[Bug]: Mismatch of the file size between the diffusers and onetrainer format for the Pixart-Sigma model

[frutiemax92]

What happened?

I developped a tool to easily merge checkpoints together and I've noticed that OneTrainer is generating checkpoints a little bit bigger than the one generated in the diffusers format.

What did you expect would happen?

After investigation, this key in the OneTrainer checkpoint should not be used : pos_embed. When you remove that key, the save state dictionnary becomes the same size as the diffusers format.

Relevant log output

No response

Output of pip freeze

No response