LLama3ChatSession example fails to work with LLama 2 model

[asmirnov82]

Description

LLama3ChatSession example is said to "supports models such as llama3, llama2, phi3, qwen1.5, etc".
However it fails due to several reasons:

1. Null reference exception (as in ;lama2 model.Tokens.EndOfTurnToken is not defined in metadata and that's why is null)
2. PromtTemplateTransformer doesn't work correctly with llama2. According to https://www.llama.com/docs/model-cards-and-prompt-formats/meta-llama-2 and https://huggingface.co/blog/llama2#how-to-prompt-llama-2, llama2 should be prompted using this template:

<s>[INST]
{{ system_prompt }}
<</SYS>>

{{ user_message_1 }} [/INST] {{ model_answer_1 }} </s>
<s>[INST] {{ user_message_2 }} [/INST]

Llama2 doesn't provide chat template in metadata, so by default different template is used by llama.cpp:

<|im_start|>user
hello<|im_end|>
<|im_start|>assistant
response<|im_end|>
<|im_start|>user
again<|im_end|>
<|im_start|>assistant
response<|im_end|>

As a result, conversation using LLama3 example after fixing NullRefence exception looks like this:

It would be nice to have an example that uses custom HistoryTransform for llama2

[martindevans]

Llama2 doesn't provide chat template in metadata

I'm not sure, but I think llama.cpp has tools to modify GGUF files to add new metadata, the easiest fix would probably be to modify your llama2 file to include the correct template.

Edit: I did just merge your PR though, so that'll be in the next release :)