app.py

import datetime
import json
import os
import shutil
from typing import Optional
from typing import Tuple

import gradio as gr
import torch
from fastchat.serve.inference import compress_module
from fastchat.serve.inference import raise_warning_for_old_weights
from huggingface_hub import Repository
from huggingface_hub import hf_hub_download
from huggingface_hub import snapshot_download
from peft import LoraConfig
from peft import get_peft_model
from peft import set_peft_model_state_dict
from transformers import AutoModelForCausalLM
from transformers import GenerationConfig
from transformers import LlamaTokenizer

print(datetime.datetime.now())

NUM_THREADS = 1

print(NUM_THREADS)

print("starting server ...")

BASE_MODEL = "decapoda-research/llama-13b-hf"
LORA_WEIGHTS = "izumi-lab/llama-13b-japanese-lora-v0-1ep"
HF_TOKEN = os.environ.get("HF_TOKEN", None)
DATASET_REPOSITORY = os.environ.get("DATASET_REPOSITORY", None)

repo = None
LOCAL_DIR = "/home/user/data/"
PROMPT_LANG = "en"
assert PROMPT_LANG in ["ja", "en"]

if HF_TOKEN and DATASET_REPOSITORY:
    try:
        shutil.rmtree(LOCAL_DIR)
    except Exception:
        pass

    repo = Repository(
        local_dir=LOCAL_DIR,
        clone_from=DATASET_REPOSITORY,
        use_auth_token=HF_TOKEN,
        repo_type="dataset",
    )
    repo.git_pull()

tokenizer = LlamaTokenizer.from_pretrained(BASE_MODEL)

if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"

try:
    if torch.backends.mps.is_available():
        device = "mps"
except Exception:
    pass

resume_from_checkpoint = snapshot_download(
    repo_id=LORA_WEIGHTS, use_auth_token=HF_TOKEN
)
checkpoint_name = hf_hub_download(
    repo_id=LORA_WEIGHTS, filename="adapter_model.bin", use_auth_token=HF_TOKEN
)
if device == "cuda":
    model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL, load_in_8bit=True, device_map="auto", torch_dtype=torch.float16
    )
elif device == "mps":
    model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL,
        device_map={"": device},
        load_in_8bit=True,
        torch_dtype=torch.float16,
    )
else:
    model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL,
        device_map={"": device},
        load_in_8bit=True,
        low_cpu_mem_usage=True,
        torch_dtype=torch.float16,
    )

config = LoraConfig.from_pretrained(resume_from_checkpoint)
model = get_peft_model(model, config)
adapters_weights = torch.load(checkpoint_name)
set_peft_model_state_dict(model, adapters_weights)
raise_warning_for_old_weights(BASE_MODEL, model)
compress_module(model, device)
# if device == "cuda" or device == "mps":
#     model = model.to(device)


def generate_prompt(instruction: str, input: Optional[str] = None):
    if input:
        if PROMPT_LANG == "ja":
            return f"以下はタスクを説明する指示とさらなる文脈を適用する入力の組み合わせです。\n\n### 指示:\n{instruction}\n\n### 入力:\n{input}\n\n### Response:\n"
        elif PROMPT_LANG == "en":
            return f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
{instruction}
### Input:
{input}
### Response:"""
        else:
            raise ValueError("PROMPT_LANG")
    else:
        if PROMPT_LANG == "ja":
            return f"以下はタスクを説明する指示とさらなる文脈を適用する入力の組み合わせです。\n\n### 指示:\n{instruction}\n\n### 返答:\n"
        elif PROMPT_LANG == "en":
            return f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{instruction}
### Response:"""
        else:
            raise ValueError("PROMPT_LANG")


if device != "cpu":
    model.half()
model.eval()
if torch.__version__ >= "2":
    model = torch.compile(model)


def save_inputs_and_outputs(now, inputs, outputs, generate_kwargs):
    current_hour = now.strftime("%Y-%m-%d_%H")
    file_name = f"prompts_{LORA_WEIGHTS.split('/')[-1]}_{current_hour}.jsonl"

    if repo is not None:
        repo.git_pull(rebase=True)
        with open(os.path.join(LOCAL_DIR, file_name), "a", encoding="utf-8") as f:
            json.dump(
                {
                    "inputs": inputs,
                    "outputs": outputs,
                    "generate_kwargs": generate_kwargs,
                },
                f,
                ensure_ascii=False,
            )
            f.write("\n")
        repo.push_to_hub()


