Despite that this is a Pynq-Z2, is another option to leverage DPU-PYNQ in conjunction with TVM, assuming that nfmzl can port DPU-PYNQ to the Z2?
https://github.com/Xilinx/Vitis-AI/tree/master/tvm
You should also be able to consider use of the TVM runtime / compiler with no DPU. I would assume that in this context you would use the TVM main repository (see tvm.ai) to directly target the A53 for inference acceleration.
TVM solves the problem of deployment of models that originate from framework versions that are not supported directly in Vitis-AI, and should likewise solve the problem of TF compatibility with PYNQ.
My 2 cents…
–Quenton