In general, hardware vendors have their own tool chains, which are neither open source nor complete. For example, the vendor toolchain only supports TensorFlow model as input, and only supports part of TensorFlow ops(typical ops). We want to use TVM to support more front-end model inputs and generate code automatically. Problem: Taking the TensorFlow model as an example, most part of network are off-loaded to the vendor’s toolchain, and the remaining ops are automatically generate code by the TVM. I can use Graph Partition to divide the network. (The key issue） Because the vendor’s toolchain only supports TensorFlow model as input, the off-load part needs to convert the Relay IR to Tensorflow model. Any good suggestions? Are there any examples? Thank you ！
Hi @zfhn ,
I have landed up in a similar situation. Were you able to figure out or are there any inputs from the community?