unsloth_compiled_cache/UnslothRewardTrainer.py

"""
2025.3.17
2025.3.19
4.50.3
0.15.2
__UNSLOTH_VERSIONING__
"""
from torch import Tensor
import torch
import torch.nn as nn
from torch.nn import functional as F
from trl.trainer.reward_trainer import (Any, BaseImageProcessor, Callable, DataCollator, Dataset, EvalPrediction, FeatureExtractionMixin, FrozenInstanceError, Optional, PartialState, PeftModel, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, RewardConfig, RewardDataCollatorWithPadding, RewardTrainer, Trainer, TrainerCallback, Union, _tokenize, compute_accuracy, decode_and_strip_padding, defaultdict, disable_dropout_in_model, gather_object, generate_model_card, get_comet_experiment_url, inspect, is_peft_available, is_wandb_available, log_table_to_comet_experiment, maybe_apply_chat_template, nested_detach, nn, os, pd, prepare_model_for_kbit_training, print_rich_table, replace, torch, wandb, warnings)


import os
from typing import *
from dataclasses import dataclass, field
from packaging.version import Version
import torch
import numpy as np
from contextlib import nullcontext
from torch.nn import functional as F
from transformers import DataCollatorForSeq2Seq, DataCollatorForLanguageModeling

torch_compile_options = {
    "epilogue_fusion"   : True,
    "max_autotune"      : False,
    "shape_padding"     : True,
    "trace.enabled"     : False,
    "triton.cudagraphs" : False,
}

@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
def selective_log_softmax(logits, index):
    logits = logits.to(torch.float32)
    selected_logits = torch.gather(logits, dim = -1, index = index.unsqueeze(-1)).squeeze(-1)
    # loop to reduce peak mem consumption
    # logsumexp_values = torch.stack([torch.logsumexp(lg, dim=-1) for lg in logits])
    logsumexp_values = torch.logsumexp(logits, dim = -1)
    per_token_logps = selected_logits - logsumexp_values  # log_softmax(x_i) = x_i - logsumexp(x)
    return per_token_logps
@dataclass
class UnslothRewardConfig(RewardConfig):
    """
    
    Configuration class for the [`RewardTrainer`].

    Using [`~transformers.HfArgumentParser`] we can turn this class into
    [argparse](https://docs.python.org/3/library/argparse#module-argparse) arguments that can be specified on the
    command line.

    Parameters:
        max_length (`int` or `None`, *optional*, defaults to `1024`):
            Maximum length of the sequences (prompt + completion) in the batch, filters out entries that exceed the
            limit. This argument is required if you want to use the default data collator.
        disable_dropout (`bool`, *optional*, defaults to `True`):
            Whether to disable dropout in the model.
        dataset_num_proc (`int`, *optional*, defaults to `None`):
            Number of processes to use for processing the dataset.
        center_rewards_coefficient (`float`, *optional*, defaults to `None`):
            Coefficient to incentivize the reward model to output mean-zero rewards (proposed by
            https://huggingface.co/papers/2312.09244, Eq. 2). Recommended value: `0.01`.
        remove_unused_columns (`bool`, *optional*, defaults to `False`):
            Whether to remove the columns that are not used by the model's forward pass. Can be `True` only if
            the dataset is pretokenized.
    
