diff --git a/tests/kernels/attention/test_deepgemm_attention.py b/tests/kernels/attention/test_deepgemm_attention.py
index f4b4fac840151..74a5d81179623 100644
--- a/tests/kernels/attention/test_deepgemm_attention.py
+++ b/tests/kernels/attention/test_deepgemm_attention.py
@@ -6,7 +6,7 @@ import pytest
 import torch
 
 from vllm.platforms import current_platform
-from vllm.utils import cdiv, has_deep_gemm
+from vllm.utils import cdiv
 from vllm.utils.deep_gemm import (
     _ceil_to_ue8m0,
     calc_diff,
@@ -15,6 +15,7 @@ from vllm.utils.deep_gemm import (
     get_num_sms,
     get_paged_mqa_logits_metadata,
 )
+from vllm.utils.import_utils import has_deep_gemm
 
 
 def kv_cache_cast_to_fp8(x: torch.Tensor) -> torch.Tensor:
diff --git a/tests/kernels/moe/modular_kernel_tools/common.py b/tests/kernels/moe/modular_kernel_tools/common.py
index 94a305a063c3a..c517e5c026b46 100644
--- a/tests/kernels/moe/modular_kernel_tools/common.py
+++ b/tests/kernels/moe/modular_kernel_tools/common.py
@@ -23,7 +23,7 @@ from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEQuantConfig,
 )
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
-from vllm.utils import has_deep_ep, has_deep_gemm, has_pplx
+from vllm.utils.import_utils import has_deep_ep, has_deep_gemm, has_pplx
 
 from .mk_objects import (
     TestMoEQuantConfig,
diff --git a/tests/kernels/moe/modular_kernel_tools/mk_objects.py b/tests/kernels/moe/modular_kernel_tools/mk_objects.py
index aa41f89cae7dc..21eeffb1c7264 100644
--- a/tests/kernels/moe/modular_kernel_tools/mk_objects.py
+++ b/tests/kernels/moe/modular_kernel_tools/mk_objects.py
@@ -35,9 +35,9 @@ from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     cutlass_fp8_supported,
 )
 from vllm.platforms import current_platform
-from vllm.utils import has_deep_ep, has_deep_gemm, has_pplx
 from vllm.utils.deep_gemm import is_deep_gemm_supported
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
+from vllm.utils.import_utils import has_deep_ep, has_deep_gemm, has_pplx
 
 
 @dataclass
diff --git a/tests/kernels/moe/parallel_utils.py b/tests/kernels/moe/parallel_utils.py
index 43fbd05775a28..90728c1e30a46 100644
--- a/tests/kernels/moe/parallel_utils.py
+++ b/tests/kernels/moe/parallel_utils.py
@@ -15,7 +15,7 @@ from torch.distributed import ProcessGroup
 from torch.multiprocessing import spawn  # pyright: ignore[reportPrivateImportUsage]
 from typing_extensions import ParamSpec
 
-from vllm.utils import has_deep_ep
+from vllm.utils.import_utils import has_deep_ep
 from vllm.utils.network_utils import get_open_port
 
 if has_deep_ep():
diff --git a/tests/kernels/moe/test_block_fp8.py b/tests/kernels/moe/test_block_fp8.py
index 11b1e2ff3c27e..60f9f14b7f6f1 100644
--- a/tests/kernels/moe/test_block_fp8.py
+++ b/tests/kernels/moe/test_block_fp8.py
@@ -21,11 +21,11 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
     modular_triton_fused_moe,
 )
 from vllm.platforms import current_platform
-from vllm.utils import has_deep_gemm
 from vllm.utils.deep_gemm import (
     get_mk_alignment_for_contiguous_layout,
     is_deep_gemm_e8m0_used,
 )
+from vllm.utils.import_utils import has_deep_gemm
 
 dg_available = has_deep_gemm()
 
diff --git a/tests/kernels/moe/test_deepep_deepgemm_moe.py b/tests/kernels/moe/test_deepep_deepgemm_moe.py
index 65cd3e110a0fa..d46f453488a98 100644
--- a/tests/kernels/moe/test_deepep_deepgemm_moe.py
+++ b/tests/kernels/moe/test_deepep_deepgemm_moe.py
@@ -21,8 +21,8 @@ from vllm.model_executor.layers.fused_moe.config import (
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
 from vllm.model_executor.layers.fused_moe.modular_kernel import FusedMoEModularKernel
 from vllm.platforms import current_platform
-from vllm.utils import has_deep_ep, has_deep_gemm
 from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used, is_deep_gemm_supported
+from vllm.utils.import_utils import has_deep_ep, has_deep_gemm
 
