i wanted to use TVM in production env as inference serving service, because i want to leverage the optimazition provided by TVM, such as OP fusion and autotvm.
i tried searching information about model serving using TVM, but i think there were really few people doing this, and also i searched related topics about TVM Serving in community, found:
topic 1:
accord to it’s tutorial, i think the author modified TF serving to accommordating TVM models
topic 2:
which actually done a good job on embeding tvm op in TF quickly, and solved part of my question about serving, and the disscussion metioned the futher way may be tf-tvm.
so my question is, is there any chance in developing TVM own serving for fully leveraging the optimization provided by TVM?
or just try to embeding TVM into frameworks such as TF, pytorch, MXNet, caffee …, especially in inference serving domain.