8
-# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team.
-# All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Qwen2VL model configuration"""
-
-import os
-from typing import Union
-
-from transformers import PretrainedConfig
-
-
-class Qwen2VLVisionConfig(PretrainedConfig):
-    model_type = "qwen2_vl"
-
-    def __init__(
-        self,
-        depth=32,
-        embed_dim=1280,
-        hidden_size=3584,
-        hidden_act="quick_gelu",
-        mlp_ratio=4,
-        num_heads=16,
-        in_channels=3,
-        patch_size=14,
-        spatial_merge_size=2,
-        temporal_patch_size=2,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
-
-        self.depth = depth
-        self.embed_dim = embed_dim
-        self.hidden_size = hidden_size
-        self.hidden_act = hidden_act
-        self.mlp_ratio = mlp_ratio
-        self.num_heads = num_heads
-        self.in_channels = in_channels
-        self.patch_size = patch_size
-        self.spatial_merge_size = spatial_merge_size
-        self.temporal_patch_size = temporal_patch_size
-
-    @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path: Union[str,
-                                                                  os.PathLike],
-                        **kwargs) -> "PretrainedConfig":
-        cls._set_token_in_kwargs(kwargs)
-
-        config_dict, kwargs = cls.get_config_dict(
-            pretrained_model_name_or_path, **kwargs)
-
-        if config_dict.get("model_type") == "qwen2_vl":
-            config_dict = config_dict["vision_config"]
-
-        return cls.from_dict(config_dict, **kwargs)
-
-
-class Qwen2VLConfig(PretrainedConfig):
-
-    def __init__(
-        self,
-        vocab_size=152064,
-        hidden_size=8192,
-        intermediate_size=29568,
-        num_hidden_layers=80,
-        num_attention_heads=64,
-        num_key_value_heads=8,
-        hidden_act="silu",
-        max_position_embeddings=32768,
-        initializer_range=0.02,
-        rms_norm_eps=1e-05,
-        use_cache=True,
-        tie_word_embeddings=False,
-        rope_theta=1000000.0,
-        use_sliding_window=False,
-        sliding_window=4096,
-        max_window_layers=80,
-        attention_dropout=0.0,
-        vision_config=None,
-        rope_scaling=None,
-        **kwargs,
-    ):
-        if isinstance(vision_config, dict):
-            self.vision_config = Qwen2VLVisionConfig(**vision_config)
-        elif vision_config is None:
-            self.vision_config = Qwen2VLVisionConfig()
-
-        self.vocab_size = vocab_size
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.use_sliding_window = use_sliding_window
-        self.sliding_window = sliding_window
-        self.max_window_layers = max_window_layers
-
-        # for backward compatibility
-        if num_key_value_heads is None:
-            num_key_value_heads = num_attention_heads
-
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.initializer_range = initializer_range
-        self.rms_norm_eps = rms_norm_eps
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.attention_dropout = attention_dropout
-        self.rope_scaling = rope_scaling
-
-        # NOTE: the following section from original transformers config
-        # for Qwen2-VL is commented out to address rope config loading issue
-        #
-        # if self.rope_scaling is not None and "type" in self.rope_scaling:
-        #     if self.rope_scaling["type"] == "mrope":
-        #         self.rope_scaling["type"] = "default"
-        #     self.rope_scaling["rope_type"] = self.rope_scaling["type"]
-        # rope_config_validation(self)
-
-        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
diff --git a/vllm/worker/cpu_model_runner.py b/vllm/worker/cpu_model_runner.py
index 795511aea6754..dd38b550eb011 100644
--- a/vllm/worker/cpu_model_runner.py
+++ b/vllm/worker/cpu_model_runner.py
@@ -19,6 +19,7 @@
                              MultiModalInputs)
 from vllm.sequence import (IntermediateTensors, SequenceData,
                            SequenceGroupMetadata)
+from vllm.transformers_utils.config import uses_mrope
 from vllm.utils import make_tensor_with_pad
 from vllm.worker.model_runner_base import (
     ModelRunnerBase, ModelRunnerInputBase, ModelRunnerInputBuilderBase,
@@ -439,10 +440,7 @@ def __init__(
     def model_is_mrope(self) -> bool:
         """Detect if the model has "mrope" rope_scaling type.
         mrope requires keep "rope_deltas" between prompt and decoding phases."""
-        rope_scaling = getattr(self.model_config.hf_config, "rope_scaling", {})
-        if rope_scaling is None:
-            return False
-        return rope_scaling.get("type", None) == "mrope"
+        return uses_mrope(self.model_config.hf_config)
 
     def load_model(self) -> None:
         self.model = get_model(model_config=self.model_config,
diff --git a/vllm/worker/model_runner.py b/vllm/worker/model_runner.py
index f88b1d84fbcd1..0f3c379cee8f0 100644
--- a/vllm/worker/model_runner.py
+++ b/vllm/worker/model_runner.py
@@ -47,6 +47,7 @@
     LRUCacheWorkerPromptAdapterManager)
 from vllm.sampling_params import SamplingParams
 from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
+from vllm.transformers_utils.config import uses_mrope
 from vllm.utils import (DeviceMemoryProfiler, PyObjectCache, async_tensor_h2d,
                         flatten_2d_lists, is_hip, is_pin_memory_available,
                         supports_dynamo)
@@ -1379,10 +1380,7 @@ def list_prompt_adapters(self) -> Set[int]:
     def model_is_mrope(self) -> bool:
         """Detect if the model has "mrope" rope_scaling type.
         mrope requires keep "rope_deltas" between prompt and decoding phases."""
-        rope_scaling = getattr(self.model_config.hf_config, "rope_scaling", {})
-        if rope_scaling is None:
-            return False
-        return rope_scaling.get("type", None) == "mrope"
+        return uses_mrope(self.model_config.hf_config)
 
     @torch.inference_mode()
     def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None: