src/training/trainer.py

"""
Custom HuggingFace Trainer subclass.
Uses the model's built-in cross-entropy loss (computed during forward pass)
instead of recomputing it, saving ~60MB of VRAM.
"""

from transformers import Trainer
import torch
from loguru import logger


class CorrectionTrainer(Trainer):
    """Custom trainer — uses model's built-in loss directly."""

    def __init__(self, loss_fn, fingerprinter, tokenizer, **kwargs):
        super().__init__(**kwargs)
        self.loss_fn = loss_fn  # Kept for API compat, not actually used
        self.fingerprinter = fingerprinter
        self.correction_tokenizer = tokenizer

    def _strip_custom_fields(self, inputs):
        """Remove dataset fields that T5 doesn't accept."""
        inputs.pop("style_vector", None)
        inputs.pop("input_text", None)
        inputs.pop("target_text", None)
        return {k: v for k, v in inputs.items() if k in ("input_ids", "attention_mask", "labels")}

    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
        """Use model's built-in CE loss — avoids double-computing logits loss."""
        model_inputs = self._strip_custom_fields(inputs)

        outputs = model(**model_inputs)
        # T5 computes CE loss internally when labels are provided — use it directly
        # This avoids keeping the full logits tensor (batch × seq × 32128) alive
        loss = outputs.loss

        return (loss, outputs) if return_outputs else loss

    def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys=None):
        """Compute eval loss directly — strips custom fields and runs forward.

        The parent's prediction_step doesn't return eval_loss when custom
        fields are present, so we handle it ourselves.
        """
        model_inputs = self._strip_custom_fields(inputs)
        model_inputs = self._prepare_inputs(model_inputs)

        with torch.no_grad():
            outputs = model(**model_inputs)
            loss = outputs.loss.detach()

        return (loss, None, None)