trl/trainer/dpo_trainer.py

# DPO Authors: Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn 2023
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import inspect
import random
import warnings
from collections import defaultdict
from copy import deepcopy
from typing import Any, Callable, Dict, List, Literal, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from accelerate.utils import is_deepspeed_available
from datasets import Dataset
from torch.utils.data import DataLoader
from transformers import (
    AutoModelForCausalLM,
    DataCollator,
    PreTrainedModel,
    PreTrainedTokenizerBase,
    Trainer,
    TrainingArguments,
)
from transformers.trainer_callback import TrainerCallback
from transformers.trainer_utils import EvalLoopOutput

from ..import_utils import is_peft_available, is_wandb_available
from ..models import PreTrainedModelWrapper, create_reference_model
from .utils import DPODataCollatorWithPadding, disable_dropout_in_model, pad_to_length


if is_peft_available():
    from peft import PeftModel, get_peft_model, prepare_model_for_kbit_training


if is_wandb_available():
    import wandb

if is_deepspeed_available():
    import deepspeed


class DPOTrainer(Trainer):
    r"""
    Initialize DPOTrainer.

    Args:
        model (`transformers.PreTrainedModel`):
            The model to train, preferably an `AutoModelForSequenceClassification`.
        ref_model (`PreTrainedModelWrapper`):
            Hugging Face transformer model with a casual language modelling head. Used for implicit reward computation and loss. If no
            reference model is provided, the trainer will create a reference model with the same architecture as the model to be optimized.
        beta (`float`, defaults to 0.1):
            The beta factor in DPO loss. Higher beta means less divergence from the initial policy. For the IPO loss, beta is the regularization parameter denoted by tau in the paper.
        label_smoothing (`float`, defaults to 0):
            The robust DPO label smoothing parameter from the [cDPO](https://ericmitchell.ai/cdpo.pdf) report that should be between 0 and 0.5.
        loss_type (`str`, defaults to `"sigmoid"`):
            The type of DPO loss to use. Either `"sigmoid"` the default DPO loss,`"hinge"` loss from [SLiC](https://arxiv.org/abs/2305.10425) paper, `"ipo"` from [IPO](https://arxiv.org/abs/2310.12036) paper, or `"kto"` from the HALOs [report](https://github.com/ContextualAI/HALOs/blob/main/assets/report.pdf).
        args (`transformers.TrainingArguments`):
            The arguments to use for training.
        data_collator (`transformers.DataCollator`):
            The data collator to use for training. If None is specified, the default data collator (`DPODataCollatorWithPadding`) will be used
            which will pad the sequences to the maximum length of the sequences in the batch, given a dataset of paired sequences.
        label_pad_token_id (`int`, defaults to `-100`):
            The label pad token id. This argument is required if you want to use the default data collator.
        padding_value (`int`, defaults to `0`):
            The padding value. This argument is required if you want to use the default data collator.
        truncation_mode (`str`, defaults to `keep_end`):
            The truncation mode to use, either `keep_end` or `keep_start`. This argument is required if you want to use the default data collator.
        train_dataset (`datasets.Dataset`):
            The dataset to use for training.
        eval_dataset (`datasets.Dataset`):
            The dataset to use for evaluation.
        tokenizer (`transformers.PreTrainedTokenizerBase`):
            The tokenizer to use for training. This argument is required if you want to use the default data collator.
        model_init (`Callable[[], transformers.PreTrainedModel]`):
            The model initializer to use for training. If None is specified, the default model initializer will be used.
        callbacks (`List[transformers.TrainerCallback]`):
            The callbacks to use for training.
        optimizers (`Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`):
            The optimizer and scheduler to use for training.
        preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`):
            The function to use to preprocess the logits before computing the metrics.
        max_length (`int`, defaults to `None`):
            The maximum length of the sequences in the batch. This argument is required if you want to use the default data collator.
        max_prompt_length (`int`, defaults to `None`):
            The maximum length of the prompt. This argument is required if you want to use the default data collator.
        max_target_length (`int`, defaults to `None`):
            The maximum length of the target. This argument is required if you want to use the default data collator and your model is an encoder-decoder.
        peft_config (`Dict`, defaults to `None`):
            The PEFT configuration to use for training. If you pass a PEFT configuration, the model will be wrapped in a PEFT model.
        is_encoder_decoder (`Optional[bool]`, `optional`, defaults to `None`):
            If no model is provided, we need to know if the model_init returns an encoder-decoder.
        disable_dropout (`bool`, defaults to `True`):
            Whether or not to disable dropouts in `model` and `ref_model`.
        generate_during_eval (`bool`, defaults to `False`):
            Whether to sample and log generations during evaluation step.
        compute_metrics (`Callable[[EvalPrediction], Dict]`, *optional*):
            The function to use to compute the metrics. Must take a `EvalPrediction` and return
            a dictionary string to metric values.
        model_init_kwargs: (`Optional[Dict]`, *optional*):
            Dict of Optional kwargs to pass when instantiating the model from a string
        ref_model_init_kwargs: (`Optional[Dict]`, *optional*):
            Dict of Optional kwargs to pass when instantiating the ref model from a string

