awesome_chat.py

import base64
import copy
from io import BytesIO
import io
import os
import random
import time
import traceback
import uuid
import requests
import re
import json
import logging
import argparse
import yaml
from PIL import Image, ImageDraw
from diffusers.utils import load_image
from pydub import AudioSegment
import threading
from queue import Queue
from get_token_ids import get_token_ids_for_task_parsing, get_token_ids_for_choose_model, count_tokens, get_max_context_length
from huggingface_hub.inference_api import InferenceApi
from huggingface_hub.inference_api import ALL_TASKS
from models_server import models, status
from functools import partial

parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, default="config.yaml.dev")
parser.add_argument("--mode", type=str, default="cli")
args = parser.parse_args()

if __name__ != "__main__":
    args.config = "config.gradio.yaml"

config = yaml.load(open(args.config, "r"), Loader=yaml.FullLoader)

if not os.path.exists("logs"):
    os.mkdir("logs")

logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)

handler = logging.StreamHandler()
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
handler.setFormatter(formatter)
if not config["debug"]:
    handler.setLevel(logging.INFO)
logger.addHandler(handler)

log_file = config["log_file"]
if log_file:
    filehandler = logging.FileHandler(log_file)
    filehandler.setLevel(logging.DEBUG)
    filehandler.setFormatter(formatter)
    logger.addHandler(filehandler)

LLM = config["model"]
use_completion = config["use_completion"]

# consistent: wrong msra model name 
LLM_encoding = LLM
if LLM == "gpt-3.5-turbo":
    LLM_encoding = "text-davinci-003"
task_parsing_highlight_ids = get_token_ids_for_task_parsing(LLM_encoding)
choose_model_highlight_ids = get_token_ids_for_choose_model(LLM_encoding)

# ENDPOINT	MODEL NAME	
# /v1/chat/completions	gpt-4, gpt-4-0314, gpt-4-32k, gpt-4-32k-0314, gpt-3.5-turbo, gpt-3.5-turbo-0301	
# /v1/completions	text-davinci-003, text-davinci-002, text-curie-001, text-babbage-001, text-ada-001, davinci, curie, babbage, ada

if use_completion:
    api_name = "completions"
else:
    api_name = "chat/completions"

if not config["dev"]:
    if not config["openai"]["key"].startswith("sk-") and not config["openai"]["key"]=="gradio":
        raise ValueError("Incrorrect OpenAI key. Please check your config.yaml file.")
    OPENAI_KEY = config["openai"]["key"]
    endpoint = f"https://api.openai.com/v1/{api_name}"
    if OPENAI_KEY.startswith("sk-"):
        HEADER = {
            "Authorization": f"Bearer {OPENAI_KEY}"
        }
    else:
        HEADER = None
else:
    endpoint = f"{config['local']['endpoint']}/v1/{api_name}"
    HEADER = None

PROXY = None
if config["proxy"]:
    PROXY = {
        "https": config["proxy"],
    }

inference_mode = config["inference_mode"]

parse_task_demos_or_presteps = open(config["demos_or_presteps"]["parse_task"], "r").read()
choose_model_demos_or_presteps = open(config["demos_or_presteps"]["choose_model"], "r").read()
response_results_demos_or_presteps = open(config["demos_or_presteps"]["response_results"], "r").read()

parse_task_prompt = config["prompt"]["parse_task"]
choose_model_prompt = config["prompt"]["choose_model"]
response_results_prompt = config["prompt"]["response_results"]

parse_task_tprompt = config["tprompt"]["parse_task"]
choose_model_tprompt = config["tprompt"]["choose_model"]
response_results_tprompt = config["tprompt"]["response_results"]

