app.py

import json
from threading import Thread

import gradio as gr
import spaces
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    TextIteratorStreamer,
)

MAX_NEW_TOKENS = 8192
MODEL_NAME = "Azure99/Blossom-V6.2-14B"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)


def get_messages(user, history):
    try:
        parsed_body = json.loads(user)
        if parsed_body.get("by_json_str"):
            return parsed_body["messages"]
    except:
        pass

    messages = []
    messages.extend(history or [])
    messages.append({"role": "user", "content": user})
    return messages


@spaces.GPU(duration=120)
def chat(user, history, temperature, top_p, repetition_penalty):
    streamer = TextIteratorStreamer(
        tokenizer, skip_prompt=True, skip_special_tokens=True
    )

    messages = get_messages(user, history)
    print(f"Messages: {messages}")
    input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to(
        model.device
    )

    generation_kwargs = dict(
        input_ids=input_ids,
        streamer=streamer,
        do_sample=True,
        max_new_tokens=MAX_NEW_TOKENS,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
    )

    Thread(target=model.generate, kwargs=generation_kwargs).start()

    outputs = ""
    for new_text in streamer:
        outputs += new_text
        yield outputs


additional_inputs = [
    gr.Slider(
        label="Temperature",
        value=0.5,
        minimum=0.0,
        maximum=1.0,
        step=0.05,
        interactive=True,
        info="Controls randomness in choosing words.",
    ),
    gr.Slider(
        label="Top-P",
        value=0.85,
        minimum=0.0,
        maximum=1.0,
        step=0.05,
        interactive=True,
        info="Picks words until their combined probability is at least top_p.",
    ),
    gr.Slider(
        label="Repetition penalty",
        value=1.05,
        minimum=1.0,
        maximum=1.2,
        step=0.01,
        interactive=True,
        info="Repetition Penalty: Controls how much repetition is penalized.",
    ),
]

gr.ChatInterface(
    chat,
    type="messages",
    chatbot=gr.Chatbot(
        show_label=False,
        height=500,
        show_copy_button=True,
        render_markdown=True,
        type="messages",
        latex_delimiters=[{"left": "\\[", "right": "\\]", "display": True}],
    ),
    textbox=gr.Textbox(placeholder="", container=False, scale=7),
    title=f"{MODEL_NAME} Demo",
    description="Hello, I am Blossom, an open source conversational large language model.🌠"
    '<a href="https://github.com/Azure99/BlossomLM">GitHub</a>',
    theme="soft",
    examples=[
        ["Hello"],
        ["What is MBTI"],
        ["用Python实现二分查找"],
        ["为switch写一篇小红书种草文案，带上emoji"],
    ],
    cache_examples=False,
    additional_inputs=additional_inputs,
    additional_inputs_accordion=gr.Accordion(label="Config", open=True),
).queue().launch()