llm-foundry
llm-foundry copied to clipboard
[WIP] Refactor logging
This PR refactors the logging to:
- centralize verbosity controls into the python logging levels, instead of passing
verbosearguments to many helper methods - downgrades most warnings to info level.
warningsshould only be used for errors or unexpected behavior that do not require raising an exceptions, not informative messages such as reporting expected behavior. - removes the warnings around usage of torch vs. triton vs. flash. These have led to users mistakenly believing this is the source of other errors.
Note: This PR is API-breaking. Should I keep the verbose arguments to the methods, and add a deprecation warnings instead that this argument is being ignored?
TODO:
- [ ] add python log level control in scripts
- [ ] move the attention implementation information to an FAQ
- [ ] add some protection for invalid choices (e.g. alibi and attn_impl conflicts)
- [ ] remove remaining
verboseandwarnings