    """

    def __init__(
        self,
        model: Union[PreTrainedModel, nn.Module, str] = None,
        ref_model: Optional[Union[PreTrainedModel, nn.Module, str]] = None,
        beta: float = 0.1,
        label_smoothing: float = 0,
        loss_type: Literal["sigmoid", "hinge", "ipo", "kto"] = "sigmoid",
        args: TrainingArguments = None,
        data_collator: Optional[DataCollator] = None,
        label_pad_token_id: int = -100,
        padding_value: int = 0,
        truncation_mode: str = "keep_end",
        train_dataset: Optional[Dataset] = None,
        eval_dataset: Optional[Union[Dataset, Dict[str, Dataset]]] = None,
        tokenizer: Optional[PreTrainedTokenizerBase] = None,
        model_init: Optional[Callable[[], PreTrainedModel]] = None,
        callbacks: Optional[List[TrainerCallback]] = None,
        optimizers: Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (
            None,
            None,
        ),
        preprocess_logits_for_metrics: Optional[Callable[[torch.Tensor, torch.Tensor], torch.Tensor]] = None,
        max_length: Optional[int] = None,
        max_prompt_length: Optional[int] = None,
        max_target_length: Optional[int] = None,
        peft_config: Optional[Dict] = None,
        is_encoder_decoder: Optional[bool] = None,
        disable_dropout: bool = True,
        generate_during_eval: bool = False,
        compute_metrics: Optional[Callable[[EvalLoopOutput], Dict]] = None,
        model_init_kwargs: Optional[Dict] = None,
        ref_model_init_kwargs: Optional[Dict] = None,
    ):
        if model_init_kwargs is None:
            model_init_kwargs = {}
        elif not isinstance(model, str):
            raise ValueError("You passed model_kwargs to the DPOTrainer. But your model is already instantiated.")

        if ref_model_init_kwargs is None:
            ref_model_init_kwargs = {}
        elif not isinstance(ref_model, str):
            raise ValueError(
                "You passed ref_model_kwargs to the DPOTrainer. But your ref_model is already instantiated."
            )

        if isinstance(model, str):
            warnings.warn(
                "You passed a model_id to the DPOTrainer. This will automatically create an "
                "`AutoModelForCausalLM` or a `PeftModel` (if you passed a `peft_config`) for you."
            )
            model = AutoModelForCausalLM.from_pretrained(model, **model_init_kwargs)

        if isinstance(ref_model, str):
            warnings.warn(
                "You passed a ref model_id to the DPOTrainer. This will automatically create an "
                "`AutoModelForCausalLM`"
            )
            ref_model = AutoModelForCausalLM.from_pretrained(ref_model, **ref_model_init_kwargs)

        if not is_peft_available() and peft_config is not None:
            raise ValueError(
                "PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it to use the PEFT models"
            )
        elif is_peft_available() and peft_config is not None:
            # if model is a peft model and we have a peft_config, we merge and unload it first
            if isinstance(model, PeftModel):
                model = model.merge_and_unload()

