This is a great development, as-is allowing to use TVM in TF models that cannot be fully translated. I have some clarifying questions, also along the lines that jwfromm@ was thinking.
Suppose I want to implement a tool that takes a tf graph as input (be it from a saved model or some other input) and also writes a tf graph as output, where the difference is that the graph has been converted to use TVM for as much of the graph that can be supported, leaving behind only the pieces of TF that could not be converted to TVM. Also, any necessary compiled ops from TVM would be embedded in the output, so that a TF runtime can run it without having any TVM ops shipped with the runtime. I think the work you've done is partway there to such a tool, allowing to represent TVM subgraphs in TF, leaving some other parts like automatically identifying the pieces of the TF graph that can be converted, and automatically exercising TVM to generate implementations of those subgraphs, and storing those compiled TVM ops alongside the model so that a plain vanilla TF runtime with no TVM ops shipped with it can run the model. Did I get that right? @jwfromm I understand from the online and your description that pytorch-tvm is closer to enabling such a tool. *Is* it already such a tool, for PT, or is there still a distance remaining to that? (I didn't spot an ahead-of-time compilation mode) -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/apache/incubator-tvm/issues/4464#issuecomment-564742256
