Update app.py
Browse files
app.py
CHANGED
|
@@ -44,7 +44,7 @@ def check_and_import_dependencies():
|
|
| 44 |
"""Vérifie et importe toutes les dépendances"""
|
| 45 |
global numpy, torch, NUMPY_AVAILABLE, TORCH_AVAILABLE, TRANSFORMERS_AVAILABLE
|
| 46 |
global DATASETS_AVAILABLE, HF_HUB_AVAILABLE, PIL_AVAILABLE, LIBROSA_AVAILABLE, CV2_AVAILABLE
|
| 47 |
-
global AutoTokenizer, AutoModel, AutoProcessor, AutoModelForCausalLM
|
| 48 |
global TrainingArguments, Trainer, DataCollatorForLanguageModeling
|
| 49 |
global Dataset, load_dataset, concatenate_datasets, HfApi, Image, librosa, cv2
|
| 50 |
|
|
@@ -67,14 +67,14 @@ def check_and_import_dependencies():
|
|
| 67 |
# Transformers
|
| 68 |
try:
|
| 69 |
from transformers import (
|
| 70 |
-
AutoTokenizer, AutoModel, AutoProcessor,
|
| 71 |
AutoModelForCausalLM, TrainingArguments, Trainer,
|
| 72 |
DataCollatorForLanguageModeling
|
| 73 |
)
|
| 74 |
TRANSFORMERS_AVAILABLE = True
|
| 75 |
except ImportError:
|
| 76 |
TRANSFORMERS_AVAILABLE = False
|
| 77 |
-
AutoTokenizer = AutoModel = AutoProcessor = None
|
| 78 |
AutoModelForCausalLM = TrainingArguments = Trainer = None
|
| 79 |
DataCollatorForLanguageModeling = None
|
| 80 |
|
|
@@ -144,16 +144,17 @@ class MultimodalTrainer:
|
|
| 144 |
"""Installe les dépendances manquantes"""
|
| 145 |
installation_results = []
|
| 146 |
|
| 147 |
-
# Mapping des packages
|
| 148 |
package_mapping = {
|
| 149 |
-
"torch": "torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu",
|
| 150 |
-
"transformers": "transformers",
|
| 151 |
-
"datasets": "datasets",
|
| 152 |
-
"accelerate": "accelerate",
|
| 153 |
-
"pillow": "pillow",
|
| 154 |
-
"librosa": "librosa",
|
| 155 |
-
"opencv": "opencv-python",
|
| 156 |
-
"huggingface_hub": "huggingface_hub"
|
|
|
|
| 157 |
}
|
| 158 |
|
| 159 |
for package in packages_to_install:
|
|
@@ -167,7 +168,7 @@ class MultimodalTrainer:
|
|
| 167 |
try:
|
| 168 |
subprocess.check_call([
|
| 169 |
sys.executable, "-m", "pip", "install",
|
| 170 |
-
"torch", "torchvision", "torchaudio",
|
| 171 |
"--index-url", "https://download.pytorch.org/whl/cpu",
|
| 172 |
"--quiet"
|
| 173 |
])
|
|
@@ -230,145 +231,222 @@ class MultimodalTrainer:
|
|
| 230 |
status += f"🚀 GPU: {torch.cuda.get_device_name()}\n"
|
| 231 |
status += f"🔋 VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f}GB\n"
|
| 232 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 233 |
return status
|
| 234 |
|
| 235 |
-
def
|
| 236 |
-
"""
|
| 237 |
if not TRANSFORMERS_AVAILABLE:
|
| 238 |
-
return "❌ Transformers non installé! Utilisez l'outil d'installation."
|
| 239 |
|
| 240 |
if not TORCH_AVAILABLE or not torch:
|
| 241 |
-
return "❌ PyTorch non installé! Utilisez l'outil d'installation."
