app.py

import gradio as gr
from gradio_client import Client
from huggingface_hub import InferenceClient
from deep_translator import GoogleTranslator
import random

ss_client = Client("https://omnibus-html-image-current-tab.hf.space/")

models = [
    "google/gemma-7b",
    "google/gemma-7b-it",
    "google/gemma-2b",
    "google/gemma-2b-it"
]

clients = [
    InferenceClient(models[0]),
    InferenceClient(models[1]),
    InferenceClient(models[2]),
    InferenceClient(models[3]),
]

VERBOSE = False

def translate_to_english(prompt):
    translated_prompt = GoogleTranslator(source='auto', target='en').translate(prompt)
    return translated_prompt

def translate_to_persian_text(response):
    translated_response = GoogleTranslator(source='auto', target='fa').translate(response)
    return translated_response

def load_models(inp):
    if VERBOSE == True:
        print(type(inp))
        print(inp)
        print(models[inp])
    return gr.update(label=models[inp])

def format_prompt(message, history, cust_p):
    prompt = "<s>"
    if history:
        for user_prompt, bot_response in history:
            prompt += f"<start_of_turn>user{user_prompt}<end_of_turn>"
            prompt += f"<start_of_turn>model{bot_response}<end_of_turn></s>"
            if VERBOSE == True:
                print(prompt)
    prompt += cust_p.replace("USER_INPUT", message)
    return prompt

def chat_inf(system_prompt, prompt, history, memory, client_choice, seed, temp, tokens, top_p, rep_p, chat_mem, custom_prompt, translate_to_persian):
    hist_len = 0
    client = clients[int(client_choice) - 1]
    if not history:
        history = []
    if not memory:
        memory = []

    if memory:
        for ea in memory[0 - chat_mem:]:
            hist_len += len(str(ea))
    in_len = len(system_prompt + prompt) + hist_len

    if (in_len + tokens) > 8000:
        history.append((prompt, "Wait, that's too many tokens, please reduce the 'Chat Memory' value, or reduce the 'Max new tokens' value"))
        yield history, memory
    else:
        generate_kwargs = dict(
            max_new_tokens=tokens,
        )
        if system_prompt:
            formatted_prompt = format_prompt(f"{system_prompt}, {prompt}", memory[0 - chat_mem:], custom_prompt)
        else:
            formatted_prompt = format_prompt(prompt, memory[0 - chat_mem:], custom_prompt)

        translated_prompt = translate_to_english(formatted_prompt)

        chat = [
            {"role": "user", "content": f"{translated_prompt}"},
        ]

        stream = client.text_generation(translated_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True)
        output = ""
        for response in stream:
            output += response.token.text
            if translate_to_persian:
                output = translate_to_persian_text(output)
            yield [(prompt, output)], memory
        history.append((prompt, output))
        memory.append((prompt, output))
        yield history, memory

def clear_fn():
    return None, None, None, None

rand_val = random.randint(1, 1111111111111111)

def check_rand(inp, val):
    if inp == True:
        return gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, value=random.randint(1, 1111111111111111))
    else:
        return gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, value=int(val))

def chat_wrapper(sys_inp, inp, chat_b, memory, client_choice, seed, temp, tokens, top_p, rep_p, chat_mem, custom_prompt, translate_to_persian_checkbox):
    return chat_inf(sys_inp, inp, chat_b, memory, client_choice, seed, temp, tokens, top_p, rep_p, chat_mem, custom_prompt, translate_to_persian_checkbox)

with gr.Blocks() as app:
    memory = gr.State()
    chat_b = gr.Chatbot(height=500)
    with gr.Group():
        with gr.Row():
            with gr.Column(scale=3):
                inp = gr.Textbox(label="Prompt")
                sys_inp = gr.Textbox(label="System Prompt (optional)")
                with gr.Row():
                    with gr.Column(scale=2):
                        btn = gr.Button("Chat")
                    with gr.Column(scale=1):
                        with gr.Group():
                            stop_btn = gr.Button("Stop")
                            clear_btn = gr.Button("Clear")
                client_choice = gr.Dropdown(label="Models", type='index', choices=[c for c in models], value=models[0], interactive=True)
                with gr.Accordion("Prompt Format", open=False):
                    custom_prompt = gr.Textbox(label="Modify Prompt Format", info="For testing purposes. 'USER_INPUT' is where 'SYSTEM_PROMPT, PROMPT' will be placed", lines=5, value="<start_of_turn>userUSER_INPUT<end_of_turn><start_of_turn>model")
            with gr.Column(scale=1):
                with gr.Group():
                    translate_to_persian_checkbox = gr.Checkbox(label="Translate to Persian", value=True)
                    rand = gr.Checkbox(label="Random Seed", value=True)
                    seed = gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, step=1, value=rand_val)
                    tokens = gr.Slider(label="Max new tokens", value=1600, minimum=0, maximum=8000, step=64, interactive=True, visible=True, info="The maximum number of tokens")
                    temp = gr.Slider(label="Temperature", step=0.01, minimum=0.01, maximum=1.0, value=0.9)
                    top_p = gr.Slider(label="Top-P", step=0.01, minimum=0.01, maximum=1.0, value=0.9)
                    rep_p = gr.Slider(label="Repetition Penalty", step=0.1, minimum=0.1, maximum=2.0, value=1.0)
                    chat_mem = gr.Number(label="Chat Memory", info="Number of previous chats to retain", value=4)

    client_choice.change(load_models, client_choice, [chat_b])
    app.load(load_models, client_choice, [chat_b])

    chat_sub = inp.submit(check_rand, [rand, seed], seed).then(chat_wrapper, [sys_inp, inp, chat_b, memory, client_choice, seed, temp, tokens, top_p, rep_p, chat_mem, custom_prompt, translate_to_persian_checkbox]).then(chat_b.display, memory)
    go = btn.click(check_rand, [rand, seed], seed).then(chat_wrapper, [sys_inp, inp, chat_b, memory, client_choice, seed, temp, tokens, top_p, rep_p, chat_mem, custom_prompt, translate_to_persian_checkbox]).then(chat_b.display, memory)

    clear_btn.click(clear_fn, None, [inp, sys_inp, chat_b, memory])
app.queue(default_concurrency_limit=10).launch()