Update app.py from falcon-7b-gglm-m app.py"

.ruff.toml

@@ -8,10 +8,12 @@ line-length = 300
 select = ["F", "E", "W", "I001", "YTT", "D", "PLC"]
 # select = ["ALL"]
 
+# D100 Missing docstring in public module
+# E501 Line too long
 # D103 Missing docstring in public function
 # D101 Missing docstring in public class
 # `multi-line-summary-first-line` (D212)
 # `one-blank-line-before-class` (D203)
-extend-ignore = ["D103", "D101", "D212", "D203"]
+extend-ignore = ["E501", "D100", "D103", "D101", "D212", "D203"]
 
 exclude = [".venv"]

app.py

@@ -1,510 +1,264 @@
-"""Run codes."""
-# pylint: disable=line-too-long, broad-exception-caught, invalid-name, missing-function-docstring, too-many-instance-attributes, missing-class-docstring
-import os
-import time
-from dataclasses import asdict, dataclass
 from pathlib import Path
-from types import SimpleNamespace
 from urllib.parse import urlparse
 
 import gradio as gr
 import psutil
-from about_time import about_time
 from ctransformers import AutoModelForCausalLM
-from huggingface_hub import hf_hub_download, snapshot_download
-from loguru import logger
-
-filename_list = [
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q2_K.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_L.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_M.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q3_K_S.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_0.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_1.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_M.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_S.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_0.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_1.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_K_M.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q5_K_S.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q6_K.bin",
-    "Wizard-Vicuna-7B-Uncensored.ggmlv3.q8_0.bin",
-]
-
-URL = "https://huggingface.co/TheBloke/Wizard-Vicuna-7B-Uncensored-GGML/raw/main/Wizard-Vicuna-7B-Uncensored.ggmlv3.q4_K_M.bin"  # 4.05G
-
-URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/blob/main/30b-Lazarus.ggmlv3.q4_0.bin"
-URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/blob/main/30b-Lazarus.ggmlv3.q4_1.bin"
-URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/resolve/main/30b-Lazarus.ggmlv3.q4_K_M.bin"
-URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/resolve/main/30b-Lazarus.ggmlv3.q4_K_S.bin"  # 18GB
-
-URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/blob/main/30b-Lazarus.ggmlv3.q3_K_S.bin"  # 14GB
-
-MODEL_FILENAME = Path(URL).name
-
-# MODEL_FILENAME = filename_list[0]  # q2_K 4.05G
-# MODEL_FILENAME = filename_list[5]  # q4_1 4.21
-
-REPO_ID = "/".join(
-    urlparse(URL).path.strip("/").split("/")[:2]
-)
-# TheBloke/30B-Lazarus-GGML
-# # TheBloke/Wizard-Vicuna-7B-Uncensored-GGML
-
-DESTINATION_FOLDER = "models"
+from huggingface_hub import hf_hub_download
 
-os.environ["TZ"] = "Asia/Shanghai"
-try:
-    time.tzset()  # type: ignore # pylint: disable=no-member
-except Exception:
-    # Windows
-    logger.warning("Windows, cant run time.tzset()")
-
-ns = SimpleNamespace(
-    response="",
-    generator=[],
+_ = """
+snapshot_download(
+  repo_id="TheBloke/falcon-7b-instruct-GGML",
+  allow_patterns="falcon7b-instruct.ggmlv3.q4_0.bin",
+  revision="ggmlv3",
+  local_dir="models",
+  local_dir_use_symlinks=False,  # default "auto"
 )
 
-default_system_prompt = "A conversation between a user and an LLM-based AI assistant named Local Assistant. Local Assistant gives helpful and honest answers."
-
-user_prefix = "[user]: "
-assistant_prefix = "[assistant]: "
-
-
-def predict_str(prompt, bot):  # bot is in fact bot_history
-    # logger.debug(f"{prompt=}, {bot=}, {timeout=}")
-
-    if bot is None:
-        bot = []
-
-    logger.debug(f"{prompt=}, {bot=}")
-
-    try:
-        # user_prompt = prompt
-        generator = generate(
-            LLM,
-            GENERATION_CONFIG,
-            system_prompt=default_system_prompt,
-            user_prompt=prompt.strip(),
-        )
-
-        ns.generator = generator  # for .then
-
-    except Exception as exc:
-        logger.error(exc)
-
-    # bot.append([prompt, f"{response} {_}"])
-    # return prompt, bot
-
-    _ = bot + [[prompt, None]]
-    logger.debug(f"{prompt=}, {_=}")
-
-    return prompt, _
-
-
-def bot_str(bot):
-    if bot:
-        bot[-1][1] = ""
-    else:
-        bot = [["Something is wrong", ""]]
-
-    print(assistant_prefix, end=" ", flush=True)
-
-    response = ""
-
-    flag = 1
-    then = time.time()
-    for word in ns.generator:
-        # record first response time
-        if flag:
-            logger.debug(f"\t {time.time() - then:.1f}s")
-            flag = 0
-        print(word, end="", flush=True)
-        # print(word, flush=True)  # vertical stream
-        response += word
-        bot[-1][1] = response
-        yield bot
-
-
-def predict(prompt, bot):
-    # logger.debug(f"{prompt=}, {bot=}, {timeout=}")
-    logger.debug(f"{prompt=}, {bot=}")
-
-    ns.response = ""
-    then = time.time()
-    with about_time() as atime:  # type: ignore
-        try:
-            # user_prompt = prompt
-            generator = generate(
-                LLM,
-                GENERATION_CONFIG,
-                system_prompt=default_system_prompt,
-                user_prompt=prompt.strip(),
-            )
-
-            ns.generator = generator  # for .then
-
-            print(assistant_prefix, end=" ", flush=True)
-
-            response = ""
-            buff.update(value="diggin...")
-
-            flag = 1
-            for word in generator:
-                # record first response time
-                if flag:
-                    logger.debug(f"\t {time.time() - then:.1f}s")
-                    flag = 0
-                # print(word, end="", flush=True)
-                print(word, flush=True)  # vertical stream
-                response += word
-                ns.response = response
-                buff.update(value=response)
-            print("")
-            logger.debug(f"{response=}")
-        except Exception as exc:
-            logger.error(exc)
-            response = f"{exc=}"
-
-    # bot = {"inputs": [response]}
-    _ = (
-        f"(time elapsed: {atime.duration_human}, "  # type: ignore
-        f"{atime.duration/(len(prompt) + len(response)):.1f}s/char)"  # type: ignore
-    )
-
-    bot.append([prompt, f"{response} {_}"])
-
-    return prompt, bot
-
-
-def predict_api(prompt):
-    logger.debug(f"{prompt=}")
-    ns.response = ""
-    try:
-        # user_prompt = prompt
-        _ = GenerationConfig(
-            temperature=0.2,
-            top_k=0,
-            top_p=0.9,
-            repetition_penalty=1.0,
-            max_new_tokens=512,  # adjust as needed
-            seed=42,
-            reset=False,  # reset history (cache)
-            stream=True,  # TODO stream=False and generator
-            threads=os.cpu_count() // 2,  # type: ignore  # adjust for your CPU
-            stop=["<|im_end|>", "|<"],
-        )
-
-        # TODO: stream does not make sense in api?
-        generator = generate(
-            LLM, _, system_prompt=default_system_prompt, user_prompt=prompt.strip()
-        )
-        print(assistant_prefix, end=" ", flush=True)
-
-        response = ""
-        buff.update(value="diggin...")
-        for word in generator:
-            print(word, end="", flush=True)
-            response += word
-            ns.response = response
-            buff.update(value=response)
-        print("")
-        logger.debug(f"{response=}")
-    except Exception as exc:
-        logger.error(exc)
-        response = f"{exc=}"
-    # bot = {"inputs": [response]}
-    # bot = [(prompt, response)]
-
-    return response
-
-
-def download_quant(destination_folder: str, repo_id: str, model_filename: str):
-    local_path = os.path.abspath(destination_folder)
-    return hf_hub_download(
-        repo_id=repo_id,
-        filename=model_filename,
-        local_dir=local_path,
-        local_dir_use_symlinks=True,
-    )
-
-
-@dataclass
-class GenerationConfig:
-    temperature: float
-    top_k: int
-    top_p: float
-    repetition_penalty: float
-    max_new_tokens: int
-    seed: int
-    reset: bool
-    stream: bool
-    threads: int
-    stop: list[str]
-
-
-def format_prompt(system_prompt: str, user_prompt: str):
-    """Format prompt based on: https://huggingface.co/spaces/mosaicml/mpt-30b-chat/blob/main/app.py."""
-    # TODO: fix prompts
-
-    system_prompt = f"<|im_start|>system\n{system_prompt}<|im_end|>\n"
-    user_prompt = f"<|im_start|>user\n{user_prompt}<|im_end|>\n"
-    assistant_prompt = "<|im_start|>assistant\n"
-
-    return f"{system_prompt}{user_prompt}{assistant_prompt}"
-
-
-def generate(
-    llm: AutoModelForCausalLM,
-    generation_config: GenerationConfig,
-    system_prompt: str = default_system_prompt,
-    user_prompt: str = "",
-):
-    """Run model inference, will return a Generator if streaming is true."""
-    # if not user_prompt.strip():
-    return llm(
-        format_prompt(
-            system_prompt,
-            user_prompt,
-        ),
-        **asdict(generation_config),
-    )
-
-
-# if "mpt" in model_filename:
-#     config = AutoConfig.from_pretrained("mosaicml/mpt-30b-cha t", context_length=8192)
-#     llm = AutoModelForCausalLM.from_pretrained(
-#         os.path.abspath(f"models/{model_filename}"),
-#         model_type="mpt",
-#         config=config,
-#     )
+hf_hub_download(
+    repo_id=repo_id,
+    filename=model_filename,
+    local_dir=local_path,
+    local_dir_use_symlinks=True,
+)
+# """
+# 4.06G
 
