telco_app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM
import time
import torch
from pynvml import * # needs restart of IDE to install, from nvidia-ml-py3

# Get device
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Streamlit setup
st.title("Telco Chat Bot")
st.page_link("https://github.com/Ali-maatouk/Tele-LLMs", label="Tele-LLMs backend", icon="📱")
# Add text giving credit
col1, col2 = st.columns(2)
if 'conversation' not in st.session_state:
    st.session_state.conversation = []
user_input = st.text_input("You:", "") # user input

# Resource monitoring:
def print_gpu_utilization():
    nvmlInit()
    handle = nvmlDeviceGetHandleByIndex(0)
    info = nvmlDeviceGetMemoryInfo(handle)
    print(f"GPU memory occupied: {info.used//1024**2} MB.")


# Model functions:
@st.cache_resource(show_spinner=False)
def load_model():
    """ Load model from Hugging face."""
    print_gpu_utilization()
    success_placeholder = st.empty()
    with st.spinner("Loading model... please wait"):
        #model_name = "AliMaatouk/TinyLlama-1.1B-Tele"  # Replace with the correct model name
        #model_name = "AliMaatouk/LLama-3-8B-Tele-it"
        model_name = "AliMaatouk/Gemma-2B-Tele"
        if str(DEVICE) == "cuda:0": # may not need this, need to test on CPU if device map is okay anyway
            tokenizer = AutoTokenizer.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
        else:
            tokenizer = AutoTokenizer.from_pretrained(model_name, torch_dtype="auto")
        model = AutoModelForCausalLM.from_pretrained(model_name).to(DEVICE)

    success_placeholder.success("Model loaded successfully!", icon="🔥")
    time.sleep(2)
    success_placeholder.empty()
    return model, tokenizer

def generate_response(user_input):
    """ Query the model. """
    success_placeholder = st.empty()
    with st.spinner("Thinking..."):
        inputs = tokenizer(user_input, return_tensors="pt").to(DEVICE)
        #outputs = model.generate(**inputs, max_length=1000, pad_token_id=tokenizer.eos_token_id)
        outputs = model.generate(**inputs, max_new_tokens=750)
        print_gpu_utilization()
        generated_tokens = outputs[0, len(inputs['input_ids'][0]):]
    success_placeholder.success("Response generated!", icon="✅")
    time.sleep(2)
    success_placeholder.empty()
    text = tokenizer.decode(generated_tokens, skip_special_tokens=True)
    return text

# RUNTIME EVENTS:

# Load model and tokenizer
model, tokenizer = load_model()

# Submit button to send the query
with col1:
    if st.button("send"):
        if user_input:
            st.session_state.conversation.append({"role": "user", "content": user_input})
            # Querying model
            # Add a loading spinner during model loading
            response = generate_response(user_input)
            # Display bot response
            st.session_state.conversation.append({"role": "bot", "content": response})

# Clear button to reset
with col2:
    if st.button("clear chat"):
        if user_input:
            st.session_state.conversation = []

# Display conversation history
for chat in st.session_state.conversation:
    if chat['role'] == 'user':
        st.write(f"You: {chat['content']}")
    else:
        st.write(f"Bot: {chat['content']}")