Update app.py

app.py

@@ -5,8 +5,6 @@ from huggingface_hub import login
 import os
 import matplotlib.pyplot as plt
 import seaborn as sns
-import numpy as np
-import time
 
 # Authentification
 login(token=os.environ["HF_TOKEN"])
@@ -35,19 +33,23 @@ tokenizer = None
 def load_model(model_name, progress=gr.Progress()):
     global model, tokenizer
     try:
-        for i in progress.tqdm(range(100)):
-            time.sleep(0.01)  # Simuler le chargement
-            if i == 25:
-                tokenizer = AutoTokenizer.from_pretrained(model_name)
-            elif i == 75:
-                model = AutoModelForCausalLM.from_pretrained(
-                    model_name, 
-                    torch_dtype=torch.float32,
-                    device_map="cpu",
-                    attn_implementation="eager"
-                )
-                if tokenizer.pad_token is None:
-                    tokenizer.pad_token = tokenizer.eos_token
+        progress(0, desc="Chargement du tokenizer")
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+        progress(0.3, desc="Tokenizer chargé")
+        
+        progress(0.3, desc="Chargement du modèle")
+        model = AutoModelForCausalLM.from_pretrained(
+            model_name, 
+            torch_dtype=torch.bfloat16, 
+            device_map="auto", 
+            attn_implementation="eager"
+        )
+        progress(0.9, desc="Modèle chargé")
+        
+        if tokenizer.pad_token is None:
+            tokenizer.pad_token = tokenizer.eos_token
+        
+        progress(1.0, desc="Chargement terminé")
         return f"Modèle {model_name} chargé avec succès."
     except Exception as e:
         return f"Erreur lors du chargement du modèle : {str(e)}"
@@ -58,7 +60,7 @@ def analyze_next_token(input_text, temperature, top_p, top_k):
     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)
+    inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512).to(model.device)
     
     try:
         with torch.no_grad():
@@ -74,9 +76,12 @@ def analyze_next_token(input_text, temperature, top_p, top_k):
         
         prob_text = "\n".join([f"{word}: {prob:.4f}" for word, prob in prob_data.items()])
         
-        attention_heatmap = plot_attention_alternative(inputs["input_ids"][0], last_token_logits)
+        # Simplification de l'affichage de l'attention
+        attention_text = "Attention non disponible pour ce modèle"
+        if hasattr(outputs, 'attentions') and outputs.attentions is not None:
+            attention_text = "Attention disponible"
         
-        return prob_text, attention_heatmap, prob_plot
+        return prob_text, attention_text, prob_plot
     except Exception as e:
         return f"Erreur lors de l'analyse : {str(e)}", None, None
 
@@ -86,20 +91,20 @@ def generate_text(input_text, temperature, top_p, top_k):
     if model is None or tokenizer is None:
         return "Veuillez d'abord charger un modèle."
 
-    inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512)
+    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=1,
+                max_new_tokens=1,  # Génère seulement le prochain mot
                 temperature=temperature,
                 top_p=top_p,
                 top_k=top_k
             )
         
         generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
-        return generated_text  # Retourne l'input + le nouveau mot
+        return generated_text  # Retourne l'input + le nouveau mot généré
     except Exception as e:
         return f"Erreur lors de la génération : {str(e)}"
 
@@ -116,20 +121,6 @@ def plot_probabilities(prob_data):
     plt.tight_layout()
     return fig
 
-def plot_attention_alternative(input_ids, last_token_logits):
-    input_tokens = tokenizer.convert_ids_to_tokens(input_ids)
-    attention_scores = torch.nn.functional.softmax(last_token_logits, dim=-1)
-    top_k = min(len(input_tokens), 10)  # Limiter à 10 tokens pour la lisibilité
-    top_attention_scores, _ = torch.topk(attention_scores, top_k)
-    
-    fig, ax = plt.subplots(figsize=(12, 6))
-    sns.heatmap(top_attention_scores.unsqueeze(0).numpy(), annot=True, cmap="YlOrRd", cbar=False, ax=ax)
-    ax.set_xticklabels(input_tokens[-top_k:], rotation=45, ha="right")
-    ax.set_yticklabels(["Attention"], rotation=0)
-    ax.set_title("Scores d'attention pour les derniers tokens")
-    plt.tight_layout()
-    return fig
-
 def reset():
     global model, tokenizer
     model = None
@@ -153,10 +144,9 @@ with gr.Blocks() as demo:
     analyze_button = gr.Button("Analyser le prochain token")
     
     next_token_probs = gr.Textbox(label="Probabilités du prochain token")
+    attention_info = gr.Textbox(label="Information sur l'attention")
     
-    with gr.Row():
-        attention_plot = gr.Plot(label="Visualisation de l'attention")
-        prob_plot = gr.Plot(label="Probabilités des tokens suivants")
+    prob_plot = gr.Plot(label="Probabilités des tokens suivants")
     
     generate_button = gr.Button("Générer le prochain mot")
     generated_text = gr.Textbox(label="Texte généré")
@@ -166,12 +156,12 @@ with gr.Blocks() as demo:
     load_button.click(load_model, inputs=[model_dropdown], outputs=[load_output])
     analyze_button.click(analyze_next_token, 
                          inputs=[input_text, temperature, top_p, top_k], 
-                         outputs=[next_token_probs, attention_plot, prob_plot])
+                         outputs=[next_token_probs, attention_info, prob_plot])
     generate_button.click(generate_text, 
                           inputs=[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])
+                       outputs=[input_text, temperature, top_p, top_k, next_token_probs, attention_info, prob_plot, generated_text])
 
 if __name__ == "__main__":
     demo.launch()