    """
    vllm_sampling_params: Optional[Any] = field(
        default = None,
        metadata = {'help': 'vLLM SamplingParams'},
    )
    unsloth_num_chunks : Optional[int] = field(
        default = -1,
        metadata = {'help': 'Chunk size to reduce memory usage. -1 is most efficient.'},
    )
    def __init__(
        self,
        output_dir = None,
        overwrite_output_dir = None,
        do_train = False,
        do_eval = False,
        do_predict = False,
        eval_strategy = 'no',
        prediction_loss_only = False,
        per_device_train_batch_size = 4,
        per_device_eval_batch_size = 4,
        per_gpu_train_batch_size = None,
        per_gpu_eval_batch_size = None,
        gradient_accumulation_steps = 2,
        eval_accumulation_steps = 2,
        eval_delay = 0,
        torch_empty_cache_steps = 250,
        learning_rate = 5e-05,
        weight_decay = 0.01,
        adam_beta1 = 0.9,
        adam_beta2 = 0.999,
        adam_epsilon = 1e-08,
        max_grad_norm = 1.0,
        num_train_epochs = 3.0,
        max_steps = -1,
        lr_scheduler_type = 'linear',
        warmup_ratio = 0.1,
        warmup_steps = 0,
        log_level = 'passive',
        log_level_replica = 'warning',
        log_on_each_node = True,
        logging_dir = None,
        logging_strategy = 'steps',
        logging_first_step = False,
        logging_steps = 1,
        logging_nan_inf_filter = False,
        save_strategy = 'steps',
        save_steps = 500,
        save_total_limit = None,
        save_safetensors = True,
        save_on_each_node = False,
        save_only_model = False,
        restore_callback_states_from_checkpoint = False,
        no_cuda = False,
        use_cpu = False,
        use_mps_device = False,
        seed = 3407,
        data_seed = 3407,
        jit_mode_eval = False,
        use_ipex = False,
        bf16 = False,
        fp16 = False,
        fp16_opt_level = 'O1',
        half_precision_backend = 'auto',
        bf16_full_eval = False,
        fp16_full_eval = False,
        tf32 = None,
        local_rank = -1,
        ddp_backend = None,
        tpu_num_cores = None,
        tpu_metrics_debug = False,
        debug = '',
        dataloader_drop_last = False,
        eval_steps = None,
        dataloader_num_workers = 0,
        dataloader_prefetch_factor = None,
        past_index = -1,
        run_name = None,
        disable_tqdm = None,
        remove_unused_columns = False,
        label_names = None,
        load_best_model_at_end = False,
        metric_for_best_model = None,
        greater_is_better = None,
        ignore_data_skip = False,
        fsdp = '',
        fsdp_min_num_params = 0,
        fsdp_config = None,
        tp_size = 0,
        fsdp_transformer_layer_cls_to_wrap = None,
        accelerator_config = None,
        deepspeed = None,
        label_smoothing_factor = 0.0,
        optim = 'adamw_8bit',
        optim_args = None,
        adafactor = False,
        group_by_length = False,
        length_column_name = 'length',
        report_to = None,
        ddp_find_unused_parameters = None,
        ddp_bucket_cap_mb = None,
        ddp_broadcast_buffers = None,
        dataloader_pin_memory = True,
        dataloader_persistent_workers = False,
        skip_memory_metrics = True,
        use_legacy_prediction_loop = False,
        push_to_hub = False,
        resume_from_checkpoint = None,
        hub_model_id = None,
        hub_strategy = 'every_save',
        hub_token = None,
        hub_private_repo = None,
        hub_always_push = False,
        gradient_checkpointing = False,
        gradient_checkpointing_kwargs = None,
        include_inputs_for_metrics = False,
        eval_do_concat_batches = True,
        fp16_backend = 'auto',
        evaluation_strategy = None,
        push_to_hub_model_id = None,
        push_to_hub_organization = None,
        push_to_hub_token = None,
        mp_parameters = '',
        auto_find_batch_size = False,
        full_determinism = False,
        torchdynamo = None,
        ray_scope = 'last',
        ddp_timeout = 1800,
        torch_compile = False,
        torch_compile_backend = None,
        torch_compile_mode = None,
        dispatch_batches = None,
        split_batches = None,
        include_tokens_per_second = False,
        include_num_input_tokens_seen = False,
        neftune_noise_alpha = None,
        optim_target_modules = None,
        batch_eval_metrics = False,
        eval_on_start = False,
        use_liger_kernel = False,
        eval_use_gather_object = False,
        average_tokens_across_devices = False,
        max_length = 1024,
        disable_dropout = True,
        dataset_num_proc = None,
        center_rewards_coefficient = None,
        vllm_sampling_params = None,
        unsloth_num_chunks = -1,
        **kwargs,
    ):
        if learning_rate < 1e-7: raise FloatingPointError(f'Unsloth: Your learning rate of `{learning_rate}` is too small and less than 1e-7! Consider increasing it, otherwise gradient updates will be close to 0!')
        if learning_rate > 1: raise OverflowError(f'Unsloth: Your learning rate of `{learning_rate}` is way too larger > 1! Consider decreasing it to 1e-1, otherwise gradient updates will explode!')
        if output_dir is None and save_strategy == 'steps' and save_steps == 500:
            output_dir = 'unsloth_training_checkpoints'
            save_strategy = 'no'
        if dataset_num_proc is None:
            from multiprocessing import cpu_count
            dataset_num_proc = cpu_count()
        
        super().__init__(
            output_dir = output_dir,
            overwrite_output_dir = overwrite_output_dir,
            do_train = do_train,
            do_eval = do_eval,
            do_predict = do_predict,
            eval_strategy = eval_strategy,
            prediction_loss_only = prediction_loss_only,
            per_device_train_batch_size = per_device_train_batch_size,
            per_device_eval_batch_size = per_device_eval_batch_size,
            per_gpu_train_batch_size = per_gpu_train_batch_size,
            per_gpu_eval_batch_size = per_gpu_eval_batch_size,
            gradient_accumulation_steps = gradient_accumulation_steps,
            eval_accumulation_steps = eval_accumulation_steps,
            eval_delay = eval_delay,
            torch_empty_cache_steps = torch_empty_cache_steps,
            learning_rate = learning_rate,
            weight_decay = weight_decay,
            adam_beta1 = adam_beta1,
            adam_beta2 = adam_beta2,
            adam_epsilon = adam_epsilon,
            max_grad_norm = max_grad_norm,
            num_train_epochs = num_train_epochs,
            max_steps = max_steps,
            lr_scheduler_type = lr_scheduler_type,
            warmup_ratio = warmup_ratio,
            warmup_steps = warmup_steps,
            log_level = log_level,
            log_level_replica = log_level_replica,
            log_on_each_node = log_on_each_node,
            logging_dir = logging_dir,
            logging_strategy = logging_strategy,
            logging_first_step = logging_first_step,
            logging_steps = logging_steps,
            logging_nan_inf_filter = logging_nan_inf_filter,
            save_strategy = save_strategy,
            save_steps = save_steps,
            save_total_limit = save_total_limit,
            save_safetensors = save_safetensors,
            save_on_each_node = save_on_each_node,
            save_only_model = save_only_model,
            restore_callback_states_from_checkpoint = restore_callback_states_from_checkpoint,
            no_cuda = no_cuda,
            use_cpu = use_cpu,
            use_mps_device = use_mps_device,
            seed = seed,
            data_seed = data_seed,
            jit_mode_eval = jit_mode_eval,
            use_ipex = use_ipex,
            bf16 = bf16,
            fp16 = fp16,
            fp16_opt_level = fp16_opt_level,
            half_precision_backend = half_precision_backend,
            bf16_full_eval = bf16_full_eval,
            fp16_full_eval = fp16_full_eval,
            tf32 = tf32,
            local_rank = local_rank,
            ddp_backend = ddp_backend,
            tpu_num_cores = tpu_num_cores,
            tpu_metrics_debug = tpu_metrics_debug,
            debug = debug,
            dataloader_drop_last = dataloader_drop_last,
            eval_steps = eval_steps,
            dataloader_num_workers = dataloader_num_workers,
            dataloader_prefetch_factor = dataloader_prefetch_factor,
            past_index = past_index,
            run_name = run_name,
            disable_tqdm = disable_tqdm,
            remove_unused_columns = remove_unused_columns,
            label_names = label_names,
            load_best_model_at_end = load_best_model_at_end,
            metric_for_best_model = metric_for_best_model,
            greater_is_better = greater_is_better,
            ignore_data_skip = ignore_data_skip,
            fsdp = fsdp,
            fsdp_min_num_params = fsdp_min_num_params,
            fsdp_config = fsdp_config,
            tp_size = tp_size,
            fsdp_transformer_layer_cls_to_wrap = fsdp_transformer_layer_cls_to_wrap,
            accelerator_config = accelerator_config,
            deepspeed = deepspeed,
            label_smoothing_factor = label_smoothing_factor,
            optim = optim,
            optim_args = optim_args,
            adafactor = adafactor,
            group_by_length = group_by_length,
            length_column_name = length_column_name,
            report_to = report_to,
            ddp_find_unused_parameters = ddp_find_unused_parameters,
            ddp_bucket_cap_mb = ddp_bucket_cap_mb,
            ddp_broadcast_buffers = ddp_broadcast_buffers,
            dataloader_pin_memory = dataloader_pin_memory,
            dataloader_persistent_workers = dataloader_persistent_workers,
            skip_memory_metrics = skip_memory_metrics,
            use_legacy_prediction_loop = use_legacy_prediction_loop,
            push_to_hub = push_to_hub,
            resume_from_checkpoint = resume_from_checkpoint,
            hub_model_id = hub_model_id,
            hub_strategy = hub_strategy,
            hub_token = hub_token,
            hub_private_repo = hub_private_repo,
            hub_always_push = hub_always_push,
            gradient_checkpointing = gradient_checkpointing,
            gradient_checkpointing_kwargs = gradient_checkpointing_kwargs,
            include_inputs_for_metrics = include_inputs_for_metrics,
            eval_do_concat_batches = eval_do_concat_batches,
            fp16_backend = fp16_backend,
            evaluation_strategy = evaluation_strategy,
            push_to_hub_model_id = push_to_hub_model_id,
            push_to_hub_organization = push_to_hub_organization,
            push_to_hub_token = push_to_hub_token,
            mp_parameters = mp_parameters,
            auto_find_batch_size = auto_find_batch_size,
            full_determinism = full_determinism,
            torchdynamo = torchdynamo,
            ray_scope = ray_scope,
            ddp_timeout = ddp_timeout,
            torch_compile = torch_compile,
            torch_compile_backend = torch_compile_backend,
            torch_compile_mode = torch_compile_mode,
            dispatch_batches = dispatch_batches,
            split_batches = split_batches,
            include_tokens_per_second = include_tokens_per_second,
            include_num_input_tokens_seen = include_num_input_tokens_seen,
            neftune_noise_alpha = neftune_noise_alpha,
            optim_target_modules = optim_target_modules,
            batch_eval_metrics = batch_eval_metrics,
            eval_on_start = eval_on_start,
            use_liger_kernel = use_liger_kernel,
            eval_use_gather_object = eval_use_gather_object,
            average_tokens_across_devices = average_tokens_across_devices,
            max_length = max_length,
            disable_dropout = disable_dropout,
            dataset_num_proc = dataset_num_proc,
            center_rewards_coefficient = center_rewards_coefficient,**kwargs)
        self.vllm_sampling_params = vllm_sampling_params
        self.unsloth_num_chunks = unsloth_num_chunks
pass

