In the latest version of TVM tvm-0.16 .0, relax is already supported. However, the source code has only some optimized pass and related tests, and there is only one vmbuild compilation. There is no complete compilation function like relay that can directly invoke the compilation model to the executable file of a platform. I want to ask whether relax only implements some independent pass optimization for our own invocation.Or is it still under development? What is missing compared with relay will be gradually added later? I am a TVM beginner. Thank you very much for your discussion.
We are working on documents on how to use relax to compose an end to end flow. One thing to note is that we are moving towards a model where each domain can bring its own set of compilations.
As of now the set of features are very useful to support emerging genAI workloads like LLM
1 Like