diff --git a/sgl-kernel/python/sgl_kernel/elementwise.py b/sgl-kernel/python/sgl_kernel/elementwise.py
index 3b402366e..5684800be 100644
--- a/sgl-kernel/python/sgl_kernel/elementwise.py
+++ b/sgl-kernel/python/sgl_kernel/elementwise.py
@@ -369,7 +369,7 @@ def downcast_fp8(
     )
 
 
-def copy_to_gpu_no_ce(input: List[int], output: torch.Tensor):
+def copy_to_gpu_no_ce(input: torch.Tensor, output: torch.Tensor):
     torch.ops.sgl_kernel.copy_to_gpu_no_ce(input, output)
 
 
diff --git a/sgl-kernel/tests/test_copy.py b/sgl-kernel/tests/test_copy.py
new file mode 100644
index 000000000..70ed864c1
--- /dev/null
+++ b/sgl-kernel/tests/test_copy.py
@@ -0,0 +1,16 @@
+import pytest
+import sgl_kernel
+import torch
+from sgl_kernel.elementwise import copy_to_gpu_no_ce
+
+
+@pytest.mark.parametrize("size", [64, 72])
+def test_copy_to_gpu_no_ce(size):
+    tensor_cpu = torch.randint(0, 1000000, (size,), dtype=torch.int32, device="cpu")
+    tensor_gpu = torch.empty_like(tensor_cpu, device="cuda")
+    copy_to_gpu_no_ce(tensor_cpu, tensor_gpu)
+    assert torch.all(tensor_cpu.cuda() == tensor_gpu)
+
+
+if __name__ == "__main__":
+    pytest.main([__file__])