evaluator / models /model_loader.py
kleervoyans's picture
Create model_loader.py
ff2d1f9 verified
import logging
from transformers import AutoTokenizer, pipeline, BitsAndBytesConfig
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
class ModelLoader:
"""
Loads a single translation model + tokenizer, with optional 8-bit quantization.
"""
def __init__(self, quantize: bool = True):
self.quantize = quantize
def load(self, model_name: str):
# 1) Tokenizer
logger.info(f"Loading tokenizer for {model_name}")
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
if not hasattr(tokenizer, "lang_code_to_id"):
raise AttributeError(f"Tokenizer for {model_name} has no lang_code_to_id mapping")
# 2) Pipeline
try:
bnb_cfg = BitsAndBytesConfig(load_in_8bit=self.quantize)
pipe = pipeline(
"translation",
model=model_name,
tokenizer=tokenizer,
device_map="auto",
quantization_config=bnb_cfg,
)
logger.info(f"Loaded {model_name} with 8-bit quantization")
except Exception as e:
logger.warning(f"8-bit quantization failed ({e}), loading full-precision")
pipe = pipeline(
"translation",
model=model_name,
tokenizer=tokenizer,
device_map="auto",
)
logger.info(f"Loaded {model_name} in full precision")
return tokenizer, pipe