eos_token_id discrepency
Hello, I don't know if this was intended, but model.generation_config.eos_token_id != tokenizer.eos_token_id. This discrepency can create a problem when training the model while adding the tokenizer's eos token and expecting the generation to stop with a different token.
Hi, thanks for your interest to imp. In our tests, model.generation_config.eos_token_idand tokenizer.eos_token_id are using the same token id </s>.
And can you tell me the version of your tokneizer package, this could be help for finding the problem.
No thank you you, it works great for me, and still even Nvidia cannot beat you https://twitter.com/Jacoed/status/1770795726877990953.
tok.eos_token_id
50256
llm.generation_config.eos_token_id
50295
from transformers import ViltProcessor, AutoModel, AutoModelForCausalLM, AutoTokenizer
import transformers
transformers.version
'4.39.0'
The tokenizer comes from transformers for me, does that answer your question @Oyoy1235 ?
Great, I think we don't have other questions in eos_token_id.