Adamliu1 / SNLP_GCW

3 stars 0 forks source link

Add precomputing of normal responses (only 1 model loaded at a time). #142

Closed Willmish closed 2 months ago

Willmish commented 2 months ago

As per #137, this PR introduces a potential major gpu memory saving benefit, by loading a single reference model, precomputing the "pretrained/reference model" loss components of the normal loss for ALL batches, and unlearning on the same model.

This is increases the RAM usage of the unlearning (depending on unlearning sample size), but requires only a single model loaded to memory

Closes #137