app.py

import gradio as gr
from gradio_client import Client
from huggingface_hub import InferenceClient
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]),
]

def compress_history(history,client_choice,seed,temp,tokens,top_p,rep_p,chat_mem):
    client=clients[int(client_choice)-1]
    COMPRESS_HISTORY="""You are an Information Summarizer Agent. Your duty is to summarize the following information into a more concise format with far less words.
    Retain all the main points and provide a brief and concise summary of the conversation.
    Converstion:
    {history}"""
    print("COMPRESSING")
    formatted_prompt=f"{COMPRESS_HISTORY.format(history=history[0-chat_mem:])}"
    generate_kwargs = dict(
        temperature=temp,
        max_new_tokens=1024,
        top_p=top_p,
        repetition_penalty=rep_p,
        do_sample=True,
        seed=seed,
    )
    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
    output = ""
    for response in stream:
        output += response.token.text
    return output
    
def format_prompt(message, history):
    prompt = ""
    if history:
        #<start_of_turn>userHow does the brain work?<end_of_turn><start_of_turn>model
        for user_prompt, bot_response in history:
            prompt += f"{user_prompt}\n"
            #print(prompt)
            prompt += f"{bot_response}\n"
            #print(prompt)
    prompt += f"<start_of_turn>user{message}<end_of_turn><start_of_turn>model"
    #print(prompt)
    return prompt
result = []


def chat_inf(system_prompt,prompt,history,memory,client_choice,seed,temp,tokens,top_p,rep_p,chat_mem):
    #token max=8192
    hist_len=0
    client=clients[int(client_choice)-1]
    if not history:
        history = []
        hist_len=0
    if memory:
        for ea in memory[0-chat_mem:]:
            hist_len+=len(str(ea))
        print(hist_len)
    in_len=len(system_prompt+prompt)+hist_len
    
    
    print("\n######### HIST "+str(in_len))
    print("\n######### TOKENS "+str(tokens))
    if (in_len+tokens) > 8000:
        history.append((prompt,"Wait, that's too many tokens, please reduce the Chat Memory value"))
        yield history,memory
        #hist=compress_history(history,client_choice,seed,temp,tokens,top_p,rep_p,chat_mem)
        #yield [(prompt,"History has been compressed, processing request...")]
        #history.append((prompt,hist))
    else:
        generate_kwargs = dict(
            temperature=temp,
            max_new_tokens=tokens,
            top_p=top_p,
            repetition_penalty=rep_p,
            do_sample=True,
            seed=seed,
        )
        #formatted_prompt=prompt   
        formatted_prompt = format_prompt(f"{system_prompt}, {prompt}", memory[0-chat_mem:])
        print("\n######### PROMPT "+str(len(formatted_prompt)))
        stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
        output = ""
        for response in stream:
            output += response.token.text
            yield [(prompt,output)],memory
        history.append((prompt,output))
        memory.append((prompt,output))
        yield history,memory

def get_screenshot(chat: list,height=5000,width=600,chatblock=[],theme="light",wait=3000,header=True):
    print(chatblock)
    tog = 0
    if chatblock:
        tog = 3
    result = ss_client.predict(str(chat),height,width,chatblock,header,theme,wait,api_name="/run_script")
    out = f'https://omnibus-html-image-current-tab.hf.space/file={result[tog]}'
    print(out)
    return out



def clear_fn():
    return 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))


    
with gr.Blocks() as app:
    memory=gr.State()
    gr.HTML("""<center><h1 style='font-size:xx-large;'>Google Gemma Models</h1><br><h3>running on Huggingface Inference Client</h3><br><h7>EXPERIMENTAL""")
    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.Column(scale=1):
                with gr.Group():
                    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=3840,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=5)
        with gr.Accordion(label="Screenshot",open=False):
            with gr.Row():
                with gr.Column(scale=3):
                    im_btn=gr.Button("Screenshot")
                    img=gr.Image(type='filepath')
                with gr.Column(scale=1):
                    with gr.Row():
                        im_height=gr.Number(label="Height",value=5000)
                        im_width=gr.Number(label="Width",value=500)
                    wait_time=gr.Number(label="Wait Time",value=3000)
                    theme=gr.Radio(label="Theme", choices=["light","dark"],value="light")
                    chatblock=gr.Dropdown(label="Chatblocks",info="Choose specific blocks of chat",choices=[c for c in range(1,40)],multiselect=True)
            


    im_go=im_btn.click(get_screenshot,[chat_b,im_height,im_width,chatblock,theme,wait_time],img)
    chat_sub=inp.submit(check_rand,[rand,seed],seed).then(chat_inf,[sys_inp,inp,chat_b,memory,client_choice,seed,temp,tokens,top_p,rep_p,chat_mem],[chat_b,memory])
    
    go=btn.click(check_rand,[rand,seed],seed).then(chat_inf,[sys_inp,inp,chat_b,memory,client_choice,seed,temp,tokens,top_p,rep_p,chat_mem],[chat_b,memory])
    
    stop_btn.click(None,None,None,cancels=[go,im_go,chat_sub])
    clear_btn.click(clear_fn,None,[inp,sys_inp,chat_b])
app.queue(default_concurrency_limit=10).launch()