app.py

from threading import Thread
import gradio as gr
import inspect
from gradio import routes
from typing import List, Type
from petals import AutoDistributedModelForCausalLM
from transformers import AutoTokenizer
import npc_data
import requests, os, re, asyncio, json

loop = asyncio.get_event_loop()

# init code
def get_types(cls_set: List[Type], component: str):
    docset = []
    types = []
    if component == "input":
        for cls in cls_set:
            doc = inspect.getdoc(cls)
            doc_lines = doc.split("\n")
            docset.append(doc_lines[1].split(":")[-1])
            types.append(doc_lines[1].split(")")[0].split("(")[-1])
    else:
        for cls in cls_set:
            doc = inspect.getdoc(cls)
            doc_lines = doc.split("\n")
            docset.append(doc_lines[-1].split(":")[-1])
            types.append(doc_lines[-1].split(")")[0].split("(")[-1])
    return docset, types
routes.get_types = get_types

# App code

model_name = "daekeun-ml/Llama-2-ko-instruct-13B"

#daekeun-ml/Llama-2-ko-instruct-13B
#quantumaikr/llama-2-70b-fb16-korean
tokenizer = AutoTokenizer.from_pretrained(model_name)

model = None

def check(model_name):
    data = requests.get("https://health.petals.dev/api/v1/state").json()
    out = []
    for d in data['model_reports']:
        if d['name'] == model_name:
            if d['state']=="healthy":
                return True
    return False

def init():
    global model
    if check(model_name):
        model = AutoDistributedModelForCausalLM.from_pretrained(model_name)


def chat(id, npc, prompt):
    if model == None:
        init()
        return "no model"
    # get_coin endpoint
    response = requests.post("https://ldhldh-api-for-unity.hf.space/run/predict_6", json={
      "data": [
        id,
    ]}).json()
    
    coin = response["data"][0]
    if int(coin) == 0:
        return "no coin"
    
    # model inference

    if check(model_name):
        prom = prompt
        inputs = tokenizer(prom, return_tensors="pt")["input_ids"]
        outputs = model.generate(inputs, max_new_tokens=100)
        print(tokenizer.decode(outputs[0]))
        output = outputs[len(prom):-1]
        print(output)
    else:
        output = "no model"


    # add_transaction endpoint
    response = requests.post("https://ldhldh-api-for-unity.hf.space/run/predict_5", json={
      "data": [
        id, 
        "inference", 
        "### input:\n" + prompt + "\n\n### output:\n" + output
    ]}).json()
    
    d = response["data"][0]
    
    return output


with gr.Blocks() as demo:
    count = 0
    aa = gr.Interface(
      fn=chat,
      inputs=["text","text","text"],
      outputs="text",
      description="chat, ai 응답을 반환합니다. 내부적으로 트랜잭션 생성. \n /run/predict",
    )
    
    demo.queue(max_size=32).launch(enable_queue=True)