rename package

src/llamafactory/train/rm/__init__.py

@@ -0,0 +1,4 @@
+from .workflow import run_rm
+
+
+__all__ = ["run_rm"]

src/llamafactory/train/rm/metric.py

@@ -0,0 +1,8 @@
+from typing import Dict, Sequence, Tuple, Union
+
+import numpy as np
+
+
+def compute_accuracy(eval_preds: Sequence[Union[np.ndarray, Tuple[np.ndarray]]]) -> Dict[str, float]:
+    preds, _ = eval_preds
+    return {"accuracy": (preds[0] > preds[1]).sum() / len(preds[0])}

src/llamafactory/train/rm/trainer.py

@@ -0,0 +1,128 @@
+import json
+import os
+from types import MethodType
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import Trainer
+
+from ...extras.logging import get_logger
+from ..utils import create_custom_optimzer, create_custom_scheduler
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, ProcessorMixin
+    from transformers.trainer import PredictionOutput
+
+    from ...hparams import FinetuningArguments
+
+
+logger = get_logger(__name__)
+
+
+class PairwiseTrainer(Trainer):
+    r"""
+    Inherits Trainer to compute pairwise loss.
+    """
+
+    def __init__(
+        self, finetuning_args: "FinetuningArguments", processor: Optional["ProcessorMixin"], **kwargs
+    ) -> None:
+        super().__init__(**kwargs)
+        self.finetuning_args = finetuning_args
+        self.processor = processor
+        self.can_return_loss = True  # override property to return eval_loss
+        if finetuning_args.use_badam:
+            from badam import clip_grad_norm_for_sparse_tensor
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
+
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimzer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    def create_scheduler(
+        self, num_training_steps: int, optimizer: Optional["torch.optim.Optimizer"] = None
+    ) -> "torch.optim.lr_scheduler.LRScheduler":
+        create_custom_scheduler(self.args, num_training_steps, optimizer)
+        return super().create_scheduler(num_training_steps, optimizer)
+
+    def _save(self, output_dir: Optional[str] = None, state_dict: Optional[Dict[str, "torch.Tensor"]] = None) -> None:
+        super()._save(output_dir, state_dict)
+        if self.processor is not None:
+            output_dir = output_dir if output_dir is not None else self.args.output_dir
+            getattr(self.processor, "image_processor").save_pretrained(output_dir)
+
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, torch.Tensor], return_outputs: bool = False
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]]]:
+        r"""
+        Computes pairwise loss. The first n examples are chosen and the last n examples are rejected.
+
+        Subclass and override to inject custom behavior.
+
+        Note that the first element will be removed from the output tuple.
+        See: https://github.com/huggingface/transformers/blob/v4.39.1/src/transformers/trainer.py#L3777
+        """
+        # Compute rewards
+        _, _, values = model(**inputs, output_hidden_states=True, return_dict=True)
+
+        unwrapped_model: "PreTrainedModel" = self.accelerator.unwrap_model(self.model)
+        if getattr(unwrapped_model.config, "model_type", None) == "chatglm":
+            values = torch.transpose(values, 0, 1)
+
+        # Split the inputs and rewards into two parts, chosen and rejected
+        batch_size = inputs["input_ids"].size(0) // 2
+        chosen_input_ids, rejected_input_ids = inputs["input_ids"][:batch_size], inputs["input_ids"][batch_size:]
+        chosen_rewards, rejected_rewards = values[:batch_size], values[batch_size:]
+        chosen_scores, rejected_scores = [], []
+
+        # Compute pairwise loss. Only backprop on the different tokens before padding
+        # Inspired by: https://github.com/CarperAI/trlx/blob/main/examples/summarize_rlhf/reward_model/reward_model.py
+        loss = 0
+        for i in range(batch_size):
+            chosen_length = (chosen_input_ids[i] != self.tokenizer.pad_token_id).nonzero()[-1] + 1
+            rejected_length = (rejected_input_ids[i] != self.tokenizer.pad_token_id).nonzero()[-1] + 1
+            check_divergence = (chosen_input_ids[i] != rejected_input_ids[i]).nonzero()
+
+            if len(check_divergence) == 0:
+                end_index = chosen_length
+                div_index = end_index - 1
+            else:
+                end_index = max(chosen_length, rejected_length)
+                div_index = check_divergence[0]
+
+            assert div_index > 0
+            chosen_trunc_rewards = chosen_rewards[i, div_index:end_index]
+            rejected_trunc_rewards = rejected_rewards[i, div_index:end_index]
+            if return_outputs:  # use the score on the last token except pad token for inference
+                chosen_scores.append(chosen_rewards[i, chosen_length - 1])
+                rejected_scores.append(rejected_rewards[i, rejected_length - 1])
+            loss += -torch.nn.functional.logsigmoid(chosen_trunc_rewards - rejected_trunc_rewards).mean()
+
+        loss = loss / batch_size
+        if return_outputs:
+            chosen_scores, rejected_scores = torch.stack(chosen_scores), torch.stack(rejected_scores)
+            return loss, [loss, chosen_scores, rejected_scores]
+
+        return loss
+
+    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}")
+        chosen_scores, rejected_scores = predict_results.predictions
+
+        with open(output_prediction_file, "w", encoding="utf-8") as writer:
+            res: List[str] = []
+            for c_score, r_score in zip(chosen_scores, rejected_scores):
+                res.append(json.dumps({"chosen": round(float(c_score), 2), "rejected": round(float(r_score), 2)}))
+            writer.write("\n".join(res))

src/llamafactory/train/rm/workflow.py

@@ -0,0 +1,75 @@
+# Inspired by: https://github.com/CarperAI/trlx/blob/main/examples/summarize_rlhf/reward_model/train_reward_model_gptj.py
+
+from typing import TYPE_CHECKING, List, Optional
+
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
+from ...extras.callbacks import FixValueHeadModelCallback
+from ...extras.misc import fix_valuehead_checkpoint
+from ...extras.ploting import plot_loss
+from ...model import load_model, load_tokenizer
+from ..utils import create_modelcard_and_push
+from .metric import compute_accuracy
+from .trainer import PairwiseTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments, ModelArguments
+
+
+def run_rm(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer_module = load_tokenizer(model_args)
+    tokenizer = tokenizer_module["tokenizer"]
+    dataset = get_dataset(model_args, data_args, training_args, stage="rm", **tokenizer_module)
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train, add_valuehead=True)
+    data_collator = PairwiseDataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
+
+    # Update arguments
+    training_args.remove_unused_columns = False  # important for pairwise dataset
+
+    # Initialize our Trainer
+    trainer = PairwiseTrainer(
+        model=model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        data_collator=data_collator,
+        callbacks=callbacks + [FixValueHeadModelCallback()],
+        compute_metrics=compute_accuracy,
+        **tokenizer_module,
+        **split_dataset(dataset, data_args, training_args),
+    )
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        trainer.save_model()
+        if training_args.should_save:
+            fix_valuehead_checkpoint(model, training_args.output_dir, training_args.save_safetensors)
+        trainer.log_metrics("train", train_result.metrics)
+        trainer.save_metrics("train", train_result.metrics)
+        trainer.save_state()
+        if trainer.is_world_process_zero() and finetuning_args.plot_loss:
+            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "eval_accuracy"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval")
+        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")
+        trainer.log_metrics("predict", predict_results.metrics)
+        trainer.save_metrics("predict", predict_results.metrics)
+        trainer.save_predictions(predict_results)
+
+    # Create model card
+    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)