Pr eval lora

[danbider]

Modified from @SingL3 PR to pin down peft version.

Previously saw conflicts between the huggingface/peft/main and the accelerate library that was pinned down in llm-foundry.

Tested this on an interactive session with:

• hf lora module as tested by @SingL3.
• a Composer checkpoint of a trained peft/lora model (trained on pile-of-law).
  eval works fine in both cases.

[codestar12]

LGTM