class _UnslothRewardTrainer(Trainer):
    _tag_names = ["trl", "reward-trainer"]

    def __init__(
        self,
        model: Optional[Union[PreTrainedModel, nn.Module]] = None,
        args: Optional[RewardConfig] = None,
        data_collator: Optional[DataCollator] = None,
        train_dataset: Optional[Dataset] = None,
        eval_dataset: Optional[Union[Dataset, dict[str, Dataset]]] = None,
        processing_class: Optional[
            Union[PreTrainedTokenizerBase, BaseImageProcessor, FeatureExtractionMixin, ProcessorMixin]
        ] = None,
        model_init: Optional[Callable[[], PreTrainedModel]] = None,
        compute_metrics: Optional[Callable[[EvalPrediction], dict]] = 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,
        peft_config: Optional[dict] = None,
    ):
        """
        Initialize RewardTrainer.

        Args:
            model (`transformers.PreTrainedModel`):
                The model to train, preferably an `AutoModelForSequenceClassification`.
            args (`RewardConfig`):
                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 (`RewardDataCollatorWithPadding`) will be used
                which will pad the sequences to the maximum length of the sequences in the batch, given a dataset of paired sequences.
            train_dataset (`datasets.Dataset`):
                The dataset to use for training.
            eval_dataset (`datasets.Dataset`):
                The dataset to use for evaluation.
            processing_class (`PreTrainedTokenizerBase` or `BaseImageProcessor` or `FeatureExtractionMixin` or `ProcessorMixin`, *optional*):
                Processing class used to process the data. If provided, will be used to automatically process the inputs
                for the model, and it will be saved along the model to make it easier to rerun an interrupted training or
                reuse the fine-tuned model.
            model_init (`Callable[[], transformers.PreTrainedModel]`):
                The model initializer to use for training. If None is specified, the default model initializer will be used.
            compute_metrics (`Callable[[transformers.EvalPrediction], dict]`, *optional* defaults to `compute_accuracy`):
                The metrics to use for evaluation. If no metrics are specified, the default metric (`compute_accuracy`) 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.
            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.
        """
        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 not isinstance(model, PeftModel):
                if getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_quantized", False):
                    _supports_gc_kwargs = "gradient_checkpointing_kwargs" in list(
                        inspect.signature(prepare_model_for_kbit_training).parameters
                    )

                    prepare_model_kwargs = {"use_gradient_checkpointing": args.gradient_checkpointing}

                    if not _supports_gc_kwargs and args.gradient_checkpointing_kwargs is not None:
                        warnings.warn(
                            "You passed `gradient_checkpointing_kwargs` in the trainer's kwargs, but your peft version does not support it. "
                            "please update to the latest version of peft to use `gradient_checkpointing_kwargs`.",
                            UserWarning,
                        )
                    elif _supports_gc_kwargs and args.gradient_checkpointing_kwargs is not None:
                        prepare_model_kwargs["gradient_checkpointing_kwargs"] = args.gradient_checkpointing_kwargs

                    model = prepare_model_for_kbit_training(model, **prepare_model_kwargs)

                model = model

        # Disable dropout in the model
        if args.disable_dropout:
            disable_dropout_in_model(model)

        if compute_metrics is None:
            compute_metrics = compute_accuracy

        if data_collator is None:
            if processing_class is None:
                raise ValueError(
                    "A processing_class must be specified when using the default RewardDataCollatorWithPadding"
                )

            max_length = args.max_length

            data_collator = RewardDataCollatorWithPadding(processing_class)

            if args.remove_unused_columns:
                try:  # for bc before https://github.com/huggingface/transformers/pull/25435
                    args.remove_unused_columns = False
                except FrozenInstanceError:
                    args = replace(args, remove_unused_columns=False)
                # warn users
                warnings.warn(
                    "When using RewardDataCollatorWithPadding, you should set `remove_unused_columns=False` in your RewardConfig"
                    " we have set it for you, but you should do it yourself in the future.",
                    UserWarning,
                )

            self.use_reward_data_collator = True
        else:
            self.use_reward_data_collator = False

        # The trainer estimates the number of FLOPs (floating-point operations) using the number of elements in the
        # input tensor associated with the key "input_ids". However, in Reward, the sampled data does not include the
        # "input_ids" key. Instead, the available keys are "input_ids_chosen" and "input_ids_rejected". As a result,
        # the trainer issues the warning: "Could not estimate the number of tokens of the input, floating-point
        # operations will not be computed." To suppress this warning, we set the "estimate_tokens" key in the model's
        # "warnings_issued" dictionary to True. This acts as a flag to indicate that the warning has already been
        # issued.
        model.warnings_issued["estimate_tokens"] = True

        if "input_ids_chosen" not in train_dataset.column_names:
            with PartialState().local_main_process_first():
                fn_kwargs = {"tokenizer": processing_class}
                train_dataset = train_dataset.map(maybe_apply_chat_template, fn_kwargs={"tokenizer": processing_class})
                train_dataset = train_dataset.map(
                    _tokenize,
                    batched=True,
                    fn_kwargs=fn_kwargs,
                    num_proc=args.dataset_num_proc,
                )
                # This filter is important because otherwise you get samples that exceed the model's context length and
                # get truncated => noisy signal the chosen/rejected label gets lost. The downside is that the
                # user might get surprised if N samples are missing from training.
                train_dataset = train_dataset.filter(
                    lambda x: len(x["input_ids_chosen"]) <= max_length and len(x["input_ids_rejected"]) <= max_length,
                    num_proc=args.dataset_num_proc,
                )
                if eval_dataset is not None:
                    eval_dataset = eval_dataset.map(
                        maybe_apply_chat_template, fn_kwargs={"tokenizer": processing_class}
                    )
                    eval_dataset = eval_dataset.map(
                        _tokenize,
                        fn_kwargs=fn_kwargs,
                        batched=True,
                        num_proc=args.dataset_num_proc,
                    )
                    # This filter is important because otherwise you get samples that exceed the model's context length and
                    # get truncated => noisy signal the chosen/rejected label gets lost. The downside is that the
                    # user might get surprised if N samples are missing from training.
                    eval_dataset = eval_dataset.filter(
                        lambda x: len(x["input_ids_chosen"]) <= max_length
                        and len(x["input_ids_rejected"]) <= max_length,
                        num_proc=args.dataset_num_proc,
                    )