-# https://huggingface.co/spaces/matthoffner/wizardcoder-ggml/blob/main/main.py
 _ = """
 llm = AutoModelForCausalLM.from_pretrained(
-    "TheBloke/WizardCoder-15B-1.0-GGML",
-    model_file="WizardCoder-15B-1.0.ggmlv3.q4_0.bin",
-    model_type="starcoder",
-    threads=8
+    "TheBloke/falcon-7b-instruct-GGML",
+    model_file="falcon7b-instruct.ggmlv3.q4_0.bin",
+    model_type="falcon", gpu_layers=32, threads=2,
 )
 # """
-
-logger.info(f"start dl, {REPO_ID=}, {MODEL_FILENAME=}, {DESTINATION_FOLDER=}")
-
-# download_quant(DESTINATION_FOLDER, REPO_ID, MODEL_FILENAME)
-snapshot_download(
-  repo_id=REPO_ID,  # TheBloke/30B-Lazarus-GGML
-  allow_patterns=MODEL_FILENAME,  # 30b-Lazarus.ggmlv3.q4_K_S.bin 18.3G
-  #   revision="ggmlv3",
-  local_dir="models",
+# _ = Path("models", "falcon7b-instruct.ggmlv3.q4_0.bin").absolute().as_posix()
+# assert Path(_).exists(), f"{_} does not exist, perhaps snapshot_download failed?"
+
+URL = "https://huggingface.co/TheBloke/falcon-7b-instruct-GGML/blob/main/falcon-7b-instruct.ggccv1.q4_1.bin"
+URL = "https://huggingface.co/TheBloke/falcon-7b-instruct-GGML/blob/ggmlv3/falcon7b-instruct.ggmlv3.q4_1.bin"
+repo_id = "/".join(urlparse(URL).path.strip("/").split("/")[:2])
+URL = "https://huggingface.co/TheBloke/30B-Lazarus-GGML/blob/main/30b-Lazarus.ggmlv3.q2_K.bin"  # 13.6 G
+
+model_file = Path(URL).name
+
+_ = hf_hub_download(
+    repo_id=repo_id,
+    revision="ggmlv3",
+    filename=model_file,
+    local_dir="models",
+    # local_dir_use_symlinks=True,
 )
 
