fix(mgb/cuda): fix conv error when the input tensor is too large

GitOrigin-RevId: 1b1d693795

dnn/src/cuda/conv_bias/cudnn_conv.cpp

@@ -40,6 +40,12 @@ bool ConvBiasForwardImpl::AlgoCUDNNConv::is_available(const SizeArgs& args) cons
         return false;
     }
 
+    // In conv_args.init_conv_desc will call cudnnSetTensor4dDescriptorEx(),which can't
+    // been supported when total_nr_elems() > 2 ^ 31
+    if (args.src_layout->total_nr_elems() > INT_MAX ||
+        args.dst_layout->total_nr_elems() > INT_MAX) {
+        return false;
+    }
     auto dst_layout = *args.dst_layout;
     if (dst_layout.dtype.enumv() != args.bias_layout->dtype.enumv()) {
         dst_layout.dtype = DType();