Changing tensor addresses for stream captured IExecutionContext::enqueueV3

After performing stream capture of an enqueueV3, cudaGraphLaunch seems to only read from the addresses specified before the capture. This differs from the behavior of directly calling enqueueV3, in which case the tensors most recently set via setInputTensorAddress and setTensorAddress are read from.

Is there any way of updating cudaGraphInst to read from new sets of addresses after instantiation? Or some other solution to this problem?