add moe debug flag

pytorch/cuda/moe_cuda_kernel.cu

@@ -166,8 +166,10 @@ std::vector<torch::Tensor> moe_cuda_backward(
     const auto num_expert = weight.size(0);
     const auto out_feat = weight.size(1);
     const auto in_feat = weight.size(2);
+    
+#ifdef MOE_DEBUG
     printf("[backward] b=%ld, expert=%ld, in_feat (d_model)=%ld, out_feat (d_ffn)=%ld\n", batch_size, num_expert, in_feat, out_feat);
-
+#endif
     auto grad_input = grad_output.new_zeros({batch_size, in_feat});  // batch_size x in_feat
     auto grad_weight = grad_output.new_zeros({num_expert, out_feat, in_feat}); // num_expert x out_feat x in_feat