[New Model]: Llama 3 8B Instruct

[K-Mistele]

The model to consider.

https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct

The closest model vllm already supports.

LLama 1 & 2

What's your difficulty of supporting the model you want?

LLama 3 instruct requires a different stop token than is specified in the tokenizer.json file.
The tokenizer.json specifies <|end_of_text|> as the end of string token which works for the base LLama 3 model, but this is not the right token for the instruct tune. The instruct tune uses <|eot_id|>.

You can see this in the inference code for the model on the llama 3 8B instruct model card, where this token is added:

import transformers
import torch

model_id = "meta-llama/Meta-Llama-3-8B-Instruct"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

prompt = pipeline.tokenizer.apply_chat_template(
        messages, 
        tokenize=False, 
        add_generation_prompt=True
)

# HERE is where they add the `<|eot_id|>` token, which is not the default end of string token, to the list of terminators.
terminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = pipeline(
    prompt,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
print(outputs[0]["generated_text"][len(prompt):])

Here is a discussion of this topic in the llama.cpp repository:
ggml-org/llama.cpp#6751

[agt]

This will be resolved via #4182 which has been merged and will be released with 0.4.1 .

In the meantime, #4180 has suggestions on workarounds including manually editing config.json to set the correct stop token.

[JPonsa]

@agt is still open? is llama3 still not supported in vLLM?

[hmellor]

Llama-3 is supported