diff --git a/colossalai/shardformer/layer/attn.py b/colossalai/shardformer/layer/attn.py
index 2f8e4d677..5f0e9261c 100644
--- a/colossalai/shardformer/layer/attn.py
+++ b/colossalai/shardformer/layer/attn.py
@@ -121,7 +121,8 @@ class ColoAttention:
             )
 
         if size >= MEMORY_BOUND:
-            ColoAttention._flash_kernel_dispatch = ColoAttention._flash_kernel_dispatch.load()
+            if isinstance(ColoAttention._flash_kernel_dispatch, KernelLoader):
+                ColoAttention._flash_kernel_dispatch = ColoAttention._flash_kernel_dispatch.load()
         # lazy load
         if isinstance(ColoAttention._kernel_dispatch_map[dtype][mask_type], KernelLoader):
             ColoAttention._kernel_dispatch_map[dtype][mask_type] = ColoAttention._kernel_dispatch_map[dtype][