# we cant add typing now
# https://github.com/gradio-app/gradio/issues/3514
def evaluate(
    instruction,
    input=None,
    temperature=0.7,
    max_tokens=384,
    repetition_penalty=1.0,
):
    num_beams: int = 1
    top_p: float = 0.75
    top_k: int = 40
    prompt = generate_prompt(instruction, input)
    inputs = tokenizer(prompt, return_tensors="pt")
    if len(inputs["input_ids"][0]) > max_tokens + 10:
        if HF_TOKEN and DATASET_REPOSITORY:
            try:
                now = datetime.datetime.now()
                current_time = now.strftime("%Y-%m-%d %H:%M:%S")
                print(f"[{current_time}] Pushing prompt and completion to the Hub")
                save_inputs_and_outputs(
                    now,
                    prompt,
                    "",
                    {
                        "temperature": temperature,
                        "top_p": top_p,
                        "top_k": top_k,
                        "num_beams": num_beams,
                        "max_tokens": max_tokens,
                        "repetition_penalty": repetition_penalty,
                    },
                )
            except Exception as e:
                print(e)
        return (
            f"please reduce the input length. Currently, {len(inputs['input_ids'][0])} tokens are used.",
            gr.update(interactive=True),
            gr.update(interactive=True),
        )
    input_ids = inputs["input_ids"].to(device)
    generation_config = GenerationConfig(
        temperature=temperature,
        top_p=top_p,
        top_k=top_k,
        repetition_penalty=repetition_penalty,
        num_beams=num_beams,
        pad_token_id=tokenizer.pad_token_id,
        eos_token=tokenizer.eos_token_id,
    )
    with torch.no_grad():
        generation_output = model.generate(
            input_ids=input_ids,
            generation_config=generation_config,
            return_dict_in_generate=True,
            output_scores=True,
            max_new_tokens=max_tokens - len(input_ids),
        )
    s = generation_output.sequences[0]
    output = tokenizer.decode(s, skip_special_tokens=True)
    if prompt.endswith("Response:"):
        output = output.split("### Response:")[1].strip()
    elif prompt.endswith("返答:"):
        output = output.split("### 返答:")[1].strip()
    else:
        raise ValueError(f"No valid prompt ends. {prompt}")
    if HF_TOKEN and DATASET_REPOSITORY:
        try:
            now = datetime.datetime.now()
            current_time = now.strftime("%Y-%m-%d %H:%M:%S")
            print(f"[{current_time}] Pushing prompt and completion to the Hub")
            save_inputs_and_outputs(
                now,
                prompt,
                output,
                {
                    "temperature": temperature,
                    "top_p": top_p,
                    "top_k": top_k,
                    "num_beams": num_beams,
                    "max_tokens": max_tokens,
                    "repetition_penalty": repetition_penalty,
                },
            )
        except Exception as e:
            print(e)
    return output, gr.update(interactive=True), gr.update(interactive=True)


def reset_textbox():
    return gr.update(value=""), gr.update(value=""), gr.update(value="")


def no_interactive() -> Tuple[gr.Request, gr.Request]:
    return gr.update(interactive=False), gr.update(interactive=False)


title = """<h1 align="center">LLaMA-13B Japanese LoRA</h1>"""

