From b10d47e0e099b29aecc20740daaaa1afe3476c59 Mon Sep 17 00:00:00 2001
From: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Date: Tue, 23 Dec 2025 11:41:49 +0000
Subject: [PATCH] Add util function for checking nesting of rope parameters
 (#31146)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
---
 vllm/config/model.py                          |  6 ++----
 .../models/transformers/utils.py              |  4 ++--
 vllm/transformers_utils/config.py             | 21 +++++++++++++++++--
 3 files changed, 23 insertions(+), 8 deletions(-)

diff --git a/vllm/config/model.py b/vllm/config/model.py
index dd2b7b9d7a786..f2f39ac6af022 100644
--- a/vllm/config/model.py
+++ b/vllm/config/model.py
@@ -11,7 +11,6 @@ import torch
 from pydantic import ConfigDict, Field, field_validator, model_validator
 from pydantic.dataclasses import dataclass
 from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
-from transformers.configuration_utils import ALLOWED_LAYER_TYPES
 
 import vllm.envs as envs
 from vllm.attention.backends.registry import AttentionBackendEnum
@@ -29,6 +28,7 @@ from vllm.transformers_utils.config import (
     get_pooling_config,
     get_sentence_transformer_tokenizer_config,
     is_encoder_decoder,
+    is_rope_parameters_nested,
     try_get_dense_modules,
     try_get_generation_config,
     try_get_safetensors_metadata,
@@ -2125,9 +2125,7 @@ def _get_and_verify_max_len(
     # In Transformers v5 rope_parameters could be TypedDict or dict[str, TypedDict].
     # To simplify the verification, we convert it to dict[str, TypedDict].
     rope_parameters = getattr(hf_config, "rope_parameters", None)
-    if rope_parameters and not set(rope_parameters.keys()).issubset(
-        ALLOWED_LAYER_TYPES
-    ):
+    if rope_parameters and not is_rope_parameters_nested(rope_parameters):
         rope_parameters = {"": rope_parameters}
 
     # NOTE(woosuk): Gemma3's max_model_len (128K) is already scaled by RoPE
diff --git a/vllm/model_executor/models/transformers/utils.py b/vllm/model_executor/models/transformers/utils.py
index b807f45b5d52b..c7844381eb633 100644
--- a/vllm/model_executor/models/transformers/utils.py
+++ b/vllm/model_executor/models/transformers/utils.py
@@ -22,7 +22,6 @@ from typing import TYPE_CHECKING, Literal
 
 import torch
 from torch import nn
-from transformers.configuration_utils import ALLOWED_LAYER_TYPES
 
 from vllm.config.utils import getattr_iter
 from vllm.logger import init_logger
@@ -32,6 +31,7 @@ from vllm.model_executor.layers.linear import (
     ReplicatedLinear,
     RowParallelLinear,
 )
+from vllm.transformers_utils.config import is_rope_parameters_nested
 
 if TYPE_CHECKING:
     from vllm.config import VllmConfig
@@ -207,7 +207,7 @@ def can_enable_torch_compile(vllm_config: "VllmConfig") -> bool:
     rope_parameters: dict | None = getattr(text_config, "rope_parameters", None) or {}
     if rope_parameters:
         # Nest rope_parameters if not nested already to simplify logic
-        if not set(rope_parameters.keys()).issubset(ALLOWED_LAYER_TYPES):
+        if not is_rope_parameters_nested(rope_parameters):
             rope_parameters = {"": rope_parameters}
         return all(rp["rope_type"] != "dynamic" for rp in rope_parameters.values())
     return True
diff --git a/vllm/transformers_utils/config.py b/vllm/transformers_utils/config.py
index 887f936a2d8ae..4635277c83585 100644
--- a/vllm/transformers_utils/config.py
+++ b/vllm/transformers_utils/config.py
@@ -15,7 +15,6 @@ from huggingface_hub import (
 )
 from packaging.version import Version
 from transformers import GenerationConfig, PretrainedConfig
-from transformers.configuration_utils import ALLOWED_LAYER_TYPES
 from transformers.models.auto.image_processing_auto import get_image_processor_config
 from transformers.models.auto.modeling_auto import (
     MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
@@ -44,6 +43,16 @@ from .repo_utils import (
     with_retry,
 )
 
+try:
+    # Transformers v5
+    from transformers.configuration_utils import ALLOWED_ATTENTION_LAYER_TYPES
+except ImportError:
+    # Transformers v4
+    from transformers.configuration_utils import (
+        ALLOWED_LAYER_TYPES as ALLOWED_ATTENTION_LAYER_TYPES,
+    )
+
+
 if envs.VLLM_USE_MODELSCOPE:
     from modelscope import AutoConfig
 else:
@@ -104,6 +113,14 @@ _AUTO_CONFIG_KWARGS_OVERRIDES: dict[str, dict[str, Any]] = {
 }
 
 
+def is_rope_parameters_nested(rope_parameters: dict[str, Any]) -> bool:
+    """Check if rope_parameters is nested by layer types."""
+    # Cannot be nested if rope_parameters is empty
+    if not rope_parameters:
+        return False
+    return set(rope_parameters.keys()).issubset(ALLOWED_ATTENTION_LAYER_TYPES)
+
+
 class HFConfigParser(ConfigParserBase):
     def parse(
         self,
@@ -346,7 +363,7 @@ def patch_rope_parameters(config: PretrainedConfig) -> None:
         config.rope_parameters["original_max_position_embeddings"] = ompe
 
     # Handle nested rope_parameters in interleaved sliding attention models
-    if set(config.rope_parameters.keys()).issubset(ALLOWED_LAYER_TYPES):
+    if is_rope_parameters_nested(config.rope_parameters):
         for rope_parameters_layer_type in config.rope_parameters.values():
             patch_rope_parameters_dict(rope_parameters_layer_type)
     else: