cciwon-code-review-cli 2.0.2 → 2.0.3

package/unsloth_compiled_cache/UnslothXPOTrainer.py

@@ -0,0 +1,1255 @@
+"""
+2025.12.6
+2025.12.7
+4.57.1
+0.24.0
+__UNSLOTH_VERSIONING__
+"""
+
+# Unsloth auto generated code
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU Lesser General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+from torch import Tensor
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
+from trl.trainer.xpo_trainer import (Any, BaseImageProcessor, BasePairwiseJudge, Callable, Dataset, EvalPrediction, F, FeatureExtractionMixin, IterableDataset, OnlineDPOTrainer, OptimizerNames, Optional, PeftModel, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, SIMPLE_CHAT_TEMPLATE, TrainerCallback, Union, XPOConfig, XPOTrainer, empty_cache, get_reward, is_conversational, is_peft_available, jinja2, maybe_apply_chat_template, nn, selective_log_softmax, textwrap, torch, truncate_right, unwrap_model_for_generation)
+
+
+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
+import inspect
+from transformers import DataCollatorForSeq2Seq, DataCollatorForLanguageModeling as TransformersDataCollatorForLanguageModeling
+from transformers.training_args import ParallelMode
+
+# Wrap trainer with padding to right and enable training mode
+# Also patches W&B since multiple runs must use wandb.finish()
+import functools
+from types import MethodType
+def prepare_for_training_mode(f):
+    @functools.wraps(f)
+    def wrapper(self, *args, **kwargs):
+        # Enable training mode
+        if hasattr(self, 'model') and hasattr(self.model, "for_training"):
+            self.model.for_training()
+        output = f(self, *args, **kwargs)
+        # Return inference mode
+        if hasattr(self, 'model') and hasattr(self.model, "for_inference"):
+            self.model.for_inference()
+        # Patch W&B to enable logging on future runs, otherwise it'll overwrite the first run
+        try:
+            import wandb
+            wandb.finish()
+        except:
+            pass
+        return output
+    return wrapper
+pass
+
+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 chunked_selective_log_softmax(logits, index):
+    # Split into 4 chunks only
+    chunked_logits = torch.chunk(logits.reshape(-1, logits.shape[-1]), chunks = 4, dim = 0)
+    chunked_index  = torch.chunk(index.reshape(-1), chunks = 4, dim = 0)
+    all_per_token_logps = []
+    # Below loop does the same as selective_log_softmax(chunk_logits, chunk_index)
+    for chunk_logits, chunk_index in zip(chunked_logits, chunked_index):
+        chunk_logits = chunk_logits.to(torch.float32)
+        selected_logits = torch.gather(chunk_logits, dim = -1, index = chunk_index.unsqueeze(-1)).squeeze(-1)
+        logsumexp_values = torch.logsumexp(chunk_logits, dim = -1)
+        per_token_logps = selected_logits - logsumexp_values
+        all_per_token_logps.append(per_token_logps)
+    pass
+    all_per_token_logps = torch.concat(all_per_token_logps)
+    all_per_token_logps = all_per_token_logps.reshape((logits.shape[0], logits.shape[1]))
+    return all_per_token_logps
+
+def calculate_pad_tokens_in_prompt(
+    input_ids: torch.Tensor,
+    logits_to_keep: int,
+    pad_token_id: int
+) -> torch.Tensor:
+    """
+    Given prompt tensor, it returns all the left padded tokens in that sequence. so [pad, pad, pad, cat] = 3 tokens
+    """
+    if logits_to_keep >= input_ids.shape[1]:
+        raise ValueError("logits_to_keep must be smaller than the sequence length.")
+
+    prompt_section = input_ids[:, :-logits_to_keep]
+
+    padding_mask = (prompt_section == pad_token_id)
+
+    pad_token_counts = padding_mask.sum(dim=1)
+
+    return pad_token_counts
+
+def create_completion_attention_mask(
+    completion_input_ids: torch.Tensor,
+    left_pad_tokens_per_prompt: torch.Tensor,
+    max_left_pad: int,
+    pad_token_id: int
+) -> torch.Tensor:
+    """
+    Given that we have a sequence, [p,p,p,c,c,c,pad,pad,pad]
+
+    Where p are extra prompt tokens we got from slicing the torch tensor, c is completion tokens
+    and pad are pad tokens, this function would make a completion mask that would 0 out the pad
+    and p tokens. so in this example [0,0,0,1,1,1,0,0,0]
+    """
+    batch_size, completion_len = completion_input_ids.shape
+    device = completion_input_ids.device
+
+    num_tokens_to_mask = max_left_pad - left_pad_tokens_per_prompt
+
+    indices = torch.arange(completion_len, device=device).unsqueeze(0)
+    shift_mask = indices >= num_tokens_to_mask.unsqueeze(1)
+
+    non_padding_mask = (completion_input_ids != pad_token_id)
+
+    final_mask = shift_mask & non_padding_mask
+
+    return final_mask
+
+def left_pack_padding(tensor: torch.Tensor, pad_id: int) -> torch.Tensor:
+    """
+    Moves all padding tokens in each sequence of a batch to the right.
+    """
+    mask = (tensor != pad_id)
+    # Must do stable=True since binary mark is unordered
+    sorted_indices = torch.argsort(mask, dim=1, descending=True, stable=True)
+    packed_tensor = torch.gather(tensor, 1, sorted_indices)
+    return packed_tensor
+
+def align_logprobs_with_mask(
+    logprob_tensor: torch.Tensor,
+    attention_mask: torch.Tensor,
+    pad_value: float = 0.0
+) -> torch.Tensor:
+    """
+    Aligns a log probability tensor with a given attention mask.
+    """
+
+    device = logprob_tensor.device
+    batch_size, logprob_seq_len = logprob_tensor.shape
+    mask_seq_len = attention_mask.shape[1]
+
+    padded_logprobs = torch.full(
+        attention_mask.shape,
+        fill_value=pad_value,
+        dtype=logprob_tensor.dtype,
+        device=device
+    )
+
+    left_pad_counts = torch.argmax(attention_mask, dim=1)
+
+    cols = torch.arange(logprob_seq_len, device=device)
+    dest_indices = left_pad_counts.unsqueeze(1) + cols
+
+    # Create destination row indices
+    # Shape: [batch_size, logprob_seq_len]
+    row_indices = torch.arange(batch_size, device=device).unsqueeze(1).expand_as(dest_indices)
+
+    # --- 4. Filter out-of-bounds indices and perform assignment ---
+    # Create a mask to identify only the indices that are within the bounds
+    # of the target tensor's sequence length.
+    valid_mask = dest_indices < mask_seq_len
+
+    # Use this mask to select only the valid row indices, column indices,
+    # and the corresponding values from the logprob tensor.
+    # This flattens the selected elements into 1D tensors.
+    valid_rows = row_indices[valid_mask]
+    valid_cols = dest_indices[valid_mask]
+    valid_vals = logprob_tensor[valid_mask]
+
+    # Place the valid values into their correct positions in the padded tensor
+    # using a single, efficient advanced indexing operation.
+    padded_logprobs[valid_rows, valid_cols] = valid_vals
+
+    return padded_logprobs
+@dataclass
+class UnslothXPOConfig(XPOConfig):
+    """
+    
+    Configuration class for the [`XPOTrainer`].
+
+    Subclass of [`OnlineDPOConfig`] we can use all its arguments and add the following:
+
+    Parameters:
+        alpha (`float` or `list[float]`, *optional*, defaults to `1e-5`):
+            Weight of the XPO loss term. If a list of floats is provided then the alpha is selected for each new epoch
+            and the last alpha is used for the rest of the epochs.
+    
+    """
+    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.'},
+    )
+    max_seq_length : Optional[int] = field(
+        default = None,
+        metadata = {'help': 'Maximum sequence length to truncate to.'},
+    )
+    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,
+        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 = True,
+        label_names = None,
+        load_best_model_at_end = False,
+        metric_for_best_model = None,
+        greater_is_better = None,
+        ignore_data_skip = False,
+        fsdp = None,
+        fsdp_min_num_params = 0,
+        fsdp_config = None,
+        fsdp_transformer_layer_cls_to_wrap = None,
+        accelerator_config = None,
+        parallelism_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',
+        project = 'huggingface',
+        trackio_space_id = 'trackio',
+        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,
+        hub_revision = None,
+        gradient_checkpointing = True,
+        gradient_checkpointing_kwargs = None,
+        include_inputs_for_metrics = False,
+        eval_do_concat_batches = True,
+        fp16_backend = 'auto',
+        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,
+        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,
+        liger_kernel_config = None,
+        eval_use_gather_object = False,
+        average_tokens_across_devices = True,
+        reward_model_path = None,
+        judge = None,
+        max_new_tokens = 64,
+        max_length = 512,
+        temperature = 0.9,
+        top_p = 1.0,
+        top_k = None,
+        min_p = None,
+        repetition_penalty = 1.0,
+        generation_kwargs = {},
+        use_transformers_paged = False,
+        cache_implementation = None,
+        missing_eos_penalty = None,
+        loss_type = 'sigmoid',
+        disable_dropout = True,
+        use_vllm = False,
+        vllm_model_impl = 'vllm',
+        vllm_guided_decoding_regex = None,
+        vllm_gpu_memory_utilization = 0.55,
+        vllm_mode = 'colocate',
+        vllm_server_base_url = None,
+        vllm_server_host = '0.0.0.0',
+        vllm_server_port = 8000,
+        vllm_server_timeout = 240.0,
+        vllm_tensor_parallel_size = 1,
+        ds3_gather_for_generation = True,
+        model_init_kwargs = None,
+        reward_weights = None,
+        dataset_num_proc = None,
+        gpu_memory_utilization = None,
+        vllm_sampling_params = None,
+        unsloth_num_chunks = -1,
+        max_seq_length = None,
+        **kwargs,
+    ):
+        if learning_rate < 1e-7: print(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: print(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 = min(max(cpu_count()+4, 2), 64)
+        if temperature <= 0:
+            raise MathError('Unsloth: Please set a positive non-zero temperature since your results will be wrong.')
