import gradio as gr import torch from transformers import AutoTokenizer, AutoModelForCausalLM from huggingface_hub import login import os import matplotlib.pyplot as plt import seaborn as sns import numpy as np import asyncio import gc # Authentification login(token=os.environ["HF_TOKEN"]) # Restructuration des modèles et de leurs informations models_info = { "Meta-llama": { "Llama 2": { "7B": {"name": "meta-llama/Llama-2-7b-hf", "languages": ["en"]}, "13B": {"name": "meta-llama/Llama-2-13b-hf", "languages": ["en"]}, }, "Llama 3": { "8B": {"name": "meta-llama/Llama-3-8B", "languages": ["en"]}, "3.2-3B": {"name": "meta-llama/Llama-3.2-3B", "languages": ["en", "de", "fr", "it", "pt", "hi", "es", "th"]}, }, }, "Mistral AI": { "Mistral": { "7B-v0.1": {"name": "mistralai/Mistral-7B-v0.1", "languages": ["en"]}, "7B-v0.3": {"name": "mistralai/Mistral-7B-v0.3", "languages": ["en"]}, }, "Mixtral": { "8x7B-v0.1": {"name": "mistralai/Mixtral-8x7B-v0.1", "languages": ["en", "fr", "it", "de", "es"]}, }, }, "Google": { "Gemma": { "2B": {"name": "google/gemma-2-2b", "languages": ["en"]}, "7B": {"name": "google/gemma-2-7b", "languages": ["en"]}, }, }, "CroissantLLM": { "CroissantLLMBase": { "Base": {"name": "croissantllm/CroissantLLMBase", "languages": ["en", "fr"]}, }, }, } # Paramètres recommandés pour chaque modèle model_parameters = { "meta-llama/Llama-2-7b-hf": {"temperature": 0.8, "top_p": 0.9, "top_k": 40}, "meta-llama/Llama-2-13b-hf": {"temperature": 0.8, "top_p": 0.9, "top_k": 40}, "meta-llama/Llama-3-8B": {"temperature": 0.75, "top_p": 0.9, "top_k": 50}, "meta-llama/Llama-3.2-3B": {"temperature": 0.75, "top_p": 0.9, "top_k": 50}, "mistralai/Mistral-7B-v0.1": {"temperature": 0.7, "top_p": 0.9, "top_k": 50}, "mistralai/Mistral-7B-v0.3": {"temperature": 0.7, "top_p": 0.9, "top_k": 50}, "mistralai/Mixtral-8x7B-v0.1": {"temperature": 0.8, "top_p": 0.95, "top_k": 50}, "google/gemma-2-2b": {"temperature": 0.7, "top_p": 0.95, "top_k": 40}, "google/gemma-2-7b": {"temperature": 0.7, "top_p": 0.95, "top_k": 40}, "croissantllm/CroissantLLMBase": {"temperature": 0.8, "top_p": 0.92, "top_k": 50} } # Variables globales model_cache = {} # Fonctions utilitaires def update_model_type(family): return gr.Dropdown(choices=list(models_info[family].keys()), value=None, interactive=True) def update_model_variation(family, model_type): if family and model_type: return gr.Dropdown(choices=list(models_info[family][model_type].keys()), value=None, interactive=True) return gr.Dropdown(choices=[], value=None, interactive=False) def update_selected_model(family, model_type, variation): if family and model_type and variation: model_name = models_info[family][model_type][variation]["name"] return model_name, gr.Dropdown(choices=models_info[family][model_type][variation]["languages"], value=models_info[family][model_type][variation]["languages"][0], visible=True, interactive=True) return "", gr.Dropdown(visible=False) async def load_model_async(model_name, progress=gr.Progress()): try: if model_name not in model_cache: progress(0.1, f"Chargement du tokenizer pour {model_name}...") tokenizer = await asyncio.to_thread(AutoTokenizer.from_pretrained, model_name) progress(0.4, f"Chargement du modèle {model_name}...") model = await asyncio.to_thread(AutoModelForCausalLM.from_pretrained, model_name, torch_dtype=torch.float16, device_map="auto", low_cpu_mem_usage=True) if tokenizer.pad_token is None: tokenizer.pad_token = tokenizer.eos_token model_cache[model_name] = (model, tokenizer) progress(1.0, f"Modèle {model_name} chargé avec succès") return f"Modèle {model_name} chargé avec succès" except Exception as e: return