            if getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_loaded_in_4bit", False):
                _support_gc_kwargs = hasattr(
                    args, "gradient_checkpointing_kwargs"
                ) and "gradient_checkpointing_kwargs" in list(
                    inspect.signature(prepare_model_for_kbit_training).parameters
                )

                preprare_model_kwargs = {"use_gradient_checkpointing": args.gradient_checkpointing}

                if _support_gc_kwargs:
                    preprare_model_kwargs["gradient_checkpointing_kwargs"] = args.gradient_checkpointing_kwargs

                model = prepare_model_for_kbit_training(model, **preprare_model_kwargs)
            elif getattr(args, "gradient_checkpointing", False):
                # For backward compatibility with older versions of transformers
                if hasattr(model, "enable_input_require_grads"):
                    model.enable_input_require_grads()
                else:

                    def make_inputs_require_grad(module, input, output):
                        output.requires_grad_(True)

                    model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)

            # get peft model with the given config
            model = get_peft_model(model, peft_config)

        # For models that use gradient_checkpoiting, we need to attach a hook that enables input
        # to explicitly have `requires_grad=True`, otherwise training will either silently
        # fail or completely fail.
        elif getattr(args, "gradient_checkpointing", False):
            # For backward compatibility with older versions of transformers
            if hasattr(model, "enable_input_require_grads"):
                model.enable_input_require_grads()
            else:

                def make_inputs_require_grad(module, input, output):
                    output.requires_grad_(True)

                model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)

        if generate_during_eval and not is_wandb_available():
            raise ValueError(
                "`generate_during_eval=True` requires Weights and Biases to be installed."
                " Please install `wandb` to resolve."
            )

        if model is not None:
            self.is_encoder_decoder = model.config.is_encoder_decoder
        elif is_encoder_decoder is None:
            raise ValueError("When no model is provided, you need to pass the parameter is_encoder_decoder.")
        else:
            self.is_encoder_decoder = is_encoder_decoder

        self.is_peft_model = is_peft_available() and isinstance(model, PeftModel)

        if ref_model:
            self.ref_model = ref_model
        elif self.is_peft_model:
            # The `model` with adapters turned off will be used as the reference model
            self.ref_model = None
        else:
            self.ref_model = create_reference_model(model)

        if data_collator is None:
            if tokenizer is None:
                raise ValueError(
                    "max_length or a tokenizer must be specified when using the default DPODataCollatorWithPadding"
                )
            if max_length is None:
                warnings.warn(
                    "When using DPODataCollatorWithPadding, you should set `max_length` in the DPOTrainer's init"
                    " it will be set to `512` by default, but you should do it yourself in the future.",
                    UserWarning,
                )
                max_length = 512
            if max_prompt_length is None:
                warnings.warn(
                    "When using DPODataCollatorWithPadding, you should set `max_prompt_length` in the DPOTrainer's init"
                    " it will be set to `128` by default, but you should do it yourself in the future.",
                    UserWarning,
                )
                max_prompt_length = 128

            if max_target_length is None and self.is_encoder_decoder:
                warnings.warn(
                    "When using DPODataCollatorWithPadding with an encoder decoder architecture, you should set `max_target_length` in the DPOTrainer's init"
                    " it will be set to `128` by default, but you should do it yourself in the future.",
                    UserWarning,
                )
                max_target_length = 128

            data_collator = DPODataCollatorWithPadding(
                tokenizer,
                max_length=max_length,
                max_prompt_length=max_prompt_length,
                label_pad_token_id=label_pad_token_id,
                padding_value=padding_value,
                truncation_mode=truncation_mode,
                is_encoder_decoder=self.is_encoder_decoder,
                max_target_length=max_target_length,
            )

            if args.remove_unused_columns:
                args.remove_unused_columns = False
                # warn users
                warnings.warn(
                    "When using DPODataCollatorWithPadding, you should set `remove_unused_columns=False` in your TrainingArguments"
                    " we have set it for you, but you should do it yourself in the future.",
                    UserWarning,
                )

