Update app.py
Browse files
app.py
CHANGED
|
@@ -246,20 +246,96 @@ class MultimodalTrainer:
|
|
| 246 |
try:
|
| 247 |
logger.info(f"Chargement du modèle: {model_name}")
|
| 248 |
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 263 |
|
| 264 |
# Charge le tokenizer
|
| 265 |
try:
|
|
@@ -270,6 +346,14 @@ class MultimodalTrainer:
|
|
| 270 |
self.current_tokenizer.pad_token = self.current_tokenizer.eos_token
|
| 271 |
except Exception as e:
|
| 272 |
logger.warning(f"Tokenizer non trouvé: {e}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 273 |
|
| 274 |
# Charge le processor
|
| 275 |
try:
|
|
@@ -279,7 +363,7 @@ class MultimodalTrainer:
|
|
| 279 |
except Exception as e:
|
| 280 |
logger.warning(f"Processor non trouvé: {e}")
|
| 281 |
|
| 282 |
-
return f"✅ Modèle {model_name} chargé avec succès!\nType: {type(self.current_model).__name__}"
|
| 283 |
|
| 284 |
except Exception as e:
|
| 285 |
error_msg = f"❌ Erreur lors du chargement: {str(e)}"
|
|
@@ -354,7 +438,55 @@ class MultimodalTrainer:
|
|
| 354 |
info += f"📈 Exemples: {len(self.training_data):,}\n"
|
| 355 |
info += f"📝 Colonnes: {list(self.training_data.column_names)}\n"
|
| 356 |
|
| 357 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 358 |
|
| 359 |
# Initialisation
|
| 360 |
trainer = MultimodalTrainer()
|
|
@@ -428,6 +560,7 @@ def create_interface():
|
|
| 428 |
value="causal"
|
| 429 |
)
|
| 430 |
load_model_btn = gr.Button("🔄 Charger le modèle", variant="primary")
|
|
|
|
| 431 |
|
| 432 |
with gr.Column():
|
| 433 |
model_status = gr.Textbox(
|
|
@@ -449,6 +582,12 @@ def create_interface():
|
|
| 449 |
outputs=model_status
|
| 450 |
)
|
| 451 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 452 |
info_btn.click(trainer.get_model_info, outputs=model_info)
|
| 453 |
|
| 454 |
with gr.Tab("📊 Données"):
|
|
|
|
| 246 |
try:
|
| 247 |
logger.info(f"Chargement du modèle: {model_name}")
|
| 248 |
|
| 249 |
+
# Stratégies de chargement multiples
|
| 250 |
+
model_loaded = False
|
| 251 |
+
error_messages = []
|
| 252 |
+
|
| 253 |
+
# Stratégie 1: AutoModelForCausalLM avec trust_remote_code
|
| 254 |
+
if model_type == "causal" and not model_loaded:
|
| 255 |
+
try:
|
| 256 |
+
self.current_model = AutoModelForCausalLM.from_pretrained(
|
| 257 |
+
model_name,
|
| 258 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 259 |
+
device_map="auto" if torch.cuda.is_available() else None,
|
| 260 |
+
trust_remote_code=True
|
| 261 |
+
)
|
| 262 |
+
model_loaded = True
|
| 263 |
+
except Exception as e:
|
| 264 |
+
error_messages.append(f"AutoModelForCausalLM: {str(e)}")
|
| 265 |
+
|
| 266 |
+
# Stratégie 2: AutoModel générique
|
| 267 |
+
if not model_loaded:
|
| 268 |
+
try:
|
| 269 |
+
self.current_model = AutoModel.from_pretrained(
|
| 270 |
+
model_name,
|
| 271 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 272 |
+
device_map="auto" if torch.cuda.is_available() else None,
|
| 273 |
+
trust_remote_code=True
|
| 274 |
+
)
|
| 275 |
+
model_loaded = True
|
| 276 |
+
except Exception as e:
|
| 277 |
+
error_messages.append(f"AutoModel: {str(e)}")
|
| 278 |
+
|
| 279 |
+
# Stratégie 3: Détection automatique basée sur le nom
|
| 280 |
+
if not model_loaded and any(x in model_name.lower() for x in ['llama', 'mistral', 'qwen', 'phi']):
|
| 281 |
+
try:
|
| 282 |
+
# Pour les modèles de type LLaMA/Mistral/Qwen
|
| 283 |
+
from transformers import LlamaForCausalLM, MistralForCausalLM
|
| 284 |
+
|
| 285 |
+
if 'llama' in model_name.lower():
|
| 286 |
+
self.current_model = LlamaForCausalLM.from_pretrained(
|
| 287 |
+
model_name,
|
| 288 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 289 |
+
device_map="auto" if torch.cuda.is_available() else None,
|
| 290 |
+
trust_remote_code=True
|
| 291 |
+
)
|
| 292 |
+
elif 'mistral' in model_name.lower():
|
| 293 |
+
self.current_model = MistralForCausalLM.from_pretrained(
|
| 294 |
+
model_name,
|
| 295 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 296 |
+
device_map="auto" if torch.cuda.is_available() else None,
|
| 297 |
+
trust_remote_code=True
|
| 298 |
+
)
|
| 299 |
+
model_loaded = True
|
| 300 |
+
except Exception as e:
|
| 301 |
+
error_messages.append(f"Modèle spécifique: {str(e)}")
|
| 302 |
+
|
| 303 |
+
# Stratégie 4: Configuration manuelle
|
| 304 |
+
if not model_loaded:
|
| 305 |
+
try:
|
| 306 |
+
# Télécharge la configuration d'abord
|
| 307 |
+
from transformers import AutoConfig
|
| 308 |
+
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
|
| 309 |
+
|
| 310 |
+
# Force le model_type si manquant
|
| 311 |
+
if not hasattr(config, 'model_type') or config.model_type is None:
|
| 312 |
+
# Détection basée sur l'architecture
|
| 313 |
+
if hasattr(config, 'architectures') and config.architectures:
|
| 314 |
+
arch = config.architectures[0].lower()
|
| 315 |
+
if 'llama' in arch:
|
| 316 |
+
config.model_type = 'llama'
|
| 317 |
+
elif 'mistral' in arch:
|
| 318 |
+
config.model_type = 'mistral'
|
| 319 |
+
elif 'qwen' in arch:
|
| 320 |
+
config.model_type = 'qwen2'
|
| 321 |
+
elif 'phi' in arch:
|
| 322 |
+
config.model_type = 'phi'
|
| 323 |
+
else:
|
| 324 |
+
config.model_type = 'llama' # Par défaut
|
| 325 |
+
|
| 326 |
+
self.current_model = AutoModelForCausalLM.from_pretrained(
|
| 327 |
+
model_name,
|
| 328 |
+
config=config,
|
| 329 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 330 |
+
device_map="auto" if torch.cuda.is_available() else None,
|
| 331 |
+
trust_remote_code=True
|
| 332 |
+
)
|
| 333 |
+
model_loaded = True
|
| 334 |
+
except Exception as e:
|
| 335 |
+
error_messages.append(f"Configuration manuelle: {str(e)}")
|
| 336 |
+
|
| 337 |
+
if not model_loaded:
|
| 338 |
+
return f"❌ Impossible de charger le modèle. Erreurs:\n" + "\n".join(error_messages)
|
| 339 |
|
| 340 |
# Charge le tokenizer
|
| 341 |
try:
|
|
|
|
| 346 |
self.current_tokenizer.pad_token = self.current_tokenizer.eos_token
|
| 347 |
except Exception as e:
|
| 348 |
logger.warning(f"Tokenizer non trouvé: {e}")
|
| 349 |
+
try:
|
| 350 |
+
# Essaye avec un tokenizer générique
|
| 351 |
+
from transformers import LlamaTokenizer
|
| 352 |
+
self.current_tokenizer = LlamaTokenizer.from_pretrained(
|
| 353 |
+
model_name, trust_remote_code=True
|
| 354 |
+
)
|
| 355 |
+
except:
|
| 356 |
+
logger.warning("Aucun tokenizer trouvé")
|
| 357 |
|
| 358 |
# Charge le processor
|
| 359 |
try:
|
|
|
|
| 363 |
except Exception as e:
|
| 364 |
logger.warning(f"Processor non trouvé: {e}")
|
| 365 |
|
| 366 |
+
return f"✅ Modèle {model_name} chargé avec succès!\nType: {type(self.current_model).__name__}\nArchitecture: {getattr(self.current_model.config, 'architectures', ['Inconnue'])[0] if hasattr(self.current_model, 'config') else 'Inconnue'}"