MODELS = [json.loads(line) for line in open("data/p0_models.jsonl", "r").readlines()]
MODELS_MAP = {}
for model in MODELS:
    tag = model["task"]
    if tag not in MODELS_MAP:
        MODELS_MAP[tag] = []
    MODELS_MAP[tag].append(model)
METADATAS = {}
for model in MODELS:
    METADATAS[model["id"]] = model

def convert_chat_to_completion(data):
    messages = data.pop('messages', [])
    tprompt = ""
    if messages[0]['role'] == "system":
        tprompt = messages[0]['content']
        messages = messages[1:]
    final_prompt = ""
    for message in messages:
        if message['role'] == "user":
            final_prompt += ("<im_start>"+ "user" + "\n" + message['content'] + "<im_end>\n")
        elif message['role'] == "assistant":
            final_prompt += ("<im_start>"+ "assistant" + "\n" + message['content'] + "<im_end>\n")
        else:
            final_prompt += ("<im_start>"+ "system" + "\n" + message['content'] + "<im_end>\n")
    final_prompt = tprompt + final_prompt
    final_prompt = final_prompt + "<im_start>assistant"
    data["prompt"] = final_prompt
    data['stop'] = data.get('stop', ["<im_end>"])
    data['max_tokens'] = data.get('max_tokens', max(get_max_context_length(LLM) - count_tokens(LLM_encoding, final_prompt), 1))
    return data

def send_request(data):
    global HEADER
    openaikey = data.pop("openaikey")
    if use_completion:
        data = convert_chat_to_completion(data)
    if openaikey and openaikey.startswith("sk-"):
        HEADER = {
            "Authorization": f"Bearer {openaikey}"
        }
    
    response = requests.post(endpoint, json=data, headers=HEADER, proxies=PROXY)
    logger.debug(response.text.strip())
    if "choices" not in response.json():
        return response.json()
    if use_completion:
        return response.json()["choices"][0]["text"].strip()
    else:
        return response.json()["choices"][0]["message"]["content"].strip()

def replace_slot(text, entries):
    for key, value in entries.items():
        if not isinstance(value, str):
            value = str(value)
        text = text.replace("{{" + key +"}}", value.replace('"', "'").replace('\n', ""))
    return text

def find_json(s):
    s = s.replace("\'", "\"")
    start = s.find("{")
    end = s.rfind("}")
    res = s[start:end+1]
    res = res.replace("\n", "")
    return res

def field_extract(s, field):
    try:
        field_rep = re.compile(f'{field}.*?:.*?"(.*?)"', re.IGNORECASE)
        extracted = field_rep.search(s).group(1).replace("\"", "\'")
    except:
        field_rep = re.compile(f'{field}:\ *"(.*?)"', re.IGNORECASE)
        extracted = field_rep.search(s).group(1).replace("\"", "\'")
    return extracted

def get_id_reason(choose_str):
    reason = field_extract(choose_str, "reason")
    id = field_extract(choose_str, "id")
    choose = {"id": id, "reason": reason}
    return id.strip(), reason.strip(), choose

def record_case(success, **args):
    if success:
        f = open("logs/log_success.jsonl", "a")
    else:
        f = open("logs/log_fail.jsonl", "a")
    log = args
    f.write(json.dumps(log) + "\n")
    f.close()

def image_to_bytes(img_url):
    img_byte = io.BytesIO()
    type = img_url.split(".")[-1]
    load_image(img_url).save(img_byte, format="png")
    img_data = img_byte.getvalue()
    return img_data

def resource_has_dep(command):
    args = command["args"]
    for _, v in args.items():
        if "<GENERATED>" in v:
            return True
    return False

def fix_dep(tasks):
    for task in tasks:
        args = task["args"]
        task["dep"] = []
        for k, v in args.items():
            if "<GENERATED>" in v:
                dep_task_id = int(v.split("-")[1])
                if dep_task_id not in task["dep"]:
                    task["dep"].append(dep_task_id)
        if len(task["dep"]) == 0:
            task["dep"] = [-1]
    return tasks

def unfold(tasks):
    flag_unfold_task = False
    try:
        for task in tasks:
            for key, value in task["args"].items():
                if "<GENERATED>" in value:
                    generated_items = value.split(",")
                    if len(generated_items) > 1:
                        flag_unfold_task = True
                        for item in generated_items:
                            new_task = copy.deepcopy(task)
                            dep_task_id = int(item.split("-")[1])
                            new_task["dep"] = [dep_task_id]
                            new_task["args"][key] = item
                            tasks.append(new_task)
                        tasks.remove(task)
    except Exception as e:
        print(e)
        traceback.print_exc()
        logger.debug("unfold task failed.")

    if flag_unfold_task:
        logger.debug(f"unfold tasks: {tasks}")
        
    return tasks

def chitchat(messages, openaikey=None):
    data = {
        "model": LLM,
        "messages": messages,
        "openaikey": openaikey
    }
    return send_request(data)

def parse_task(context, input, openaikey=None):
    demos_or_presteps = parse_task_demos_or_presteps
    messages = json.loads(demos_or_presteps)
    messages.insert(0, {"role": "system", "content": parse_task_tprompt})

    # cut chat logs
    start = 0
    while start <= len(context):
        history = context[start:]
        prompt = replace_slot(parse_task_prompt, {
            "input": input,
            "context": history 
        })
        messages.append({"role": "user", "content": prompt})
        history_text = "<im_end>\nuser<im_start>".join([m["content"] for m in messages])
        num = count_tokens(LLM_encoding, history_text)
        if get_max_context_length(LLM) - num > 800:
            break
        messages.pop()
        start += 2
    
    logger.debug(messages)
    data = {
        "model": LLM,
        "messages": messages,
        "temperature": 0,
        "logit_bias": {item: config["logit_bias"]["parse_task"] for item in task_parsing_highlight_ids},
        "openaikey": openaikey
    }
    return send_request(data)

def choose_model(input, task, metas, openaikey = None):
    prompt = replace_slot(choose_model_prompt, {
        "input": input,
        "task": task,
        "metas": metas,
    })
    demos_or_presteps = replace_slot(choose_model_demos_or_presteps, {
        "input": input,
        "task": task,
        "metas": metas
    })
    messages = json.loads(demos_or_presteps)
    messages.insert(0, {"role": "system", "content": choose_model_tprompt})
    messages.append({"role": "user", "content": prompt})
    logger.debug(messages)
    data = {
        "model": LLM,
        "messages": messages,
        "temperature": 0,
        "logit_bias": {item: config["logit_bias"]["choose_model"] for item in choose_model_highlight_ids}, # 5
        "openaikey": openaikey
    }
    return send_request(data)


def response_results(input, results, openaikey=None):
    results = [v for k, v in sorted(results.items(), key=lambda item: item[0])]
    prompt = replace_slot(response_results_prompt, {
        "input": input,
    })
    demos_or_presteps = replace_slot(response_results_demos_or_presteps, {
        "input": input,
        "processes": results
    })
    messages = json.loads(demos_or_presteps)
    messages.insert(0, {"role": "system", "content": response_results_tprompt})
    messages.append({"role": "user", "content": prompt})
    logger.debug(messages)
    data = {
        "model": LLM,
        "messages": messages,
        "temperature": 0,
        "openaikey": openaikey
    }
    return send_request(data)

def huggingface_model_inference(model_id, data, task, huggingfacetoken=None):
    if huggingfacetoken is None:
        HUGGINGFACE_HEADERS = {}
    else:
        HUGGINGFACE_HEADERS = {
            "Authorization": f"Bearer {huggingfacetoken}",
    }
    task_url = f"https://api-inference.huggingface.co/models/{model_id}" # InferenceApi does not yet support some tasks
    inference = InferenceApi(repo_id=model_id, token=huggingfacetoken)
    
    # NLP tasks
    if task == "question-answering":
        inputs = {"question": data["text"], "context": (data["context"] if "context" in data else "" )}
        result = inference(inputs)
    if task == "sentence-similarity":
        inputs = {"source_sentence": data["text1"], "target_sentence": data["text2"]}
        result = inference(inputs)
    if task in ["text-classification",  "token-classification", "text2text-generation", "summarization", "translation", "conversational", "text-generation"]:
        inputs = data["text"]
        result = inference(inputs)
    
