fix: freeze to specific set of versions with reliable tighter lower and upper bounds.

foundation-model-stack / fms-hf-tuning

🚀 Collection of tuning recipes with HuggingFace SFTTrainer and PyTorch FSDP.

Apache License 2.0

9 stars 30 forks source link

fix: freeze to specific set of versions with reliable tighter lower and upper bounds. #196

Open kmehant opened 2 weeks ago

kmehant commented 2 weeks ago

From various collaborations and experience, we faced issues for not having strict version bounds given quickly updating HF dependencies some introducing breaking changes. For instance, there were issues with not freezing to certain versions when we had to start with the existing conda environment and install on top of it. It would be of great help if we can narrow down error surface by having tighter bounds over the versions.

VassilisVassiliadis commented 2 weeks ago

I think this is related to https://github.com/foundation-model-stack/fms-hf-tuning/issues/60

kmehant commented 2 weeks ago

@VassilisVassiliadis Thanks for referencing, will close this once the attached issue is closed as well. Will keep this open for reference related to IL work. Thanks