 from ...utils import multi_gpu_test
 from .parallel_utils import ProcessGroupInfo, parallel_launch
diff --git a/tests/kernels/moe/test_deepep_moe.py b/tests/kernels/moe/test_deepep_moe.py
index 527c20fe6f80b..b49319a7e6f54 100644
--- a/tests/kernels/moe/test_deepep_moe.py
+++ b/tests/kernels/moe/test_deepep_moe.py
@@ -21,7 +21,7 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8,
 )
 from vllm.platforms import current_platform
-from vllm.utils import has_deep_ep
+from vllm.utils.import_utils import has_deep_ep
 
 from ...utils import multi_gpu_test
 from .parallel_utils import ProcessGroupInfo, parallel_launch
diff --git a/tests/kernels/moe/test_gpt_oss_triton_kernels.py b/tests/kernels/moe/test_gpt_oss_triton_kernels.py
index f78596d220bfa..b8f213d33c963 100644
--- a/tests/kernels/moe/test_gpt_oss_triton_kernels.py
+++ b/tests/kernels/moe/test_gpt_oss_triton_kernels.py
@@ -6,7 +6,7 @@ import pytest
 import torch
 import torch.nn.functional as F
 
-from vllm.utils import has_triton_kernels
+from vllm.utils.import_utils import has_triton_kernels
 
 if not has_triton_kernels():
     pytest.skip(
diff --git a/tests/kernels/moe/test_modular_kernel_combinations.py b/tests/kernels/moe/test_modular_kernel_combinations.py
index a7beb313011ae..a46b0053e75a3 100644
--- a/tests/kernels/moe/test_modular_kernel_combinations.py
+++ b/tests/kernels/moe/test_modular_kernel_combinations.py
@@ -13,8 +13,8 @@ import torch
 import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from vllm.config import VllmConfig, set_current_vllm_config
 from vllm.platforms import current_platform
-from vllm.utils import has_deep_ep, has_deep_gemm, has_pplx
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
+from vllm.utils.import_utils import has_deep_ep, has_deep_gemm, has_pplx
 from vllm.utils.torch_utils import cuda_device_count_stateless
 
 from .modular_kernel_tools.common import (
diff --git a/tests/kernels/quantization/test_block_fp8.py b/tests/kernels/quantization/test_block_fp8.py
index a6dfb5428c52e..55f092e7ea694 100644
--- a/tests/kernels/quantization/test_block_fp8.py
+++ b/tests/kernels/quantization/test_block_fp8.py
@@ -18,12 +18,12 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     w8a8_triton_block_scaled_mm,
 )
 from vllm.platforms import current_platform
-from vllm.utils import has_deep_gemm
 from vllm.utils.deep_gemm import (
     fp8_gemm_nt,
     get_col_major_tma_aligned_tensor,
     per_block_cast_to_fp8,
 )
+from vllm.utils.import_utils import has_deep_gemm
 
 if current_platform.get_device_capability() < (9, 0):
     pytest.skip("FP8 Triton requires CUDA 9.0 or higher", allow_module_level=True)
diff --git a/tests/models/quantization/test_fp8.py b/tests/models/quantization/test_fp8.py
index bac613913e91f..2a6f34a9c4823 100644
--- a/tests/models/quantization/test_fp8.py
+++ b/tests/models/quantization/test_fp8.py
@@ -12,7 +12,6 @@ from tests.quantization.utils import is_quant_method_supported
 from vllm.attention.utils.fa_utils import flash_attn_supports_fp8
 from vllm.platforms import current_platform
 from vllm.utils import STR_BACKEND_ENV_VAR
-
 from ..utils import check_logprobs_close
 
 
diff --git a/vllm/distributed/device_communicators/all2all.py b/vllm/distributed/device_communicators/all2all.py
index fae48cbe33744..10db1c2c4e670 100644
--- a/vllm/distributed/device_communicators/all2all.py
+++ b/vllm/distributed/device_communicators/all2all.py
@@ -9,8 +9,8 @@ import vllm.envs as envs
 from vllm.distributed import get_dp_group, get_ep_group
 from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
-from vllm.utils import has_deep_ep, has_pplx
 from vllm.utils.flashinfer import has_flashinfer_all2all
+from vllm.utils.import_utils import has_deep_ep, has_pplx
 