    # CV tasks
    if task == "visual-question-answering" or task == "document-question-answering":
        img_url = data["image"]
        text = data["text"]
        img_data = image_to_bytes(img_url)
        img_base64 = base64.b64encode(img_data).decode("utf-8")
        json_data = {}
        json_data["inputs"] = {}
        json_data["inputs"]["question"] = text
        json_data["inputs"]["image"] = img_base64
        result = requests.post(task_url, headers=HUGGINGFACE_HEADERS, json=json_data).json()
        # result = inference(inputs) # not support

    if task == "image-to-image":
        img_url = data["image"]
        img_data = image_to_bytes(img_url)
        # result = inference(data=img_data) # not support
        HUGGINGFACE_HEADERS["Content-Length"] = str(len(img_data))
        r = requests.post(task_url, headers=HUGGINGFACE_HEADERS, data=img_data)
        result = r.json()
        if "path" in result:
            result["generated image"] = result.pop("path")
    
    if task == "text-to-image":
        inputs = data["text"]
        img = inference(inputs)
        name = str(uuid.uuid4())[:4]
        img.save(f"public/images/{name}.png")
        result = {}
        result["generated image"] = f"/images/{name}.png"

    if task == "image-segmentation":
        img_url = data["image"]
        img_data = image_to_bytes(img_url)
        image = Image.open(BytesIO(img_data))
        predicted = inference(data=img_data)
        colors = []
        for i in range(len(predicted)):
            colors.append((random.randint(100, 255), random.randint(100, 255), random.randint(100, 255), 155))
        for i, pred in enumerate(predicted):
            label = pred["label"]
            mask = pred.pop("mask").encode("utf-8")
            mask = base64.b64decode(mask)
            mask = Image.open(BytesIO(mask), mode='r')
            mask = mask.convert('L')

            layer = Image.new('RGBA', mask.size, colors[i])
            image.paste(layer, (0, 0), mask)
        name = str(uuid.uuid4())[:4]
        image.save(f"public/images/{name}.jpg")
        result = {}
        result["generated image with segmentation mask"] = f"/images/{name}.jpg"
        result["predicted"] = predicted

    if task == "object-detection":
        img_url = data["image"]
        img_data = image_to_bytes(img_url)
        predicted = inference(data=img_data)
        image = Image.open(BytesIO(img_data))
        draw = ImageDraw.Draw(image)
        labels = list(item['label'] for item in predicted)
        color_map = {}
        for label in labels:
            if label not in color_map:
                color_map[label] = (random.randint(0, 255), random.randint(0, 100), random.randint(0, 255))
        for label in predicted:
            box = label["box"]
            draw.rectangle(((box["xmin"], box["ymin"]), (box["xmax"], box["ymax"])), outline=color_map[label["label"]], width=2)
            draw.text((box["xmin"]+5, box["ymin"]-15), label["label"], fill=color_map[label["label"]])
        name = str(uuid.uuid4())[:4]
        image.save(f"public/images/{name}.jpg")
        result = {}
        result["generated image with predicted box"] = f"/images/{name}.jpg"
        result["predicted"] = predicted

    if task in ["image-classification"]:
        img_url = data["image"]
        img_data = image_to_bytes(img_url)
        result = inference(data=img_data)
 
    if task == "image-to-text":
        img_url = data["image"]
        img_data = image_to_bytes(img_url)
        HUGGINGFACE_HEADERS["Content-Length"] = str(len(img_data))
        r = requests.post(task_url, headers=HUGGINGFACE_HEADERS, data=img_data)
        result = {}
        if "generated_text" in r.json()[0]:
            result["generated text"] = r.json()[0].pop("generated_text")
    
    # AUDIO tasks
    if task == "text-to-speech":
        inputs = data["text"]
        response = inference(inputs, raw_response=True)
        # response = requests.post(task_url, headers=HUGGINGFACE_HEADERS, json={"inputs": text})
        name = str(uuid.uuid4())[:4]
        with open(f"public/audios/{name}.flac", "wb") as f:
            f.write(response.content)
        result = {"generated audio": f"/audios/{name}.flac"}
    if task in ["automatic-speech-recognition", "audio-to-audio", "audio-classification"]:
        audio_url = data["audio"]
        audio_data = requests.get(audio_url, timeout=10).content
        response = inference(data=audio_data, raw_response=True)
        result = response.json()
        if task == "audio-to-audio":
            content = None
            type = None
            for k, v in result[0].items():
                if k == "blob":
                    content = base64.b64decode(v.encode("utf-8"))
                if k == "content-type":
                    type = "audio/flac".split("/")[-1]
            audio = AudioSegment.from_file(BytesIO(content))
            name = str(uuid.uuid4())[:4]
            audio.export(f"public/audios/{name}.{type}", format=type)
            result = {"generated audio": f"/audios/{name}.{type}"}
    return result

