As of Disable TORCH_MLIR_ENABLE_JIT_IR_IMPORTER and TORCH_MLIR_ENABLE_PYTOR… · llvm/torch-mlir@fbb0db1 · GitHub, we are no longer building features that natively link to pytorch via c++. Concretely, this includes all old TorchScript based interop and LTC (although LTC was already disabled by default). If this doesn’t mean anything to you, then a way to tell is by looking to see if projects/pt1 is in the path (we still have some developer tooling like the test driver housed there but there shouldn’t be any user features or API). A first pass through the docs has been done to excise obsolete content but more is in progress.
The author of the patch is still working on migrating some developer features:
- ODS code generation relies on a c++ entry point in the c++ extension and it just be enabled to use it.
- The test framework code lives under the pt1 directory because it used to be entangled. But it no longer requires the extension to be built to use it and will be moved to the top level.
Until the first issue is resolved via a pure python path (ie. From existing python APIs or parsing torch op yaml files), the CI is still testing with the extension enabled. If this produces issues with getting PRs landed, speak up: we can likely disable troublesome tests at this point.
With this change, torch-mlir is not tightly coupled to a specific pytorch version and will work with any version that does not break APIs that are in use by the project. I’m the future, we should be able to build it against the stable cpython API, also making any wheels binary compatible with all Python versions. This will make testing and CI significantly cheaper to operate.
We have no immediate plans to delete any of the gated code. If anyone would like to step up and maintain any of it, please speak up and we can work out a plan to mainline any desired feature before the eventual deletion of the projects/pt1 directory.
Thanks to the long list of people who made this migration possible, most notably by helping to migrate everything to the pure python FX based flow. Special thanks to @penguin-wwy who spent several weeks porting tests and infra earlier in the year.
For anyone wondering what this is all about, this represents the effective completion of the roadmap. Always good to see a plan finished, especially when it is a multi year, multi company effort.