backwards compatibility breaking change from sharded checkpoints

[bghira]

Describe the bug

Since #7830 we have SDXL checkpoints sharded automatically, which did not happen before. These checkpoints are not loadable on earlier versions.

Reproduction

I think all one has to do is load from_pretrained() and then save_pretrained()

Logs

$ ls /home/user/training/lite-models/checkpoint-400/unet/
config.json  diffusion_pytorch_model-00001-of-00002.safetensors  diffusion_pytorch_model-00002-of-00002.safetensors  diffusion_pytorch_model.safetensors.index.json

System Info

main git branch

Who can help?

@sayakpaul as discussed

[DN6]

cc: @sayakpaul

[sayakpaul]

@DN6 please also see the linked PR :-)