Matt Watson

Closing, this was landed with some edits on https://github.com/keras-team/keras-nlp/pull/389 by @chenmoneygithub. But huge props on writing this, this is a big deal for the library!!

Thank! Yeah was wondering about that too. Currently it only increases the test time from ~7 minutes to ~9 minutes, which is not a big deal, but I think maybe...

@jbischof thanks! Replies below... > Isn't some code required to link the new flag to our GCP infra? Yes, the single line in `.cloudbuild/unit_test_jobs.jsonnet` does this. > Why is only...

@jbischof `--run_slow` is still comprehensive. It will run everything tagged with `--run_gcp` as well. I wonder if there is a better way I can preset that so it is more...

I've updated this to move testing inside the `keras_nlp/models/xx` directory. And changed the flag names so we just have the notion of test sizes (currently `large` and `extra_large`), which we...

I guess my thinking here was that it is nice if the presets and preset testing could be delivered as a unit. We can definitely see today how we will...

Just talked with @jbischof. I think the laziest approach we can have here is to just keep a forked and unexported layer for the Roberta style packing inside the roberta...

Hopefully, without the need for segment ids, the roberta layer can be a bit simpler anyway.

I'm not actually too opinionated on wether we add the defaults or not. There will probably be some models like gpt-2, where having a default is really natural (there is...

This all sounds good to me! SST as the first offering sounds great. > Each classifier_checkpoints entry will need to specify num_classes. This sounds good to me to! I think...