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 = "hanzla/Falcon3-Mamba-R1-v0"

TITLE = "<h1><center>Falcon3-Mamba-R1-v0 playground</center></h1>"

SUB_TITLE = """<center>Falcon3 Mamba R1 is a Selective State Space model (Mamba) that scales on test time compute for reasoning.</center>"""
SYSTEM_PROMPT = os.getenv('SYSTEM_PROMPT')

print(SYSTEM_PROMPT)

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})

    print(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(theme="JohnSmith9982/small_and_pretty") as demo:
    gr.HTML(TITLE)
    gr.HTML(SUB_TITLE)
    
    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=4096, 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=[
            ["""Consider the following statements:

1. If it rains, then the ground will be wet.
2. It is raining.

Using propositional logic, determine whether the conclusion "The ground is wet" is valid.  
Also, identify the rule of inference used to reach the conclusion.
"""],
            ["""A satellite is in a circular orbit around Earth at an altitude of 500 km above the surface. Calculate:

1. The orbital velocity of the satellite.  
2. The orbital period of the satellite.

Given:
- Radius of Earth, R_E = 6.37 × 10^6 m  
- Gravitational constant, G = 6.674 × 10^−11 Nm²/kg²  
- Mass of Earth, M_E = 5.97 × 10^24 kg"""],
        ],
        cache_examples=False,
    )

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