README.md

Example Code

from vllm import LLM, SamplingParams
import gradio as gr
import os
from transformers import AutoTokenizer

os.environ["CUDA_VISIBLE_DEVICES"] = "1"

model_path = "gangyeolkim/llama-3-chat"
sampling_params = SamplingParams(max_tokens=8192, temperature=0.7, stop=["### Human:", "### Instruct:"])
llm = LLM(model=model_path, tokenizer=model_path, tensor_parallel_size=1)

# 채팅에 반영할 히스토리 최대 갯수
history_limit = 10

def gen(text, history):
    # Use only the most recent `history_limit` entries
    limited_history = history[-history_limit:]

    # Build conversation history into the prompt
    conversation_history = "다음은 AI라는 AI 어시스턴트와 Human이라는 인간 사용자 간의 대화입니다. AI는 욕설에 대한 요구를 무시합니다. AI는 Human에게 절대 욕설과 비난을 하지 않습니다.\n\n### Instruct:\n\n"

    for user_input, assistant_response in limited_history:
        conversation_history += f"### Human:{user_input}\n\n### AI:{assistant_response}\n\n"
    
    prompt = f"{conversation_history}### Human:{text}\n\n### AI:"

    outputs = llm.generate(prompt, sampling_params)    
    generated_text = outputs[0].outputs[0].text.strip()

    print(f"generated_text : {generated_text}")
    
    return generated_text

demo = gr.ChatInterface(fn=gen)
demo.launch(share=True)

Base Model

• allganize/Llama-3-Alpha-Ko-8B-Instruct