def local_model_inference(model_id, data, task):
    inference = partial(models, model_id)
    # contronlet
    if model_id.startswith("lllyasviel/sd-controlnet-"):
        img_url = data["image"]
        text = data["text"]
        results = inference({"img_url": img_url, "text": text})
        if "path" in results:
            results["generated image"] = results.pop("path")
        return results
    if model_id.endswith("-control"):
        img_url = data["image"]
        results = inference({"img_url": img_url})
        if "path" in results:
            results["generated image"] = results.pop("path")
        return results
        
    if task == "text-to-video":
        results = inference(data)
        if "path" in results:
            results["generated video"] = results.pop("path")
        return results

    # NLP tasks
    if task == "question-answering" or task == "sentence-similarity":
        results = inference(json=data)
        return results
    if task in ["text-classification",  "token-classification", "text2text-generation", "summarization", "translation", "conversational", "text-generation"]:
        results = inference(json=data)
        return results

    # CV tasks
    if task == "depth-estimation":
        img_url = data["image"]
        results = inference({"img_url": img_url})
        if "path" in results:
            results["generated depth image"] = results.pop("path")
        return results
    if task == "image-segmentation":
        img_url = data["image"]
        results = inference({"img_url": img_url})
        results["generated image with segmentation mask"] = results.pop("path")
        return results
    if task == "image-to-image":
        img_url = data["image"]
        results = inference({"img_url": img_url})
        if "path" in results:
            results["generated image"] = results.pop("path")
        return results
    if task == "text-to-image":
        results = inference(data)
        if "path" in results:
            results["generated image"] = results.pop("path")
        return results
    if task == "object-detection":
        img_url = data["image"]
        predicted = inference({"img_url": img_url})
        if "error" in predicted:
            return predicted
        image = load_image(img_url)
        draw = ImageDraw.Draw(image)
        labels = list(item['label'] for item in predicted)
        color_map = {}
        for label in labels:
            if label not in color_map:
                color_map[label] = (random.randint(0, 255), random.randint(0, 100), random.randint(0, 255))
        for label in predicted:
            box = label["box"]
            draw.rectangle(((box["xmin"], box["ymin"]), (box["xmax"], box["ymax"])), outline=color_map[label["label"]], width=2)
            draw.text((box["xmin"]+5, box["ymin"]-15), label["label"], fill=color_map[label["label"]])
        name = str(uuid.uuid4())[:4]
        image.save(f"public/images/{name}.jpg")
        results = {}
        results["generated image with predicted box"] = f"/images/{name}.jpg"
        results["predicted"] = predicted
        return results
    if task in ["image-classification", "image-to-text", "document-question-answering", "visual-question-answering"]:
        img_url = data["image"]
        text = None
        if "text" in data:
            text = data["text"]
        results = inference({"img_url": img_url, "text": text})
        return results
    # AUDIO tasks
    if task == "text-to-speech":
        results = inference(data)
        if "path" in results:
            results["generated audio"] = results.pop("path")
        return results
    if task in ["automatic-speech-recognition", "audio-to-audio", "audio-classification"]:
        audio_url = data["audio"]
        results = inference({"audio_url": audio_url})
        return results


def model_inference(model_id, data, hosted_on, task, huggingfacetoken=None):
    if huggingfacetoken:
        HUGGINGFACE_HEADERS = {
            "Authorization": f"Bearer {huggingfacetoken}",
        }
    else:
        HUGGINGFACE_HEADERS = None
    if hosted_on == "unknown":
        r = status(model_id)
        logger.debug("Local Server Status: " + str(r))
        if "loaded" in r and r["loaded"]:
            hosted_on = "local"
        else:
            huggingfaceStatusUrl = f"https://api-inference.huggingface.co/status/{model_id}"
            r = requests.get(huggingfaceStatusUrl, headers=HUGGINGFACE_HEADERS, proxies=PROXY)
            logger.debug("Huggingface Status: " + str(r.json()))
            if "loaded" in r and r["loaded"]:
                hosted_on = "huggingface"
    try:
        if hosted_on == "local":
            inference_result = local_model_inference(model_id, data, task)
        elif hosted_on == "huggingface":
            inference_result = huggingface_model_inference(model_id, data, task, huggingfacetoken)
    except Exception as e:
        print(e)
        traceback.print_exc()
        inference_result = {"error":{"message": str(e)}}
    return inference_result