-logger.info("done dl")
-
-logger.debug(f"{os.cpu_count()=} {psutil.cpu_count(logical=False)=}")
-cpu_count = os.cpu_count() // 2  # type: ignore
-cpu_count = psutil.cpu_count(logical=False)
-
-logger.debug(f"{cpu_count=}")
-
-logger.info("load llm")
-
-# from ctransformers import AutoConfig
-# AutoConfig(REPO_ID)
-# AutoConfig(config='TheBloke/30B-Lazarus-GGML', model_type=None)
-
-_ = Path("models", MODEL_FILENAME).absolute().as_posix()
-logger.debug(f"model_file: {_}, exists: {Path(_).exists()}")
-LLM = AutoModelForCausalLM.from_pretrained(
-    # "TheBloke/WizardCoder-15B-1.0-GGML",
-    # REPO_ID,  # DESTINATION_FOLDER,  # model_path_or_repo_id: str required
+llm = AutoModelForCausalLM.from_pretrained(
+    # repo_id,  # "TheBloke/falcon-7b-instruct-GGML",
+    # model_file=model_file,
     # model_file=_,
     _,
-    model_type="llama",  #AutoConfig.from_pretrained(REPO_ID).model_type,
-    threads=cpu_count,
-)
-
-logger.info("done load llm")
-
-GENERATION_CONFIG = GenerationConfig(
-    temperature=0.2,
-    top_k=0,
-    top_p=0.9,
-    repetition_penalty=1.0,
-    max_new_tokens=512,  # adjust as needed
-    seed=42,
-    reset=False,  # reset history (cache)
-    stream=True,  # streaming per word/token
-    threads=cpu_count,
-    stop=["<|im_end|>", "|<"],  # TODO possible fix of stop
+    # model_type="falcon",
+    model_type="llama",
+    threads=psutil.cpu_count(logical=False),
 )
 
