matthiasgeihs

**Update** The behavior described below was due to an error in the causal attention mask... Now it works as expected. --- Converted to `draft` because of a curious observation: The...

Introduced a bug with the last commit, hence corrected and force pushed.

> @matthiasgeihs Any plans for [rotary embeddings](https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/model.py#L64)? @KendallPark not on my priority list currently, but feel free to give it a shot ;)

> @matthiasgeihs - is your implementation working now? Yes, you can test with: ``` $ python train.py config/train_shakespeare_char.py --out_dir=out-shakespeare-char-alibi --pos_encoding=alibi ```

If I see it correctly, this fails for all `.ml` files. Concretely, I was testing with the [German CSCA Master List](https://www.bsi.bund.de/SharedDocs/Downloads/DE/BSI/ElekAusweise/CSCA/GermanMasterList.html).

Feel free to close the issue whenever you feel like it. 🙂

> It seems like the extra check is unnecessary, because the passphrase is hashed inside the zkvm, which means we can be sure that it was computed properly. I understand...

hey, does there exist a prototype implementation of this?

@Wind4Greg have you found the time to code up a prototype?

Hi @Wind4Greg, thank you for your reply. Can the holder secret be used to create signatures on other documents as well? Use case: A credential holder wants to sign a...