| |
| |
|
|
| |
|
|
| class LocalInference: |
| def __init__(self): |
| self.model = None |
| self.tokenizer = None |
| self.model_name = Config.MODEL_NAME |
| self.is_loading = True |
| self.load_error = None |
| self._load_task = None |
|
|
| async def initialize(self): |
| self._load_task = asyncio.create_task(self._load_model()) |
| await self._load_task |
|
|
| async def _load_model(self): |
| device = Config.DEVICE |
| |
| models_to_try = [ |
| self.model_name, |
| "Qwen/Qwen1.5-1.8B-Chat", |
| "TinyLlama/TinyLlama-1.1B-Chat-v1.0", |
| ] |
| for model_name in models_to_try: |
| try: |
| logger.info(f"Tentative de chargement: {model_name}") |
| loop = asyncio.get_running_loop() |
| def _load(): |
| |
| try: |
| tokenizer = AutoTokenizer.from_pretrained( |
| model_name, |
| token=Config.HF_TOKEN, |
| trust_remote_code=True, |
| use_fast=False, |
| ) |
| except Exception as e: |
| |
| logger.warning(f"Tokenizer avec use_fast=False échoué: {e}, tentative par défaut") |
| tokenizer = AutoTokenizer.from_pretrained( |
| model_name, |
| token=Config.HF_TOKEN, |
| trust_remote_code=True, |
| ) |
| |
| model = AutoModelForCausalLM.from_pretrained( |
| model_name, |
| token=Config.HF_TOKEN, |
| trust_remote_code=True, |
| dtype=torch.float16 if device == "cuda" else torch.float32, |
| low_cpu_mem_usage=True, |
| ) |
| model.to(device) |
| model.eval() |
| return tokenizer, model |
|
|
| self.tokenizer, self.model = await loop.run_in_executor(None, _load) |
| self.model_name = model_name |
| self.is_loading = False |
| logger.info(f"Modele charge avec succes: {model_name}") |
| return |
| except Exception as e: |
| logger.warning(f"Echec chargement {model_name}: {e}") |
| continue |
| self.is_loading = False |
| self.load_error = "Aucun modele n'a pu etre charge" |
| logger.error(self.load_error) |
|
|
| |
