initial commit

.gitignore

@@ -0,0 +1,173 @@
+logs/
+.idea/
+models
+public/*
+*.pyc
+!public/examples
+
+.idea/
+
+# ALL
+*.dev
+
+# for server
+server/models/*
+!server/models/download.sh
+!server/models/download.ps1
+server/logs/
+server/models_dev
+server/public/*
+!server/public/examples/
+server/public/examples/*
+!server/public/examples/a.jpg
+!server/public/examples/b.jpg
+!server/public/examples/c.jpg
+!server/public/examples/d.jpg
+!server/public/examples/e.jpg
+!server/public/examples/f.jpg
+!server/public/examples/g.jpg
+
+# docker
+Dockerfile
+docker-compose.yml
+
+# for gradio
+# server/run_gradio.py
+
+# for web
+web/node_modules
+web/package-lock.json
+web/dist
+web/electron-dist
+web/yarn.lock
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+

app.py

@@ -0,0 +1,202 @@
+import uuid
+import gradio as gr
+import re
+from diffusers.utils import load_image
+import requests
+from awesome_chat import chat_huggingface
+import os
+
+os.makedirs("public/images", exist_ok=True)
+os.makedirs("public/audios", exist_ok=True)
+os.makedirs("public/videos", exist_ok=True)
+
+class Client:
+    def __init__(self) -> None:
+        self.OPENAI_KEY = ""
+        self.HUGGINGFACE_TOKEN = ""
+        self.all_messages = []
+
+    def set_key(self, openai_key):
+        self.OPENAI_KEY = openai_key
+        if len(self.HUGGINGFACE_TOKEN)>0:
+            gr.update(visible = True)
+        return self.OPENAI_KEY
+
+    def set_token(self, huggingface_token):
+        self.HUGGINGFACE_TOKEN = huggingface_token
+        if len(self.OPENAI_KEY)>0:
+            gr.update(visible = True)
+        return self.HUGGINGFACE_TOKEN
+    
+    def add_message(self, content, role):
+        message = {"role":role, "content":content}
+        self.all_messages.append(message)
+
+    def extract_medias(self, message):
+        # url_pattern = re.compile(r"(http(s?):|\/)?([\.\/_\w:-])*?")
+        urls = []
+        # for match in url_pattern.finditer(message):
+        #     if match.group(0) not in urls:
+        #         urls.append(match.group(0))
+
+        image_pattern = re.compile(r"(http(s?):|\/)?([\.\/_\w:-])*?\.(jpg|jpeg|tiff|gif|png)")
+        image_urls = []
+        for match in image_pattern.finditer(message):
+            if match.group(0) not in image_urls:
+                image_urls.append(match.group(0))
+
+        audio_pattern = re.compile(r"(http(s?):|\/)?([\.\/_\w:-])*?\.(flac|wav)")
+        audio_urls = []
+        for match in audio_pattern.finditer(message):
+            if match.group(0) not in audio_urls:
+                audio_urls.append(match.group(0))
+
+        video_pattern = re.compile(r"(http(s?):|\/)?([\.\/_\w:-])*?\.(mp4)")
+        video_urls = []
+        for match in video_pattern.finditer(message):
+            if match.group(0) not in video_urls:
+                video_urls.append(match.group(0))
+
+        return urls, image_urls, audio_urls, video_urls
+
+    def add_text(self, messages, message):
+        if len(self.OPENAI_KEY) == 0 or not self.OPENAI_KEY.startswith("sk-") or len(self.HUGGINGFACE_TOKEN) == 0 or not self.HUGGINGFACE_TOKEN.startswith("hf_"):
+            return messages, "Please set your OpenAI API key and Hugging Face token first!!!"
+        self.add_message(message, "user")
+        messages = messages + [(message, None)]
+        urls, image_urls, audio_urls, video_urls = self.extract_medias(message)
+
+        for image_url in image_urls:
+            if not image_url.startswith("http") and not image_url.startswith("public"):
+                image_url = "public/" + image_url
+            image = load_image(image_url)
+            name = f"public/images/{str(uuid.uuid4())[:4]}.jpg" 
+            image.save(name)
+            messages = messages + [((f"{name}",), None)]
+        for audio_url in audio_urls and not audio_url.startswith("public"):
+            if not audio_url.startswith("http"):
+                audio_url = "public/" + audio_url
+            ext = audio_url.split(".")[-1]
+            name = f"public/audios/{str(uuid.uuid4()[:4])}.{ext}"
+            response = requests.get(audio_url)
+            with open(name, "wb") as f:
+                f.write(response.content)
+            messages = messages + [((f"{name}",), None)]
+        for video_url in video_urls and not video_url.startswith("public"):
+            if not video_url.startswith("http"):
+                video_url = "public/" + video_url
+            ext = video_url.split(".")[-1]
+            name = f"public/audios/{str(uuid.uuid4()[:4])}.{ext}"
+            response = requests.get(video_url)
+            with open(name, "wb") as f:
+                f.write(response.content)
+            messages = messages + [((f"{name}",), None)]
+        return messages, ""
+
+    def bot(self, messages):
+        if len(self.OPENAI_KEY) == 0 or not self.OPENAI_KEY.startswith("sk-") or len(self.HUGGINGFACE_TOKEN) == 0 or not self.HUGGINGFACE_TOKEN.startswith("hf_"):
+            return messages, {}
+        message, results = chat_huggingface(self.all_messages, self.OPENAI_KEY, self.HUGGINGFACE_TOKEN)
+        urls, image_urls, audio_urls, video_urls = self.extract_medias(message)
+        self.add_message(message, "assistant")
+        messages[-1][1] = message
+        for image_url in image_urls:
+            if not image_url.startswith("http"):
+                image_url = image_url.replace("public/", "")
+                messages = messages + [((None, (f"public/{image_url}",)))]
+            # else:
+            #     messages = messages + [((None, (f"{image_url}",)))]
+        for audio_url in audio_urls:
+            if not audio_url.startswith("http"):
+                audio_url = audio_url.replace("public/", "")
+                messages = messages + [((None, (f"public/{audio_url}",)))]
+            # else:
+            #     messages = messages + [((None, (f"{audio_url}",)))]
+        for video_url in video_urls:
+            if not video_url.startswith("http"):
+                video_url = video_url.replace("public/", "")
+                messages = messages + [((None, (f"public/{video_url}",)))]
+            # else:
+            #     messages = messages + [((None, (f"{video_url}",)))]
+        # replace int key to string key
+        results = {str(k): v for k, v in results.items()}
+        return messages, results
+    
+css = ".json {height: 527px; overflow: scroll;} .json-holder {height: 527px; overflow: scroll;}"
+with gr.Blocks(css=css) as demo:
+    state = gr.State(value={"client": Client()})
+    gr.Markdown("<h1><center>HuggingGPT</center></h1>")
+    gr.Markdown("<p align='center'><img src='https://i.ibb.co/qNH3Jym/logo.png' height='25' width='95'></p>")
+    gr.Markdown("<p align='center' style='font-size: 20px;'>A system to connect LLMs with ML community. See our <a href='https://github.com/microsoft/JARVIS'>Project</a> and <a href='http://arxiv.org/abs/2303.17580'>Paper</a>.</p>")
+    gr.HTML('''<center><a href="https://huggingface.co/spaces/microsoft/HuggingGPT?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>Duplicate the Space and run securely with your OpenAI API Key and Hugging Face Token</center>''')
+    with gr.Row().style():
+        with gr.Column(scale=0.85):
+            openai_api_key = gr.Textbox(
+                show_label=False,
+                placeholder="Set your OpenAI API key here and press Enter",
+                lines=1,
+                type="password"
+            ).style(container=False)
+        with gr.Column(scale=0.15, min_width=0):
+            btn1 = gr.Button("Submit").style(full_height=True)
+
+    with gr.Row().style():
+        with gr.Column(scale=0.85):
+            hugging_face_token = gr.Textbox(
+                show_label=False,
+                placeholder="Set your Hugging Face Token here and press Enter",
+                lines=1,
+                type="password"
+            ).style(container=False)
+        with gr.Column(scale=0.15, min_width=0):
+            btn3 = gr.Button("Submit").style(full_height=True)
+    
+
+    with gr.Row().style():
+        with gr.Column(scale=0.6):
+            chatbot = gr.Chatbot([], elem_id="chatbot").style(height=500)
+        with gr.Column(scale=0.4):
+            results = gr.JSON(elem_classes="json")
+
+
+    with gr.Row().style():
+        with gr.Column(scale=0.85):
+            txt = gr.Textbox(
+                show_label=False,
+                placeholder="Enter text and press enter. The url of the multimedia resource must contain the extension name.",
+                lines=1,
+            ).style(container=False)
+        with gr.Column(scale=0.15, min_width=0):
+            btn2 = gr.Button("Send").style(full_height=True)
+        
+    def set_key(state, openai_api_key):
+        return state["client"].set_key(openai_api_key)
+
+    def add_text(state, chatbot, txt):
+        return state["client"].add_text(chatbot, txt)
+    
+    def set_token(state, hugging_face_token):
+        return state["client"].set_token(hugging_face_token)
+    
+    def bot(state, chatbot):
+        return state["client"].bot(chatbot)
+
+    openai_api_key.submit(set_key, [state, openai_api_key], [openai_api_key])
+    txt.submit(add_text, [state, chatbot, txt], [chatbot, txt]).then(bot, [state, chatbot], [chatbot, results])
+    hugging_face_token.submit(set_token, [state, hugging_face_token], [hugging_face_token])
+    btn1.click(set_key, [state, openai_api_key], [openai_api_key])
+    btn2.click(add_text, [state, chatbot, txt], [chatbot, txt]).then(bot, [state, chatbot], [chatbot, results])
+    btn3.click(set_token, [state, hugging_face_token], [hugging_face_token])
+
+    gr.Examples(
+        examples=["Given a collection of image A: /examples/a.jpg, B: /examples/b.jpg, C: /examples/c.jpg, please tell me how many zebras in these picture?",
+                "Please generate a canny image based on /examples/f.jpg",
+                "show me a joke and an image of cat",
+                "what is in the examples/a.jpg",
+                "based on the /examples/a.jpg, please generate a video and audio",
+                "based on pose of /examples/d.jpg and content of /examples/e.jpg, please show me a new image",
+                ],
+        inputs=txt
+    )
+
+demo.launch()

awesome_chat.py

@@ -0,0 +1,920 @@
+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 performs well on some NLP tasks as well."
+        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 str(task_str), {}
+    else:
+        task_str = task_str.strip()
+
+    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 response, {}
+
+    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 response, {}
+
+    if len(tasks)==1 and tasks[0]["task"] in ["summarization", "translation", "conversational", "text-generation", "text2text-generation"]:
+        record_case(success=True, **{"input": input, "task": tasks, "reason": "task parsing fail: empty", "op": "chitchat"})
+        response = chitchat(messages, openaikey)
+        best_model_id = "ChatGPT"
+        reason = "ChatGPT performs well on some NLP tasks as well."
+        choose = {"id": best_model_id, "reason": reason}
+        return response, collect_result(tasks[0], choose, {"response": 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 response, results

config.gradio.yaml

@@ -0,0 +1,34 @@
+openai:
+  key: gradio # "gradio" (set when request) or your_personal_key
+huggingface:
+  token: # required: huggingface token @ https://huggingface.co/settings/tokens
+dev: false
+debug: true
+log_file: logs/debug.log
+model: text-davinci-003 # text-davinci-003
+use_completion: true
+inference_mode: hybrid # local, huggingface or hybrid
+local_deployment: standard # minimal, standard or full
+num_candidate_models: 5
+max_description_length: 100
+proxy: 
+logit_bias:
+  parse_task: 0.5
+  choose_model: 5
+tprompt:
+  parse_task: >-
+    #1 Task Planning Stage: The AI assistant can parse user input to several tasks: [{"task": task, "id": task_id, "dep": dependency_task_id, "args": {"text": text or <GENERATED>-dep_id, "image": image_url or <GENERATED>-dep_id, "audio": audio_url or <GENERATED>-dep_id}}]. The special tag "<GENERATED>-dep_id" refer to the one genereted text/image/audio in the dependency task (Please consider whether the dependency task generates resources of this type.) and "dep_id" must be in "dep" list. The "dep" field denotes the ids of the previous prerequisite tasks which generate a new resource that the current task relies on. The "args" field must in ["text", "image", "audio"], nothing else. The task MUST be selected from the following options: "token-classification", "text2text-generation", "summarization", "translation", "question-answering", "conversational", "text-generation", "sentence-similarity", "tabular-classification", "object-detection", "image-classification", "image-to-image", "image-to-text", "text-to-image", "text-to-video", "visual-question-answering", "document-question-answering", "image-segmentation", "depth-estimation", "text-to-speech", "automatic-speech-recognition", "audio-to-audio", "audio-classification", "canny-control", "hed-control", "mlsd-control", "normal-control", "openpose-control", "canny-text-to-image", "depth-text-to-image", "hed-text-to-image", "mlsd-text-to-image", "normal-text-to-image", "openpose-text-to-image", "seg-text-to-image". There may be multiple tasks of the same type. Think step by step about all the tasks needed to resolve the user's request. Parse out as few tasks as possible while ensuring that the user request can be resolved. Pay attention to the dependencies and order among tasks. If the user input can't be parsed, you need to reply empty JSON []. 
+  choose_model: >-
+    #2 Model Selection Stage: Given the user request and the parsed tasks, the AI assistant helps the user to select a suitable model from a list of models to process the user request. The assistant should focus more on the description of the model and find the model that has the most potential to solve requests and tasks. Also, prefer models with local inference endpoints for speed and stability.
+  response_results: >-
+    #4 Response Generation Stage: With the task execution logs, the AI assistant needs to describe the process and inference results.
+demos_or_presteps:
+  parse_task: demos/demo_parse_task.json
+  choose_model: demos/demo_choose_model.json
+  response_results: demos/demo_response_results.json 
+prompt:
+  parse_task: The chat log [ {{context}} ] may contain the resources I mentioned. Now I input { {{input}} }. Pay attention to the input and output types of tasks and the dependencies between tasks.
+  choose_model: >-
+    Please choose the most suitable model from {{metas}} for the task {{task}}. The output must be in a strict JSON format: {"id": "id", "reason": "your detail reasons for the choice"}.
+  response_results: >-
+    Yes. Please first think carefully and directly answer my request based on the inference results. Some of the inferences may not always turn out to be correct and require you to make careful consideration in making decisions. Then please detail your workflow including the used models and inference results for my request in your friendly tone. Please filter out information that is not relevant to my request. Tell me the complete path or urls of files in inference results. If there is nothing in the results, please tell me you can't make it. }

demos/demo_choose_model.json

@@ -0,0 +1,10 @@
+[
+    {
+        "role": "user",
+        "content": "{{input}}"
+    },
+    {
+        "role": "assistant",
+        "content": "{{task}}"
+    }
+]

demos/demo_parse_task.json

@@ -0,0 +1,65 @@
+[
+    {
+        "role": "user",
+        "content": "Give you some pictures e1.jpg, e2.png, e3.jpg, help me count the number of sheep?"
+    },
+    {
+        "role": "assistant",
+        "content": "[{\"task\": \"image-to-text\", \"id\": 0, \"dep\": [-1], \"args\": {\"image\": \"e1.jpg\" }}, {\"task\": \"object-detection\", \"id\": 1, \"dep\": [-1], \"args\": {\"image\": \"e1.jpg\" }}, {\"task\": \"visual-question-answering\", \"id\": 2, \"dep\": [1], \"args\": {\"image\": \"<GENERATED>-1\", \"text\": \"How many sheep in the picture\"}} }}, {\"task\": \"image-to-text\", \"id\": 3, \"dep\": [-1], \"args\": {\"image\": \"e2.png\" }}, {\"task\": \"object-detection\", \"id\": 4, \"dep\": [-1], \"args\": {\"image\": \"e2.png\" }}, {\"task\": \"visual-question-answering\", \"id\": 5, \"dep\": [4], \"args\": {\"image\": \"<GENERATED>-4\", \"text\": \"How many sheep in the picture\"}} }}, {\"task\": \"image-to-text\", \"id\": 6, \"dep\": [-1], \"args\": {\"image\": \"e3.jpg\" }},  {\"task\": \"object-detection\", \"id\": 7, \"dep\": [-1], \"args\": {\"image\": \"e3.jpg\" }}, {\"task\": \"visual-question-answering\", \"id\": 8, \"dep\": [7], \"args\": {\"image\": \"<GENERATED>-7\", \"text\": \"How many sheep in the picture\"}}]"
+    },
+
+    {
+        "role":"user",
+        "content":"Look at /e.jpg, can you tell me how many objects in the picture? Give me a picture and video similar to this one."
+    },
+    {
+        "role":"assistant",
+        "content":"[{\"task\": \"image-to-text\", \"id\": 0, \"dep\": [-1], \"args\": {\"image\": \"/e.jpg\" }}, {\"task\": \"object-detection\", \"id\": 1, \"dep\": [-1], \"args\": {\"image\": \"/e.jpg\" }}, {\"task\": \"visual-question-answering\", \"id\": 2, \"dep\": [1], \"args\": {\"image\": \"<GENERATED>-1\", \"text\": \"how many objects in the picture?\" }}, {\"task\": \"text-to-image\", \"id\": 3, \"dep\": [0], \"args\": {\"text\": \"<GENERATED-0>\" }}, {\"task\": \"image-to-image\", \"id\": 4, \"dep\": [-1], \"args\": {\"image\": \"/e.jpg\" }}, {\"task\": \"text-to-video\", \"id\": 5, \"dep\": [0], \"args\": {\"text\": \"<GENERATED-0>\" }}]"
+    },
+
+    {
+        "role":"user",
+        "content":"given a document /images/e.jpeg, answer me what is the student amount? And describe the image with your voice"
+    },
+    {
+        "role":"assistant",
+        "content":"{\"task\": \"document-question-answering\", \"id\": 0, \"dep\": [-1], \"args\": {\"image\": \"/images/e.jpeg\", \"text\": \"what is the student amount?\" }}, {\"task\": \"visual-question-answering\", \"id\": 1, \"dep\": [-1], \"args\": {\"image\": \"/images/e.jpeg\", \"text\": \"what is the student amount?\" }}, {\"task\": \"image-to-text\", \"id\": 2, \"dep\": [-1], \"args\": {\"image\": \"/images/e.jpg\" }}, {\"task\": \"text-to-speech\", \"id\": 3, \"dep\": [2], \"args\": {\"text\": \"<GENERATED>-2\" }}]"
+    },
+
+    {
+        "role": "user",
+        "content": "Given an image /example.jpg, first generate a hed image, then based on the hed image generate a new image where a girl is reading a book"
+    },
+    {
+        "role": "assistant",
+        "content": "[{\"task\": \"openpose-control\", \"id\": 0, \"dep\": [-1], \"args\": {\"image\": \"/example.jpg\" }},  {\"task\": \"openpose-text-to-image\", \"id\": 1, \"dep\": [0], \"args\": {\"text\": \"a girl is reading a book\", \"image\": \"<GENERATED>-0\" }}]"
+    },
+
+    {
+        "role": "user",
+        "content": "please show me a video and an image of (based on the text) 'a boy is running' and dub it"
+    },
+    {
+        "role": "assistant",
+        "content": "[{\"task\": \"text-to-video\", \"id\": 0, \"dep\": [-1], \"args\": {\"text\": \"a boy is running\" }}, {\"task\": \"text-to-speech\", \"id\": 1, \"dep\": [-1], \"args\": {\"text\": \"a boy is running\" }}, {\"task\": \"text-to-image\", \"id\": 2, \"dep\": [-1], \"args\": {\"text\": \"a boy is running\" }}]"
+    },
+
+
+    {
+        "role": "user",
+        "content": "please show me a joke and an image of cat"
+    },
+    {
+        "role": "assistant",
+        "content": "[{\"task\": \"conversational\", \"id\": 0, \"dep\": [-1], \"args\": {\"text\": \"please show me a joke of cat\" }}, {\"task\": \"text-to-image\", \"id\": 1, \"dep\": [-1], \"args\": {\"text\": \"a photo of cat\" }}]"
+    },
+
+    {
+        "role": "user",
+        "content": "give me a picture about a cut dog, then describe the image to me and tell a story about it"
+    },
+    {
+        "role": "assistant",
+        "content": "[{\"task\": \"text-to-image\", \"id\": 0, \"dep\": [-1], \"args\": {\"text\": \"a picture of a cut dog\" }}, {\"task\": \"image-to-text\", \"id\": 1, \"dep\": [0], \"args\": {\"image\": \"<GENERATED>-0\" }}, {\"task\": \"text-generation\", \"id\": 2, \"dep\": [1], \"args\": {\"text\": \"<GENERATED>-1\" }}, {\"task\": \"text-to-speech\", \"id\": 3, \"dep\": [2], \"args\": {\"text\": \"<GENERATED>-2\" }}]"
+    }
+]

demos/demo_response_results.json

@@ -0,0 +1,10 @@
+[
+    {
+        "role": "user",
+        "content": "{{input}}"
+    },
+    {
+        "role": "assistant",
+        "content": "Before give you a response, I want to introduce my workflow for your request, which is shown in the following JSON data: {{processes}}. Do you have any demands regarding my response?"
+    }
+]

get_token_ids.py

@@ -0,0 +1,49 @@
+import tiktoken
+
+encodings = {
+    "gpt-3.5-turbo": tiktoken.get_encoding("cl100k_base"),
+    "gpt-3.5-turbo-0301": tiktoken.get_encoding("cl100k_base"),
+    "text-davinci-003": tiktoken.get_encoding("p50k_base"),
+    "text-davinci-002": tiktoken.get_encoding("p50k_base"),
+    "text-davinci-001": tiktoken.get_encoding("r50k_base"),
+    "text-curie-001": tiktoken.get_encoding("r50k_base"),
+    "text-babbage-001": tiktoken.get_encoding("r50k_base"),
+    "text-ada-001": tiktoken.get_encoding("r50k_base"),
+    "davinci": tiktoken.get_encoding("r50k_base"),
+    "curie": tiktoken.get_encoding("r50k_base"),
+    "babbage": tiktoken.get_encoding("r50k_base"),
+    "ada": tiktoken.get_encoding("r50k_base"),
+}
+
+max_length = {
+    "gpt-3.5-turbo": 4096,
+    "gpt-3.5-turbo-0301": 4096,
+    "text-davinci-003": 4096,
+    "text-davinci-002": 4096,
+    "text-davinci-001": 2049,
+    "text-curie-001": 2049,
+    "text-babbage-001": 2049,
+    "text-ada-001": 2049,
+    "davinci": 2049,
+    "curie": 2049,
+    "babbage": 2049,
+    "ada": 2049
+}
+
+def count_tokens(model_name, text):
+    return len(encodings[model_name].encode(text))
+
+def get_max_context_length(model_name):
+    return max_length[model_name]
+
+def get_token_ids_for_task_parsing(model_name):
+    text = '''{"task": "text-classification",  "token-classification", "text2text-generation", "summarization", "translation",  "question-answering", "conversational", "text-generation", "sentence-similarity", "tabular-classification", "object-detection", "image-classification", "image-to-image", "image-to-text", "text-to-image", "visual-question-answering", "document-question-answering", "image-segmentation", "text-to-speech", "text-to-video", "automatic-speech-recognition", "audio-to-audio", "audio-classification", "canny-control", "hed-control", "mlsd-control", "normal-control", "openpose-control", "canny-text-to-image", "depth-text-to-image", "hed-text-to-image", "mlsd-text-to-image", "normal-text-to-image", "openpose-text-to-image", "seg-text-to-image", "args", "text", "path", "dep", "id", "<GENERATED>-"}'''
+    res = encodings[model_name].encode(text)
+    res = list(set(res))
+    return res
+
+def get_token_ids_for_choose_model(model_name):
+    text = '''{"id": "reason"}'''
+    res = encodings[model_name].encode(text)
+    res = list(set(res))
+    return res

models_server.py

@@ -0,0 +1,618 @@
+import argparse
+import logging
+import random
+import uuid
+import numpy as np
+from transformers import pipeline
+from diffusers import DiffusionPipeline, StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler
+from diffusers.utils import load_image
+from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
+from diffusers.utils import export_to_video
+from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5ForSpeechToSpeech
+from transformers import BlipProcessor, BlipForConditionalGeneration
+from transformers import TrOCRProcessor, VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
+from datasets import load_dataset
+from PIL import Image
+import io
+from torchvision import transforms
+import torch
+import torchaudio
+from speechbrain.pretrained import WaveformEnhancement
+import joblib
+from huggingface_hub import hf_hub_url, cached_download
+from transformers import AutoImageProcessor, TimesformerForVideoClassification
+from transformers import MaskFormerFeatureExtractor, MaskFormerForInstanceSegmentation, AutoFeatureExtractor
+from controlnet_aux import OpenposeDetector, MLSDdetector, HEDdetector, CannyDetector, MidasDetector
+from controlnet_aux.open_pose.body import Body
+from controlnet_aux.mlsd.models.mbv2_mlsd_large import MobileV2_MLSD_Large
+from controlnet_aux.hed import Network
+from transformers import DPTForDepthEstimation, DPTFeatureExtractor
+import warnings
+import time
+from espnet2.bin.tts_inference import Text2Speech
+import soundfile as sf
+from asteroid.models import BaseModel
+import traceback
+import os
+import yaml
+
+warnings.filterwarnings("ignore")
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--config", type=str, default="config.yaml")
+args = parser.parse_args()
+
+if __name__ != "__main__":
+    args.config = "config.gradio.yaml"
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.INFO)
+handler = logging.StreamHandler()
+handler.setLevel(logging.INFO)
+formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
+handler.setFormatter(formatter)
+logger.addHandler(handler)
+
+config = yaml.load(open(args.config, "r"), Loader=yaml.FullLoader)
+
+local_deployment = config["local_deployment"]
+if config["inference_mode"] == "huggingface":
+    local_deployment = "none"
+
+PROXY = None
+if config["proxy"]:
+    PROXY = {
+        "https": config["proxy"],
+    }
+
+start = time.time()
+
+# local_models = "models/"
+local_models = ""
+
+
+def load_pipes(local_deployment):
+    other_pipes = {}
+    standard_pipes = {}
+    controlnet_sd_pipes = {}
+    if local_deployment in ["full"]:
+        other_pipes = {
+
+            # "Salesforce/blip-image-captioning-large": {
+            #     "model": BlipForConditionalGeneration.from_pretrained(f"Salesforce/blip-image-captioning-large"),
+            #     "processor": BlipProcessor.from_pretrained(f"Salesforce/blip-image-captioning-large"),
+            #     "device": "cuda:0"
+            # },
+            "damo-vilab/text-to-video-ms-1.7b": {
+                "model": DiffusionPipeline.from_pretrained(f"{local_models}damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16"),
+                "device": "cuda:0"
+            },
+            # "facebook/maskformer-swin-large-ade": {
+            #     "model": MaskFormerForInstanceSegmentation.from_pretrained(f"facebook/maskformer-swin-large-ade"),
+            #     "feature_extractor" : AutoFeatureExtractor.from_pretrained("facebook/maskformer-swin-large-ade"),
+            #     "device": "cuda:0"
+            # },
+            # "microsoft/trocr-base-printed": {
+            #     "processor": TrOCRProcessor.from_pretrained(f"microsoft/trocr-base-printed"),
+            #     "model": VisionEncoderDecoderModel.from_pretrained(f"microsoft/trocr-base-printed"),
+            #     "device": "cuda:0"
+            # },
+            # "microsoft/trocr-base-handwritten": {
+            #     "processor": TrOCRProcessor.from_pretrained(f"microsoft/trocr-base-handwritten"),
+            #     "model": VisionEncoderDecoderModel.from_pretrained(f"microsoft/trocr-base-handwritten"),
+            #     "device": "cuda:0"
+            # },
+            "JorisCos/DCCRNet_Libri1Mix_enhsingle_16k": {
+                "model": BaseModel.from_pretrained("JorisCos/DCCRNet_Libri1Mix_enhsingle_16k"),
+                "device": "cuda:0"
+            },
+
+            # "CompVis/stable-diffusion-v1-4": {
+            #     "model": DiffusionPipeline.from_pretrained(f"CompVis/stable-diffusion-v1-4"),
+            #     "device": "cuda:0"
+            # },
+            # "stabilityai/stable-diffusion-2-1": {
+            #     "model": DiffusionPipeline.from_pretrained(f"stabilityai/stable-diffusion-2-1"),
+            #     "device": "cuda:0"
+            # },
+
+            # "microsoft/speecht5_tts":{
+            #     "processor": SpeechT5Processor.from_pretrained(f"microsoft/speecht5_tts"),
+            #     "model": SpeechT5ForTextToSpeech.from_pretrained(f"microsoft/speecht5_tts"),
+            #     "vocoder":  SpeechT5HifiGan.from_pretrained(f"microsoft/speecht5_hifigan"),
+            #     "embeddings_dataset": load_dataset(f"Matthijs/cmu-arctic-xvectors", split="validation"),
+            #     "device": "cuda:0"
+            # },
+            # "speechbrain/mtl-mimic-voicebank": {
+            #     "model": WaveformEnhancement.from_hparams(source="speechbrain/mtl-mimic-voicebank", savedir="models/mtl-mimic-voicebank"),
+            #     "device": "cuda:0"
+            # },
+            "microsoft/speecht5_vc":{
+                "processor": SpeechT5Processor.from_pretrained(f"{local_models}microsoft/speecht5_vc"),
+                "model": SpeechT5ForSpeechToSpeech.from_pretrained(f"{local_models}microsoft/speecht5_vc"),
+                "vocoder": SpeechT5HifiGan.from_pretrained(f"{local_models}microsoft/speecht5_hifigan"),
+                "embeddings_dataset": load_dataset(f"{local_models}Matthijs/cmu-arctic-xvectors", split="validation"),
+                "device": "cuda:0"
+            },
+            # "julien-c/wine-quality": {
+            #     "model": joblib.load(cached_download(hf_hub_url("julien-c/wine-quality", "sklearn_model.joblib")))
+            # },
+            # "facebook/timesformer-base-finetuned-k400": {
+            #     "processor": AutoImageProcessor.from_pretrained(f"facebook/timesformer-base-finetuned-k400"),
+            #     "model": TimesformerForVideoClassification.from_pretrained(f"facebook/timesformer-base-finetuned-k400"),
+            #     "device": "cuda:0"
+            # },
+            "facebook/maskformer-swin-base-coco": {
+                "feature_extractor": MaskFormerFeatureExtractor.from_pretrained(f"{local_models}facebook/maskformer-swin-base-coco"),
+                "model": MaskFormerForInstanceSegmentation.from_pretrained(f"{local_models}facebook/maskformer-swin-base-coco"),
+                "device": "cuda:0"
+            },
+            "Intel/dpt-hybrid-midas": {
+                "model": DPTForDepthEstimation.from_pretrained(f"{local_models}Intel/dpt-hybrid-midas", low_cpu_mem_usage=True),
+                "feature_extractor": DPTFeatureExtractor.from_pretrained(f"{local_models}Intel/dpt-hybrid-midas"),
+                "device": "cuda:0"
+            }
+        }
+
+    if local_deployment in ["full", "standard"]:
+        standard_pipes = {
+            # "nlpconnect/vit-gpt2-image-captioning":{
+            #     "model": VisionEncoderDecoderModel.from_pretrained(f"{local_models}nlpconnect/vit-gpt2-image-captioning"),
+            #     "feature_extractor": ViTImageProcessor.from_pretrained(f"{local_models}nlpconnect/vit-gpt2-image-captioning"),
+            #     "tokenizer": AutoTokenizer.from_pretrained(f"{local_models}nlpconnect/vit-gpt2-image-captioning"),
+            #     "device": "cuda:0"
+            # },
+            "espnet/kan-bayashi_ljspeech_vits": {
+                "model": Text2Speech.from_pretrained("espnet/kan-bayashi_ljspeech_vits"),
+                "device": "cuda:0"
+            },
+            # "lambdalabs/sd-image-variations-diffusers": {
+            #     "model": DiffusionPipeline.from_pretrained(f"{local_models}lambdalabs/sd-image-variations-diffusers"), #torch_dtype=torch.float16
+            #     "device": "cuda:0"
+            # },
+            "runwayml/stable-diffusion-v1-5": {
+                "model": DiffusionPipeline.from_pretrained(f"{local_models}runwayml/stable-diffusion-v1-5"),
+                "device": "cuda:0"
+            },
+            # "superb/wav2vec2-base-superb-ks": {
+            #     "model": pipeline(task="audio-classification", model=f"superb/wav2vec2-base-superb-ks"), 
+            #     "device": "cuda:0"
+            # },
+            "openai/whisper-base": {
+                "model": pipeline(task="automatic-speech-recognition", model=f"{local_models}openai/whisper-base"), 
+                "device": "cuda:0"
+            },
+            # "microsoft/speecht5_asr": {
+            #     "model": pipeline(task="automatic-speech-recognition", model=f"{local_models}microsoft/speecht5_asr"), 
+            #     "device": "cuda:0"
+            # },
+            "Intel/dpt-large": {
+                "model": pipeline(task="depth-estimation", model=f"{local_models}Intel/dpt-large"), 
+                "device": "cuda:0"
+            },
+            # "microsoft/beit-base-patch16-224-pt22k-ft22k": {
+            #     "model": pipeline(task="image-classification", model=f"microsoft/beit-base-patch16-224-pt22k-ft22k"), 
+            #     "device": "cuda:0"
+            # },
+            "facebook/detr-resnet-50-panoptic": {
+                "model": pipeline(task="image-segmentation", model=f"{local_models}facebook/detr-resnet-50-panoptic"), 
+                "device": "cuda:0"
+            },
+            "facebook/detr-resnet-101": {
+                "model": pipeline(task="object-detection", model=f"{local_models}facebook/detr-resnet-101"), 
+                "device": "cuda:0"
+            },
+            # "openai/clip-vit-large-patch14": {
+            #     "model": pipeline(task="zero-shot-image-classification", model=f"openai/clip-vit-large-patch14"), 
+            #     "device": "cuda:0"
+            # },
+            # "google/owlvit-base-patch32": {
+            #     "model": pipeline(task="zero-shot-object-detection", model=f"{local_models}google/owlvit-base-patch32"), 
+            #     "device": "cuda:0"
+            # },
+            # "microsoft/DialoGPT-medium": {
+            #     "model": pipeline(task="conversational", model=f"microsoft/DialoGPT-medium"), 
+            #     "device": "cuda:0"
+            # },
+            # "bert-base-uncased": {
+            #     "model": pipeline(task="fill-mask", model=f"bert-base-uncased"), 
+            #     "device": "cuda:0"
+            # },
+            # "deepset/roberta-base-squad2": {
+            #     "model": pipeline(task = "question-answering", model=f"deepset/roberta-base-squad2"), 
+            #     "device": "cuda:0"
+            # },
+            # "facebook/bart-large-cnn": {
+            #     "model": pipeline(task="summarization", model=f"facebook/bart-large-cnn"), 
+            #     "device": "cuda:0"
+            # },
+            # "google/tapas-base-finetuned-wtq": {
+            #     "model": pipeline(task="table-question-answering", model=f"google/tapas-base-finetuned-wtq"), 
+            #     "device": "cuda:0"
+            # },
+            # "distilbert-base-uncased-finetuned-sst-2-english": {
+            #     "model": pipeline(task="text-classification", model=f"distilbert-base-uncased-finetuned-sst-2-english"), 
+            #     "device": "cuda:0"
+            # },
+            # "gpt2": {
+            #     "model": pipeline(task="text-generation", model="gpt2"), 
+            #     "device": "cuda:0"
+            # },
+            # "mrm8488/t5-base-finetuned-question-generation-ap": {
+            #     "model": pipeline(task="text2text-generation", model=f"mrm8488/t5-base-finetuned-question-generation-ap"), 
+            #     "device": "cuda:0"
+            # },
+            # "Jean-Baptiste/camembert-ner": {
+            #     "model": pipeline(task="token-classification", model=f"Jean-Baptiste/camembert-ner", aggregation_strategy="simple"), 
+            #     "device": "cuda:0"
+            # },
+            # "t5-base": {
+            #     "model": pipeline(task="translation", model=f"t5-base"), 
+            #     "device": "cuda:0"
+            # },
+            "impira/layoutlm-document-qa": {
+                "model": pipeline(task="document-question-answering", model=f"{local_models}impira/layoutlm-document-qa"), 
+                "device": "cuda:0"
+            },
+            "ydshieh/vit-gpt2-coco-en": {
+                "model": pipeline(task="image-to-text", model=f"{local_models}ydshieh/vit-gpt2-coco-en"), 
+                "device": "cuda:0"
+            },
+            "dandelin/vilt-b32-finetuned-vqa": {
+                "model": pipeline(task="visual-question-answering", model=f"{local_models}dandelin/vilt-b32-finetuned-vqa"), 
+                "device": "cuda:0"
+            }
+        }
+
+    if local_deployment in ["full", "standard", "minimal"]:
+        
+        controlnet = ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
+        controlnetpipe = StableDiffusionControlNetPipeline.from_pretrained(
+            f"{local_models}runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
+        )
+
+
+        hed_network = HEDdetector.from_pretrained('lllyasviel/ControlNet')
+
+        controlnet_sd_pipes = {
+            "openpose-control": {
+                "model": OpenposeDetector.from_pretrained('lllyasviel/ControlNet')
+            },
+            "mlsd-control": {
+                "model": MLSDdetector.from_pretrained('lllyasviel/ControlNet')
+            },
+            "hed-control": {
+                "model": hed_network
+            },
+            "scribble-control": {
+                "model": hed_network
+            },
+            "midas-control": {
+                "model": MidasDetector.from_pretrained('lllyasviel/ControlNet')
+            },
+            "canny-control": {
+                "model": CannyDetector()
+            },
+            "lllyasviel/sd-controlnet-canny":{
+                "control": controlnet, 
+                "model": controlnetpipe,
+                "device": "cuda:0"
+            },
+            "lllyasviel/sd-controlnet-depth":{
+                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-depth", torch_dtype=torch.float16),
+                "model": controlnetpipe,
+                "device": "cuda:0"
+            },
+            "lllyasviel/sd-controlnet-hed":{
+                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-hed", torch_dtype=torch.float16), 
+                "model": controlnetpipe,
+                "device": "cuda:0"
+            },
+            "lllyasviel/sd-controlnet-mlsd":{
+                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-mlsd", torch_dtype=torch.float16), 
+                "model": controlnetpipe,
+                "device": "cuda:0"
+            },