            self.use_dpo_data_collator = True
        else:
            self.use_dpo_data_collator = False

        if disable_dropout:
            disable_dropout_in_model(model)
            if self.ref_model is not None:
                disable_dropout_in_model(self.ref_model)

        self.max_length = max_length
        self.generate_during_eval = generate_during_eval
        self.label_pad_token_id = label_pad_token_id
        self.padding_value = padding_value

        if loss_type in ["hinge", "ipo", "kto"] and label_smoothing > 0:
            warnings.warn(
                "You are using a loss type that does not support label smoothing. Ignoring label_smoothing parameter."
            )

        self.beta = beta
        self.label_smoothing = label_smoothing
        self.loss_type = loss_type

        self._stored_metrics = defaultdict(lambda: defaultdict(list))

        super().__init__(
            model=model,
            args=args,
            data_collator=data_collator,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            tokenizer=tokenizer,
            model_init=model_init,
            compute_metrics=compute_metrics,
            callbacks=callbacks,
            optimizers=optimizers,
            preprocess_logits_for_metrics=preprocess_logits_for_metrics,
        )

        if not hasattr(self, "accelerator"):
            raise AttributeError(
                "Your `Trainer` does not have an `accelerator` object. Consider upgrading `transformers`."
            )

        if self.ref_model is None:
            if not hasattr(self.accelerator.unwrap_model(self.model), "disable_adapter"):
                raise ValueError(
                    "You are using a `peft` version that does not support `disable_adapter`. Please update your `peft` version to the latest version."
                )
        else:
            if self.is_deepspeed_enabled:
                self.ref_model = self._prepare_deepspeed(self.ref_model)
            else:
                self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)

    def _prepare_deepspeed(self, model: PreTrainedModelWrapper):
        # Adapted from accelerate: https://github.com/huggingface/accelerate/blob/739b135f8367becb67ffaada12fe76e3aa60fefd/src/accelerate/accelerator.py#L1473
        deepspeed_plugin = self.accelerator.state.deepspeed_plugin
        config_kwargs = deepcopy(deepspeed_plugin.deepspeed_config)

        if model is not None:
            if hasattr(model, "config"):
                hidden_size = (
                    max(model.config.hidden_sizes)
                    if getattr(model.config, "hidden_sizes", None)
                    else getattr(model.config, "hidden_size", None)
                )
                if hidden_size is not None and config_kwargs["zero_optimization"]["stage"] == 3:
                    # Note that `stage3_prefetch_bucket_size` can produce DeepSpeed messages like: `Invalidate trace cache @ step 0: expected module 1, but got module 0`
                    # This is expected and is not an error, see: https://github.com/microsoft/DeepSpeed/discussions/4081
                    config_kwargs.update(
                        {
                            "zero_optimization.reduce_bucket_size": hidden_size * hidden_size,
                            "zero_optimization.stage3_param_persistence_threshold": 10 * hidden_size,
                            "zero_optimization.stage3_prefetch_bucket_size": 0.9 * hidden_size * hidden_size,
                        }
                    )

        # If ZeRO-3 is used, we shard both the active and reference model.
        # Otherwise, we assume the reference model fits in memory and is initialized on each device with ZeRO disabled (stage 0)
        if config_kwargs["zero_optimization"]["stage"] != 3:
            config_kwargs["zero_optimization"]["stage"] = 0
        model, *_ = deepspeed.initialize(model=model, config=config_kwargs)
        model.eval()
        return model

    def concatenated_inputs(self, batch: Dict[str, Union[List, torch.LongTensor]]) -> Dict[str, torch.LongTensor]:
        """Concatenate the chosen and rejected inputs into a single tensor.

        Args:
            batch: A batch of data. Must contain the keys 'chosen_input_ids' and 'rejected_input_ids', which are tensors of shape (batch_size, sequence_length).

