disentangle model from tuner and rename modules

Former-commit-id: 02cbf91e7e424f8379c1fed01b82a5f7a83b6947

src/llmtuner/train/sft/trainer.py

@@ -0,0 +1,92 @@
+import os
+import json
+import torch
+import numpy as np
+import torch.nn as nn
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
+from transformers import Seq2SeqTrainer
+
+from llmtuner.extras.constants import IGNORE_INDEX
+from llmtuner.extras.logging import get_logger
+
+if TYPE_CHECKING:
+    from transformers.trainer import PredictionOutput
+
+
+logger = get_logger(__name__)
+
+
+class CustomSeq2SeqTrainer(Seq2SeqTrainer):
+    r"""
+    Inherits PeftTrainer to compute generative metrics such as BLEU and ROUGE.
+    """
+
+    def prediction_step(
+        self,
+        model: nn.Module,
+        inputs: Dict[str, Union[torch.Tensor, Any]],
+        prediction_loss_only: bool,
+        ignore_keys: Optional[List[str]] = None,
+    ) -> Tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]:
+        r"""
+        Removes the prompt part in the generated tokens.
+
+        Subclass and override to inject custom behavior.
+        """
+        labels = inputs["labels"].detach().clone() if "labels" in inputs else None # backup labels
+        if self.args.predict_with_generate:
+            assert self.tokenizer.padding_side == "left", "This method only accepts left-padded tensor."
+            prompt_len, label_len = inputs["input_ids"].size(-1), inputs["labels"].size(-1)
+            if prompt_len > label_len:
+                inputs["labels"] = self._pad_tensors_to_target_len(inputs["labels"], inputs["input_ids"])
+            if label_len > prompt_len:
+                inputs["labels"] = inputs["labels"][:, :prompt_len] # truncate the labels instead of padding the inputs
+
+        loss, generated_tokens, _ = super().prediction_step(
+            model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys
+        )
+        if generated_tokens is not None and self.args.predict_with_generate:
+            generated_tokens[:, :prompt_len] = self.tokenizer.pad_token_id
+            generated_tokens = generated_tokens.contiguous()
+
+        return loss, generated_tokens, labels
+
+    def _pad_tensors_to_target_len(
+        self,
+        src_tensor: torch.Tensor,
+        tgt_tensor: torch.Tensor
+    ) -> torch.Tensor:
+        r"""
+        Pads the tensor to the same length as the target tensor.
+        """
+        assert self.tokenizer.pad_token_id is not None, "Pad token is required."
+        padded_tensor = self.tokenizer.pad_token_id * torch.ones_like(tgt_tensor)
+        padded_tensor[:, -src_tensor.shape[-1]:] = src_tensor # adopt left-padding
+        return padded_tensor.contiguous() # in contiguous memory
+
+    def save_predictions(
+        self,
+        predict_results: "PredictionOutput"
+    ) -> None:
+        r"""
+        Saves model predictions to `output_dir`.
+
+        A custom behavior that not contained in Seq2SeqTrainer.
+        """
+        if not self.is_world_process_zero():
+            return
+
+        output_prediction_file = os.path.join(self.args.output_dir, "generated_predictions.jsonl")
+        logger.info(f"Saving prediction results to {output_prediction_file}")
+
+        preds = np.where(predict_results.predictions != IGNORE_INDEX, predict_results.predictions, self.tokenizer.pad_token_id)
+        labels = np.where(predict_results.label_ids != IGNORE_INDEX, predict_results.label_ids, self.tokenizer.pad_token_id)
+
+        decoded_preds = self.tokenizer.batch_decode(preds, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+        decoded_labels = self.tokenizer.batch_decode(labels, skip_special_tokens=True, clean_up_tokenization_spaces=True)
+
+        with open(output_prediction_file, "w", encoding="utf-8") as writer:
+            res: List[str] = []
+            for pred, label in zip(decoded_preds, decoded_labels):
+                res.append(json.dumps({"label": label, "predict": pred}, ensure_ascii=False))
+            writer.write("\n".join(res))