+            "lllyasviel/sd-controlnet-openpose":{
+                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-openpose", torch_dtype=torch.float16), 
+                "model": controlnetpipe,
+                "device": "cuda:0"
+            },
+            "lllyasviel/sd-controlnet-scribble":{
+                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-scribble", torch_dtype=torch.float16), 
+                "model": controlnetpipe,
+                "device": "cuda:0"
+            },
+            "lllyasviel/sd-controlnet-seg":{
+                "control": ControlNetModel.from_pretrained(f"{local_models}lllyasviel/sd-controlnet-seg", torch_dtype=torch.float16), 
+                "model": controlnetpipe,
+                "device": "cuda:0"
+            }    
+        }
+    pipes = {**standard_pipes, **other_pipes, **controlnet_sd_pipes}
+    return pipes
+
+pipes = load_pipes(local_deployment)
+
+end = time.time()
+during = end - start
+
+print(f"[ ready ] {during}s")
+
+def running():
+    return {"running": True}
+
+def status(model_id):
+    disabled_models = ["microsoft/trocr-base-printed", "microsoft/trocr-base-handwritten"]
+    if model_id in pipes.keys() and model_id not in disabled_models:
+        print(f"[ check {model_id} ] success")
+        return {"loaded": True}
+    else:
+        print(f"[ check {model_id} ] failed")
+        return {"loaded": False}
+
+def models(model_id, data):
+    while "using" in pipes[model_id] and pipes[model_id]["using"]:
+        print(f"[ inference {model_id} ] waiting")
+        time.sleep(0.1)
+    pipes[model_id]["using"] = True
+    print(f"[ inference {model_id} ] start")
+
+    start = time.time()
+
+    pipe = pipes[model_id]["model"]
+    
+    if "device" in pipes[model_id]:
+        try:
+            pipe.to(pipes[model_id]["device"])
+        except:
+            pipe.device = torch.device(pipes[model_id]["device"])
+            pipe.model.to(pipes[model_id]["device"])
+    
+    result = None
+    try:
+        # text to video
+        if model_id == "damo-vilab/text-to-video-ms-1.7b":
+            pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+            # pipe.enable_model_cpu_offload()
+            prompt = data["text"]
+            video_frames = pipe(prompt, num_inference_steps=50, num_frames=40).frames
+            file_name = str(uuid.uuid4())[:4]
+            video_path = export_to_video(video_frames, f"public/videos/{file_name}.mp4")
+            
+            new_file_name = str(uuid.uuid4())[:4]
+            os.system(f"ffmpeg -i {video_path} -vcodec libx264 public/videos/{new_file_name}.mp4")
+
+            if os.path.exists(f"public/videos/{new_file_name}.mp4"): 
+                result = {"path": f"/videos/{new_file_name}.mp4"}
+            else:
+                result = {"path": f"/videos/{file_name}.mp4"}
+
+        # controlnet
+        if model_id.startswith("lllyasviel/sd-controlnet-"):
+            pipe.controlnet.to('cpu')
+            pipe.controlnet = pipes[model_id]["control"].to(pipes[model_id]["device"])
+            pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+            control_image = load_image(data["img_url"])
+            # generator = torch.manual_seed(66)
+            out_image: Image = pipe(data["text"], num_inference_steps=20, image=control_image).images[0]
+            file_name = str(uuid.uuid4())[:4]
+            out_image.save(f"public/images/{file_name}.png")
+            result = {"path": f"/images/{file_name}.png"}
+
+        if model_id.endswith("-control"):
+            image = load_image(data["img_url"])
+            if "scribble" in model_id:
+                control = pipe(image, scribble = True)
+            elif "canny" in model_id:
+                control = pipe(image, low_threshold=100, high_threshold=200)
+            else:
+                control = pipe(image)
+            file_name = str(uuid.uuid4())[:4]
+            control.save(f"public/images/{file_name}.png")
+            result = {"path": f"/images/{file_name}.png"}
+
+        # image to image
+        if model_id == "lambdalabs/sd-image-variations-diffusers":
+            im = load_image(data["img_url"])
+            file_name = str(uuid.uuid4())[:4]
+            with open(f"public/images/{file_name}.png", "wb") as f:
+                f.write(data)
+            tform = transforms.Compose([
+                transforms.ToTensor(),
+                transforms.Resize(
+                    (224, 224),
+                    interpolation=transforms.InterpolationMode.BICUBIC,
+                    antialias=False,
+                    ),
+                transforms.Normalize(
+                [0.48145466, 0.4578275, 0.40821073],
+                [0.26862954, 0.26130258, 0.27577711]),
+            ])
+            inp = tform(im).to(pipes[model_id]["device"]).unsqueeze(0)
+            out = pipe(inp, guidance_scale=3)
+            out["images"][0].save(f"public/images/{file_name}.jpg")
+            result = {"path": f"/images/{file_name}.jpg"}
+
+        # image to text
+        if model_id == "Salesforce/blip-image-captioning-large":
+            raw_image = load_image(data["img_url"]).convert('RGB')
+            text = data["text"]
+            inputs = pipes[model_id]["processor"](raw_image, return_tensors="pt").to(pipes[model_id]["device"])
+            out = pipe.generate(**inputs)
+            caption = pipes[model_id]["processor"].decode(out[0], skip_special_tokens=True)
+            result = {"generated text": caption}
+        if model_id == "ydshieh/vit-gpt2-coco-en":
+            img_url = data["img_url"]
+            generated_text = pipe(img_url)[0]['generated_text']
+            result = {"generated text": generated_text}
+        if model_id == "nlpconnect/vit-gpt2-image-captioning":
+            image = load_image(data["img_url"]).convert("RGB")
+            pixel_values = pipes[model_id]["feature_extractor"](images=image, return_tensors="pt").pixel_values
+            pixel_values = pixel_values.to(pipes[model_id]["device"])
+            generated_ids = pipe.generate(pixel_values, **{"max_length": 200, "num_beams": 1})
+            generated_text = pipes[model_id]["tokenizer"].batch_decode(generated_ids, skip_special_tokens=True)[0]
+            result = {"generated text": generated_text}
+        # image to text: OCR
+        if model_id == "microsoft/trocr-base-printed" or  model_id == "microsoft/trocr-base-handwritten":
+            image = load_image(data["img_url"]).convert("RGB")
+            pixel_values = pipes[model_id]["processor"](image, return_tensors="pt").pixel_values
+            pixel_values = pixel_values.to(pipes[model_id]["device"])
+            generated_ids = pipe.generate(pixel_values)
+            generated_text = pipes[model_id]["processor"].batch_decode(generated_ids, skip_special_tokens=True)[0]
+            result = {"generated text": generated_text}
+
+        # text to image
+        if model_id == "runwayml/stable-diffusion-v1-5":
+            file_name = str(uuid.uuid4())[:4]
+            text = data["text"]
+            out = pipe(prompt=text)
+            out["images"][0].save(f"public/images/{file_name}.jpg")
+            result = {"path": f"/images/{file_name}.jpg"}
+
+        # object detection
+        if model_id == "google/owlvit-base-patch32" or model_id == "facebook/detr-resnet-101":
+            img_url = data["img_url"]