|
| 242 |
|
| 243 |
-
if not model_name.strip():
|
| 244 |
-
return "❌ Veuillez entrer un nom de modèle"
|
| 245 |
-
|
| 246 |
try:
|
| 247 |
-
logger.info(f"Chargement du modèle: {model_name}")
|
| 248 |
|
| 249 |
-
#
|
| 250 |
-
|
| 251 |
-
|
|
|
|
|
|
|
|
|
|
| 252 |
|
| 253 |
-
#
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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 |
-
|
| 263 |
-
|
| 264 |
-
|
| 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 |
-
|
| 276 |
-
|
| 277 |
-
|
| 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 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 334 |
except Exception as e:
|
| 335 |
-
|
|
|
|
|
|
|
| 336 |
|
| 337 |
-
if
|
| 338 |
-
return f"❌
|
| 339 |
|
| 340 |
-
#
|
|
|
|
| 341 |
try:
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
)
|
| 345 |
-
if self.current_tokenizer.pad_token is None:
|
| 346 |
-
self.current_tokenizer.pad_token = self.current_tokenizer.eos_token
|
| 347 |
except Exception as e:
|
| 348 |
-
logger.warning(f"
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 359 |
try:
|
| 360 |
-
|
| 361 |
-
|
| 362 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 363 |
except Exception as e:
|
| 364 |
-
|
| 365 |
|
| 366 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 367 |
|
| 368 |
except Exception as e:
|
| 369 |
-
|
| 370 |
-
logger.error(error_msg)
|
| 371 |
-
return error_msg
|
| 372 |
|
| 373 |
def load_single_dataset(self, dataset_name: str, split: str = "train"):
|
| 374 |
"""Charge un dataset individuel"""
|
|
@@ -437,56 +515,8 @@ class MultimodalTrainer:
|
|
| 437 |
info += f"\n📊 DONNÉES:\n"
|
| 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 |
-
|
| 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()
|
|
@@ -497,17 +527,18 @@ def create_interface():
|
|
| 497 |
|
| 498 |
gr.Markdown("""
|
| 499 |
# 🔥 Multimodal Training Hub
|
| 500 |
-
### Plateforme d'entraînement de modèles multimodaux
|
| 501 |
|
| 502 |
🤖 Modèles • 📊 Datasets • 🏋️ Training • 🛠️ Outils
|
| 503 |
""")
|
| 504 |
|
| 505 |
with gr.Tab("🔧 Diagnostic"):
|
| 506 |
-
gr.Markdown("### 🩺 Vérification du système")
|
| 507 |
|
| 508 |
with gr.Row():
|
| 509 |
check_deps_btn = gr.Button("🔍 Vérifier dépendances", variant="primary")
|
| 510 |
install_core_btn = gr.Button("📦 Installer packages critiques", variant="secondary")
|
|
|
|
| 511 |
|
| 512 |
deps_status = gr.Textbox(
|
| 513 |
label="État des dépendances",
|
|
@@ -545,28 +576,41 @@ def create_interface():
|
|
| 545 |
lambda: trainer.install_dependencies(["torch", "transformers", "datasets", "accelerate"]),
|
| 546 |
outputs=install_status
|
| 547 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 548 |
|
| 549 |
with gr.Tab("🤖 Modèle"):
|
| 550 |
with gr.Row():
|
| 551 |
with gr.Column():
|
| 552 |
model_input = gr.Textbox(
|
| 553 |
label="Nom du modèle HuggingFace",
|
| 554 |
-
placeholder="
|
| 555 |
-
value="
|
| 556 |
)
|
| 557 |
model_type = gr.Dropdown(
|
| 558 |
label="Type de modèle",
|
| 559 |
choices=["causal", "base"],
|
| 560 |
value="causal"
|
| 561 |
)
|
| 562 |
-
|