f"Erreur lors du chargement du modèle {model_name} : {str(e)}" def set_language(lang): return f"Langue sélectionnée : {lang}" def ensure_token_display(token, tokenizer): if token.isdigit() or (token.startswith('-') and token[1:].isdigit()): return tokenizer.decode([int(token)]) return token async def analyze_next_token(model_name, input_text, temperature, top_p, top_k, progress=gr.Progress()): if model_name not in model_cache: return "Veuillez d'abord charger le modèle", None, None model, tokenizer = model_cache[model_name] inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512).to(model.device) try: progress(0.5, "Analyse en cours...") with torch.no_grad(): outputs = model(**inputs) last_token_logits = outputs.logits[0, -1, :] probabilities = torch.nn.functional.softmax(last_token_logits, dim=-1) top_k = min(10, top_k) top_probs, top_indices = torch.topk(probabilities, top_k) top_words = [ensure_token_display(tokenizer.decode([idx.item()]), tokenizer) for idx in top_indices] prob_data = {word: prob.item() for word, prob in zip(top_words, top_probs)} prob_text = "Prochains tokens les plus probables :\n\n" for word, prob in prob_data.items(): prob_text += f"{word}: {prob:.2%}\n" prob_plot = plot_probabilities(prob_data) attention_plot = plot_attention(inputs["input_ids"][0].cpu(), last_token_logits.cpu(), tokenizer) progress(1.0, "Analyse terminée") return prob_text, attention_plot, prob_plot except Exception as e: return f"Erreur lors de l'analyse : {str(e)}", None, None async def generate_text(model_name, input_text, temperature, top_p, top_k, progress=gr.Progress()): if model_name not in model_cache: return "Veuillez d'abord charger le modèle" model, tokenizer = model_cache[model_name] inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512).to(model.device) try: progress(0.5, "Génération en cours...") with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=50, temperature=temperature, top_p=top_p, top_k=top_k ) generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True) progress(1.0, "Génération terminée") return generated_text except Exception as e: return f"Erreur lors de la génération : {str(e)}" def plot_probabilities(prob_data): try: words = list(prob_data.keys()) probs = list(prob_data.values()) fig, ax = plt.subplots(figsize=(12, 6)) bars = ax.bar(range(len(words)), probs, color='lightgreen') ax.set_title("Probabilités des tokens suivants les plus probables") ax.set_xlabel("Tokens") ax.set_ylabel("Probabilité") ax.set_xticks(range(len(words))) ax.set_xticklabels(words, rotation=45, ha='right') for i, (bar, word) in enumerate(zip(bars, words)): height = bar.get_height() ax.text(i, height, f'{height:.2%}', ha='center', va='bottom', rotation=0) plt.tight_layout() return fig except Exception as e: print(f"Erreur lors de la création du graphique : {str(e)}") return None def plot_attention(input_ids, last_token_logits, tokenizer): try: input_tokens = [ensure_token_display(tokenizer.decode([id]), tokenizer) for id in input_ids] attention_scores = torch.nn.functional.softmax(last_token_logits, dim=-1) top_k = min(len(input_tokens), 10) top_attention_scores, _ = torch.topk(attention_scores, top_k) fig, ax = plt.subplots(figsize=(14, 7)) sns.heatmap(top_attention_scores.unsqueeze(0).numpy(), annot=True, cmap="YlOrRd", cbar=True, ax=ax, fmt='.2%') ax.set_xticklabels(input_tokens[-top_k:], rotation=45, ha="right", fontsize=10) ax.set_yticklabels(["Attention"], rotation=0, fontsize=10) ax.set_title("Scores d'attention pour les derniers tokens", fontsize=16) cbar = ax.collections[0].colorbar cbar.set_label("Score d'attention", fontsize=12) cbar.ax.tick_params(labelsize=10) plt.tight_layout() return