KeyError: 'llama' in /recover_model_weights.py about alpaca_farm

[rtaori]

Which version of transformers was used when running the weights conversion script? We converted the raw llama weights into huggingface format using transformers==4.29.2 (as documented in the README)

[mensch72]

4.26.0, as required per requirements.txt. If I upgrade to 4.32.0, I get an even faster error:

Traceback (most recent call last):
  File "/p/projects/ou/labs/gane/rlhf/git/alpaca_farm-collective/pretrained_models/recover_model_weights.py", line 11, in <module>
    from alpaca_farm.models.reward_model import RewardConfig, RewardModel
  File "/p/projects/ou/labs/gane/rlhf/envs/alpaca-env/lib/python3.10/site-packages/alpaca_farm/models/reward_model.py", line 20, in <module>
    from .. import common
  File "/p/projects/ou/labs/gane/rlhf/envs/alpaca-env/lib/python3.10/site-packages/alpaca_farm/common.py", line 32, in <module>
    from transformers.trainer import WEIGHTS_NAME, is_deepspeed_zero3_enabled
ImportError: cannot import name 'is_deepspeed_zero3_enabled' from 'transformers.trainer' (/p/projects/ou/labs/gane/rlhf/envs/alpaca-env/lib/python3.10/site-packages/transformers/trainer.py)

With 4.29.2 it now seems starts downloading stuff! So apparently one needs this exact version, not much older but also not much newer... Thanks for your help!