transformers
transformers copied to clipboard
[WIP] [deepspeed] from_pretrained deal with ignore_mismatched_sizes
An attempt to fix the issue reported https://github.com/huggingface/transformers/issues/17336
Fixes: https://github.com/huggingface/transformers/issues/17336
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.
After creating a test I discovered it breaks on tied variables since they get ignored in model.named_parameters - so back to the drawing table.