model_loader.py

from transformers import AutoModelForTokenClassification, PretrainedConfig
import torch

class ModernBertForTokenClassificationCRF(AutoModelForTokenClassification):
    def __init__(self, config):
        super().__init__(config)
        # Add CRF layer if needed
        if config.use_crf:
            from torchcrf import CRF
            self.crf = CRF(num_tags=config.num_labels, batch_first=True)
            
    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
        config = kwargs.pop('config', None)
        if config is None:
            config = PretrainedConfig.from_pretrained(pretrained_model_name_or_path)
        return super().from_pretrained(pretrained_model_name_or_path, *args, config=config, **kwargs)