step_train/src/model/base_model.py

"""Base model loading utilities for Qwen3-8B."""

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

from src.utils import load_yaml, setup_logger

logger = setup_logger(__name__)


def load_base_model(config_path: str = "configs/model/base_model.yaml"):
    """Load the base LLM (Qwen3-8B).

    Args:
        config_path: Path to base model config YAML.

    Returns:
        (model, tokenizer) tuple.
    """
    config = load_yaml(config_path)

    logger.info(f"Loading base model from {config['model_name_or_path']}")

    dtype_map = {
        "bfloat16": torch.bfloat16,
        "float16": torch.float16,
        "float32": torch.float32,
    }
    torch_dtype = dtype_map.get(config.get("torch_dtype", "bfloat16"), torch.bfloat16)

    model = AutoModelForCausalLM.from_pretrained(
        config["model_name_or_path"],
        torch_dtype=torch_dtype,
        device_map=config.get("device_map", "auto"),
        trust_remote_code=config.get("trust_remote_code", True),
        tp_plan=None,
    )

    tokenizer = load_tokenizer(config["model_name_or_path"])

    logger.info(f"Model loaded: {model.__class__.__name__}, dtype={torch_dtype}")
    return model, tokenizer


def load_tokenizer(model_name_or_path: str):
    """Load the tokenizer.

    Args:
        model_name_or_path: Model path.

    Returns:
        Tokenizer instance.
    """
    tokenizer = AutoTokenizer.from_pretrained(
        model_name_or_path,
        trust_remote_code=True,
    )
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"
    return tokenizer