| 563 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 564 |
|
| 565 |
with gr.Column():
|
| 566 |
model_status = gr.Textbox(
|
| 567 |
label="Status du modèle",
|
| 568 |
interactive=False,
|
| 569 |
-
lines=
|
| 570 |
)
|
| 571 |
|
| 572 |
info_btn = gr.Button("ℹ️ Info modèle")
|
|
@@ -659,23 +703,4 @@ def create_interface():
|
|
| 659 |
|
| 660 |
with gr.Column():
|
| 661 |
training_status = gr.Textbox(
|
| 662 |
-
label="Status
|
| 663 |
-
interactive=False,
|
| 664 |
-
lines=12
|
| 665 |
-
)
|
| 666 |
-
|
| 667 |
-
train_btn.click(
|
| 668 |
-
trainer.simulate_training,
|
| 669 |
-
inputs=[output_dir, num_epochs, learning_rate, batch_size],
|
| 670 |
-
outputs=training_status
|
| 671 |
-
)
|
| 672 |
-
|
| 673 |
-
# Auto-check au démarrage
|
| 674 |
-
app.load(trainer.check_dependencies, outputs=deps_status)
|
| 675 |
-
|
| 676 |
-
return app
|
| 677 |
-
|
| 678 |
-
# Lancement
|
| 679 |
-
if __name__ == "__main__":
|
| 680 |
-
app = create_interface()
|
| 681 |
-
app.launch(share=True, server_name="0.0.0.0", server_port=7860)
|
|
|
|
| 44 |
"""Vérifie et importe toutes les dépendances"""
|
| 45 |
global numpy, torch, NUMPY_AVAILABLE, TORCH_AVAILABLE, TRANSFORMERS_AVAILABLE
|
| 46 |
global DATASETS_AVAILABLE, HF_HUB_AVAILABLE, PIL_AVAILABLE, LIBROSA_AVAILABLE, CV2_AVAILABLE
|
| 47 |
+
global AutoTokenizer, AutoModel, AutoProcessor, AutoModelForCausalLM, AutoConfig
|
| 48 |
global TrainingArguments, Trainer, DataCollatorForLanguageModeling
|
| 49 |
global Dataset, load_dataset, concatenate_datasets, HfApi, Image, librosa, cv2
|
| 50 |
|
|
|
|
| 67 |
# Transformers
|
| 68 |
try:
|
| 69 |
from transformers import (
|
| 70 |
+
AutoTokenizer, AutoModel, AutoProcessor, AutoConfig,
|
| 71 |
AutoModelForCausalLM, TrainingArguments, Trainer,
|
| 72 |
DataCollatorForLanguageModeling
|
| 73 |
)
|
| 74 |
TRANSFORMERS_AVAILABLE = True
|
| 75 |
except ImportError:
|
| 76 |
TRANSFORMERS_AVAILABLE = False
|
| 77 |
+
AutoTokenizer = AutoModel = AutoProcessor = AutoConfig = None
|
| 78 |
AutoModelForCausalLM = TrainingArguments = Trainer = None
|
| 79 |
DataCollatorForLanguageModeling = None
|
| 80 |
|
|
|
|
| 144 |
"""Installe les dépendances manquantes"""
|
| 145 |
installation_results = []
|
| 146 |
|
| 147 |
+
# Mapping des packages avec versions spécifiques
|
| 148 |
package_mapping = {
|
| 149 |
+
"torch": "torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cpu",
|
| 150 |
+
"transformers": "transformers>=4.46.2",
|
| 151 |
+
"datasets": "datasets>=2.21.0",
|
| 152 |
+
"accelerate": "accelerate>=1.1.0",
|
| 153 |
+
"pillow": "pillow>=10.1.0",
|
| 154 |
+
"librosa": "librosa>=0.10.1",
|
| 155 |
+
"opencv": "opencv-python-headless>=4.8.1.78",
|
| 156 |
+
"huggingface_hub": "huggingface_hub>=0.26.0",
|
| 157 |
+
"qwen": "qwen-vl-utils>=0.0.8"
|
| 158 |
}
|
| 159 |
|
| 160 |
for package in packages_to_install:
|
|
|
|
| 168 |
try:
|
| 169 |
subprocess.check_call([
|
| 170 |
sys.executable, "-m", "pip", "install",
|
| 171 |
+
"torch==2.1.0", "torchvision==0.16.0", "torchaudio==2.1.0",
|
| 172 |
"--index-url", "https://download.pytorch.org/whl/cpu",
|
| 173 |
"--quiet"
|
| 174 |