+            open_types = ["cat", "couch", "person", "car", "dog", "horse", "sheep", "cow", "elephant", "bear", "zebra", "giraffe", "backpack", "umbrella", "handbag", "tie", "suitcase", "frisbee", "skis", "snowboard", "sports ball", "kite", "baseball bat", "baseball glove", "skateboard", "surfboard", "tennis racket", "bottle", "wine glass", "cup", "fork", "knife", "spoon", "bowl", "banana", "apple", "sandwich", "orange", "broccoli", "carrot", "hot dog", "pizza", "donut", "cake", "chair", "couch", "potted plant", "bed", "dining table", "toilet", "tv", "laptop", "mouse", "remote", "keyboard", "cell phone", "microwave", "oven", "toaster", "sink", "refrigerator", "book", "clock", "vase", "scissors", "teddy bear", "hair drier", "toothbrush", "traffic light", "fire hydrant", "stop sign", "parking meter", "bench", "bird"]
+            result = pipe(img_url, candidate_labels=open_types)
+        
+        # VQA
+        if model_id == "dandelin/vilt-b32-finetuned-vqa":
+            question = data["text"]
+            img_url = data["img_url"]
+            result = pipe(question=question, image=img_url)
+        
+        #DQA
+        if model_id == "impira/layoutlm-document-qa":
+            question = data["text"]
+            img_url = data["img_url"]
+            result = pipe(img_url, question)
+
+        # depth-estimation
+        if model_id == "Intel/dpt-large":
+            output = pipe(data["img_url"])
+            image = output['depth']
+            name = str(uuid.uuid4())[:4]
+            image.save(f"public/images/{name}.jpg")
+            result = {"path": f"/images/{name}.jpg"}
+
+        if model_id == "Intel/dpt-hybrid-midas" and model_id == "Intel/dpt-large":
+            image = load_image(data["img_url"])
+            inputs = pipes[model_id]["feature_extractor"](images=image, return_tensors="pt")
+            with torch.no_grad():
+                outputs = pipe(**inputs)
+                predicted_depth = outputs.predicted_depth
+            prediction = torch.nn.functional.interpolate(
+                predicted_depth.unsqueeze(1),
+                size=image.size[::-1],
+                mode="bicubic",
+                align_corners=False,
+            )
+            output = prediction.squeeze().cpu().numpy()
+            formatted = (output * 255 / np.max(output)).astype("uint8")
+            image = Image.fromarray(formatted)
+            name = str(uuid.uuid4())[:4]
+            image.save(f"public/images/{name}.jpg")
+            result = {"path": f"/images/{name}.jpg"}
+
+        # TTS
+        if model_id == "espnet/kan-bayashi_ljspeech_vits":
+            text = data["text"]
+            wav = pipe(text)["wav"]
+            name = str(uuid.uuid4())[:4]
+            sf.write(f"public/audios/{name}.wav", wav.cpu().numpy(), pipe.fs, "PCM_16")
+            result = {"path": f"/audios/{name}.wav"}
+
+        if model_id == "microsoft/speecht5_tts":
+            text = data["text"]
+            inputs = pipes[model_id]["processor"](text=text, return_tensors="pt")
+            embeddings_dataset = pipes[model_id]["embeddings_dataset"]
+            speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0).to(pipes[model_id]["device"])
+            pipes[model_id]["vocoder"].to(pipes[model_id]["device"])
+            speech = pipe.generate_speech(inputs["input_ids"].to(pipes[model_id]["device"]), speaker_embeddings, vocoder=pipes[model_id]["vocoder"])
+            name = str(uuid.uuid4())[:4]
+            sf.write(f"public/audios/{name}.wav", speech.cpu().numpy(), samplerate=16000)
+            result = {"path": f"/audios/{name}.wav"}
+
+        # ASR
+        if model_id == "openai/whisper-base" or model_id == "microsoft/speecht5_asr":
+            audio_url = data["audio_url"]
+            result = { "text": pipe(audio_url)["text"]}
+
+        # audio to audio
+        if model_id == "JorisCos/DCCRNet_Libri1Mix_enhsingle_16k":
+            audio_url = data["audio_url"]
+            wav, sr = torchaudio.load(audio_url)
+            with torch.no_grad():
+                result_wav = pipe(wav.to(pipes[model_id]["device"]))
+            name = str(uuid.uuid4())[:4]
+            sf.write(f"public/audios/{name}.wav", result_wav.cpu().squeeze().numpy(), sr)
+            result = {"path": f"/audios/{name}.wav"}
+        
+        if model_id == "microsoft/speecht5_vc":
+            audio_url = data["audio_url"]
+            wav, sr = torchaudio.load(audio_url)
+            inputs = pipes[model_id]["processor"](audio=wav, sampling_rate=sr, return_tensors="pt")
+            embeddings_dataset = pipes[model_id]["embeddings_dataset"]
+            speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+            pipes[model_id]["vocoder"].to(pipes[model_id]["device"])
+            speech = pipe.generate_speech(inputs["input_ids"].to(pipes[model_id]["device"]), speaker_embeddings, vocoder=pipes[model_id]["vocoder"])
+            name = str(uuid.uuid4())[:4]
+            sf.write(f"public/audios/{name}.wav", speech.cpu().numpy(), samplerate=16000)
+            result = {"path": f"/audios/{name}.wav"}
+        
+        # segmentation
+        if model_id == "facebook/detr-resnet-50-panoptic":
+            result = []
+            segments = pipe(data["img_url"])
+            image = load_image(data["img_url"])
+
+            colors = []
+            for i in range(len(segments)):
+                colors.append((random.randint(100, 255), random.randint(100, 255), random.randint(100, 255), 50))
+
+            for segment in segments:
+                mask = segment["mask"]
+                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 = {"path": f"/images/{name}.jpg"}
+
+        if model_id == "facebook/maskformer-swin-base-coco" or model_id == "facebook/maskformer-swin-large-ade":
+            image = load_image(data["img_url"])
+            inputs = pipes[model_id]["feature_extractor"](images=image, return_tensors="pt").to(pipes[model_id]["device"])
+            outputs = pipe(**inputs)
+            result = pipes[model_id]["feature_extractor"].post_process_panoptic_segmentation(outputs, target_sizes=[image.size[::-1]])[0]
+            predicted_panoptic_map = result["segmentation"].cpu().numpy()
+            predicted_panoptic_map = Image.fromarray(predicted_panoptic_map.astype(np.uint8))
+            name = str(uuid.uuid4())[:4]
+            predicted_panoptic_map.save(f"public/images/{name}.jpg")
+            result = {"path": f"/images/{name}.jpg"}
+
+    except Exception as e:
+        print(e)
+        traceback.print_exc()
+        result = {"error": {"message": "Error when running the model inference."}}
+
+    if "device" in pipes[model_id]:
+        try:
+            pipe.to("cpu")
+            torch.cuda.empty_cache()
+        except:
+            pipe.device = torch.device("cpu")
+            pipe.model.to("cpu")
+            torch.cuda.empty_cache()
+
+    pipes[model_id]["using"] = False
+
+    if result is None:
+        result = {"error": {"message": "model not found"}}
+    
+    end = time.time()
+    during = end - start
+    print(f"[ complete {model_id} ] {during}s")
+    print(f"[ result {model_id} ] {result}")
+
+    return result

packages.txt

@@ -0,0 +1 @@
+tesseract-ocr

requirements.txt

@@ -0,0 +1,18 @@
+git+https://github.com/huggingface/diffusers.git@8c530fc2f6a76a2aefb6b285dce6df1675092ac6#egg=diffusers
+git+https://github.com/huggingface/transformers@c612628045822f909020f7eb6784c79700813eda#egg=transformers
+git+https://github.com/patrickvonplaten/controlnet_aux@78efc716868a7f5669c288233d65b471f542ce40#egg=controlnet_aux
+tiktoken==0.3.3
+pydub==0.25.1
+espnet==202301
+espnet_model_zoo==0.1.7
+flask==2.2.3
+flask_cors==3.0.10
+waitress==2.1.2
+datasets==2.11.0
+asteroid==0.6.0
+speechbrain==0.5.14
+timm==0.6.13
+typeguard==2.13.3
+accelerate==0.18.0
+pytesseract==0.3.10
+basicsr==1.4.2