Add some debug prints

apex/contrib/csrc/bottleneck/bottleneck.cpp

@@ -2129,6 +2129,7 @@ at::Tensor bottleneck_backward_grad_out2(bool explicit_nhwc, int stride_1X1, std
             dw3,
             dy3,
             CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR);
+  DEBUG_MSG("[DEBUG] new wgrad3 : " << wgrad3.to(at::kFloat).sum().item<float>());
 
   // dgrad
   auto grad_out2 = at::empty(backward_state.outdim2, inputs[0].type(), output_format);
@@ -2262,6 +2263,7 @@ at::Tensor bottleneck_backward_wgrad2(bool explicit_nhwc, int stride_1X1, std::v
             dw2,
             dy2,
             CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR);
+  DEBUG_MSG("[DEBUG] new wgrad2 : " << wgrad2.to(at::kFloat).sum().item<float>());
 
   return wgrad2;
 }
@@ -2460,8 +2462,6 @@ void bottleneck_backward_rest(bool explicit_nhwc, int stride_1X1, std::vector<at
 
   DEBUG_MSG("[DEBUG] new dx : " << grad_x.to(at::kFloat).sum().item<float>());
   DEBUG_MSG("[DEBUG] new wgrad1 : " << wgrad1.to(at::kFloat).sum().item<float>());
-  DEBUG_MSG("[DEBUG] new wgrad2 : " << wgrad2.to(at::kFloat).sum().item<float>());
-  DEBUG_MSG("[DEBUG] new wgrad3 : " << wgrad3.to(at::kFloat).sum().item<float>());
 
   if (stride_1X1 != 1 || backward_state.filterdimA3[0] != backward_state.dimA[1]) {
     DEBUG_MSG("[DEBUG] new wgrad4 : " << wgrad4.to(at::kFloat).sum().item<float>());