])
|
|
|
|
| 231 |
status += f"🚀 GPU: {torch.cuda.get_device_name()}\n"
|
| 232 |
status += f"🔋 VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f}GB\n"
|
| 233 |
|
| 234 |
+
# Versions spécifiques
|
| 235 |
+
if TRANSFORMERS_AVAILABLE:
|
| 236 |
+
import transformers
|
| 237 |
+
status += f"🤗 Transformers: {transformers.__version__}\n"
|
| 238 |
+
|
| 239 |
return status
|
| 240 |
|
| 241 |
+
def load_model_safe(self, model_name: str):
|
| 242 |
+
"""Chargement sécurisé du modèle avec gestion d'erreurs avancée"""
|
| 243 |
if not TRANSFORMERS_AVAILABLE:
|
| 244 |
+
return "❌ Transformers non installé! Utilisez l'outil d'installation.", None, None
|
| 245 |
|
| 246 |
if not TORCH_AVAILABLE or not torch:
|
| 247 |
+
return "❌ PyTorch non installé! Utilisez l'outil d'installation.", None, None
|
| 248 |
|
|
|
|
|
|
|
|
|
|
| 249 |
try:
|
| 250 |
+
logger.info(f"Chargement sécurisé du modèle: {model_name}")
|
| 251 |
|
| 252 |
+
# Étape 1: Vérification de la configuration
|
| 253 |
+
try:
|
| 254 |
+
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
|
| 255 |
+
logger.info(f"Configuration chargée: {config.model_type}")
|
| 256 |
+
except Exception as e:
|
| 257 |
+
return f"❌ Erreur configuration: {str(e)}", None, None
|
| 258 |
|
| 259 |
+
# Étape 2: Chargement du tokenizer
|
| 260 |
+
tokenizer = None
|
| 261 |
+
try:
|
| 262 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 263 |
+
model_name,
|
| 264 |
+
trust_remote_code=True,
|
| 265 |
+
use_fast=False
|
| 266 |
+
)
|
| 267 |
+
if tokenizer.pad_token is None:
|
| 268 |
+
tokenizer.pad_token = tokenizer.eos_token
|
| 269 |
+
logger.info("Tokenizer chargé avec succès")
|
| 270 |
+
except Exception as e:
|
| 271 |
+
logger.warning(f"Tokenizer non trouvé: {e}")
|
| 272 |
+
return f"❌ Erreur tokenizer: {str(e)}", None, None
|
| 273 |
+
|
| 274 |
+
# Étape 3: Chargement du modèle avec stratégies multiples
|
| 275 |
+
model = None
|
| 276 |
+
loading_strategies = [
|
| 277 |
+
{
|
| 278 |
+
"name": "AutoModelForCausalLM standard",
|
| 279 |
+
"loader": lambda: AutoModelForCausalLM.from_pretrained(
|
| 280 |
model_name,
|
| 281 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 282 |
device_map="auto" if torch.cuda.is_available() else None,
|
| 283 |
+
trust_remote_code=True,
|
| 284 |
+
low_cpu_mem_usage=True
|
| 285 |
)
|
| 286 |
+
},
|
| 287 |
+
{
|
| 288 |
+
"name": "AutoModelForCausalLM avec config explicite",
|
| 289 |
+
"loader": lambda: AutoModelForCausalLM.from_pretrained(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 290 |
model_name,
|
| 291 |
+
config=config,
|
| 292 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 293 |
device_map="auto" if torch.cuda.is_available() else None,
|
| 294 |
+
trust_remote_code=True,
|
| 295 |
+
low_cpu_mem_usage=True,
|
| 296 |
+
attn_implementation="eager"
|
| 297 |
)
|
| 298 |
+
},
|
| 299 |
+
{
|
| 300 |
+
"name": "AutoModel générique",
|
| 301 |
+
"loader": lambda: AutoModel.from_pretrained(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 302 |
model_name,
|
|
|
|
| 303 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 304 |