+        elif temperature >= 10:
+            raise MathError('Unsloth: Please set a positive non-zero temperature less than 10, since sampling will be quite erratic.')
+        
+        
+        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,
+            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,
+            fsdp_transformer_layer_cls_to_wrap = fsdp_transformer_layer_cls_to_wrap,
+            accelerator_config = accelerator_config,
+            parallelism_config = parallelism_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,
+            project = project,
+            trackio_space_id = trackio_space_id,
+            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,
+            hub_revision = hub_revision,
+            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,
+            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,
+            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,
+            liger_kernel_config = liger_kernel_config,
+            eval_use_gather_object = eval_use_gather_object,
+            average_tokens_across_devices = average_tokens_across_devices,
+            reward_model_path = reward_model_path,
+            judge = judge,
+            max_new_tokens = max_new_tokens,
+            max_length = max_length,
+            temperature = temperature,
+            top_p = top_p,
+            top_k = top_k,
+            min_p = min_p,
+            repetition_penalty = repetition_penalty,
+            generation_kwargs = generation_kwargs,
+            use_transformers_paged = use_transformers_paged,
+            cache_implementation = cache_implementation,
+            missing_eos_penalty = missing_eos_penalty,
+            loss_type = loss_type,
+            disable_dropout = disable_dropout,
+            use_vllm = use_vllm,
+            vllm_model_impl = vllm_model_impl,
+            vllm_guided_decoding_regex = vllm_guided_decoding_regex,
+            vllm_gpu_memory_utilization = vllm_gpu_memory_utilization,
+            vllm_mode = vllm_mode,
+            vllm_server_base_url = vllm_server_base_url,
+            vllm_server_host = vllm_server_host,
+            vllm_server_port = vllm_server_port,
+            vllm_server_timeout = vllm_server_timeout,
+            vllm_tensor_parallel_size = vllm_tensor_parallel_size,
+            ds3_gather_for_generation = ds3_gather_for_generation,
+            model_init_kwargs = model_init_kwargs,
+            reward_weights = reward_weights,
+            dataset_num_proc = dataset_num_proc,
+            gpu_memory_utilization = gpu_memory_utilization,**kwargs)
+        self.vllm_sampling_params = vllm_sampling_params
+        self.unsloth_num_chunks = unsloth_num_chunks
+        self.max_seq_length = max_seq_length
+pass
+
+class _UnslothXPOTrainer(OnlineDPOTrainer):
+    """"""
+
+    _tag_names = ["trl", "xpo"]
+    _name = "XPO"
+    _paper = {
+        "title": "Exploratory Preference Optimization: Harnessing Implicit Q*-Approximation for Sample-Efficient RLHF",
+        "id": "2405.21046",
+        # docstyle-ignore
+        "citation": textwrap.dedent("""\
+            @article{jung2024binary,
+                title        = {{Exploratory Preference Optimization: Harnessing Implicit Q*-Approximation for Sample-Efficient RLHF}},
+                author       = {Tengyang Xie and Dylan J. Foster and Akshay Krishnamurthy and Corby Rosset and Ahmed Awadallah and Alexander Rakhlin},
+                year         = 2024,
+                eprint       = {arXiv:2405.21046}
+            }"""),
+    }
+
+    def __init__(
+        self,
+        model: Union[PreTrainedModel, nn.Module] = None,
+        ref_model: Union[PreTrainedModel, nn.Module] = None,
+        reward_funcs: Optional[nn.Module] = None,
+        judge: Optional[BasePairwiseJudge] = None,
+        args: Optional[XPOConfig] = None,
+        data_collator: Optional[Callable] = None,
+        train_dataset: Optional[Union[Dataset, IterableDataset]] = None,
+        eval_dataset: Optional[Union[Dataset, dict[str, Dataset]]] = None,
+        processing_class: Optional[
+            Union[PreTrainedTokenizerBase, BaseImageProcessor, FeatureExtractionMixin, ProcessorMixin]
+        ] = None,
+        reward_processing_classes: Optional[Union[PreTrainedTokenizerBase, list[PreTrainedTokenizerBase]]] = None,
+        peft_config: Optional[dict] = 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,
+        # Deprecated parameters