def get_model_status(model_id, url, headers, queue = None):
    endpoint_type = "huggingface" if "huggingface" in url else "local"
    if "huggingface" in url:
        r = requests.get(url, headers=headers, proxies=PROXY)
    else:
        r = status(model_id)
    if "loaded" in r and r["loaded"]:
        if queue:
            queue.put((model_id, True, endpoint_type))
        return True
    else:
        if queue:
            queue.put((model_id, False, None))
        return False

def get_avaliable_models(candidates, topk=10, huggingfacetoken = None):
    all_available_models = {"local": [], "huggingface": []}
    threads = []
    result_queue = Queue()
    HUGGINGFACE_HEADERS = {
        "Authorization": f"Bearer {huggingfacetoken}",
    }
    for candidate in candidates:
        model_id = candidate["id"]

        if inference_mode != "local":
            huggingfaceStatusUrl = f"https://api-inference.huggingface.co/status/{model_id}"
            thread = threading.Thread(target=get_model_status, args=(model_id, huggingfaceStatusUrl, HUGGINGFACE_HEADERS, result_queue))
            threads.append(thread)
            thread.start()
        
        if inference_mode != "huggingface" and config["local_deployment"] != "minimal":
            thread = threading.Thread(target=get_model_status, args=(model_id, "", {}, result_queue))
            threads.append(thread)
            thread.start()
        
    result_count = len(threads)
    while result_count:
        model_id, status, endpoint_type = result_queue.get()
        if status and model_id not in all_available_models:
            all_available_models[endpoint_type].append(model_id)
        if len(all_available_models["local"] + all_available_models["huggingface"]) >= topk:
            break
        result_count -= 1

    for thread in threads:
        thread.join()

    return all_available_models

def collect_result(command, choose, inference_result):
    result = {"task": command}
    result["inference result"] = inference_result
    result["choose model result"] = choose
    logger.debug(f"inference result: {inference_result}")
    return result


def run_task(input, command, results, openaikey = None, huggingfacetoken = None):
    id = command["id"]
    args = command["args"]
    task = command["task"]
    deps = command["dep"]
    if deps[0] != -1:
        dep_tasks = [results[dep] for dep in deps]
    else:
        dep_tasks = []
    
    logger.debug(f"Run task: {id} - {task}")
    logger.debug("Deps: " + json.dumps(dep_tasks))

    if deps[0] != -1:
        if "image" in args and "<GENERATED>-" in args["image"]:
            resource_id = int(args["image"].split("-")[1])
            if "generated image" in results[resource_id]["inference result"]:
                args["image"] = results[resource_id]["inference result"]["generated image"]
        if "audio" in args and "<GENERATED>-" in args["audio"]:
            resource_id = int(args["audio"].split("-")[1])
            if "generated audio" in results[resource_id]["inference result"]:
                args["audio"] = results[resource_id]["inference result"]["generated audio"]
        if "text" in args and "<GENERATED>-" in args["text"]:
            resource_id = int(args["text"].split("-")[1])
            if "generated text" in results[resource_id]["inference result"]:
                args["text"] = results[resource_id]["inference result"]["generated text"]

    text = image = audio = None
    for dep_task in dep_tasks:
        if "generated text" in dep_task["inference result"]:
            text = dep_task["inference result"]["generated text"]
            logger.debug("Detect the generated text of dependency task (from results):" + text)
        elif "text" in dep_task["task"]["args"]:
            text = dep_task["task"]["args"]["text"]
            logger.debug("Detect the text of dependency task (from args): " + text)
        if "generated image" in dep_task["inference result"]:
            image = dep_task["inference result"]["generated image"]
            logger.debug("Detect the generated image of dependency task (from results): " + image)
        elif "image" in dep_task["task"]["args"]:
            image = dep_task["task"]["args"]["image"]
            logger.debug("Detect the image of dependency task (from args): " + image)
        if "generated audio" in dep_task["inference result"]:
            audio = dep_task["inference result"]["generated audio"]
            logger.debug("Detect the generated audio of dependency task (from results): " + audio)
        elif "audio" in dep_task["task"]["args"]:
            audio = dep_task["task"]["args"]["audio"]
            logger.debug("Detect the audio of dependency task (from args): " + audio)