device_map="auto" if torch.cuda.is_available() else None,
|
| 305 |
+
trust_remote_code=True,
|
| 306 |
+
low_cpu_mem_usage=True
|
| 307 |
)
|
| 308 |
+
}
|
| 309 |
+
]
|
| 310 |
+
|
| 311 |
+
last_error = None
|
| 312 |
+
for strategy in loading_strategies:
|
| 313 |
+
try:
|
| 314 |
+
logger.info(f"Tentative: {strategy['name']}")
|
| 315 |
+
model = strategy["loader"]()
|
| 316 |
+
logger.info(f"✅ Succès avec: {strategy['name']}")
|
| 317 |
+
break
|
| 318 |
except Exception as e:
|
| 319 |
+
last_error = str(e)
|
| 320 |
+
logger.warning(f"❌ Échec {strategy['name']}: {e}")
|
| 321 |
+
continue
|
| 322 |
|
| 323 |
+
if model is None:
|
| 324 |
+
return f"❌ Toutes les stratégies ont échoué. Dernière erreur: {last_error}", None, None
|
| 325 |
|
| 326 |
+
# Étape 4: Chargement du processor (optionnel)
|
| 327 |
+
processor = None
|
| 328 |
try:
|
| 329 |
+
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
|
| 330 |
+
logger.info("Processor chargé avec succès")
|
|
|
|
|
|
|
|
|
|
| 331 |
except Exception as e:
|
| 332 |
+
logger.warning(f"Processor non disponible: {e}")
|
| 333 |
+
|
| 334 |
+
return "✅ Modèle chargé avec succès!", model, tokenizer, processor
|
| 335 |
+
|
| 336 |
+
except Exception as e:
|
| 337 |
+
error_msg = f"❌ Erreur critique: {str(e)}"
|
| 338 |
+
logger.error(error_msg)
|
| 339 |
+
return error_msg, None, None
|
| 340 |
+
|
| 341 |
+
def load_model(self, model_name: str, model_type: str = "causal"):
|
| 342 |
+
"""Charge un modèle depuis Hugging Face avec gestion d'erreurs améliorée"""
|
| 343 |
+
if not model_name.strip():
|
| 344 |
+
return "❌ Veuillez entrer un nom de modèle"
|
| 345 |
+
|
| 346 |
+
# Utilise la méthode sécurisée
|
| 347 |
+
result = self.load_model_safe(model_name)
|
| 348 |
+
|
| 349 |
+
if len(result) == 4: # Succès
|
| 350 |
+
message, model, tokenizer, processor = result
|
| 351 |
+
self.current_model = model
|
| 352 |
+
self.current_tokenizer = tokenizer
|
| 353 |
+
self.current_processor = processor
|
| 354 |
+
|
| 355 |
+
# Informations détaillées
|
| 356 |
+
info = f"{message}\n"
|
| 357 |
+
info += f"🏷️ Type: {type(model).__name__}\n"
|
| 358 |
+
if hasattr(model, 'config'):
|
| 359 |
+
info += f"🏗️ Architecture: {getattr(model.config, 'architectures', ['Inconnue'])[0] if hasattr(model.config, 'architectures') else 'Inconnue'}\n"
|
| 360 |
+
info += f"📋 Model type: {getattr(model.config, 'model_type', 'Non défini')}\n"
|
| 361 |
+
|
| 362 |
+
if TORCH_AVAILABLE and torch:
|
| 363 |
+
info += f"💾 Device: {next(model.parameters()).device}\n"
|
| 364 |
+
total_params = sum(p.numel() for p in model.parameters())
|
| 365 |
+
info += f"🔢 Paramètres: {total_params:,}\n"
|
| 366 |
+
|
| 367 |
+
return info
|
| 368 |
+
else:
|
| 369 |
+
# Erreur
|
| 370 |
+
return result[0]
|
| 371 |
+
|
| 372 |
+
def diagnose_model(self, model_name: str):
|
| 373 |
+
"""Diagnostique avancé d'un modèle"""
|
| 374 |
+
if not model_name.strip():
|
| 375 |
+
return "❌ Veuillez entrer un nom de modèle"
|
| 376 |
+
|
| 377 |
+
try:
|
| 378 |
+
result = f"🔍 DIAGNOSTIC APPROFONDI: {model_name}\n\n"
|
| 379 |
+
|
