fix:add no support for cuda_arch<700 (#47509)

paddle/fluid/operators/math/bert_encoder_functor.cu

@@ -836,7 +836,13 @@ inline void MatMulWithHeadQK(const phi::GPUContext &context,
             FINAL_MASK);
       } else {
         if (bias_is_mask) {
-#ifndef __HIPCC__
+#if defined(__HIPCC__) || (defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 700)
+          PADDLE_ENFORCE_EQ(bias_is_mask,
+                            false,
+                            platform::errors::InvalidArgument(
+                                "QK_bias is mask can't be supported on rocm or "
+                                "cuda_arch<700"));
+#else
           constexpr int ITEMS_PER_THREAD = 1;
           bool is_half2 = true;
 
@@ -853,11 +859,6 @@ inline void MatMulWithHeadQK(const phi::GPUContext &context,
                                            batch_size,
                                            head_num,
                                            seq_len);
-#else
-          PADDLE_ENFORCE_EQ(bias_is_mask,
-                            false,
-                            platform::errors::InvalidArgument(
-                                "rocm can't support that QK_bias is mask"));
 #endif
         } else {
           SoftmaxKernelWithEltadd2<__half2><<<grid, block, 0, stream>>>(