app.py

import subprocess
import sys
import shlex
import spaces
import torch
import uuid
import os
import json
from pathlib import Path
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from threading import Thread


# install packages for mamba
def install_mamba():
    subprocess.run(shlex.split("pip install https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.4.0/causal_conv1d-1.4.0+cu122torch2.3cxx11abiFALSE-cp310-cp310-linux_x86_64.whl"))
    subprocess.run(shlex.split("pip install https://github.com/state-spaces/mamba/releases/download/v2.2.2/mamba_ssm-2.2.2+cu122torch2.3cxx11abiFALSE-cp310-cp310-linux_x86_64.whl"))

install_mamba()

MODEL = "tiiuae/Falcon3-Mamba-7B-Instruct"

TITLE = "<h1><center>Falcon3-Mamba-7B-Instruct playground</center></h1>"

SUB_TITLE = """<center>Playground of Falcon3-Mamba-7B-Instruct</center>"""
SYSTEM_PROMPT = os.getenv('SYSTEM_PROMPT')

CSS = """
.duplicate-button {
    margin: auto !important;
    color: white !important;
    background: black !important;
    border-radius: 100vh !important;
}
h3 {
    text-align: center;
/* Fix for chat container */
.chat-container {
    height: 600px !important;
    overflow-y: auto !important;
    flex-direction: column !important;
}
.messages-container {
    flex-grow: 1 !important;
    overflow-y: auto !important;
    padding-right: 10px !important;
}
/* Ensure consistent height */
.contain {
    height: 100% !important;
}
"""

END_MESSAGE = """
\n
**The conversation has reached to its end, please press "Clear" to restart a new conversation**
"""

device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    torch_dtype=torch.bfloat16,
).to(device)

if device == "cuda":
    model = torch.compile(model)

@spaces.GPU
def stream_chat(
    message: str, 
    history: list, 
    temperature: float = 0.3, 
    max_new_tokens: int = 100, 
    top_p: float = 1.0, 
    top_k: int = 20, 
    penalty: float = 1.2,
):
    print(f'message: {message}')
    print(f'history: {history}')

    conversation = []
    for prompt, answer in history:
        conversation.extend([
            {"role": 'system', "content": SYSTEM_PROMPT },
            {"role": "user", "content": prompt}, 
            {"role": "assistant", "content": answer},
        ])

    conversation.append({"role": "user", "content": message})
    
    input_text = tokenizer.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer.encode(input_text, return_tensors="pt").to(device)
    streamer = TextIteratorStreamer(tokenizer, timeout=40.0, skip_prompt=True, skip_special_tokens=True)
    
    generate_kwargs = dict(
        input_ids=inputs, 
        max_new_tokens=max_new_tokens,
        do_sample=False if temperature == 0 else True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        streamer=streamer,
        pad_token_id=11,
    )

    with torch.no_grad():
        thread = Thread(target=model.generate, kwargs=generate_kwargs)
        thread.start()
        
    buffer = ""
    for new_text in streamer:
        buffer += new_text
        buffer = buffer.replace("\nUser", "")
        buffer = buffer.replace("\nSystem", "")
        yield buffer

    print(f'response: {buffer}')
            
with gr.Blocks(css=CSS, theme="soft") as demo:
    gr.HTML(TITLE)
    gr.HTML(SUB_TITLE)
    gr.DuplicateButton(value="Duplicate Space for private use", elem_classes="duplicate-button")
    
    chat_interface = gr.ChatInterface(
        fn=stream_chat,
        chatbot=gr.Chatbot(
            height=600,
            container=True,
            elem_classes=["chat-container"]
        ),
        fill_height=True,
        additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False, render=False),
        additional_inputs=[
            gr.Slider(minimum=0, maximum=1, step=0.1, value=0.3, label="Temperature", render=False),
            gr.Slider(minimum=128, maximum=32768, step=1, value=1024, label="Max new tokens", render=False),
            gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1.0, label="top_p", render=False),
            gr.Slider(minimum=1, maximum=20, step=1, value=20, label="top_k", render=False),
            gr.Slider(minimum=0.0, maximum=2.0, step=0.1, value=1.2, label="Repetition penalty", render=False),
        ],
        examples=[
            ["Hello there, can you suggest few places to visit in UAE?"],
            ["What UAE is known for?"],
        ],
        cache_examples=False,
    )

if __name__ == "__main__":
    demo.launch()