main.py

from fastapi import FastAPI
from pydantic import BaseModel
from llama_cpp import Llama
import uvicorn
import prompt_style
import os

model_id = "failspy/Meta-Llama-3-8B-Instruct-abliterated-v3-GGUF"
model = Llama.from_pretrained(repo_id=model_id, filename="*-v3_q6.gguf", n_gpu_layers=-1, n_ctx=4096, verbose=False)

# model_id = "failspy/Meta-Llama-3-8B-Instruct-abliterated-v3"
# client = InferenceClient(token=os.getenv('HF_TOKEN'), model=model_id)

class Item(BaseModel):
    prompt: str
    history: list
    system_prompt: str
    temperature: float = 0.6
    max_new_tokens: int = 1024
    top_p: float = 0.95
    repetition_penalty: float = 1.0
    seed : int = 42
    
app = FastAPI()

def format_prompt(item: Item):
    messages = [
        {"role": "system", "content": prompt_style.data},
    ]
    for it in history:
        messages.append({"role" : "user", "content": it[0]})
        messages.append({"role" : "assistant", "content": it[1]})
    messages.append({"role" : "user", "content": item.prompt})
    return messages

def generate(item: Item):
    formatted_prompt = format_prompt(item)
    output = model.create_chat_completion(messages=formatted_prompt, seed=item.seed, 
                                          temperature=item.temperature,
                                          max_tokens=item.max_new_tokens)


    out = output['choices'][0]['message']['content']
    return out

@app.post("/generate/")
async def generate_text(item: Item):
    ans = generate(item)
    return {"response": ans}


@app.get("/")
def read_root():
    return {"Hello": "World!"}