        Returns:
            A dictionary containing the concatenated inputs under the key 'concatenated_input_ids'.
        """
        concatenated_batch = {}

        if self.is_encoder_decoder:
            max_length = max(batch["chosen_labels"].shape[1], batch["rejected_labels"].shape[1])
        else:
            max_length = max(batch["chosen_input_ids"].shape[1], batch["rejected_input_ids"].shape[1])

        for k in batch:
            if k.startswith("chosen") and isinstance(batch[k], torch.Tensor):
                pad_value = self.label_pad_token_id if "labels" in k or self.is_encoder_decoder else self.padding_value
                concatenated_key = k.replace("chosen", "concatenated")
                concatenated_batch[concatenated_key] = pad_to_length(batch[k], max_length, pad_value=pad_value)
        for k in batch:
            if k.startswith("rejected") and isinstance(batch[k], torch.Tensor):
                pad_value = self.label_pad_token_id if "labels" in k or self.is_encoder_decoder else self.padding_value
                concatenated_key = k.replace("rejected", "concatenated")
                concatenated_batch[concatenated_key] = torch.cat(
                    (
                        concatenated_batch[concatenated_key],
                        pad_to_length(batch[k], max_length, pad_value=pad_value),
                    ),
                    dim=0,
                ).to(self.accelerator.device)

        if self.is_encoder_decoder:
            concatenated_batch["concatenated_input_ids"] = batch["prompt_input_ids"].repeat(2, 1)
            concatenated_batch["concatenated_attention_mask"] = batch["prompt_attention_mask"].repeat(2, 1)

        return concatenated_batch

    def dpo_loss(
        self,
        policy_chosen_logps: torch.FloatTensor,
        policy_rejected_logps: torch.FloatTensor,
        reference_chosen_logps: torch.FloatTensor,
        reference_rejected_logps: torch.FloatTensor,
        reference_free: bool = False,
    ) -> Tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
        """Compute the DPO loss for a batch of policy and reference model log probabilities.

        Args:
            policy_chosen_logps: Log probabilities of the policy model for the chosen responses. Shape: (batch_size,)
            policy_rejected_logps: Log probabilities of the policy model for the rejected responses. Shape: (batch_size,)
            reference_chosen_logps: Log probabilities of the reference model for the chosen responses. Shape: (batch_size,)
            reference_rejected_logps: Log probabilities of the reference model for the rejected responses. Shape: (batch_size,)
            reference_free: If True, we ignore the _provided_ reference model and implicitly use a reference model that assigns equal probability to all responses.

        Returns:
            A tuple of three tensors: (losses, chosen_rewards, rejected_rewards).
            The losses tensor contains the DPO loss for each example in the batch.
            The chosen_rewards and rejected_rewards tensors contain the rewards for the chosen and rejected responses, respectively.
        """
        pi_logratios = policy_chosen_logps - policy_rejected_logps
        if reference_free:
            ref_logratios = 0
        else:
            ref_logratios = reference_chosen_logps - reference_rejected_logps

        logits = pi_logratios - ref_logratios

        # The beta is a temperature parameter for the DPO loss, typically something in the range of 0.1 to 0.5.
        # We ignore the reference model as beta -> 0. The label_smoothing parameter encodes our uncertainty about the labels and
        # calculates a conservative DPO loss.
        if self.loss_type == "sigmoid":
            losses = (
                -F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
                - F.logsigmoid(-self.beta * logits) * self.label_smoothing
            )
        elif self.loss_type == "hinge":
            losses = torch.relu(1 - self.beta * logits)
        elif self.loss_type == "ipo":
            # eqn (17) of the paper where beta is the regularization parameter for the IPO loss, denoted by tau in the paper.
            losses = (logits - 1 / (2 * self.beta)) ** 2
        elif self.loss_type == "kto":
            # eqn (7) of the HALOs paper
            chosen_KL = (policy_chosen_logps - reference_chosen_logps).mean().clamp(min=0)
            rejected_KL = (policy_rejected_logps - reference_rejected_logps).mean().clamp(min=0)