| 380 |
+
# Vérification de l'existence
|
| 381 |
try:
|
| 382 |
+
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
|
| 383 |
+
result += "✅ Modèle accessible sur Hugging Face\n\n"
|
| 384 |
+
|
| 385 |
+
# Analyse de la configuration
|
| 386 |
+
result += "📋 CONFIGURATION:\n"
|
| 387 |
+
result += f"🏷️ Model type: {getattr(config, 'model_type', '❌ NON DÉFINI')}\n"
|
| 388 |
+
result += f"🏗️ Architectures: {getattr(config, 'architectures', ['❌ NON DÉFINI'])}\n"
|
| 389 |
+
result += f"📚 Vocab size: {getattr(config, 'vocab_size', 'Inconnu'):,}\n"
|
| 390 |
+
result += f"🧠 Hidden size: {getattr(config, 'hidden_size', 'Inconnu')}\n"
|
| 391 |
+
result += f"🔢 Layers: {getattr(config, 'num_hidden_layers', 'Inconnu')}\n"
|
| 392 |
+
result += f"🎯 Attention heads: {getattr(config, 'num_attention_heads', 'Inconnu')}\n"
|
| 393 |
+
|
| 394 |
+
# Vérification des problèmes courants
|
| 395 |
+
result += "\n🔧 ANALYSE DES PROBLÈMES:\n"
|
| 396 |
+
|
| 397 |
+
if not hasattr(config, 'model_type') or config.model_type is None:
|
| 398 |
+
result += "⚠️ PROBLÈME: model_type manquant\n"
|
| 399 |
+
if hasattr(config, 'architectures') and config.architectures:
|
| 400 |
+
arch = config.architectures[0].lower()
|
| 401 |
+
suggested_type = None
|
| 402 |
+
if 'qwen' in arch:
|
| 403 |
+
suggested_type = 'qwen2' if 'qwen2' in arch else 'qwen'
|
| 404 |
+
elif 'llama' in arch:
|
| 405 |
+
suggested_type = 'llama'
|
| 406 |
+
elif 'mistral' in arch:
|
| 407 |
+
suggested_type = 'mistral'
|
| 408 |
+
elif 'phi' in arch:
|
| 409 |
+
suggested_type = 'phi'
|
| 410 |
+
|
| 411 |
+
if suggested_type:
|
| 412 |
+
result += f"💡 Type suggéré: {suggested_type}\n"
|
| 413 |
+
else:
|
| 414 |
+
result += f"✅ Model type défini: {config.model_type}\n"
|
| 415 |
+
|
| 416 |
+
# Vérification de la compatibilité avec Transformers
|
| 417 |
+
if hasattr(config, 'architectures') and config.architectures:
|
| 418 |
+
arch = config.architectures[0]
|
| 419 |
+
if 'Qwen2_5OmniForCausalLM' in arch:
|
| 420 |
+
result += "⚠️ Architecture Qwen2.5-Omni détectée\n"
|
| 421 |
+
result += "💡 Nécessite Transformers >= 4.45.0\n"
|
| 422 |
+
if TRANSFORMERS_AVAILABLE:
|
| 423 |
+
import transformers
|
| 424 |
+
current_version = transformers.__version__
|
| 425 |
+
result += f"📦 Version actuelle: {current_version}\n"
|
| 426 |
+
|
| 427 |
+
# Stratégies de chargement recommandées
|
| 428 |
+
result += "\n🎯 STRATÉGIES DE CHARGEMENT:\n"
|
| 429 |
+
result += "1️⃣ AutoModelForCausalLM avec trust_remote_code=True\n"
|
| 430 |
+
result += "2️⃣ Configuration explicite si model_type manquant\n"
|
| 431 |
+
result += "3️⃣ Fallback vers AutoModel générique\n"
|
| 432 |
+
|
| 433 |
+
result += "\n✅ Diagnostic terminé - Chargement possible avec adaptations"
|
| 434 |
+
|
| 435 |
except Exception as e:
|
| 436 |
+
result += f"❌ Erreur d'accès: {str(e)}\n"
|
| 437 |
|
| 438 |
+
# Suggestions basées sur l'erreur
|
| 439 |
+
if "404" in str(e):
|
| 440 |
+
result += "💡 Le modèle n'existe pas ou n'est pas public\n"
|
| 441 |
+
elif "token" in str(e).lower():
|
| 442 |
+
result += "💡 Un token d'authentification pourrait être nécessaire\n"
