Fix pp convergence to be bitwise

[3outeille]

Bitwise matching with Qwen2 but not with Llama.

Hypothesis:

• Identity() not working with DCP ?
  • one easy way to know for sure will be to manually do if toke_embeddings is None in the Transformers modeling and see if the issue persists.
• Maybe due to tie_embedding?
  • Both models has tied_embedding but when printing model structure, lm_head is showing up for HF llama3 while not for HF Qwen2)