|
| 367 |
|
| 368 |
except Exception as e:
|
| 369 |
error_msg = f"❌ Erreur lors du chargement: {str(e)}"
|
|
|
|
| 438 |
info += f"📈 Exemples: {len(self.training_data):,}\n"
|
| 439 |
info += f"📝 Colonnes: {list(self.training_data.column_names)}\n"
|
| 440 |
|
| 441 |
+
def diagnose_model(self, model_name: str):
|
| 442 |
+
"""Diagnostique un modèle avant chargement"""
|
| 443 |
+
if not model_name.strip():
|
| 444 |
+
return "❌ Veuillez entrer un nom de modèle"
|
| 445 |
+
|
| 446 |
+
try:
|
| 447 |
+
from transformers import AutoConfig
|
| 448 |
+
import requests
|
| 449 |
+
|
| 450 |
+
result = f"🔍 DIAGNOSTIC DU MODÈLE: {model_name}\n\n"
|
| 451 |
+
|
| 452 |
+
# Vérification de l'existence
|
| 453 |
+
try:
|
| 454 |
+
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
|
| 455 |
+
result += "✅ Modèle accessible\n"
|
| 456 |
+
|
| 457 |
+
# Informations sur la configuration
|
| 458 |
+
result += f"📋 Type de modèle: {getattr(config, 'model_type', 'Non défini')}\n"
|
| 459 |
+
result += f"🏗️ Architecture: {getattr(config, 'architectures', ['Inconnue'])}\n"
|
| 460 |
+
result += f"📚 Vocabulaire: {getattr(config, 'vocab_size', 'Inconnu')}\n"
|
| 461 |
+
result += f"🧠 Couches cachées: {getattr(config, 'hidden_size', 'Inconnu')}\n"
|
| 462 |
+
result += f"🔢 Nombre de couches: {getattr(config, 'num_hidden_layers', 'Inconnu')}\n"
|
| 463 |
+
|
| 464 |
+
# Recommandations
|
| 465 |
+
if not hasattr(config, 'model_type') or config.model_type is None:
|
| 466 |
+
result += "\n⚠️ PROBLÈME: model_type manquant\n"
|
| 467 |
+
result += "💡 SOLUTION: Le chargeur essaiera de détecter automatiquement\n"
|
| 468 |
+
|
| 469 |
+
if hasattr(config, 'architectures') and config.architectures:
|
| 470 |
+
arch = config.architectures[0].lower()
|
| 471 |
+
if 'llama' in arch:
|
| 472 |
+
result += "🎯 Type détecté: LLaMA\n"
|
| 473 |
+
elif 'mistral' in arch:
|
| 474 |
+
result += "🎯 Type détecté: Mistral\n"
|
| 475 |
+
elif 'qwen' in arch:
|
| 476 |
+
result += "🎯 Type détecté: Qwen\n"
|
| 477 |
+
elif 'phi' in arch:
|
| 478 |
+
result += "🎯 Type détecté: Phi\n"
|
| 479 |
+
|
| 480 |
+
result += "\n✅ Chargement possible avec les stratégies multiples"
|
| 481 |
+
|
| 482 |
+
except Exception as e:
|
| 483 |
+
result += f"❌ Erreur d'accès: {str(e)}\n"
|
| 484 |
+
result += "💡 Vérifiez que le modèle existe et est public\n"
|
| 485 |
+
|
| 486 |
+
return result
|
| 487 |
+
|
| 488 |
+
except Exception as e:
|
| 489 |
+
return f"❌ Erreur diagnostic: {str(e)}"
|
| 490 |
|
| 491 |
# Initialisation
|
| 492 |
trainer = MultimodalTrainer()
|
|
|
|
| 560 |
value="causal"
|
| 561 |
)
|
| 562 |
load_model_btn = gr.Button("🔄 Charger le modèle", variant="primary")
|
| 563 |
+
diagnose_btn = gr.Button("🔍 Diagnostiquer le modèle", variant="secondary")
|
| 564 |
|
| 565 |
with gr.Column():
|
| 566 |
model_status = gr.Textbox(
|
|
|
|
| 582 |
outputs=model_status
|
| 583 |
)
|
| 584 |
|
| 585 |
+
diagnose_btn.click(
|
| 586 |
+
trainer.diagnose_model,
|
| 587 |
+
inputs=[model_input],
|
| 588 |
+
outputs=model_status
|
| 589 |
+
)
|
| 590 |
+
|
| 591 |
info_btn.click(trainer.get_model_info, outputs=model_info)
|
| 592 |
|
| 593 |
with gr.Tab("📊 Données"):
|