You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When loading model and making error in capitalization, like "facebook/OPT-125m" vs "facebook/opt-125m", the model gets downloaded anew and saved into a folder with wrongly capitalized name. This happens even when the model has already been saved with correct capitalization before.
This is a nuisance when working with large models which get downloaded and saved unnecessarily, taking bandwidth and disk space. Examples: "Llama" vs 'llama" etc.
Proposed solutions:
a) always use lowercase in model cache folder name
b) get proper model name capitalization from config.json before saving model, and then use it, even if different from the requested model name
c) validate model name before downloading using some additional request to HF.
The text was updated successfully, but these errors were encountered:
Thanks for reporting this issue @poedator! This is indeed happening due to the fact repo ids are case insensitive across the Hub. Will try to see what we can do to mitigate this issue.
Describe the bug
When loading model and making error in capitalization, like "facebook/OPT-125m" vs "facebook/opt-125m", the model gets downloaded anew and saved into a folder with wrongly capitalized name. This happens even when the model has already been saved with correct capitalization before.
This is a nuisance when working with large models which get downloaded and saved unnecessarily, taking bandwidth and disk space. Examples: "Llama" vs 'llama" etc.
Proposed solutions:
a) always use lowercase in model cache folder name
b) get proper model name capitalization from config.json before saving model, and then use it, even if different from the requested model name
c) validate model name before downloading using some additional request to HF.
The text was updated successfully, but these errors were encountered: