Although GPT-2 supports token_type_ids, the implementation is very weird (token embeddings are also used as token type embeddings!) and in practice token_type_ids=None is used almost exclusively.

For most models, you could mimic the effects of token_type_ids=None by just passing an all-zeros array, but this does not work for GPT-2 because it completely skips the embeddings when token_type_ids=None. This means that a model exported with a token_type_ids input cannot be coerced to behave correctly.

To stop this tripping up other users, we remove token_type_ids from the GPT-2 input sig. This issue is specific to GPT-2 - most other models have more reasonable ways of handling token_type_ids and shouldn't be affected.

Fixes #26783