src/streamlit_app.py

import streamlit as st
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import random
import pandas as pd

# --- CONFIGURATION ---
st.set_page_config(page_title="FINISHA-F-SCRATCH Local Arena", layout="wide")
st.title("⚔️ SLM FROM SCRATCH Model Arena (Local Loading) ! choissez le SLM le plus unique!")

# Liste des modèles Raana-ia (Assurez-vous qu'ils tiennent en mémoire)
MODELS_LIST = [
    "Finisha-f-scratch/SMCLEM", 
    "Finisha-f-scratch/InutileGheya",
    "Finisha-f-scratch/Gheya-111m",
    "Finisha-f-scratch/Gheya-Nacid-instruct-v1",
    "Finisha-f-scratch/Mini-mistral-v1",
    "Finisha-f-scratch/Charlotte-amity",
    "Finisha-f-scratch/Charlotte-amity-v2",
    "Finisha-f-scratch/Tiny-charlotte",
    "Finisha-f-scratch/mini-gamia",
    "Finisha-f-scratch/SoraNova",
    "Finisha-f-scratch/Expedia-LLM",
    "Clem27-assistants/Learnia-Empathic-Tchat",
    "Finisha-F-scratch/Learnia-tchat-v1",
    "Finisha-F-scratch/Neko-charlotte",
    "Finisha-F-scratch/Sala",
    "Finisha-F-scratch/Charlotte-gheya",
    "Finisha-F-scratch/microBook",
    "Finisha-F-scratch/Chichalia-v1",
    "Finisha-f-scratch/Claire",
    "Finisha-F-scratch/Rosa-4M",
    "Finisha-F-scratch/Nelya",
    "Finisha-F-scratch/Nelya-neko",
    "Finisha-F-scratch/Dona-KITY-10m",
    "Finisha-F-scratch/KLA-SLM-CODING",
    "Finisha-F-scratch/Tiny-DonaKitty",
    "Finisha-F-scratch/Serena",
    "Finisha-F-scratch/Perso-SLM",
    "Finisha-F-scratch/Tiny-Rosa",
    "Finisha-F-scratch/Iris-La-guepe",
    "Finisha-F-scratch/Ilyana-lamina-Nacid",
    "Finisha-F-scratch/Ilyana-pretrain",
    "Finisha-F-scratch/Gheya-63M",
    "Finisha-F-scratch/Copina",
    "Finisha-F-scratch/Ayako-CHINESS",
    "Finisha-F-scratch/Tiny-lamina-English",
    "Finisha-F-scratch/Gheya-Nacid",
    "Finisha-F-scratch/Learnia-business",
    "Finisha-F-scratch/Lam-pest",
    "Finisha-F-scratch/ReeCi",
    "Finisha-F-scratch/melta-english",
    "Finisha-F-scratch/Maya-152M-Flowers",
    "Finisha-F-scratch/Lam-4-zero-F",
    "Finisha-F-scratch/Learnia",
    "Finisha-F-scratch/Qsana-coder-base",
    "Finisha-F-scratch/Coliria",
    "Finisha-F-scratch/Charlotte-2b",
    "Finisha-F-scratch/Natalia-pretrain",
    "Finisha-F-scratch/LilyStory",
    "Finisha-F-scratch/Nephaella"
]

# --- CHARGEMENT DES MODÈLES (CACHÉ) ---
@st.cache_resource
def load_model_pipeline(model_id):
    """Charge le modèle et le tokenizer en mémoire."""
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    # On utilise device_map="auto" pour gérer le GPU si disponible
    model = AutoModelForCausalLM.from_pretrained(
        model_id, 
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
        device_map="auto"
    )
    return pipeline("text-generation", model=model, tokenizer=tokenizer)

# --- LOGIQUE DE VOTE ET SCORE ---
if 'scores' not in st.session_state:
    st.session_state.scores = {m: 1200 for m in MODELS_LIST}

if "model_a" not in st.session_state:
    st.session_state.model_a = ""
    st.session_state.model_b = ""
    st.session_state.resp_a = ""
    st.session_state.resp_b = ""
    st.session_state.voted = False

# --- INTERFACE UTILISATEUR ---
user_prompt = st.text_area("Entrez votre question :", placeholder="Écris un poème sur l'intelligence artificielle.")

if st.button("Lancer le duel"):
    if user_prompt:
        # Sélection aléatoire
        sampled = random.sample(MODELS_LIST, 2)
        st.session_state.model_a, st.session_state.model_b = sampled
        
        with st.spinner(f"Chargement et génération en cours..."):
            # Génération Modèle A
            pipe_a = load_model_pipeline(st.session_state.model_a)
            out_a = pipe_a(user_prompt, max_new_tokens=150, do_sample=True, temperature=0.7)
            st.session_state.resp_a = out_a[0]['generated_text'].replace(user_prompt, "")

            # Génération Modèle B
            pipe_b = load_model_pipeline(st.session_state.model_b)
            out_b = pipe_b(user_prompt, max_new_tokens=150, do_sample=True, temperature=0.7)
            st.session_state.resp_b = out_b[0]['generated_text'].replace(user_prompt, "")
            
            st.session_state.voted = False
    else:
        st.error("Le prompt est vide !")

# Affichage des résultats
if st.session_state.resp_a:
    col1, col2 = st.columns(2)
    with col1:
        st.info(f"**Réponse A :**\n\n{st.session_state.resp_a}")
    with col2:
        st.info(f"**Réponse B :**\n\n{st.session_state.resp_b}")

    if not st.session_state.voted:
        c1, c2, c3 = st.columns(3)
        if c1.button("A est meilleur"):
            st.session_state.scores[st.session_state.model_a] += 20
            st.session_state.voted = True
        if c2.button("Égalité"):
            st.session_state.voted = True
        if c3.button("B est meilleur"):
            st.session_state.scores[st.session_state.model_b] += 20
            st.session_state.voted = True

if st.session_state.voted:
    st.success(f"Résultat : A était **{st.session_state.model_a}** | B était **{st.session_state.model_b}**")
    if st.button("Nouveau duel"):
        st.session_state.resp_a = ""
        st.rerun()

# Classement
st.divider()
st.subheader("📊 Leaderboard")
st.table(pd.DataFrame(st.session_state.scores.items(), columns=["Modèle", "ELO"]).sort_values("ELO", ascending=False))