theme = gr.themes.Default(primary_hue="green")
description = (
    "The official demo for **[izumi-lab/llama-13b-japanese-lora-v0-1ep](https://huggingface.co/izumi-lab/llama-13b-japanese-lora-v0-1ep)**.  "
    "It is a 13B-parameter LLaMA model finetuned to follow instructions.  "
    "It is trained on the [izumi-lab/llm-japanese-dataset](https://huggingface.co/datasets/izumi-lab/llm-japanese-dataset) dataset.  "
    "For more information, please visit [the project's website](https://llm.msuzuki.me).  "
    "This model can output up to 256 tokens, but the maximum number of tokens is 227 due to the GPU memory limit of HuggingFace Space.  "
    "Currently temperature is ineffective because num_beam is set to 1 due to lack of GPU memory.  "
    "It takes about **1 minute** to output. When access is concentrated, the operation may become slow."
)
with gr.Blocks(
    css="""#col_container { margin-left: auto; margin-right: auto;}""",
    theme=theme,
) as demo:
    gr.HTML(title)
    gr.Markdown(description)
    with gr.Column(elem_id="col_container", visible=False) as main_block:
        with gr.Row():
            with gr.Column():
                instruction = gr.Textbox(
                    lines=2, label="Instruction", placeholder="こんにちは"
                )
                inputs = gr.Textbox(lines=1, label="Input", placeholder="none")
                with gr.Row():
                    with gr.Column(scale=3):
                        clear_button = gr.Button("Clear").style(full_width=True)
                    with gr.Column(scale=5):
                        submit_button = gr.Button("Submit").style(full_width=True)
            outputs = gr.Textbox(lines=4, label="Output")

        # inputs, top_p, temperature, top_k, repetition_penalty
        with gr.Accordion("Parameters", open=True):
            temperature = gr.Slider(
                minimum=0,
                maximum=1.0,
                value=0.7,
                step=0.05,
                interactive=True,
                label="Temperature (Currently ineffective)",
            )
            max_tokens = gr.Slider(
                minimum=20,
                maximum=227,
                value=128,
                step=1,
                interactive=True,
                label="Max length (Pre-prompt + instruction + input + output))",
            )
            repetition_penalty = gr.Slider(
                minimum=1.0,
                maximum=5.0,
                value=1.2,
                step=0.1,
                interactive=True,
                label="Repetition penalty",
            )

    with gr.Column(elem_id="user_consent_container") as user_consent_block:
        # Get user consent
        gr.Markdown(
            """
            ## User Consent for Data Collection, Use, and Sharing:
            By using our app, you acknowledge and agree to the following terms regarding the data you provide:

            - **Collection**: We may collect inputs you type into our app.
            - **Use**: We may use the collected data for research purposes, to improve our services, and to develop new products or services, including commercial applications.
            - **Sharing and Publication**: Your input data may be published, shared with third parties, or used for analysis and reporting purposes.
            - **Data Retention**: We may retain your input data for as long as necessary.

            By continuing to use our app, you provide your explicit consent to the collection, use, and potential sharing of your data as described above. If you do not agree with our data collection, use, and sharing practices, please do not use our app.

            ## データ収集、利用、共有に関するユーザーの同意：
            本アプリを使用することにより、提供するデータに関する以下の条件に同意するものとします：

            - **収集**: 本アプリに入力されるテキストデータは収集される場合があります。
            - **利用**: 収集されたデータは研究や、商用アプリケーションを含むサービスの開発に使用される場合があります。
            - **共有および公開**: 入力データは第三者と共有されたり、分析や公開の目的で使用される場合があります。
            - **データ保持**: 入力データは必要な限り保持されます。

            本アプリを引き続き使用することにより、上記のようにデータの収集・利用・共有について同意します。データの利用方法に同意しない場合は、本アプリを使用しないでください。
            """
        )
        accept_button = gr.Button("I Agree")

        def enable_inputs():
            return user_consent_block.update(visible=False), main_block.update(
                visible=True
            )

    accept_button.click(
        fn=enable_inputs,
        inputs=[],
        outputs=[user_consent_block, main_block],
        queue=False,
    )
    inputs.submit(no_interactive, [], [submit_button, clear_button])
    inputs.submit(
        evaluate,
        [instruction, inputs, temperature, max_tokens, repetition_penalty],
        [outputs, submit_button, clear_button],
    )
    submit_button.click(no_interactive, [], [submit_button, clear_button])
    submit_button.click(
        evaluate,
        [instruction, inputs, temperature, max_tokens, repetition_penalty],
        [outputs, submit_button, clear_button],
    )
    clear_button.click(reset_textbox, [], [instruction, inputs, outputs], queue=False)

    demo.queue(max_size=20, concurrency_count=NUM_THREADS, api_open=False).launch(
        share=True, server_name="0.0.0.0", server_port=7860
    )