    if "image" in args and "<GENERATED>" in args["image"]:
        if image:
            args["image"] = image
    if "audio" in args and "<GENERATED>" in args["audio"]:
        if audio:
            args["audio"] = audio
    if "text" in args and "<GENERATED>" in args["text"]:
        if text:
            args["text"] = text

    for resource in ["image", "audio"]:
        if resource in args and not args[resource].startswith("public/") and len(args[resource]) > 0 and not args[resource].startswith("http"):
            args[resource] = f"public/{args[resource]}"
    
    if "-text-to-image" in command['task'] and "text" not in args:
        logger.debug("control-text-to-image task, but text is empty, so we use control-generation instead.")
        control = task.split("-")[0]
        
        if control == "seg":
            task = "image-segmentation"
            command['task'] = task
        elif control == "depth":
            task = "depth-estimation"
            command['task'] = task
        else:
            task = f"{control}-control"

    command["args"] = args
    logger.debug(f"parsed task: {command}")

    if task.endswith("-text-to-image") or task.endswith("-control"):
        if inference_mode != "huggingface":
            if task.endswith("-text-to-image"):
                control = task.split("-")[0]
                best_model_id = f"lllyasviel/sd-controlnet-{control}"
            else:
                best_model_id = task
            hosted_on = "local"
            reason = "ControlNet is the best model for this task."
            choose = {"id": best_model_id, "reason": reason}
            logger.debug(f"chosen model: {choose}")
        else:
            logger.warning(f"Task {command['task']} is not available. ControlNet need to be deployed locally.")
            record_case(success=False, **{"input": input, "task": command, "reason": f"Task {command['task']} is not available. ControlNet need to be deployed locally.", "op":"message"})
            inference_result = {"error": f"service related to ControlNet is not available."}
            results[id] = collect_result(command, "", inference_result)
            return False
    elif task in ["summarization", "translation", "conversational", "text-generation", "text2text-generation"]: # ChatGPT Can do
        best_model_id = "ChatGPT"
        reason = "ChatGPT is the best model for this task."
        choose = {"id": best_model_id, "reason": reason}
        messages = [{
            "role": "user",
            "content": f"[ {input} ] contains a task in JSON format {command}, 'task' indicates the task type and 'args' indicates the arguments required for the task. Don't explain the task to me, just help me do it and give me the result. The result must be in text form without any urls."
        }]
        response = chitchat(messages, openaikey)
        results[id] = collect_result(command, choose, {"response": response})
        return True
    else:
        if task not in MODELS_MAP:
            logger.warning(f"no available models on {task} task.")
            record_case(success=False, **{"input": input, "task": command, "reason": f"task not support: {command['task']}", "op":"message"})
            inference_result = {"error": f"{command['task']} not found in available tasks."}
            results[id] = collect_result(command, "", inference_result)
            return False

        candidates = MODELS_MAP[task][:20]
        all_avaliable_models = get_avaliable_models(candidates, config["num_candidate_models"], huggingfacetoken)
        all_avaliable_model_ids = all_avaliable_models["local"] + all_avaliable_models["huggingface"]
        logger.debug(f"avaliable models on {command['task']}: {all_avaliable_models}")

        if len(all_avaliable_model_ids) == 0:
            logger.warning(f"no available models on {command['task']}")
            record_case(success=False, **{"input": input, "task": command, "reason": f"no available models: {command['task']}", "op":"message"})
            inference_result = {"error": f"no available models on {command['task']} task."}
            results[id] = collect_result(command, "", inference_result)
            return False
            