-css = """
-    .importantButton {
-        background: linear-gradient(45deg, #7e0570,#5d1c99, #6e00ff) !important;
-        border: none !important;
-    }
-    .importantButton:hover {
-        background: linear-gradient(45deg, #ff00e0,#8500ff, #6e00ff) !important;
-        border: none !important;
-    }
-    .disclaimer {font-variant-caps: all-small-caps; font-size: xx-small;}
-    .xsmall {font-size: x-small;}
+TITLE = f"""<h2 align="center">🦅 30B-Lazarus-GGM ({model_file}) 🦅"""
+USER_NAME = "User"
+BOT_NAME = "Falcon"
+DEFAULT_INSTRUCTIONS = """The following is a conversation between a highly knowledgeable and intelligent AI assistant, called Falcon, and a human user, called User. In the following interactions, User and Falcon will converse in natural language, and Falcon will answer User's questions. Falcon was built to be respectful, polite and inclusive. Falcon was built by the Technology Innovation Institute in Abu Dhabi. Falcon will never decline to answer a question, and always attempts to give an answer that User would be satisfied with. It knows a lot, and always tells the truth. The conversation begins.
 """
-etext = """In America, where cars are an important part of the national psyche, a decade ago people had suddenly started to drive less, which had not happened since the oil shocks of the 1970s. """
-examples = [
-    ["How to pick a lock? Provide detailed steps."],
-    ["Explain the plot of Cinderella in a sentence."],
-    [
-        "How long does it take to become proficient in French, and what are the best methods for retaining information?"
-    ],
-    ["What are some common mistakes to avoid when writing code?"],
-    ["Build a prompt to generate a beautiful portrait of a horse"],
-    ["Suggest four metaphors to describe the benefits of AI"],
-    ["Write a pop song about leaving home for the sandy beaches."],
-    ["Write a summary demonstrating my ability to tame lions"],
-    ["鲁迅和周树人什么关系 说中文"],
-    ["鲁迅和周树人什么关系"],
-    ["鲁迅和周树人什么关系 用英文回答"],
-    ["从前有一头牛，这头牛后面有什么？"],
-    ["正无穷大加一大于正无穷大吗？"],
-    ["正无穷大加正无穷大大于正无穷大吗？"],
-    ["-2的平方根等于什么"],
-    ["树上有5只鸟，猎人开枪打死了一只。树上还有几只鸟？"],
-    ["树上有11只鸟，猎人开枪打死了一只。树上还有几只鸟？提示：需考虑鸟可能受惊吓飞走。"],
-    ["以红楼梦的行文风格写一张委婉的请假条。不少于320字。"],
-    [f"{etext} 翻成中文，列出3个版本"],
-    [f"{etext} \n 翻成中文，保留原意，但使用文学性的语言。不要写解释。列出3个版本"],
-    ["假定 1 + 2 = 4, 试求 7 + 8"],
-    ["判断一个数是不是质数的 javascript 码"],
-    ["实现python 里 range(10)的 javascript 码"],
-    ["实现python 里 [*(range(10)]的 javascript 码"],
-    ["Erkläre die Handlung von Cinderella in einem Satz."],
-    ["Erkläre die Handlung von Cinderella in einem Satz. Auf Deutsch"],
-]
-
-with gr.Blocks(
-    # title="mpt-30b-chat-ggml",
-    title=f"{MODEL_FILENAME}",
-    theme=gr.themes.Soft(text_size="sm", spacing_size="sm"),
-    css=css,
-) as block:
-    with gr.Accordion("🎈 Info", open=False):
-        # gr.HTML(
-        #     """<center><a href="https://huggingface.co/spaces/mikeee/mpt-30b-chat?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate"></a> and spin a CPU UPGRADE to avoid the queue</center>"""
-        # )
-        gr.Markdown(
-            f"""<h5><center><{REPO_ID}>{MODEL_FILENAME}</center></h4>
-            The bot only speaks English.
-
-            Most examples are meant for another model.
-            You probably should try to test
-            some related prompts.
-            """,
-            elem_classes="xsmall",
+RETRY_COMMAND = "/retry"
+STOP_STR = f"\n{USER_NAME}:"
+STOP_SUSPECT_LIST = [":", "\n", "User"]
+
+
+def chat_accordion():
+    with gr.Accordion("Parameters", open=False):
+        temperature = gr.Slider(
+            minimum=0.1,
