Fix Llama2 implementation: Cannot import name 'Llama2Tokenizer' from 'transformers'

[fedegr]

I created a quickstart test file which is mainly doing the same that appears on the documentation:

from llm_vm.client import Client

client = Client(
       big_model = 'neo',
       big_model_config={'model_uri':'EleutherAI/gpt-neo-1.3B'},
       small_model ='neo',
       small_model_config={'model_uri':'EleutherAI/gpt-neo-125m'})

response=client.complete(prompt='What is Anarchy?', context='') 
print(response)

But when I try to run it return it the next error:

python quickstart-test.py 
Traceback (most recent call last):
  File ".../LLM-VM/quickstart-test.py", line 2, in <module>
    from llm_vm.client import Client
  File ".../LLM-VM/src/llm_vm/client.py", line 3, in <module>
    import llm_vm.onsite_llm as llms
  File ".../LLM-VM/src/llm_vm/onsite_llm.py", line 6, in <module>
    from transformers import (
ImportError: cannot import name 'Llama2Tokenizer' from 'transformers' (.../LLM-VM/.venv/lib/python3.10/site-packages/transformers/__init__.py

If I checkout at commit 7a5877b6a everything works fine.

I checked the huggingface transformers documentation and llama2 model is using the same transformer I also checked the repo and there is no llama2 folder nor Llama2Tokenizer, Llama2ForCausalLM classes.

[fedegr]

I have created a PR were I fix the issue, but I was not able to test it because I am waiting Meta approval to have access to Llama v2 model.

[ChinmayK0607]

LLama2 uses the same tokenizers as LLama. source : https://huggingface.co/docs/transformers/main/model_doc/llama2 I have fixed this in PR #185 which was for the issue #183

[fedegr]

@abhigya-sodani shouldn't this story be closed as completed? Given that PR #162 was fixing this + adding back OpenLlama