 from .base_device_communicator import All2AllManagerBase, Cache
 
diff --git a/vllm/model_executor/layers/fused_moe/config.py b/vllm/model_executor/layers/fused_moe/config.py
index 9e84f0d00b082..6d3477cd19916 100644
--- a/vllm/model_executor/layers/fused_moe/config.py
+++ b/vllm/model_executor/layers/fused_moe/config.py
@@ -14,8 +14,9 @@ from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import (
     OCP_MX_Scheme,
 )
 from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
-from vllm.utils import cdiv, has_triton_kernels
+from vllm.utils import cdiv
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
+from vllm.utils.import_utils import has_triton_kernels
 
 logger = init_logger(__name__)
 
diff --git a/vllm/model_executor/layers/fused_moe/deep_gemm_moe.py b/vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
index 69a815a4e3a96..484b8aa9d107c 100644
--- a/vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
+++ b/vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
@@ -26,12 +26,12 @@ from vllm.model_executor.layers.fused_moe.utils import _resize_cache
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8,
 )
-from vllm.utils import has_deep_gemm
 from vllm.utils.deep_gemm import (
     get_mk_alignment_for_contiguous_layout,
     m_grouped_fp8_gemm_nt_contiguous,
 )
 from vllm.utils.func_utils import run_once
+from vllm.utils.import_utils import has_deep_gemm
 
 logger = init_logger(__name__)
 
diff --git a/vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py b/vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py
index 01fa9b99379b6..badedfc54c382 100644
--- a/vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py
+++ b/vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py
@@ -13,7 +13,7 @@ from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
     TopKWeightAndReduceNoOP,
 )
 from vllm.triton_utils import tl, triton
-from vllm.utils import has_triton_kernels
+from vllm.utils.import_utils import has_triton_kernels
 
 logger = init_logger(__name__)
 
diff --git a/vllm/model_executor/layers/fused_moe/layer.py b/vllm/model_executor/layers/fused_moe/layer.py
index feff92a162eee..6a61df739d140 100644
--- a/vllm/model_executor/layers/fused_moe/layer.py
+++ b/vllm/model_executor/layers/fused_moe/layer.py
@@ -55,8 +55,9 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
 from vllm.platforms.interface import CpuArchEnum
-from vllm.utils import cdiv, has_deep_ep, has_pplx, round_up
+from vllm.utils import cdiv, round_up
 from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
+from vllm.utils.import_utils import has_deep_ep, has_pplx
 from vllm.utils.torch_utils import current_stream, direct_register_custom_op
 from vllm.v1.worker.ubatching import dbo_current_ubatch_id
 
diff --git a/vllm/model_executor/layers/quantization/fp8.py b/vllm/model_executor/layers/quantization/fp8.py
index 447b31b92d8fa..e5681cb856258 100644
--- a/vllm/model_executor/layers/quantization/fp8.py
+++ b/vllm/model_executor/layers/quantization/fp8.py
@@ -93,7 +93,6 @@ from vllm.model_executor.parameter import (
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
 from vllm.scalar_type import scalar_types
-from vllm.utils import has_deep_gemm
 from vllm.utils.deep_gemm import (
     fp8_gemm_nt,
     get_col_major_tma_aligned_tensor,
@@ -102,6 +101,7 @@ from vllm.utils.deep_gemm import (
     should_use_deepgemm_for_fp8_linear,
 )
 from vllm.utils.flashinfer import has_flashinfer_moe
+from vllm.utils.import_utils import has_deep_gemm
 
 if TYPE_CHECKING:
     from vllm.model_executor.models.utils import WeightsMapper
diff --git a/vllm/model_executor/layers/quantization/mxfp4.py b/vllm/model_executor/layers/quantization/mxfp4.py
index cf983fcf43c99..73f911514968f 100644
--- a/vllm/model_executor/layers/quantization/mxfp4.py
+++ b/vllm/model_executor/layers/quantization/mxfp4.py
@@ -48,11 +48,9 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import is_layer_s
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
 from vllm.scalar_type import scalar_types
-from vllm.utils import (
-    has_triton_kernels,
-    round_up,
-)
+from vllm.utils import round_up
 from vllm.utils.flashinfer import has_flashinfer
+from vllm.utils.import_utils import has_triton_kernels
 from vllm.utils.torch_utils import is_torch_equal_or_newer
 
 logger = init_logger(__name__)
diff --git a/vllm/utils/__init__.py b/vllm/utils/__init__.py
index 6704668c95265..230a6cfc52e30 100644
--- a/vllm/utils/__init__.py
+++ b/vllm/utils/__init__.py
@@ -5,7 +5,6 @@ import contextlib
 import datetime
 import enum
 import getpass
-import importlib
 import inspect
 import json
 import multiprocessing
@@ -1062,46 +1061,6 @@ def check_use_alibi(model_config: ModelConfig) -> bool:
     )
 