fig except Exception as e: print(f"Erreur lors de la création du graphique d'attention : {str(e)}") return None def reset(): global model_cache for model in model_cache.values(): del model model_cache.clear() torch.cuda.empty_cache() gc.collect() return ( "", 1.0, 1.0, 50, None, None, None, None, gr.Dropdown(choices=list(models_info.keys()), value=None, interactive=True), gr.Dropdown(choices=[], value=None, interactive=False), gr.Dropdown(choices=[], value=None, interactive=False), "", gr.Dropdown(visible=False), "" ) def reset_comparison(): return [gr.Dropdown(choices=[], value=None) for _ in range(4)] + ["", "", gr.Dropdown(choices=[], value=None), 1.0, 1.0, 50, "", "", None, None, None, None] async def compare_models(model1, model2, input_text, temp, top_p, top_k, progress=gr.Progress()): if model1 not in model_cache or model2 not in model_cache: return "Veuillez d'abord charger les deux modèles", "", None, None, None, None progress(0.1, "Analyse du premier modèle...") results1 = await analyze_next_token(model1, input_text, temp, top_p, top_k) progress(0.4, "Analyse du second modèle...") results2 = await analyze_next_token(model2, input_text, temp, top_p, top_k) progress(1.0, "Comparaison terminée") return ( results1[0], results2[0], # Probabilités du prochain token results1[2], results2[2], # Graphiques de probabilités results1[1], results2[1] # Graphiques d'attention ) with gr.Blocks() as demo: gr.Markdown("# LLM&BIAS") with gr.Tabs(): with gr.Tab("Analyse individuelle"): with gr.Accordion("Sélection du modèle", open=True): with gr.Row(): model_family = gr.Dropdown(choices=list(models_info.keys()), label="Famille de modèle", interactive=True) model_type = gr.Dropdown(choices=[], label="Type de modèle", interactive=False) model_variation = gr.Dropdown(choices=[], label="Variation du modèle", interactive=False) selected_model = gr.Textbox(label="Modèle sélectionné", interactive=False) load_button = gr.Button("Charger le modèle") load_output = gr.Textbox(label="Statut du chargement") language_dropdown = gr.Dropdown(label="Choisissez une langue", visible=False) language_output = gr.Textbox(label="Langue sélectionnée") 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", lines=3) analyze_button = gr.Button("Analyser le prochain token") next_token_probs = gr.Textbox(label="Probabilités du prochain token") with gr.Row(): attention_plot = gr.Plot(label="Visualisation de l'attention") prob_plot = gr.Plot(label="Probabilités des tokens suivants") generate_button = gr.Button("Générer le texte") generated_text = gr.Textbox(label="Texte généré") reset_button = gr.Button("Réinitialiser") with gr.Tab("Comparaison de modèles"): with gr.Row(): model1_family = gr.Dropdown(choices=list(models_info.keys()), label="Famille du modèle 1", interactive=True) model1_type = gr.Dropdown(choices=[], label="Type du modèle 1", interactive=False) model1_variation = gr.Dropdown(choices=[], label="Variation du modèle 1", interactive=False) with gr.Row(): model2_family = gr.Dropdown(choices=list(models_info.keys()), label="Famille du modèle 2", interactive=True) model2_type = gr.Dropdown(choices=[], label="Type du modèle 2", interactive=False) model2_variation = gr.Dropdown(choices=[], label="Variation du modèle 2", interactive=False) model1_selected = gr.Textbox(label="Modèle 1 sélectionné", interactive=False) model2_selected = gr.Textbox(label="Modèle 2 sélectionné", interactive=False) load_models_button = gr.Button("Charger les modèles") load_models_output = gr.Textbox(label="Statut du chargement des modèles") comparison_language = gr.Dropdown(label="Langue pour la