app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login
import os
import matplotlib.pyplot as plt
import numpy as np

# Authentification
login(token=os.environ["HF_TOKEN"])

# Liste des modèles
models = [
    "meta-llama/Llama-2-13b-hf",
    "meta-llama/Llama-2-7b-hf",
    "meta-llama/Llama-2-70b-hf",
    "meta-llama/Meta-Llama-3-8B",
    "meta-llama/Llama-3.2-3B",
    "meta-llama/Llama-3.1-8B",
    "mistralai/Mistral-7B-v0.1",
    "mistralai/Mixtral-8x7B-v0.1",
    "mistralai/Mistral-7B-v0.3",
    "google/gemma-2-2b",
    "google/gemma-2-9b",
    "google/gemma-2-27b",
    "croissantllm/CroissantLLMBase"
]

# Variables globales
model = None
tokenizer = None

def load_model(model_name):
    global model, tokenizer
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="eager")
        if tokenizer.pad_token is None:
            tokenizer.pad_token = tokenizer.eos_token
        return f"Modèle {model_name} chargé avec succès."
    except Exception as e:
        return f"Erreur lors du chargement du modèle : {str(e)}"

def generate_text(input_text, temperature, top_p, top_k):
    global model, tokenizer
    
    if model is None or tokenizer is None:
        return "Veuillez d'abord charger un modèle.", None, None

    inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512).to(model.device)
    
    try:
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=50,
                temperature=temperature,
                top_p=top_p,
                top_k=top_k,
                output_attentions=True,
                return_dict_in_generate=True,
                output_scores=True
            )
        
        generated_text = tokenizer.decode(outputs.sequences[0], skip_special_tokens=True)
        
        # Obtenir les logits pour le dernier token généré
        if outputs.scores:
            last_token_logits = outputs.scores[-1][0]
            
            # Appliquer softmax pour obtenir les probabilités
            probabilities = torch.nn.functional.softmax(last_token_logits, dim=-1)
            
            # Obtenir les top 5 tokens les plus probables
            top_k = 5
            top_probs, top_indices = torch.topk(probabilities, top_k)
            top_words = [tokenizer.decode([idx.item()]) for idx in top_indices]
            
            # Préparer les données pour le graphique des probabilités
            prob_data = {word: prob.item() for word, prob in zip(top_words, top_probs)}
            
            # Extraire les attentions (moyenne sur toutes les couches et têtes d'attention)
            if outputs.attentions:
                attentions = torch.mean(torch.stack(outputs.attentions), dim=(0, 1)).cpu().numpy()
                attention_plot = plot_attention(attentions, tokenizer.convert_ids_to_tokens(inputs["input_ids"][0]))
            else:
                attention_plot = None
            
            prob_plot = plot_probabilities(prob_data)
        else:
            attention_plot = None
            prob_plot = None
        
        return generated_text, attention_plot, prob_plot
    except Exception as e:
        return f"Erreur lors de la génération : {str(e)}", None, None

def plot_attention(attention, tokens):
    fig, ax = plt.subplots(figsize=(10, 10))
    im = ax.imshow(attention, cmap='viridis')
    ax.set_xticks(range(len(tokens)))
    ax.set_yticks(range(len(tokens)))
    ax.set_xticklabels(tokens, rotation=90)
    ax.set_yticklabels(tokens)
    plt.colorbar(im)
    plt.title("Carte d'attention")
    plt.tight_layout()
    return fig

def plot_probabilities(prob_data):
    words = list(prob_data.keys())
    probs = list(prob_data.values())
    
    fig, ax = plt.subplots(figsize=(10, 5))
    ax.bar(words, probs)
    ax.set_title("Probabilités des tokens suivants les plus probables")
    ax.set_xlabel("Tokens")
    ax.set_ylabel("Probabilité")
    plt.xticks(rotation=45)
    plt.tight_layout()
    return fig

def reset():
    global model, tokenizer
    model = None
    tokenizer = None
    return "", 1.0, 1.0, 50, None, None, None

with gr.Blocks() as demo:
    gr.Markdown("# Générateur de texte avec visualisation d'attention")
    
    with gr.Accordion("Sélection du modèle"):
        model_dropdown = gr.Dropdown(choices=models, label="Choisissez un modèle")
        load_button = gr.Button("Charger le modèle")
        load_output = gr.Textbox(label="Statut du chargement")
    
    with gr.Row():
        temperature = gr.Slider(0.1, 2.0, value=1.0, label="Température")
        top_p = gr.Slider(0.1, 1.0, value=1.0, label="Top-p")
        top_k = gr.Slider(1, 100, value=50, step=1, label="Top-k")
    
    input_text = gr.Textbox(label="Texte d'entrée")
    generate_button = gr.Button("Générer")
    
    output_text = gr.Textbox(label="Texte généré")
    
    with gr.Row():
        attention_plot = gr.Plot(label="Visualisation de l'attention")
        prob_plot = gr.Plot(label="Probabilités des tokens suivants")
    
    reset_button = gr.Button("Réinitialiser")
    
    load_button.click(load_model, inputs=[model_dropdown], outputs=[load_output])
    generate_button.click(generate_text, 
                          inputs=[input_text, temperature, top_p, top_k], 
                          outputs=[output_text, attention_plot, prob_plot])
    reset_button.click(reset, 
                       outputs=[input_text, temperature, top_p, top_k, output_text, attention_plot, prob_plot])

demo.launch()