cpu-only environment

[Kevin-chen-sheng]

Can I run it in a cpu-only environment?

[nickbhat]

Could you say more about what you are trying to run cpu-only? Pretraining, fine-tuning, embedding, etc.

[rmrao]

I actually had added support for this throughout, but there were a few bugs. Just fixed them in #3 so if you clone the repository, the commands (except for tape-train-distributed) can be run with the --no_cuda flag, which should work in a cpu-only environment.

It is much, much slower to do this than to run via gpu, obviously (60-100x slower for embedding). We're thinking of adding in some of the smaller unirep models (e.g. babbler-60) since a lot of people seem to like using those instead of the large ones, and those should be at least reasonably quick on cpu.

[Kevin-chen-sheng]

thanks @rmrao.