Hi community, I am currently working on a project where I train a model (e.g., bert-base-uncased) and push it to my private Hugging Face Hub repository, along with its optimizer state, using the Trainer class. In my workflow, I save model checkpoints and later want to resume training from a specif…

I see. So, the current method is probably correct. If you want to deviate from the 1 repo 1 model (revision management is possible), it is more reliable to manage it yourself. If you use hf_hub_download or snapshot_download, there should not be much difference in speed from_pretrained…

Resuming Training from Checkpoints Stored on Hugging Face Hub (without downloading manually)

🤗Transformers

gcheronflex February 10, 2025, 10:32am 7

Thanks for the clarification! I’ll stick with my approach then.
Appreciate your help!

1 Like

Topic		Replies	Views
Trainer .train (resume _from _checkpoint =True) Beginners	9	15435	May 16, 2024
How to resume from checkpoint on the hub? Not using trainer api, I'm using TF api Beginners	1	347	February 28, 2023
Checkpoint missing Optimizer.pt? How to Resume? 🤗Transformers	7	5527	May 18, 2021
Resume training from checkpoint Beginners	1	3084	January 5, 2023
Trainer API not pushing checkpoints to HUB 🤗Transformers	0	309	December 7, 2022

Resuming Training from Checkpoints Stored on Hugging Face Hub (without downloading manually)

Related topics