fix cudnn workspace size problem during inference. (#26021)

test=develop

paddle/fluid/operators/fused/conv_fusion_op.cu

@@ -216,6 +216,12 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
               perf_results.get()));
       algo = (perf_results.get())[best_algo_idx].algo;
       VLOG(3) << "cuDNN forward algo " << algo;
+      PADDLE_ENFORCE_CUDA_SUCCESS(
+          platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
+              handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc,
+              cudnn_output_desc, algo, &workspace_size_in_bytes));
+      if (workspace_size_in_bytes > workspace_size_limit)
+        workspace_size_limit = workspace_size_in_bytes;
     } else {
       std::function<cudnnConvolutionFwdAlgo_t()> search_func =
           [&]() -> cudnnConvolutionFwdAlgo_t {