disentangle model from tuner and rename modules

Former-commit-id: 02cbf91e7e424f8379c1fed01b82a5f7a83b6947

src/llmtuner/train/sft/__init__.py

@@ -0,0 +1 @@
+from llmtuner.train.sft.workflow import run_sft

src/llmtuner/train/sft/metric.py

@@ -0,0 +1,61 @@
+import numpy as np
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Dict, Sequence, Tuple, Union
+
+from llmtuner.extras.constants import IGNORE_INDEX
+from llmtuner.extras.packages import (
+    is_jieba_available, is_nltk_available, is_rouge_available
+)
+
+if TYPE_CHECKING:
+    from transformers.tokenization_utils import PreTrainedTokenizer
+
+if is_jieba_available():
+    import jieba
+
+if is_nltk_available():
+    from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
+
+if is_rouge_available():
+    from rouge_chinese import Rouge
+
+
+@dataclass
+class ComputeMetrics:
+    r"""
+    Wraps the tokenizer into metric functions, used in Seq2SeqPeftTrainer.
+    """
+
+    tokenizer: "PreTrainedTokenizer"
+
+    def __call__(self, eval_preds: Sequence[Union[np.ndarray, Tuple[np.ndarray]]]) -> Dict[str, float]:
+        r"""
+        Uses the model predictions to compute metrics.
+        """
+        preds, labels = eval_preds
+        score_dict = {"rouge-1": [], "rouge-2": [], "rouge-l": [], "bleu-4": []}
+
+        preds = np.where(preds != IGNORE_INDEX, preds, self.tokenizer.pad_token_id)
+        labels = np.where(labels != IGNORE_INDEX, labels, self.tokenizer.pad_token_id)
+
+        decoded_preds = self.tokenizer.batch_decode(preds, skip_special_tokens=True)
+        decoded_labels = self.tokenizer.batch_decode(labels, skip_special_tokens=True)
+
+        for pred, label in zip(decoded_preds, decoded_labels):
+            hypothesis = list(jieba.cut(pred))
+            reference = list(jieba.cut(label))
+
+            if len(" ".join(hypothesis).split()) == 0 or len(" ".join(reference).split()) == 0:
+                result = {"rouge-1": {"f": 0.0}, "rouge-2": {"f": 0.0}, "rouge-l": {"f": 0.0}}
+            else:
+                rouge = Rouge()
+                scores = rouge.get_scores(" ".join(hypothesis), " ".join(reference))
+                result = scores[0]
+
+            for k, v in result.items():
+                score_dict[k].append(round(v["f"] * 100, 4))
+
+            bleu_score = sentence_bleu([list(label)], list(pred), smoothing_function=SmoothingFunction().method3)
+            score_dict["bleu-4"].append(round(bleu_score * 100, 4))
+
+        return {k: float(np.mean(v)) for k, v in score_dict.items()}

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))

src/llmtuner/train/sft/workflow.py

@@ -0,0 +1,97 @@
+# Inspired by: https://github.com/huggingface/transformers/blob/v4.34.1/examples/pytorch/summarization/run_summarization.py
+
+from typing import TYPE_CHECKING, Optional, List
+from transformers import DataCollatorForSeq2Seq, Seq2SeqTrainingArguments
+
+from llmtuner.data import get_dataset, preprocess_dataset, split_dataset
+from llmtuner.extras.constants import IGNORE_INDEX
+from llmtuner.extras.misc import get_logits_processor
+from llmtuner.extras.ploting import plot_loss
+from llmtuner.model import generate_model_card, load_model_and_tokenizer
+from llmtuner.train.sft.metric import ComputeMetrics
+from llmtuner.train.sft.trainer import CustomSeq2SeqTrainer
+
+if TYPE_CHECKING:
+    from transformers import TrainerCallback
+    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments
+
+
+def run_sft(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    generating_args: "GeneratingArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None
+):
+    dataset = get_dataset(model_args, data_args)
+    model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, training_args.do_train, stage="sft")
+    dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, stage="sft")
+
+    if training_args.predict_with_generate:
+        tokenizer.padding_side = "left" # use left-padding in generation
+
+    data_collator = DataCollatorForSeq2Seq(
+        tokenizer=tokenizer,
+        pad_to_multiple_of=4 if tokenizer.padding_side == "right" else None, # for shift short attention
+        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
+    )
+
+    # Override the decoding parameters of Seq2SeqTrainer
+    training_args_dict = training_args.to_dict()
+    training_args_dict.update(dict(
+        generation_max_length=training_args.generation_max_length or data_args.cutoff_len,
+        generation_num_beams=data_args.eval_num_beams or training_args.generation_num_beams
+    ))
+    training_args = Seq2SeqTrainingArguments(**training_args_dict)
+
+    # Initialize our Trainer
+    trainer = CustomSeq2SeqTrainer(
+        model=model,
+        args=training_args,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        compute_metrics=ComputeMetrics(tokenizer) if training_args.predict_with_generate else None,
+        **split_dataset(dataset, data_args, training_args)
+    )
+
+    # Keyword arguments for `model.generate`
+    gen_kwargs = generating_args.to_dict()
+    gen_kwargs["eos_token_id"] = [tokenizer.eos_token_id] + tokenizer.additional_special_tokens_ids
+    gen_kwargs["pad_token_id"] = tokenizer.pad_token_id
+    gen_kwargs["logits_processor"] = get_logits_processor()
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        trainer.save_model()
+        trainer.log_metrics("train", train_result.metrics)
+        trainer.save_metrics("train", train_result.metrics)
+        trainer.save_state()
+        if trainer.is_world_process_zero() and model_args.plot_loss:
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval", **gen_kwargs)
+        if training_args.predict_with_generate: # eval_loss will be wrong if predict_with_generate is enabled
+            metrics.pop("eval_loss", None)
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Predict
+    if training_args.do_predict:
+        predict_results = trainer.predict(dataset, metric_key_prefix="predict", **gen_kwargs)
+        if training_args.predict_with_generate: # predict_loss will be wrong if predict_with_generate is enabled
+            predict_results.metrics.pop("predict_loss", None)
+        trainer.log_metrics("predict", predict_results.metrics)
+        trainer.save_metrics("predict", predict_results.metrics)
+        trainer.save_predictions(predict_results)
+
+    # Create model card
+    if training_args.do_train:
+        if training_args.push_to_hub:
+            trainer.push_to_hub(**generate_model_card(model_args, data_args, finetuning_args))
+        else:
+            trainer.create_model_card(**generate_model_card(model_args, data_args, finetuning_args))