diff --git a/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp b/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp
index 05acc275b46..3bd738a372d 100644
--- a/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp
+++ b/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp
@@ -730,6 +730,8 @@ bool can_use_mem_efficient_attention(sdp_params const& params, bool debug) {
 #ifdef USE_ROCM
   constexpr auto aotriton_mem_efficient_dtypes =
       c10::array_of<at::ScalarType>(at::kHalf, at::kFloat, at::kBFloat16);
+  constexpr auto ck_mem_efficient_dtypes =
+      c10::array_of<at::ScalarType>(at::kHalf, at::kBFloat16);
 #else
   constexpr auto greater_than_or_equal_sm80_mem_efficient_dtypes =
       c10::array_of<at::ScalarType>(at::kHalf, at::kFloat, at::kBFloat16);
@@ -789,6 +791,9 @@ bool can_use_mem_efficient_attention(sdp_params const& params, bool debug) {
       return false;
     }
   }
+  if(at::globalContext().getROCmFAPreferredBackend() == at::ROCmFABackend::Ck) {
+    return check_tensor_dtype(params, ck_mem_efficient_dtypes, debug);
+  }
   return check_tensor_dtype(params, aotriton_mem_efficient_dtypes, debug);
 #else
   auto dprop = at::cuda::getCurrentDeviceProperties();