Add support for ReFT

[RonanKMcGovern]

🚀 The feature, motivation and pitch

Motivation is to allow ReFT representations to be applied on the fly during inference, which can be done in a batchwise manner.

this is much faster than applying LoRAs

Alternatives

LoRA is too slow as it requires adapter weights to be added, which increases the number of operations.

Additional context

See stanfordnlp/pyreft#63

[chris-aeviator]

as a user of pyreft I want to highlight the need for selecting subspaces into a hypothetical PyreftRequest (see stanfordnlp/pyreft#63 (comment))

[jvlinsta]

Any traction on this?

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

[chris-brightbeam]

unstale