Hi, I wanted to ask whether new releases of apache-tvm will ever by published on PyPI? The last stable release was v0.12.0 back in June 2023, with dev releases up to 0.14dev273 in November 2023. The version in the GitHub repo is now v0.20.0. Is there a policy to no longer maintain PyPI versions? If this is the case, the documentation should be updated to say that the installation available from PyPI no longer matches the API description in the documentation, e.g., tvm.relax does not exist in the PyPI installable version.
for now we focused on the source release, there are downstream mainatained packages in the meanwhile that can also be used in some of the tutorials like https://mlc.ai/
Hi @tqchen, would you be open to contributions on modernizing PyPI packaging?
I understand this to be a common point of confusion (based on experiences from myself and my colleagues), and would be happy to work on this. Also, are there are specific pitfalls (e.g., integration with external libraries like llvm, CUDA, or BYOC toolchains) that discourage the use of pre-built packages?
yes, previously one possible blocker is mainly the different deps on cuda versions etc. I think with tvm-ffi we might be able to work on a runtime restructure so device runtime deps are isolated, after that i think we can bring more formal pypi