diff --git a/fastdeploy/config.py b/fastdeploy/config.py
index c4a25a1504e..3f4d326e154 100644
--- a/fastdeploy/config.py
+++ b/fastdeploy/config.py
@@ -1807,6 +1807,15 @@ def check(self):
                 int(envs.FD_DISABLED_RECOVER) == 0
             ), "FD_DISABLED_RECOVER is not supported while ENABLE_V1_KVCACHE_SCHEDULER is turned on."
 
+        if self.eplb_config is not None and self.eplb_config.enable_eplb:
+            try:
+                import cuda  # noqa
+            except ImportError:
+                raise ImportError(
+                    "cuda-python not installed. Install the version matching your CUDA toolkit:\n"
+                    "  CUDA 12.x → pip install cuda-python==12.*\n"
+                )
+
     def print(self):
         """
         print all config
diff --git a/fastdeploy/eplb/async_expert_loader.py b/fastdeploy/eplb/async_expert_loader.py
index 6e0355276b8..9ab0bfae08a 100644
--- a/fastdeploy/eplb/async_expert_loader.py
+++ b/fastdeploy/eplb/async_expert_loader.py
@@ -22,7 +22,11 @@
 
 import numpy as np
 import paddle
-from cuda import cudart
+
+try:
+    from cuda import cudart
+except ImportError:
+    cudart = None
 
 from fastdeploy.config import EPLBConfig
 
@@ -90,6 +94,12 @@ def create_mmap(model_name: List, ep_rank: int, ep_size: int, shm_uuid: str, epl
         shm_ptr = ctypes.cast(shm_ptr, ctypes.POINTER(ctypes.c_int8))
         addr = ctypes.addressof(shm_ptr.contents)
 
+        if cudart is None:
+            raise ImportError(
+                "cuda-python not installed. Install the version matching your CUDA toolkit:\n"
+                "  CUDA 12.x → pip install cuda-python==12.*\n"
+            )
+
         # Register memory with CUDA
         (ret,) = cudart.cudaHostRegister(addr, shm_size, 0)
         if ret != cudart.cudaError_t.cudaSuccess:
diff --git a/requirements.txt b/requirements.txt
index 2e016837dd4..e5f614d8a87 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -46,4 +46,3 @@ msgspec
 einops
 setproctitle
 aistudio_sdk
-cuda-python==12.8