app.py

import streamlit as st
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from datetime import datetime

# Initialize session state variables
if 'messages' not in st.session_state:
    st.session_state.messages = []
if "user_input_widget" not in st.session_state:
    st.session_state.user_input_widget = ""

@st.cache_resource
def load_model():
    tokenizer = AutoTokenizer.from_pretrained("amd/AMD-OLMo-1B-SFT")
    model = AutoModelForCausalLM.from_pretrained("amd/AMD-OLMo-1B-SFT")
    if torch.cuda.is_available():
        model = model.to("cuda")
    return model, tokenizer

def generate_response(prompt, model, tokenizer, history):
    # Format conversation history with the template
    bos = tokenizer.eos_token
    conversation = ""
    for msg in history:
        if msg["role"] == "user":
            conversation += f"<|user|>\n{msg['content']}\n"
        else:
            conversation += f"<|assistant|>\n{msg['content']}\n"
    
    template = bos + conversation + f"<|user|>\n{prompt}\n<|assistant|>\n"
    
    inputs = tokenizer([template], return_tensors='pt', return_token_type_ids=False)
    if torch.cuda.is_available():
        inputs = inputs.to("cuda")
    
    outputs = model.generate(
        **inputs,
        max_new_tokens=1000,
        do_sample=True,
        top_k=50,
        top_p=0.95,
        temperature=0.7
    )
    
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    # Extract only the assistant's last response
    response = response.split("<|assistant|>\n")[-1].strip()
    return response

def main():
    st.set_page_config(
        page_title="AMD-OLMo Chatbot",
        layout="wide",
        initial_sidebar_state="expanded"
    )
    
    # Custom CSS
    st.markdown("""
        <style>
        .stTab {
            font-size: 20px;
        }
        .model-info {
            background-color: #f0f2f6;
            padding: 20px;
            border-radius: 10px;
            margin: 10px 0;
        }
        .chat-message {
            padding: 15px;
            border-radius: 10px;
            margin: 10px 0;
        }
        .user-message {
            background-color: #e6f3ff;
            border-left: 5px solid #2e6da4;
        }
        .assistant-message {
            background-color: #f0f2f6;
            border-left: 5px solid #5cb85c;
        }
        .stTextArea textarea {
            font-size: 16px;
        }
        .timestamp {
            font-size: 12px;
            color: #666;
            margin-top: 5px;
        }
        .st-emotion-cache-1v0mbdj.e115fcil1 {
            margin-top: 20px;
        }
        </style>
    """, unsafe_allow_html=True)

    # Create tabs
    tab1, tab2 = st.tabs(["Model Information", "Chat Interface"])

    with tab1:
        st.title("AMD-OLMo-1B-SFT Model Information")
        
        with st.container():
            st.markdown("""
            <div class="model-info">
            <h2>Model Overview</h2>
            AMD-OLMo-1B-SFT is a state-of-the-art language model developed by AMD. This model represents a significant advancement in AMD's AI capabilities.
            
            <h3>Architecture Specifications</h3>
            
            | Component | Specification |
            |-----------|---------------|
            | Parameters | 1.2B |
            | Layers | 16 |
            | Attention Heads | 16 |
            | Hidden Size | 2048 |
            | Context Length | 2048 |
            | Vocabulary Size | 50,280 |
            
            <h3>Training Details</h3>
            
            - Pre-trained on 1.3 trillion tokens from Dolma v1.7
            - Two-phase supervised fine-tuning (SFT):
                1. Tulu V2 dataset
                2. OpenHermes-2.5, WebInstructSub, and Code-Feedback datasets
            
            <h3>Key Capabilities</h3>
            
            - Natural language understanding and generation
            - Context-aware responses
            - Code understanding and generation
            - Complex reasoning tasks
            - Instruction following
            - Multi-turn conversations
            
            <h3>Hardware Optimization</h3>
            
            - Optimized for AMD Instinct™ MI250 GPUs
            - Distributed training across 16 nodes with 4 GPUs each
            - Efficient inference on consumer hardware
            </div>
            """, unsafe_allow_html=True)

    with tab2:
        st.title("Chat with AMD-OLMo")
        
        # Load model
        try:
            model, tokenizer = load_model()
            st.success("Model loaded successfully! You can start chatting.")
        except Exception as e:
            st.error(f"Error loading model: {str(e)}")
            return

        # Chat interface
        st.markdown("### Chat History")
        chat_container = st.container()
        
        with chat_container:
            for message in st.session_state.messages:
                div_class = "user-message" if message["role"] == "user" else "assistant-message"
                timestamp = message.get("timestamp", datetime.now().strftime("%Y-%m-%d %H:%M:%S"))
                st.markdown(f"""
                    <div class="chat-message {div_class}">
                        <b>{message["role"].title()}:</b> {message["content"]}
                        <div class="timestamp">{timestamp}</div>
                    </div>
                """, unsafe_allow_html=True)

        # User input section
        with st.container():
            user_input = st.text_area(
                "Your message:",
                key="user_input_widget",
                height=100,
                placeholder="Type your message here..."
            )
            
            col1, col2, col3 = st.columns([1, 1, 4])
            
            with col1:
                if st.button("Send", use_container_width=True):
                    if user_input.strip():
                        # Add user message to history with timestamp
                        st.session_state.messages.append({
                            "role": "user",
                            "content": user_input,
                            "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S")
                        })
                        
                        # Generate response
                        with st.spinner("Generating response..."):
                            response = generate_response(user_input, model, tokenizer, st.session_state.messages)
                        
                        # Add assistant response to history with timestamp
                        st.session_state.messages.append({
                            "role": "assistant",
                            "content": response,
                            "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S")
                        })
                        
                        # Clear input
                        st.session_state.user_input_widget = ""
                        st.experimental_rerun()
            
            with col2:
                if st.button("Clear History", use_container_width=True):
                    st.session_state.messages = []
                    st.session_state.user_input_widget = ""
                    st.experimental_rerun()

if __name__ == "__main__":
    main()