Eos and bos tokens can be redefined as additional tokens with other ids

[vvasily]

https://github.com/ggerganov/llama.cpp/blob/92f44ff7f778ef1b94028b2ba6d39943b5ca0ada/llama.cpp#L3251

In my implementation of llama based model we use redefined eos and bos symbols from additional tokens. In this code they are hardcoded. I think this possibility should be implemented in this code.

[HashemAlsaket]

Hello @vvasily @ggerganov . I am new to this repo and would like to start contributing. Super cool project. Thanks for any feedback!

[Sharkyii]

It looks like this issue has already been resolved in the current codebase.

• BOS/EOS token IDs are now loaded dynamically from GGUF metadata (tokenizer.ggml.bos_token_id and tokenizer.ggml.eos_token_id) via llama_vocab::load().
• The API now uses llama_vocab_bos() / llama_vocab_eos() instead of the old hardcoded functions.
• Default IDs (1 for BOS, 2 for EOS) are still provided for backward compatibility.