FIX: reduce CorDA memory consumption + docs

[5eqn]

• Added documentation about selection of run_model sample size
• Added documentation about memory consumption of preprocessing
• Pruned temporary fields after preprocessing and building LoRA adapter, so that CorDA does not consume more memory during training compared to other LoRA methods

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

Thanks for the updates to the docs and the improvements to memory consumption. The changes LGTM.

Before merging, could you please fix the following line:

peft/examples/corda_finetuning/preprocess.py

Line 24 in 3289134

from peft.mapping import get_peft_model

The import is no longer valid after merging #2282. Please always import like so: from peft import get_peft_model.

[5eqn]

Before merging, could you please fix the following line:

Sure, the line is fixed.

[BenjaminBossan]

Thanks for the updates, LGTM.