# model_loader.py from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer import logging logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) def load_model_pipeline(model_name: str = "distilgpt2", device: int = -1, max_new_tokens: int = 128, do_sample: bool = True, temperature: float = 0.7, top_k: int = 50, top_p: float = 0.95): """ Load and return a text-generation pipeline. - model_name: HF model id - device: -1 for CPU, >=0 for CUDA device Returns a callable pipeline that accepts a prompt string and returns generated text. """ logger.info(f"Loading model '{model_name}' (device={device})...") # Use Auto classes to ensure tokenizer/model compatibility tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained(model_name) gen_pipeline = pipeline( "text-generation", model=model, tokenizer=tokenizer, device=device, ) def generate(prompt: str, max_new_tokens_local: int = None, **kwargs): mnt = max_new_tokens_local if max_new_tokens_local is not None else max_new_tokens outputs = gen_pipeline( prompt, max_new_tokens=mnt, do_sample=do_sample, temperature=temperature, top_k=top_k, top_p=top_p, pad_token_id=tokenizer.eos_token_id, # avoid warnings / truncation issues **kwargs ) # pipeline returns a list of dicts with 'generated_text' return outputs[0]["generated_text"] logger.info("Model loaded.") return generate