# Dans app.py — section de chargement du modèle print("⏳ Chargement du modèle Mistral-7B...") from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline import torch model_name = "mistralai/Mistral-7B-Instruct-v0.3" tokenizer = AutoTokenizer.from_pretrained(model_name) # Détecter si GPU disponible use_gpu = torch.cuda.is_available() print(f"{'✅ GPU détecté.' if use_gpu else '⚠️ Aucun GPU — utilisation du CPU (plus lent).'}") if use_gpu: # Charger en 4-bit si GPU from transformers import BitsAndBytesConfig bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16 ) model = AutoModelForCausalLM.from_pretrained( model_name, quantization_config=bnb_config, device_map="auto", trust_remote_code=True ) device = 0 else: # Charger en FP16/FP32 sur CPU (nécessite ~14 Go de RAM) model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype=torch.float16, # ou torch.float32 si mémoire insuffisante low_cpu_mem_usage=True, trust_remote_code=True ) device = -1 pipe = pipeline( "text-generation", model=model, tokenizer=tokenizer, max_new_tokens=512, temperature=0.3, top_p=0.95, repetition_penalty=1.2, device=device ) def llm_generate(prompt: str) -> str: response = pipe(prompt) return response[0]["generated_text"].replace(prompt, "").strip()