diff --git a/vllm/entrypoints/openai/speech_to_text.py b/vllm/entrypoints/openai/speech_to_text.py
index e7589a3804c3c..09b346dcef6b8 100644
--- a/vllm/entrypoints/openai/speech_to_text.py
+++ b/vllm/entrypoints/openai/speech_to_text.py
@@ -24,7 +24,6 @@ from vllm.entrypoints.openai.serving_engine import (OpenAIServing,
 from vllm.entrypoints.openai.serving_models import OpenAIServingModels
 from vllm.inputs.data import PromptType
 from vllm.logger import init_logger
-from vllm.model_executor.model_loader import get_model_cls
 from vllm.model_executor.models import SupportsTranscription
 from vllm.outputs import RequestOutput
 from vllm.utils import PlaceholderModule
@@ -78,6 +77,7 @@ class OpenAISpeechToText(OpenAIServing):
 
     @cached_property
     def model_cls(self) -> type[SupportsTranscription]:
+        from vllm.model_executor.model_loader import get_model_cls
         model_cls = get_model_cls(self.model_config)
         return cast(type[SupportsTranscription], model_cls)
 
diff --git a/vllm/lora/utils.py b/vllm/lora/utils.py
index ee196e3f689a2..6b3291e9c92fa 100644
--- a/vllm/lora/utils.py
+++ b/vllm/lora/utils.py
@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
-from typing import Optional, Union
+from typing import TYPE_CHECKING, Optional, Union
 
 import huggingface_hub
 import regex as re
@@ -31,10 +31,14 @@ from vllm.lora.layers import (BaseLayerWithLoRA, ColumnParallelLinearWithLoRA,
                               RowParallelLinearWithLoRA,
                               VocabParallelEmbeddingWithLoRA)
 from vllm.model_executor.layers.linear import LinearBase
+
 # yapf: enable
-from vllm.model_executor.layers.logits_processor import LogitsProcessor
-from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
-from vllm.model_executor.models.utils import WeightsMapper
+
+if TYPE_CHECKING:
+    from vllm.model_executor.layers.logits_processor import LogitsProcessor
+    from vllm.model_executor.layers.vocab_parallel_embedding import (
+        ParallelLMHead)
+    from vllm.model_executor.models.utils import WeightsMapper
 
 logger = init_logger(__name__)
 
@@ -75,8 +79,8 @@ def from_layer(layer: nn.Module,
 
 
 def from_layer_logits_processor(
-    layer: LogitsProcessor,
-    lm_head: ParallelLMHead,
+    layer: "LogitsProcessor",
+    lm_head: "ParallelLMHead",
     max_loras: int,
     lora_config: LoRAConfig,
     model_config: Optional[PretrainedConfig] = None,
@@ -98,8 +102,8 @@ def replace_submodule(model: nn.Module, module_name: str,
 
 
 def parse_fine_tuned_lora_name(
-        name: str,
-        weights_mapper: Optional[WeightsMapper] = None
+    name: str,
+    weights_mapper: Optional["WeightsMapper"] = None
 ) -> tuple[str, bool, bool]:
     """Parse the name of lora weights.