diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake
index 5dacd3916c4ee911290d7a4664f6b60973adb9a2..710931d937e4277a1dbac73b66b5307d52a9cbc9 100644
--- a/cmake/cuda.cmake
+++ b/cmake/cuda.cmake
@@ -171,7 +171,7 @@ function(select_nvcc_arch_flags out_variable out_arch_bin)
     else()
       if(${CMAKE_CUDA_COMPILER_VERSION} LESS 11.1) # CUDA 11.0
         set(cuda_arch_bin "80")
-      elseif(${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0) # CUDA 11.1+
+      else()
         set(cuda_arch_bin "80 86")
       endif()
     endif()
diff --git a/paddle/phi/kernels/CMakeLists.txt b/paddle/phi/kernels/CMakeLists.txt
index 271fbca6c3f57e4cd1b3f8d896e026620dd41bd4..409c746938f0097058be49e876d5180d58136a88 100644
--- a/paddle/phi/kernels/CMakeLists.txt
+++ b/paddle/phi/kernels/CMakeLists.txt
@@ -129,7 +129,16 @@ if(WITH_CUTLASS)
     COMMAND
       ${PYTHON_EXECUTABLE}
       ${PADDLE_SOURCE_DIR}/paddle/phi/kernels/fusion/cutlass/memory_efficient_attention/generate_kernels.py
-      --cuda_arch "${NVCC_ARCH_BIN}")
+      --cuda_arch "${NVCC_ARCH_BIN}"
+    RESULT_VARIABLE memory_efficient_attention_gen_res)
+
+  if(NOT memory_efficient_attention_gen_res EQUAL 0)
+    message(
+      FATAL_ERROR
+        "The memory efficient attention kernel generation errors with NVCC_ARCH_BIN=${NVCC_ARCH_BIN}"
+    )
+  endif()
+
   file(GLOB cutlass_cu "fusion/cutlass/conv2d/generated/*.cu"
        "fusion/cutlass/conv2d/*.cu" "fusion/cutlass/*.cu"
        "fusion/cutlass/memory_efficient_attention/autogen/impl/*.cu")
diff --git a/paddle/phi/kernels/fusion/cutlass/memory_efficient_attention/generate_kernels.py b/paddle/phi/kernels/fusion/cutlass/memory_efficient_attention/generate_kernels.py
index 2baa7b07d98492b7d918a50ef61cce1ad1de7007..a512f97e48e36b95398fbec85ced3ca41cc3d935 100644
--- a/paddle/phi/kernels/fusion/cutlass/memory_efficient_attention/generate_kernels.py
+++ b/paddle/phi/kernels/fusion/cutlass/memory_efficient_attention/generate_kernels.py
@@ -44,7 +44,7 @@ def find_arch_range(min_arch, max_arch):
     assert min_arch <= max_arch
     n = len(DEFAULT_ARCH)
 
-    start_idx = 0
+    start_idx = n - 1
     for i in range(n - 1):
         if DEFAULT_ARCH[i] <= min_arch and min_arch < DEFAULT_ARCH[i + 1]:
             start_idx = i
@@ -54,6 +54,7 @@ def find_arch_range(min_arch, max_arch):
     for i in range(n - 1):
         if DEFAULT_ARCH[i] <= max_arch and max_arch < DEFAULT_ARCH[i + 1]:
             end_idx = i + 1
+
     return DEFAULT_ARCH[start_idx:end_idx]
 
 
diff --git a/python/paddle/incubate/nn/memory_efficient_attention.py b/python/paddle/incubate/nn/memory_efficient_attention.py
index 76784254e41e4061833eafcf23308080cbd5ecf8..049f00a70cf9d748b6689cce19a990284f5bf371 100644
--- a/python/paddle/incubate/nn/memory_efficient_attention.py
+++ b/python/paddle/incubate/nn/memory_efficient_attention.py
@@ -134,7 +134,7 @@ def memory_efficient_attention(
             "causal_diagonal": causal_diagonal,
             "seqlen_k": seqlen_k,
         },
-        args={
+        attrs={
             "max_seqlen_q": max_seqlen_q,
             "max_seqlen_k": max_seqlen_k,
             "causal": causal,