comparaison", choices=[], interactive=False) with gr.Row(): comp_temperature = gr.Slider(0.1, 2.0, value=1.0, label="Température") comp_top_p = gr.Slider(0.1, 1.0, value=1.0, label="Top-p") comp_top_k = gr.Slider(1, 100, value=50, step=1, label="Top-k") comp_input_text = gr.Textbox(label="Texte d'entrée pour la comparaison", lines=3) compare_button = gr.Button("Comparer les modèles") with gr.Row(): model1_output = gr.Textbox(label="Probabilités du Modèle 1", lines=10) model2_output = gr.Textbox(label="Probabilités du Modèle 2", lines=10) with gr.Row(): model1_prob_plot = gr.Plot(label="Probabilités des tokens (Modèle 1)") model2_prob_plot = gr.Plot(label="Probabilités des tokens (Modèle 2)") with gr.Row(): model1_attention_plot = gr.Plot(label="Attention (Modèle 1)") model2_attention_plot = gr.Plot(label="Attention (Modèle 2)") comp_reset_button = gr.Button("Réinitialiser la comparaison") # Événements pour l'onglet d'analyse individuelle model_family.change(update_model_type, inputs=[model_family], outputs=[model_type]) model_type.change(update_model_variation, inputs=[model_family, model_type], outputs=[model_variation]) model_variation.change(update_selected_model, inputs=[model_family, model_type, model_variation], outputs=[selected_model, language_dropdown]) load_button.click(load_model_async, inputs=[selected_model], outputs=[load_output]) language_dropdown.change(set_language, inputs=[language_dropdown], outputs=[language_output]) analyze_button.click(analyze_next_token, inputs=[selected_model, input_text, temperature, top_p, top_k], outputs=[next_token_probs, attention_plot, prob_plot]) generate_button.click(generate_text, inputs=[selected_model, input_text, temperature, top_p, top_k], outputs=[generated_text]) reset_button.click(reset, outputs=[input_text, temperature, top_p, top_k, next_token_probs, attention_plot, prob_plot, generated_text, model_family, model_type, model_variation, selected_model, language_dropdown, language_output]) # Événements pour l'onglet de comparaison model1_family.change(update_model_type, inputs=[model1_family], outputs=[model1_type]) model1_type.change(update_model_variation, inputs=[model1_family, model1_type], outputs=[model1_variation]) model1_variation.change(update_selected_model, inputs=[model1_family, model1_type, model1_variation], outputs=[model1_selected, comparison_language]) model2_family.change(update_model_type, inputs=[model2_family], outputs=[model2_type]) model2_type.change(update_model_variation, inputs=[model2_family, model2_type], outputs=[model2_variation]) model2_variation.change(update_selected_model, inputs=[model2_family, model2_type, model2_variation], outputs=[model2_selected, comparison_language]) async def load_both_models(model1, model2): result1 = await load_model_async(model1) result2 = await load_model_async(model2) return f"Modèle 1: {result1}\nModèle 2: {result2}" load_models_button.click(load_both_models, inputs=[model1_selected, model2_selected], outputs=[load_models_output]) compare_button.click( compare_models, inputs=[model1_selected, model2_selected, comp_input_text, comp_temperature, comp_top_p, comp_top_k], outputs=[model1_output, model2_output, model1_prob_plot, model2_prob_plot, model1_attention_plot, model2_attention_plot] ) comp_reset_button.click( reset_comparison, outputs=[model1_type, model1_variation, model2_type, model2_variation, model1_selected, model2_selected, comparison_language, comp_temperature, comp_top_p, comp_top_k, comp_input_text, model1_output, model2_output, model1_prob_plot, model2_prob_plot, model1_attention_plot, model2_attention_plot] ) if __name__ == "__main__": demo.launch()