app.py

import gc
import torch
import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login
import re
import os
from PIL import Image

# Load Hugging Face token
HF_TOKEN = os.getenv("HF_TOKEN")
login(token=HF_TOKEN)

# Define models
MODELS = {
    "athena-1": {
        "name": "🦁 Atlas-Flash 1215",
        "sizes": {
            "1.5B": "Spestly/Atlas-R1-1.5B-Preview",
        },
        "emoji": "🦁",
        "experimental": True,
        "is_vision": False,  # Enable vision support for this model
        "system_prompt_env": "ATLAS_FLASH_1215",  # Environment variable for system prompt
    },
}

# Profile pictures
USER_PFP = "user.png"  # Hugging Face user avatar
AI_PFP = "ai_pfp.png"  # Replace with the path to your AI's image or a URL

class AtlasInferenceApp:
    def __init__(self):
        if "current_model" not in st.session_state:
            st.session_state.current_model = {"tokenizer": None, "model": None, "config": None}
        if "chat_history" not in st.session_state:
            st.session_state.chat_history = []

        st.set_page_config(
            page_title="Atlas Model Inference",
            page_icon="🦁 ",
            layout="wide",
            menu_items={
                'Get Help': 'https://huggingface.co/collections/Spestly/athena-1-67623e58bfaadd3c2fcffb86',
                'Report a bug': 'https://huggingface.co/Spestly/Athena-1-1.5B/discussions/new',
                'About': 'Athena Model Inference Platform'
            }
        )

    def clear_memory(self):
        """Optimize memory management for CPU inference"""
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        gc.collect()

    def load_model(self, model_key, model_size):
        try:
            self.clear_memory()

            if st.session_state.current_model["model"] is not None:
                del st.session_state.current_model["model"]
                del st.session_state.current_model["tokenizer"]
                self.clear_memory()

            model_path = MODELS[model_key]["sizes"][model_size]
            
            # Load Qwen-compatible tokenizer and model
            tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
            model = AutoModelForCausalLM.from_pretrained(
                model_path,
                device_map="auto",  # Use GPU if available
                torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
                trust_remote_code=True,
                low_cpu_mem_usage=True
            )

            # Update session state
            st.session_state.current_model.update({
                "tokenizer": tokenizer,
                "model": model,
                "config": {
                    "name": f"{MODELS[model_key]['name']} {model_size}",
                    "path": model_path,
                    "system_prompt": os.getenv(MODELS[model_key]["system_prompt_env"]),  # Load system prompt from env
                }
            })
            return f"✅ {MODELS[model_key]['name']} {model_size} loaded successfully!"
        except Exception as e:
            return f"❌ Error: {str(e)}"

    def respond(self, message, max_tokens, temperature, top_p, top_k, image=None):
        if not st.session_state.current_model["model"]:
            return "⚠️ Please select and load a model first"

        try:
            # Get the system prompt for the current model
            system_prompt = st.session_state.current_model["config"]["system_prompt"]
            if not system_prompt:
                return "⚠️ System prompt not found for the selected model."

            # Add the system instruction to guide the model's behavior
            prompt = f"{system_prompt}\n\n### Instruction:\n{message}\n\n### Response:"

            inputs = st.session_state.current_model["tokenizer"](
                prompt,
                return_tensors="pt",
                max_length=512,
                truncation=True,
                padding=True
            )

            # Generate response without streaming
            with torch.no_grad():
                output = st.session_state.current_model["model"].generate(
                    input_ids=inputs.input_ids,
                    attention_mask=inputs.attention_mask,
                    max_new_tokens=max_tokens,
                    temperature=temperature,
                    top_p=top_p,
                    top_k=top_k,
                    do_sample=True,
                    pad_token_id=st.session_state.current_model["tokenizer"].pad_token_id,
                    eos_token_id=st.session_state.current_model["tokenizer"].eos_token_id,
                )
                response = st.session_state.current_model["tokenizer"].decode(output[0], skip_special_tokens=True)

            # Remove the prompt from the response
            if prompt in response:
                response = response.replace(prompt, "").strip()

            return response
        except Exception as e:
            return f"⚠️ Generation Error: {str(e)}"
        finally:
            self.clear_memory()

    def main(self):
        st.title("🦁 AtlasUI - Experimental 🧪")

        with st.sidebar:
            st.header("🛠 Model Selection")

            model_key = st.selectbox(
                "Choose Atlas Variant",
                list(MODELS.keys()),
                format_func=lambda x: f"{MODELS[x]['name']} {'🧪' if MODELS[x]['experimental'] else ''}"
            )

            model_size = st.selectbox(
                "Choose Model Size",
                list(MODELS[model_key]["sizes"].keys())
            )

            if st.button("Load Model"):
                with st.spinner("Loading model... This may take a few minutes."):
                    status = self.load_model(model_key, model_size)
                    st.success(status)

            st.header("🔧 Generation Parameters")
            max_tokens = st.slider("Max New Tokens", min_value=10, max_value=512, value=256, step=10)
            temperature = st.slider("Temperature", min_value=0.1, max_value=2.0, value=0.4, step=0.1)
            top_p = st.slider("Top-P", min_value=0.1, max_value=1.0, value=0.9, step=0.1)
            top_k = st.slider("Top-K", min_value=1, max_value=100, value=50, step=1)

            if st.button("Clear Chat History"):
                st.session_state.chat_history = []
                st.rerun()

        st.markdown("*⚠️ CAUTION: Atlas is an experimental model and this is just a preview. Responses may not be expected. Please double-check sensitive information!*")

        # Display chat history
        for message in st.session_state.chat_history:
            with st.chat_message(
                message["role"],
                avatar=USER_PFP if message["role"] == "user" else AI_PFP
            ):
                st.markdown(message["content"])
                if "image" in message:
                    st.image(message["image"], caption="Uploaded Image", use_container_width=True)  # Updated parameter

        # Input box for user messages
        if prompt := st.chat_input("Message Atlas..."):
            # Allow image upload if the model supports vision
            uploaded_image = None
            if MODELS[model_key]["is_vision"]:
                uploaded_image = st.file_uploader("Upload an image", type=["jpg", "jpeg", "png"])

            st.session_state.chat_history.append({"role": "user", "content": prompt, "image": uploaded_image})
            with st.chat_message("user", avatar=USER_PFP):
                st.markdown(prompt)
                if uploaded_image:
                    st.image(uploaded_image, caption="Uploaded Image", use_container_width=True)  # Updated parameter

            with st.chat_message("assistant", avatar=AI_PFP):
                with st.spinner("Generating response..."):
                    response = self.respond(prompt, max_tokens, temperature, top_p, top_k, image=uploaded_image)
                    st.markdown(response)

            st.session_state.chat_history.append({"role": "assistant", "content": response})

def run():
    try:
        app = AtlasInferenceApp()
        app.main()
    except Exception as e:
        st.error(f"⚠️ Application Error: {str(e)}")

if __name__ == "__main__":
    run()