+            maximum=2.0,
+            value=0.8,
+            step=0.1,
+            interactive=True,
+            label="Temperature",
         )
-
-    # chatbot = gr.Chatbot().style(height=700)  # 500
-    chatbot = gr.Chatbot(height=500)
-    buff = gr.Textbox(show_label=False, visible=False)
-    with gr.Row():
-        with gr.Column(scale=4):
-            msg = gr.Textbox(
-                label="Chat Message Box",
-                placeholder="Ask me anything (press Enter or click Submit to send)",
-                show_label=False,
-            ).style(container=False)
-        with gr.Column(scale=1, min_width=50):
-            with gr.Row():
-                submit = gr.Button("Submit", elem_classes="xsmall")
-                stop = gr.Button("Stop", visible=False)
-                clear = gr.Button("Clear History", visible=True)
-    with gr.Row(visible=False):
-        with gr.Accordion("Advanced Options:", open=False):
-            with gr.Row():
-                with gr.Column(scale=2):
-                    system = gr.Textbox(
-                        label="System Prompt",
-                        value=default_system_prompt,
-                        show_label=False,
-                    ).style(container=False)
-                with gr.Column():
-                    with gr.Row():
-                        change = gr.Button("Change System Prompt")
-                        reset = gr.Button("Reset System Prompt")
-
-    with gr.Accordion("Example Inputs", open=True):
-        examples = gr.Examples(
-            examples=examples,
-            inputs=[msg],
-            examples_per_page=20,
+        top_p = gr.Slider(
+            minimum=0.1,
+            maximum=0.99,
+            value=0.9,
+            step=0.01,
+            interactive=True,
+            label="p (nucleus sampling)",
         )
+    return temperature, top_p
+
+
+# TODO: fix prompt
+def format_chat_prompt(message: str, chat_history, instructions: str) -> str:
+    instructions = instructions.strip(" ").strip("\n")
+    prompt = instructions
+    for turn in chat_history:
+        user_message, bot_message = turn
+        prompt = f"{prompt}\n{USER_NAME}: {user_message}\n{BOT_NAME}: {bot_message}"
+    prompt = f"{prompt}\n{USER_NAME}: {message}\n{BOT_NAME}:"
+    return prompt
+
+
+def chat():
+    with gr.Column(elem_id="chat_container"):
+        with gr.Row():
+            chatbot = gr.Chatbot(elem_id="chatbot")
+        with gr.Row():
+            inputs = gr.Textbox(
+                placeholder=f"Hello {BOT_NAME} !!",
+                label="Type an input and press Enter",
+                max_lines=3,
+            )
+
+    with gr.Row(elem_id="button_container"):
+        with gr.Column():
+            submit_button = gr.Button("🚀 Submit")
+        with gr.Column():
+            retry_button = gr.Button("♻️ Retry last turn")
+        with gr.Column():
+            delete_turn_button = gr.Button("🧽 Delete last turn")
+        with gr.Column():
+            clear_chat_button = gr.Button("✨ Delete all history")
+
+    gr.Examples(
+        [
+            ["Hey! Any recommendations for my holidays in Abu Dhabi?"],
+            ["What's the Everett interpretation of quantum mechanics?"],
+            [
+                "Give me a list of the top 10 dive sites you would recommend around the world."
+            ],
+            ["Can you tell me more about deep-water soloing?"],
+            [
+                "Can you write a short tweet about 30B-Lazarus-GGM?"
+            ],
+        ],
+        inputs=inputs,
+        label="Click on any example and press Enter in the input textbox!",
+    )
 