            chosen_logratios = policy_chosen_logps - reference_chosen_logps
            rejected_logratios = policy_rejected_logps - reference_rejected_logps
            # As described in the KTO report, the KL term for chosen (rejected) is estimated using the rejected (chosen) half.
            losses = torch.cat(
                (
                    1 - F.sigmoid(self.beta * (chosen_logratios - rejected_KL)),
                    1 - F.sigmoid(self.beta * (chosen_KL - rejected_logratios)),
                ),
                0,
            )
        else:
            raise ValueError(
                f"Unknown loss type: {self.loss_type}. Should be one of ['sigmoid', 'hinge', 'ipo', 'kto']"
            )

        chosen_rewards = self.beta * (policy_chosen_logps - reference_chosen_logps).detach()
        rejected_rewards = self.beta * (policy_rejected_logps - reference_rejected_logps).detach()

        return losses, chosen_rewards, rejected_rewards

    def _get_batch_logps(
        self,
        logits: torch.FloatTensor,
        labels: torch.LongTensor,
        average_log_prob: bool = False,
    ) -> torch.FloatTensor:
        """Compute the log probabilities of the given labels under the given logits.

        Args:
            logits: Logits of the model (unnormalized). Shape: (batch_size, sequence_length, vocab_size)
            labels: Labels for which to compute the log probabilities. Label tokens with a value of label_pad_token_id are ignored. Shape: (batch_size, sequence_length)
            average_log_prob: If True, return the average log probability per (non-masked) token. Otherwise, return the sum of the log probabilities of the (non-masked) tokens.

        Returns:
            A tensor of shape (batch_size,) containing the average/sum log probabilities of the given labels under the given logits.
        """
        if logits.shape[:-1] != labels.shape:
            raise ValueError("Logits (batch and sequence length dim) and labels must have the same shape.")

        if not self.is_encoder_decoder:
            labels = labels[:, 1:].clone()
            logits = logits[:, :-1, :]
        loss_mask = labels != self.label_pad_token_id

        # dummy token; we'll ignore the losses on these tokens later
        labels[labels == self.label_pad_token_id] = 0

        per_token_logps = torch.gather(logits.log_softmax(-1), dim=2, index=labels.unsqueeze(2)).squeeze(2)

        if average_log_prob:
            return (per_token_logps * loss_mask).sum(-1) / loss_mask.sum(-1)
        else:
            return (per_token_logps * loss_mask).sum(-1)

    def concatenated_forward(
        self, model: nn.Module, batch: Dict[str, Union[List, torch.LongTensor]]
    ) -> Tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
        """Run the given model on the given batch of inputs, concatenating the chosen and rejected inputs together.

        We do this to avoid doing two forward passes, because it's faster for FSDP.
        """
        concatenated_batch = self.concatenated_inputs(batch)
        len_chosen = batch["chosen_labels"].shape[0]

        model_kwargs = (
            {
                "labels": concatenated_batch["concatenated_labels"],
                "decoder_input_ids": concatenated_batch.pop("concatenated_decoder_input_ids", None),
            }
            if self.is_encoder_decoder
            else {}
        )
        all_logits = model(
            concatenated_batch["concatenated_input_ids"],
            attention_mask=concatenated_batch["concatenated_attention_mask"],
            **model_kwargs,
        ).logits.to(torch.float32)

        all_logps = self._get_batch_logps(
            all_logits,
            concatenated_batch["concatenated_labels"],
            average_log_prob=False,
        )

        chosen_logps = all_logps[:len_chosen]
        rejected_logps = all_logps[len_chosen:]

        chosen_logits = all_logits[:len_chosen]
        rejected_logits = all_logits[len_chosen:]

        return (chosen_logps, rejected_logps, chosen_logits, rejected_logits)

    def get_batch_metrics(
        self,
        model,
        batch: Dict[str, Union[List, torch.LongTensor]],
        train_eval: Literal["train", "eval"] = "train",
    ):
        """Compute the DPO loss and other metrics for the given batch of inputs for train or test."""
        metrics = {}