|
| 443 |
+
else:
|
| 444 |
+
result += "💡 Vérifiez le nom du modèle et votre connexion\n"
|
| 445 |
+
|
| 446 |
+
return result
|
| 447 |
|
| 448 |
except Exception as e:
|
| 449 |
+
return f"❌ Erreur diagnostic: {str(e)}"
|
|
|
|
|
|
|
| 450 |
|
| 451 |
def load_single_dataset(self, dataset_name: str, split: str = "train"):
|
| 452 |
"""Charge un dataset individuel"""
|
|
|
|
| 515 |
info += f"\n📊 DONNÉES:\n"
|
| 516 |
info += f"📈 Exemples: {len(self.training_data):,}\n"
|
| 517 |
info += f"📝 Colonnes: {list(self.training_data.column_names)}\n"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 518 |
|
| 519 |
+
return info
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 520 |
|
| 521 |
# Initialisation
|
| 522 |
trainer = MultimodalTrainer()
|
|
|
|
| 527 |
|
| 528 |
gr.Markdown("""
|
| 529 |
# 🔥 Multimodal Training Hub
|
| 530 |
+
### Plateforme d'entraînement de modèles multimodaux optimisée pour Qwen2.5-Omni
|
| 531 |
|
| 532 |
🤖 Modèles • 📊 Datasets • 🏋️ Training • 🛠️ Outils
|
| 533 |
""")
|
| 534 |
|
| 535 |
with gr.Tab("🔧 Diagnostic"):
|
| 536 |
+
gr.Markdown("### 🩺 Vérification du système et installation")
|
| 537 |
|
| 538 |
with gr.Row():
|
| 539 |
check_deps_btn = gr.Button("🔍 Vérifier dépendances", variant="primary")
|
| 540 |
install_core_btn = gr.Button("📦 Installer packages critiques", variant="secondary")
|
| 541 |
+
install_qwen_btn = gr.Button("🎯 Support Qwen2.5", variant="secondary")
|
| 542 |
|
| 543 |
deps_status = gr.Textbox(
|
| 544 |
label="État des dépendances",
|
|
|
|
| 576 |
lambda: trainer.install_dependencies(["torch", "transformers", "datasets", "accelerate"]),
|
| 577 |
outputs=install_status
|
| 578 |
)
|
| 579 |
+
install_qwen_btn.click(
|
| 580 |
+
lambda: trainer.install_dependencies(["transformers", "qwen"]),
|
| 581 |
+
outputs=install_status
|
| 582 |
+
)
|
| 583 |
|
| 584 |
with gr.Tab("🤖 Modèle"):
|
| 585 |
with gr.Row():
|
| 586 |
with gr.Column():
|
| 587 |
model_input = gr.Textbox(
|
| 588 |
label="Nom du modèle HuggingFace",
|
| 589 |
+
placeholder="kvn420/Tenro_V4.1",
|
| 590 |
+
value="kvn420/Tenro_V4.1"
|
| 591 |
)
|
| 592 |
model_type = gr.Dropdown(
|
| 593 |
label="Type de modèle",
|
| 594 |
choices=["causal", "base"],
|
| 595 |
value="causal"
|
| 596 |
)
|
| 597 |
+
|
| 598 |
+
with gr.Row():
|
| 599 |
+
load_model_btn = gr.Button("🔄 Charger le modèle", variant="primary")
|
| 600 |
+
diagnose_btn = gr.Button("🔍 Diagnostiquer", variant="secondary")
|
| 601 |
+
|
| 602 |
+
gr.Markdown("""
|
| 603 |
+
💡 **Modèles testés:**
|
| 604 |
+
- `kvn420/Tenro_V4.1` (Qwen2.5-Omni)
|
| 605 |
+
- `Qwen/Qwen2.5-7B-Instruct`
|
| 606 |
+
- `microsoft/DialoGPT-medium`
|
| 607 |
+
""")
|
| 608 |
|
| 609 |
with gr.Column():
|
| 610 |
model_status = gr.Textbox(
|
| 611 |
label="Status du modèle",
|
| 612 |
interactive=False,
|
| 613 |
+
lines=10
|
| 614 |
)
|
| 615 |
|
| 616 |
info_btn = gr.Button("ℹ️ Info modèle")
|
|
|
|
| 703 |
|
| 704 |
with gr.Column():
|
| 705 |
training_status = gr.Textbox(
|
| 706 |
+
label="Status
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|