Due to momentum, the optimiser can take up a lot disk space compared to the model itself. We currently upload the full model, including the optimiser to huggingface and download that to the app. We should remove the optimiser before uploading to save on bandwidth and memory
Due to momentum, the optimiser can take up a lot disk space compared to the model itself. We currently upload the full model, including the optimiser to huggingface and download that to the app. We should remove the optimiser before uploading to save on bandwidth and memory