        (
            policy_chosen_logps,
            policy_rejected_logps,
            policy_chosen_logits,
            policy_rejected_logits,
        ) = self.concatenated_forward(model, batch)
        with torch.no_grad():
            if self.ref_model is None:
                with self.accelerator.unwrap_model(self.model).disable_adapter():
                    (
                        reference_chosen_logps,
                        reference_rejected_logps,
                        _,
                        _,
                    ) = self.concatenated_forward(self.model, batch)
            else:
                (
                    reference_chosen_logps,
                    reference_rejected_logps,
                    _,
                    _,
                ) = self.concatenated_forward(self.ref_model, batch)

        losses, chosen_rewards, rejected_rewards = self.dpo_loss(
            policy_chosen_logps,
            policy_rejected_logps,
            reference_chosen_logps,
            reference_rejected_logps,
        )
        reward_accuracies = (chosen_rewards > rejected_rewards).float()

        prefix = "eval_" if train_eval == "eval" else ""
        metrics[f"{prefix}rewards/chosen"] = chosen_rewards.cpu().mean()
        metrics[f"{prefix}rewards/rejected"] = rejected_rewards.cpu().mean()
        metrics[f"{prefix}rewards/accuracies"] = reward_accuracies.cpu().mean()
        metrics[f"{prefix}rewards/margins"] = (chosen_rewards - rejected_rewards).cpu().mean()
        metrics[f"{prefix}logps/rejected"] = policy_rejected_logps.detach().cpu().mean()
        metrics[f"{prefix}logps/chosen"] = policy_chosen_logps.detach().cpu().mean()
        metrics[f"{prefix}logits/rejected"] = policy_rejected_logits.detach().cpu().mean()
        metrics[f"{prefix}logits/chosen"] = policy_chosen_logits.detach().cpu().mean()

        return losses.mean(), metrics

    def compute_loss(
        self,
        model: Union[PreTrainedModel, nn.Module],
        inputs: Dict[str, Union[torch.Tensor, Any]],
        return_outputs=False,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, Dict[str, torch.Tensor]]]:
        if not self.use_dpo_data_collator:
            warnings.warn(
                "compute_loss is only implemented for DPODataCollatorWithPadding, and you passed a datacollator that is different than "
                "DPODataCollatorWithPadding - you might see unexpected behavior. Alternatively, you can implement your own prediction_step method if you are using a custom data collator"
            )
        loss, metrics = self.get_batch_metrics(model, inputs, train_eval="train")

        # force log the metrics
        if self.accelerator.is_main_process:
            self.store_metrics(metrics, train_eval="train")

        if return_outputs:
            return (loss, metrics)
        return loss

    def get_batch_samples(self, model, batch: Dict[str, torch.LongTensor]) -> Tuple[str, str]:
        """Generate samples from the model and reference model for the given batch of inputs."""

        policy_output = model.generate(
            input_ids=batch["prompt_input_ids"],
            attention_mask=batch["prompt_attention_mask"],
            max_length=self.max_length,
            do_sample=True,
            pad_token_id=self.tokenizer.pad_token_id,
        )

        if self.ref_model is None:
            with self.accelerator.unwrap_model(self.model).disable_adapter():
                reference_output = self.model.generate(
                    input_ids=batch["prompt_input_ids"],
                    attention_mask=batch["prompt_attention_mask"],
                    max_length=self.max_length,
                    do_sample=True,
                    pad_token_id=self.tokenizer.pad_token_id,
                )
        else:
            reference_output = self.ref_model.generate(
                input_ids=batch["prompt_input_ids"],
                attention_mask=batch["prompt_attention_mask"],
                max_length=self.max_length,
                do_sample=True,
                pad_token_id=self.tokenizer.pad_token_id,
            )

        policy_output = pad_to_length(policy_output, self.max_length, self.tokenizer.pad_token_id)
        policy_output_decoded = self.tokenizer.batch_decode(policy_output, skip_special_tokens=True)

        reference_output = pad_to_length(reference_output, self.max_length, self.tokenizer.pad_token_id)
        reference_output_decoded = self.tokenizer.batch_decode(reference_output, skip_special_tokens=True)

