Closed andrewPoulton closed 1 year ago
The tokenization_cache in the train_iterator state_dict needs to be unpickled in a metaseq-aware env.
Adding a flag to remove this so consolidated checkpoints load elsewhere.
Also fixes bug in named argument
Was testing if rebasing on main could resolve gpu test failures, which apparently adds me to all commits :( Nothing here should affect gpu tests so this is an odd failure...
The tokenization_cache in the train_iterator state_dict needs to be unpickled in a metaseq-aware env.
Adding a flag to remove this so consolidated checkpoints load elsewhere.
Also fixes bug in named argument