        if len(all_avaliable_model_ids) == 1:
            best_model_id = all_avaliable_model_ids[0]
            hosted_on = "local" if best_model_id in all_avaliable_models["local"] else "huggingface"
            reason = "Only one model available."
            choose = {"id": best_model_id, "reason": reason}
            logger.debug(f"chosen model: {choose}")
        else:
            cand_models_info = [
                {
                    "id": model["id"],
                    "inference endpoint": all_avaliable_models.get(
                        "local" if model["id"] in all_avaliable_models["local"] else "huggingface"
                    ),
                    "likes": model.get("likes"),
                    "description": model.get("description", "")[:config["max_description_length"]],
                    "language": model.get("language"),
                    "tags": model.get("tags"),
                }
                for model in candidates
                if model["id"] in all_avaliable_model_ids
            ]

            choose_str = choose_model(input, command, cand_models_info, openaikey)
            logger.debug(f"chosen model: {choose_str}")
            try:
                choose = json.loads(choose_str)
                reason = choose["reason"]
                best_model_id = choose["id"]
                hosted_on = "local" if best_model_id in all_avaliable_models["local"] else "huggingface"
            except Exception as e:
                logger.warning(f"the response [ {choose_str} ] is not a valid JSON, try to find the model id and reason in the response.")
                choose_str = find_json(choose_str)
                best_model_id, reason, choose  = get_id_reason(choose_str)
                hosted_on = "local" if best_model_id in all_avaliable_models["local"] else "huggingface"
    inference_result = model_inference(best_model_id, args, hosted_on, command['task'], huggingfacetoken)

    if "error" in inference_result:
        logger.warning(f"Inference error: {inference_result['error']}")
        record_case(success=False, **{"input": input, "task": command, "reason": f"inference error: {inference_result['error']}", "op":"message"})
        results[id] = collect_result(command, choose, inference_result)
        return False
    
    results[id] = collect_result(command, choose, inference_result)
    return True

def chat_huggingface(messages, openaikey = None, huggingfacetoken = None, return_planning = False, return_results = False):
    start = time.time()
    context = messages[:-1]
    input = messages[-1]["content"]
    logger.info("*"*80)
    logger.info(f"input: {input}")

    task_str = parse_task(context, input, openaikey)
    logger.info(task_str)

    if "error" in task_str:
        return {"message": str(task_str)}
    else:
        task_str = task_str.strip()

    if task_str == "[]":  # using LLM response for empty task
        record_case(success=False, **{"input": input, "task": [], "reason": "task parsing fail: empty", "op": "chitchat"})
        response = chitchat(messages, openaikey)
        return {"message": response}
    try:
        tasks = json.loads(task_str)
    except Exception as e:
        logger.debug(e)
        response = chitchat(messages, openaikey)
        record_case(success=False, **{"input": input, "task": task_str, "reason": "task parsing fail", "op":"chitchat"})
        return {"message": response}
    

    tasks = unfold(tasks)
    tasks = fix_dep(tasks)
    logger.debug(tasks)
    
    if return_planning:
        return tasks

    results = {}
    threads = []
    tasks = tasks[:]
    d = dict()
    retry = 0
    while True:
        num_threads = len(threads)
        for task in tasks:
            dep = task["dep"]
            # logger.debug(f"d.keys(): {d.keys()}, dep: {dep}")
            for dep_id in dep:
                if dep_id >= task["id"]:
                    task["dep"] = [-1]
                    dep = [-1]
                    break
            if len(list(set(dep).intersection(d.keys()))) == len(dep) or dep[0] == -1:
                tasks.remove(task)
                thread = threading.Thread(target=run_task, args=(input, task, d, openaikey, huggingfacetoken))
                thread.start()
                threads.append(thread)
        if num_threads == len(threads):
            time.sleep(0.5)
            retry += 1
        if retry > 160:
            logger.debug("User has waited too long, Loop break.")
            break
        if len(tasks) == 0:
            break
    for thread in threads:
        thread.join()
    
    results = d.copy()

    logger.debug(results)
    if return_results:
        return results
    
    response = response_results(input, results, openaikey).strip()

    end = time.time()
    during = end - start

    answer = {"message": response}
    record_case(success=True, **{"input": input, "task": task_str, "results": results, "response": response, "during": during, "op":"response"})
    logger.info(f"response: {response}")
    return answer