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

        # Add tags for models that have been loaded with the correct transformers version
        if hasattr(self.model, "add_model_tags"):
            self.model.add_model_tags(self._tag_names)

    def compute_loss(
        self,
        model: Union[PreTrainedModel, nn.Module],
        inputs: dict[str, Union[torch.Tensor, Any]],
        return_outputs=False,
        num_items_in_batch=None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, dict[str, torch.Tensor]]]:
        rewards_chosen = model(
            input_ids=inputs["input_ids_chosen"],
            attention_mask=inputs["attention_mask_chosen"],
            return_dict=True,
        )["logits"]
        rewards_rejected = model(
            input_ids=inputs["input_ids_rejected"],
            attention_mask=inputs["attention_mask_rejected"],
            return_dict=True,
        )["logits"]
        # calculate loss, optionally modulate with margin
        if "margin" in inputs:
            loss = -nn.functional.logsigmoid(rewards_chosen - rewards_rejected - inputs["margin"]).mean()
        else:
            loss = -nn.functional.logsigmoid(rewards_chosen - rewards_rejected).mean()

        if self.args.center_rewards_coefficient is not None:
            loss += self.args.center_rewards_coefficient * torch.mean((rewards_chosen + rewards_rejected) ** 2)

        if return_outputs:
            return loss, {
                "rewards_chosen": rewards_chosen,
                "rewards_rejected": rewards_rejected,
            }
        return loss

    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,
    ) -> tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]:
        inputs = self._prepare_inputs(inputs)
        if ignore_keys is None:
            if hasattr(self.model, "config"):
                ignore_keys = getattr(self.model.config, "keys_to_ignore_at_inference", [])
            else:
                ignore_keys = []

        with torch.no_grad():
            loss, logits_dict = self.compute_loss(model, inputs, return_outputs=True)

        if prediction_loss_only:
            return (loss, None, None)

        loss = loss.detach()
        logits = tuple(v for k, v in logits_dict.items() if k not in ignore_keys)
        logits = nested_detach(logits)
        # Stack accepted against rejected, mean over logits
        # and softmax to get preferences between accepted and rejected to sum to 1
        logits = torch.stack(logits).mean(dim=2).softmax(dim=0).T

        labels = torch.zeros(logits.shape[0])
        labels = self._prepare_inputs(labels)

        return loss, logits, labels

    def evaluate(self, *args, **kwargs):
        num_print_samples = kwargs.pop("num_print_samples", 4)
        self.visualize_samples(num_print_samples)
        return super().evaluate(*args, **kwargs)

    def visualize_samples(self, num_print_samples: int):
        """
        Visualize the reward model logits prediction

        Args:
            num_print_samples (`int`, defaults to `4`):
                The number of samples to print. Set to `-1` to print all samples.
        """
        eval_dataloader = self.get_eval_dataloader()
        table = defaultdict(list)
        for _, inputs in enumerate(eval_dataloader):
            _, logits, _ = self.prediction_step(self.model, inputs, prediction_loss_only=False)
            chosen_text = decode_and_strip_padding(inputs["input_ids_chosen"], self.processing_class)
            rejected_text = decode_and_strip_padding(inputs["input_ids_rejected"], self.processing_class)
            table["chosen_text"].extend(gather_object(chosen_text))
            table["rejected_text"].extend(gather_object(rejected_text))
            table["logits"].extend(
                gather_object([[round(inner_item, 4) for inner_item in item] for item in logits.tolist()])
            )
            if num_print_samples >= 0 and len(table["chosen_text"]) >= num_print_samples:
                break
        df = pd.DataFrame(table)
        if self.accelerator.process_index == 0:
            print_rich_table(df[:num_print_samples])
            if "wandb" in self.args.report_to:
                import wandb

