fixed a bug due to shape_change

Author: cehongwang

core/runtime/execute_engine.cpp

@@ -141,7 +141,7 @@ void setup_input_tensors(
       at::Tensor contig_input = inputs[i].view(shape).contiguous();
       formatted_inputs.emplace_back(std::move(contig_input));
 
-      if (need_cudagraphs_record) {
+      if (need_cudagraphs_record or compiled_engine->allocated_outputs.size() == 0) {
         // Create a new persistent input buffer
         compiled_engine->input_buffers[i] = std::move(formatted_inputs.back().clone());
       }

py/torch_tensorrt/dynamo/runtime/_PythonTorchTensorRTModule.py

@@ -417,7 +417,9 @@ def setup_input_tensors(
                 inputs_cpu = contiguous_inputs[i].cpu().to(torch.int64).numpy().copy()
                 self.context.set_tensor_address(input_name, inputs_cpu.ctypes.data)
             else:
-                if need_cudagraphs_record:
+                if (
+                    need_cudagraphs_record or self.output_tensors is None
+                ):  # First time execution:
                     self.context.set_input_shape(
                         input_name, tuple(contiguous_inputs[i].shape)
                     )