 
-@cache
-def _has_module(module_name: str) -> bool:
-    """Return True if *module_name* can be found in the current environment.
-
-    The result is cached so that subsequent queries for the same module incur
-    no additional overhead.
-    """
-    return importlib.util.find_spec(module_name) is not None
-
-
-def has_pplx() -> bool:
-    """Whether the optional `pplx_kernels` package is available."""
-
-    return _has_module("pplx_kernels")
-
-
-def has_deep_ep() -> bool:
-    """Whether the optional `deep_ep` package is available."""
-
-    return _has_module("deep_ep")
-
-
-def has_deep_gemm() -> bool:
-    """Whether the optional `deep_gemm` package is available."""
-
-    return _has_module("deep_gemm")
-
-
-def has_triton_kernels() -> bool:
-    """Whether the optional `triton_kernels` package is available."""
-
-    return _has_module("triton_kernels")
-
-
-def has_tilelang() -> bool:
-    """Whether the optional `tilelang` package is available."""
-
-    return _has_module("tilelang")
-
-
 def set_process_title(
     name: str, suffix: str = "", prefix: str = envs.VLLM_PROCESS_NAME_PREFIX
 ) -> None:
diff --git a/vllm/utils/deep_gemm.py b/vllm/utils/deep_gemm.py
index fce4111f19bb2..2e8cd302b0f55 100644
--- a/vllm/utils/deep_gemm.py
+++ b/vllm/utils/deep_gemm.py
@@ -16,7 +16,8 @@ import torch
 import vllm.envs as envs
 from vllm.logger import logger
 from vllm.platforms import current_platform
-from vllm.utils import cdiv, has_deep_gemm
+from vllm.utils import cdiv
+from vllm.utils.import_utils import has_deep_gemm
 
 
 @functools.cache
diff --git a/vllm/utils/import_utils.py b/vllm/utils/import_utils.py
index fdc3d356a7ebe..65f588b52e5e6 100644
--- a/vllm/utils/import_utils.py
+++ b/vllm/utils/import_utils.py
@@ -324,3 +324,39 @@ class LazyLoader(ModuleType):
         if self._module is None:
             self._module = self._load()
         return dir(self._module)
+
+
+# Optional dependency detection utilities
+@cache
+def _has_module(module_name: str) -> bool:
+    """Return True if *module_name* can be found in the current environment.
+
+    The result is cached so that subsequent queries for the same module incur
+    no additional overhead.
+    """
+    return importlib.util.find_spec(module_name) is not None
+
+
+def has_pplx() -> bool:
+    """Whether the optional `pplx_kernels` package is available."""
+    return _has_module("pplx_kernels")
+
+
+def has_deep_ep() -> bool:
+    """Whether the optional `deep_ep` package is available."""
+    return _has_module("deep_ep")
+
+
+def has_deep_gemm() -> bool:
+    """Whether the optional `deep_gemm` package is available."""
+    return _has_module("deep_gemm")
+
+
+def has_triton_kernels() -> bool:
+    """Whether the optional `triton_kernels` package is available."""
+    return _has_module("triton_kernels")
+
+
+def has_tilelang() -> bool:
+    """Whether the optional `tilelang` package is available."""
+    return _has_module("tilelang")
diff --git a/vllm/v1/worker/gpu_ubatch_wrapper.py b/vllm/v1/worker/gpu_ubatch_wrapper.py
index 3e6fd86e95d88..9de123263755b 100644
--- a/vllm/v1/worker/gpu_ubatch_wrapper.py
+++ b/vllm/v1/worker/gpu_ubatch_wrapper.py
@@ -22,7 +22,7 @@ from vllm.forward_context import (
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
-from vllm.utils import has_deep_gemm
+from vllm.utils.import_utils import has_deep_gemm
 from vllm.v1.worker.ubatching import UBatchContext, make_ubatch_contexts
 
 logger = init_logger(__name__)