                if wandb.run is not None:
                    wandb.log({"completions": wandb.Table(dataframe=df)})

            if "comet_ml" in self.args.report_to:
                log_table_to_comet_experiment(
                    name="completions.csv",
                    table=df,
                )

    def create_model_card(
        self,
        model_name: Optional[str] = None,
        dataset_name: Optional[str] = None,
        tags: Union[str, list[str], None] = None,
    ):
        """
        Creates a draft of a model card using the information available to the `Trainer`.

        Args:
            model_name (`str` or `None`, *optional*, defaults to `None`):
                Name of the model.
            dataset_name (`str` or `None`, *optional*, defaults to `None`):
                Name of the dataset used for training.
            tags (`str`, `list[str]` or `None`, *optional*, defaults to `None`):
                Tags to be associated with the model card.
        """
        if not self.is_world_process_zero():
            return

        if hasattr(self.model.config, "_name_or_path") and not os.path.isdir(self.model.config._name_or_path):
            base_model = self.model.config._name_or_path
        else:
            base_model = None

        tags = tags or []
        if isinstance(tags, str):
            tags = [tags]

        if hasattr(self.model.config, "unsloth_version"):
            tags.append("unsloth")

        model_card = generate_model_card(
            base_model=base_model,
            model_name=model_name,
            hub_model_id=self.hub_model_id,
            dataset_name=dataset_name,
            tags=tags,
            wandb_url=wandb.run.get_url() if is_wandb_available() and wandb.run is not None else None,
            comet_url=get_comet_experiment_url(),
            trainer_name="Reward",
        )

        model_card.save(os.path.join(self.args.output_dir, "README.md"))
class UnslothRewardTrainer(_UnslothRewardTrainer):
    """
    