        return policy_output_decoded, reference_output_decoded

    def prediction_step(
        self,
        model: Union[PreTrainedModel, nn.Module],
        inputs: Dict[str, Union[torch.Tensor, Any]],
        prediction_loss_only: bool,
        ignore_keys: Optional[List[str]] = None,
    ):
        if not self.use_dpo_data_collator:
            warnings.warn(
                "prediction_step is only implemented for DPODataCollatorWithPadding, and you passed a datacollator that is different than "
                "DPODataCollatorWithPadding - you might see unexpected behavior. Alternatively, you can implement your own prediction_step method if you are using a custom data collator"
            )
        if ignore_keys is None:
            if hasattr(model, "config"):
                ignore_keys = getattr(model.config, "keys_to_ignore_at_inference", [])
            else:
                ignore_keys = []

        with torch.no_grad():
            loss, metrics = self.get_batch_metrics(model, inputs, train_eval="eval")

        # force log the metrics
        if self.accelerator.is_main_process:
            self.store_metrics(metrics, train_eval="eval")

        if prediction_loss_only:
            return (loss.detach(), None, None)

        # logits for the chosen and rejected samples from model
        logits_dict = {
            "eval_logits/chosen": metrics["eval_logits/chosen"],
            "eval_logits/rejected": metrics["eval_logits/rejected"],
        }
        logits = tuple(v.unsqueeze(dim=0) for k, v in logits_dict.items() if k not in ignore_keys)
        logits = torch.stack(logits).mean(axis=1).to(self.accelerator.device)
        labels = torch.zeros(logits.shape[0], device=self.accelerator.device)

        return (loss.detach(), logits, labels)

    def store_metrics(self, metrics: Dict[str, float], train_eval: Literal["train", "eval"] = "train") -> None:
        for key, value in metrics.items():
            self._stored_metrics[train_eval][key].append(value)

    def evaluation_loop(
        self,
        dataloader: DataLoader,
        description: str,
        prediction_loss_only: Optional[bool] = None,
        ignore_keys: Optional[List[str]] = None,
        metric_key_prefix: str = "eval",
    ) -> EvalLoopOutput:
        """
        Overriding built-in evaluation loop to store metrics for each batch.
        Prediction/evaluation loop, shared by `Trainer.evaluate()` and `Trainer.predict()`.

        Works both with or without labels.
        """

        # Sample and save to game log if requested (for one batch to save time)
        if self.generate_during_eval:
            # Generate random indices within the range of the total number of samples
            num_samples = len(dataloader.dataset)
            random_indices = random.sample(range(num_samples), k=self.args.eval_batch_size)

            # Use dataloader.dataset.select to get the random batch without iterating over the DataLoader
            random_batch_dataset = dataloader.dataset.select(random_indices)
            random_batch = self.data_collator(random_batch_dataset)
            random_batch = self._prepare_inputs(random_batch)

            policy_output_decoded, ref_output_decoded = self.get_batch_samples(self.model, random_batch)

            self.log(
                {
                    "game_log": wandb.Table(
                        columns=["Prompt", "Policy", "Ref Model"],
                        rows=[
                            [prompt, pol[len(prompt) :], ref[len(prompt) :]]
                            for prompt, pol, ref in zip(
                                random_batch["prompt"], policy_output_decoded, ref_output_decoded
                            )
                        ],
                    )
                }
            )
            self.state.log_history.pop()

        # Base evaluation
        initial_output = super().evaluation_loop(
            dataloader, description, prediction_loss_only, ignore_keys, metric_key_prefix
        )

        return initial_output

    def log(self, logs: Dict[str, float]) -> None:
        """
        Log `logs` on the various objects watching training, including stored metrics.

        Args:
            logs (`Dict[str, float]`):
                The values to log.
        """
        # logs either has 'loss' or 'eval_loss'
        train_eval = "train" if "loss" in logs else "eval"
        # Add averaged stored metrics to logs
        for key, metrics in self._stored_metrics[train_eval].items():
            logs[key] = torch.tensor(metrics).mean().item()
        del self._stored_metrics[train_eval]
        return super().log(logs)