Checkpoint doesn't match

self.model = DalleBart.from_pretrained("model", dtype=dtype, abstract_init=True) I got following error message for above line looks like the checkpoint doesn’t match? ValueError: Trying to load the pretrained weight for ('model', 'decoder', 'embed_positions', 'embedding') failed: checkpoint has shape (257, 1024) which is incompatible with the model shape (256, 1024). Using ignore_mismatched_sizes=True if you really want to load this checkpoint inside this model.

After updating to self.model = DalleBart.from_pretrained("model", dtype=dtype, abstract_init=True, ignore_mismatched_sizes=True) (add ignore_mismatched_sizes=True) I got another error TypeError: Value ShapeDtypeStruct(shape=(256, 1024), dtype=float32) with type <class ‘jax._src.api.ShapeDtypeStruct’> is not a valid JAX type

Looks like both errors are from JAX, I install the latest JAX by pip install jax[cuda11_cudnn805] -f https://storage.googleapis.com/jax-releases/jax_releases.html

Any suggestions?

[shi3z]

Awesome!

I updated transformers from 14.0 to 14.16, then it work well. Thank you very much! Best

[borisdayma]

Did you try to update your installation? Make sure your transformers repo is updated too!