diff --git a/vllm/model_executor/model_loader/utils.py b/vllm/model_executor/model_loader/utils.py
index e74434e9d12cb..1db6337f4c9f9 100644
--- a/vllm/model_executor/model_loader/utils.py
+++ b/vllm/model_executor/model_loader/utils.py
@@ -19,12 +19,7 @@ from vllm.model_executor.layers.quantization.base_config import (
     QuantizationConfig,
     QuantizeMethodBase,
 )
-from vllm.model_executor.models.adapters import (
-    as_embedding_model,
-    as_reward_model,
-    as_seq_cls_model,
-    try_create_mm_pooling_model_cls,
-)
+
 from vllm.model_executor.models.interfaces import SupportsQuant, supports_multimodal
 from vllm.utils.platform_utils import is_pin_memory_available
 
@@ -172,6 +167,12 @@ _MODEL_ARCH_BY_HASH = dict[int, tuple[type[nn.Module], str]]()
 
 
 def _get_model_architecture(model_config: ModelConfig) -> tuple[type[nn.Module], str]:
+    from vllm.model_executor.models.adapters import (
+        as_embedding_model,
+        as_reward_model,
+        as_seq_cls_model,
+        try_create_mm_pooling_model_cls,
+    )
     architectures = getattr(model_config.hf_config, "architectures", [])
 
     model_cls, arch = model_config.registry.resolve_model_cls(