diff --git a/lmdeploy/pytorch/kernels/pagedattention.py b/lmdeploy/pytorch/kernels/pagedattention.py
index 39fe92204b..99a92d769a 100644
--- a/lmdeploy/pytorch/kernels/pagedattention.py
+++ b/lmdeploy/pytorch/kernels/pagedattention.py
@@ -229,21 +229,27 @@ def _reduce_split_kernel(
     tl.store(Out + out_offs, acc)
 
 
-_NV_CAP = torch.cuda.get_device_capability()
-if _NV_CAP[0] >= 8:
+def _get_convert_pv(nv_capability):
+    """lazy load convert_pv."""
+    if nv_capability[0] >= 8:
+
+        @triton.jit
+        def convert_pv(p, v):
+            """convert pv."""
+            p = p.to(v.dtype)
+            return p, v
+    else:
 
-    @triton.jit
-    def _convert_pv(p, v):
-        """convert pv."""
-        p = p.to(v.dtype)
-        return p, v
-else:
+        @triton.jit
+        def convert_pv(p, v):
+            """convert pv."""
+            v = v.to(p.dtype)
+            return p, v
 
-    @triton.jit
-    def _convert_pv(p, v):
-        """convert pv."""
-        v = v.to(p.dtype)
-        return p, v
+    return convert_pv
+
+
+_convert_pv = None
 
 
 @triton.jit
@@ -408,6 +414,10 @@ def paged_attention_fwd(
         max_seqlen (int): The max input length.
         BLOCK (int): The kernel block size.
     """
+    global _convert_pv
+    if _convert_pv is None:
+        nv_cap = torch.cuda.get_device_capability()
+        _convert_pv = _get_convert_pv(nv_cap)
 
     def _kernel_meta():
         """kernel meta."""