model.py

from torch import nn
from transformers import AutoModel, PretrainedConfig, PreTrainedModel
from transformers.modeling_outputs import QuestionAnsweringModelOutput


class CustomQAModelConfig(PretrainedConfig):
    model_type = "modernbert"

    def __init__(self, base_model_name_or_path="answerdotai/ModernBERT-base", **kwargs):
        self.base_model_name_or_path = base_model_name_or_path
        super().__init__(**kwargs)


class CustomQAModel(PreTrainedModel):
    config_class = CustomQAModelConfig

    def __init__(self, config):
        super().__init__(config)
        self.base = AutoModel.from_pretrained(config.base_model_name_or_path)
        hidden_size = self.base.config.hidden_size
        self.qa_outputs = nn.Linear(hidden_size, 2)
        self.loss_fn = nn.CrossEntropyLoss()

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        start_positions=None,
        end_positions=None,
    ):
        outputs = self.base(
            input_ids=input_ids,
            attention_mask=attention_mask,
        )
        hidden_states = outputs.last_hidden_state

        logits = self.qa_outputs(hidden_states)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1)
        end_logits = end_logits.squeeze(-1)

        loss = None
        if start_positions is not None and end_positions is not None:
            start_positions = start_positions.clamp(0, start_logits.size(1) - 1)
            end_positions = end_positions.clamp(0, end_logits.size(1) - 1)

            start_loss = self.loss_fn(start_logits, start_positions)
            end_loss = self.loss_fn(end_logits, end_positions)
            loss = (start_loss + end_loss) / 2

        return QuestionAnsweringModelOutput(
            loss=loss,
            start_logits=start_logits,
            end_logits=end_logits,
        )