You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Remove rust-cli fallback in favor of a pure-Python CLI
Summary:
All the working subcommands were falling back to Python anyways.
Moved the (currently unimplemented) subcommand stubs: `bounce` and `stop` to Python.
**Note:** couple of reasons why a Rust CLI for monarch isn't ideal:
1. Uses TorchX under the hood. TorchX is a Python library.
2. Due to #1 we have to run a Python CLI fallback anyways and the mechanics of this is meta specific (won't work for OSS).
3. Reverse pyo3 binding TorchX (call Python from Rust) doesn't work internally due to the way we package Python (hermetic PAR).
4. Any material benefits (e.g. performance?) of implementing the CLI in Rust would be negated by the effort to fix/deal-with #1-3.
**Next:**
~~[6/n] Have kd_monarch use the default component (the custom mast.py is no longer needed). Update the README with updated instructions.~~
~~[7/n] Remove rust CLI in favor of all-python (we delegate to torchx for most things anyways)~~
[8/n] Add E2E unittest using the local_cwd scheduler (actually run a mini-trainer actor)
[9/n] Write an oss hyperactor mesh-worker entrypoint binary
[10/n] Author a Dockerfile that sets up the environment (much like fbpkgs do it for internal runs)
[11/n] Author a TorchXAllocator
Reviewed By: vidhyav, suo
Differential Revision: D75176535
fbshipit-source-id: 29020f4032bd642af26b393ade74f40b868df973
0 commit comments