    """
    def __init__(
        self,
        model = None,
        args = None,
        data_collator = None,
        train_dataset = None,
        eval_dataset = None,
        processing_class = None,
        model_init = None,
        compute_metrics = None,
        callbacks = None,
        preprocess_logits_for_metrics = None,
        peft_config = None,
        **kwargs
    ):
        if args is None: args = UnslothRewardConfig()
        use_bf16 = getattr(args, 'bf16', False)
        use_fp16 = getattr(args, 'fp16', False)
        force_float32 = False
        if os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1':
            print('Unsloth: Switching to float32 training since model cannot work with float16')
            force_float32 = True
        mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')
        dtype = getattr(model.config, 'torch_dtype', None)
        if dtype is None: dtype = model.get_input_embeddings().dtype
        from unsloth_zoo.utils import _get_dtype
        dtype = _get_dtype(dtype)
        float16 = dtype == torch.float16
        if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')
        if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')
        if force_float32:
            args.fp16 = False
            args.bf16 = False
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
        elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':
            args.fp16 = float16
            args.bf16 = not float16
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'
        if getattr(args, 'eval_dataset', None) is not None and getattr(args, 'eval_strategy', 'no') == 'no':
            args.eval_strategy = 'steps'
            if getattr(args, 'eval_steps', None) is None: args.eval_steps = 0.1
        ga_steps = getattr(args, 'gradient_accumulation_steps', None)
        if ga_steps is not None and ga_steps > 1:
            from transformers import __version__ as transformers_version
            if Version(transformers_version) <= Version('4.45.2'):
                print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!\n'
                      '`pip install --upgrade --no-cache-dir --force-reinstall --no-deps unsloth transformers trl unsloth_zoo`')
        if getattr(args, 'eval_strategy', 'no') != 'no':
            eval_bsz = getattr(args, 'per_device_eval_batch_size', 8)
            if eval_bsz == 8 and args.per_device_train_batch_size < eval_bsz: args.per_device_eval_batch_size = args.per_device_train_batch_size
            if getattr(args, 'eval_accumulation_steps', None) is None and ga_steps is not None: args.eval_accumulation_steps = ga_steps
        fp16_full_eval = getattr(args, 'fp16_full_eval', False)
        bf16_full_eval = getattr(args, 'bf16_full_eval', False)
        if args.fp16 and bf16_full_eval: args.bf16_full_eval = False; args.fp16_full_eval = True
        if args.bf16 and fp16_full_eval: args.bf16_full_eval = True; args.fp16_full_eval = False
        if force_float32:
            args.bf16_full_eval = False
            args.fp16_full_eval = False
        elif os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32') == 'bfloat16':
            args.bf16_full_eval = True
            args.fp16_full_eval = False
        elif not bf16_full_eval and not fp16_full_eval:
            args.bf16_full_eval = args.bf16
            args.fp16_full_eval = args.fp16
        _output_logits = False
        if locals().get('compute_metrics', None) is not None: _output_logits = True
        if locals().get('preprocess_logits_for_metrics', None) is not None: _output_logits = True
        if _output_logits:
            os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
        if 'max_seq_length' not in locals() and not hasattr(args, 'max_seq_length'):
            pass
        else:
            model_max_seq_length = getattr(model, 'max_seq_length', None)
            args_max_seq_length  = getattr(args,  'max_seq_length', None)
            if args_max_seq_length is None and model_max_seq_length is not None:
                max_seq_length = model.max_seq_length
                if hasattr(args, 'max_seq_length'): args.max_seq_length = max_seq_length
        if model is not None and hasattr(model, 'for_training'):
            model.for_training()
        if 'tokenizer' in locals() and hasattr(tokenizer, 'padding_side'): tokenizer.padding_side = 'right'
        if 'processing_class' in locals():
            if hasattr(processing_class, 'padding_side'): processing_class.padding_side = 'right'
            if hasattr(processing_class, 'tokenizer') and hasattr(processing_class.tokenizer, 'padding_side'): processing_class.tokenizer.padding_side = 'right'
        __tokenizer = processing_class if 'processing_class' in locals() else tokenizer
        from unsloth_zoo.vision_utils import UnslothVisionDataCollator
        if not isinstance(data_collator, UnslothVisionDataCollator):
            if isinstance(data_collator, DataCollatorForSeq2Seq) and 'labels' not in train_dataset.column_names:
                data_collator = DataCollatorForLanguageModeling(__tokenizer, mlm = False)
            elif isinstance(data_collator, DataCollatorForLanguageModeling) and 'labels' in train_dataset.column_names:
                data_collator = DataCollatorForSeq2Seq(__tokenizer)
        else:
            if hasattr(args, 'remove_unused_columns'): args.remove_unused_columns = False
            if hasattr(args, 'dataset_text_field'): args.dataset_text_field = ''
            if hasattr(args, 'dataset_kwargs'): args.dataset_kwargs = {'skip_prepare_dataset': True}
        if not isinstance(data_collator, UnslothVisionDataCollator):
            if not hasattr(__tokenizer, 'pad') and hasattr(__tokenizer, 'tokenizer'):
                if isinstance(data_collator, DataCollatorForSeq2Seq):
                    data_collator = DataCollatorForSeq2Seq(__tokenizer.tokenizer)
                else:
                    data_collator = DataCollatorForLanguageModeling(__tokenizer.tokenizer, mlm = False)
        other_metrics = []
        
        from unsloth_zoo.logging_utils import PatchRLStatistics
        PatchRLStatistics('reward_trainer', other_metrics)
        
        super().__init__(
            model = model,
            args = args,
            data_collator = data_collator,
            train_dataset = train_dataset,
            eval_dataset = eval_dataset,
            processing_class = processing_class,
            model_init = model_init,
            compute_metrics = compute_metrics,
            callbacks = callbacks,
            preprocess_logits_for_metrics = preprocess_logits_for_metrics,
            peft_config = peft_config,**kwargs)
        if hasattr(self, 'neftune_hook_handle'):
            self.neftune_hook_handle.remove()
            if hasattr(self, 'neftune_hook_handle'): del self.neftune_hook_handle
        if getattr(args, 'neftune_noise_alpha', None) is not None:
            model.get_input_embeddings().neftune_noise_alpha = self.neftune_noise_alpha
        pass
        
pass