-    # with gr.Row():
-    with gr.Accordion("Disclaimer", open=False):
-        _ = "-".join(MODEL_FILENAME.split("-")[:2])
-        gr.Markdown(
-            f"Disclaimer: {_} can produce factually incorrect output, and should not be relied on to produce "
-            "factually accurate information. {_} was trained on various public datasets; while great efforts "
-            "have been taken to clean the pretraining data, it is possible that this model could generate lewd, "
-            "biased, or otherwise offensive outputs.",
-            elem_classes=["disclaimer"],
+    with gr.Row(elem_id="param_container"):
+        with gr.Column():
+            temperature, top_p = chat_accordion()
+        with gr.Column():
+            with gr.Accordion("Instructions", open=False):
+                instructions = gr.Textbox(
+                    placeholder="LLM instructions",
+                    value=DEFAULT_INSTRUCTIONS,
+                    lines=10,
+                    interactive=True,
+                    label="Instructions",
+                    max_lines=16,
+                    show_label=False,
+                )
+
+    def run_chat(
+        message: str, chat_history, instructions: str, temperature: float, top_p: float
+    ):
+        if not message or (message == RETRY_COMMAND and len(chat_history) == 0):
+            yield chat_history
+            return
+
+        if message == RETRY_COMMAND and chat_history:
+            prev_turn = chat_history.pop(-1)
+            user_message, _ = prev_turn
+            message = user_message
+
+        prompt = format_chat_prompt(message, chat_history, instructions)
+        chat_history = chat_history + [[message, ""]]
+        stream = llm(
+            prompt,
+            max_new_tokens=1024,
+            stop=[STOP_STR, "<|endoftext|>"],
+            temperature=temperature,
+            top_p=top_p,
+            stream=True,
+        )
+        acc_text = ""
+        for idx, response in enumerate(stream):
+            text_token = response
+
+            if text_token in STOP_SUSPECT_LIST:
+                acc_text += text_token
+                continue
+
+            if idx == 0 and text_token.startswith(" "):
+                text_token = text_token[1:]
+
+            acc_text += text_token
+            last_turn = list(chat_history.pop(-1))
+            last_turn[-1] += acc_text
+            chat_history = chat_history + [last_turn]
+            yield chat_history
+            acc_text = ""
+
+    def delete_last_turn(chat_history):
+        if chat_history:
+            chat_history.pop(-1)
+        return {chatbot: gr.update(value=chat_history)}
+
+    def run_retry(
+        message: str, chat_history, instructions: str, temperature: float, top_p: float
+    ):
+        yield from run_chat(
+            RETRY_COMMAND, chat_history, instructions, temperature, top_p
         )
-    _ = """
-    msg.submit(
-        # fn=conversation.user_turn,
-        fn=predict,
-        inputs=[msg, chatbot],
-        outputs=[msg, chatbot],
-        # queue=True,
-        show_progress="full",
-        api_name="predict",
+
+    def clear_chat():
+        return []
+
+    inputs.submit(
+        run_chat,
+        [inputs, chatbot, instructions, temperature, top_p],
+        outputs=[chatbot],
+        show_progress="minimal",
     )
-    submit.click(
-        fn=lambda x, y: ("",) + predict(x, y)[1:],  # clear msg
-        inputs=[msg, chatbot],
-        outputs=[msg, chatbot],
-        queue=True,
-        show_progress="full",
+    inputs.submit(lambda: "", inputs=None, outputs=inputs)
+    submit_button.click(
+        run_chat,
+        [inputs, chatbot, instructions, temperature, top_p],
+        outputs=[chatbot],
+        show_progress="minimal",
     )
-    # """
-    msg.submit(
-        # fn=conversation.user_turn,
-        fn=predict_str,
-        inputs=[msg, chatbot],
-        outputs=[msg, chatbot],
-        queue=True,
-        show_progress="full",
-        api_name="predict",
-    ).then(bot_str, chatbot, chatbot)
-    submit.click(
-        fn=lambda x, y: ("",) + predict_str(x, y)[1:],  # clear msg
-        inputs=[msg, chatbot],
-        outputs=[msg, chatbot],
-        queue=True,
-        show_progress="full",
-    ).then(bot_str, chatbot, chatbot)
-
-    clear.click(lambda: None, None, chatbot, queue=False)
-
-    # update buff Textbox, every: units in seconds)
-    # https://huggingface.co/spaces/julien-c/nvidia-smi/discussions
-    # does not work
-    # AttributeError: 'Blocks' object has no attribute 'run_forever'
-    # block.run_forever(lambda: ns.response, None, [buff], every=1)
-
-    with gr.Accordion("For Chat/Translation API", open=False, visible=False):
-        input_text = gr.Text()
-        api_btn = gr.Button("Go", variant="primary")
-        out_text = gr.Text()
-    api_btn.click(
-        predict_api,
-        input_text,
-        out_text,
-        # show_progress="full",
-        api_name="api",
+    delete_turn_button.click(delete_last_turn, inputs=[chatbot], outputs=[chatbot])
+    retry_button.click(
+        run_retry,
+        [inputs, chatbot, instructions, temperature, top_p],
+        outputs=[chatbot],
+        show_progress="minimal",
     )
-
-# concurrency_count=5, max_size=20
-# max_size=36, concurrency_count=14
-block.queue(concurrency_count=5, max_size=20).launch(debug=True)
+    clear_chat_button.click(clear_chat, [], chatbot)
+
+
+def get_demo():
+    with gr.Blocks(
+        # css=None
+        # css="""#chat_container {width: 700px; margin-left: auto; margin-right: auto;}
+        #        #button_container {width: 700px; margin-left: auto; margin-right: auto;}
+        #        #param_container {width: 700px; margin-left: auto; margin-right: auto;}"""
+        css="""#chatbot {
+    font-size: 14px;
+    min-height: 300px;
+}"""
+    ) as demo:
+        gr.HTML(TITLE)
+
+        with gr.Row():
+            with gr.Column():
+                gr.Markdown(
+                    """
+                    ⚠️ **Limitations**: the model can and will produce factually incorrect information, hallucinating facts and actions. As it has not undergone any advanced tuning/alignment, it can produce problematic outputs, especially if prompted to do so.
+                    """
+                )
+
+        chat()
+
+    return demo
+
+
+if __name__ == "__main__":
+    demo = get_demo()
+    demo.queue(max_size=64, concurrency_count=8)
+    # demo.launch(server_name="0.0.0.0", server_port=7860)
+    demo.launch()