+        reward_model: Optional[Union[PreTrainedModel, nn.Module]] = None,
+    ) -> None:
+        super().__init__(
+            model=model,
+            ref_model=ref_model,
+            judge=judge,
+            reward_funcs=reward_funcs,
+            reward_model=reward_model,
+            args=args,
+            data_collator=data_collator,
+            train_dataset=train_dataset,
+            eval_dataset=eval_dataset,
+            processing_class=processing_class,
+            reward_processing_classes=reward_processing_classes,
+            peft_config=peft_config,
+            compute_metrics=compute_metrics,
+            callbacks=callbacks,
+            optimizers=optimizers,
+            preprocess_logits_for_metrics=preprocess_logits_for_metrics,
+        )
+
+        self._alpha = self.args.alpha
+
+        # Overwrite the stats dictionary to include XPO specific statistics
+        self.stats = {
+            # Remove "non_score_reward", "rlhf_reward", "scores"
+            # Add "loss/dpo", "loss/xpo"
+            "loss/dpo": [],
+            "loss/xpo": [],
+            "objective/kl": [],
+            "objective/entropy": [],
+            "rewards/chosen": [],
+            "rewards/rejected": [],
+            "rewards/accuracies": [],
+            "rewards/margins": [],
+            "logps/chosen": [],
+            "logps/rejected": [],
+            # Replace "contain_eos_token" by "model_contain_eos_token" and "ref_contain_eos_token"
+            "val/model_contain_eos_token": [],
+            "val/ref_contain_eos_token": [],
+            "alpha": [],
+            "beta": [],
+        }
+        if self.reward_funcs is not None:
+            if len(self.reward_funcs) != 1:
+                raise ValueError("XPOTrainer only supports one reward function/model.")
+            self.reward_funcs = self.reward_funcs[0]
+            self.stats["objective/model_scores"] = []
+            self.stats["objective/ref_scores"] = []
+            self.stats["objective/scores_margin"] = []
+
+    @property
+    def alpha(self):
+        if isinstance(self._alpha, list):
+            epoch = self.state.epoch
+            return self._alpha[epoch] if epoch < len(self._alpha) else self._alpha[-1]
+        else:
+            return self._alpha
+
+    def _generate_completions(self, prompts, model):
+        with unwrap_model_for_generation(model, self.accelerator) as unwrapped_policy_model_for_gen:
+            model_output = unwrapped_policy_model_for_gen.generate(
+                input_ids=prompts["input_ids"],
+                attention_mask=prompts["attention_mask"],
+                generation_config=self.generation_config,
+            )
+
+        actual_model_for_ref_generation: torch.nn.Module
+        if self.ref_model is None:
+            unwrapped_main_model_for_ref_logic = self.accelerator.unwrap_model(model)
+
+            if is_peft_available() and isinstance(unwrapped_main_model_for_ref_logic, PeftModel):
+                actual_model_for_ref_generation = unwrapped_main_model_for_ref_logic.get_base_model()
+            else:
+                actual_model_for_ref_generation = unwrapped_main_model_for_ref_logic
+        else:
+            actual_model_for_ref_generation = self.accelerator.unwrap_model(self.ref_model)
+
+        with unwrap_model_for_generation(actual_model_for_ref_generation, self.accelerator) as final_ref_model_for_gen:
+            ref_output = final_ref_model_for_gen.generate(
+                input_ids=prompts["input_ids"],
+                attention_mask=prompts["attention_mask"],
+                generation_config=self.generation_config,
+            )
+
+        return model_output, ref_output
+
+    def _process_completions(self, model_output, ref_output, prompts):
+        context_length = prompts["input_ids"].shape[1]
+
+        # Process model completions
+        model_completion_ids = model_output[:, context_length:]
+        model_completion_ids, model_completion_mask = truncate_right(
+            model_completion_ids, self.processing_class.eos_token_id, self.processing_class.pad_token_id
+        )
+        model_data = {
+            "input_ids": torch.cat((prompts["input_ids"], model_completion_ids), dim=1),
+            "attention_mask": torch.cat((prompts["attention_mask"], model_completion_mask), dim=1),
+            "raw": prompts["raw"],
+        }
+
+        # Process reference model completions
+        ref_completion_ids = ref_output[:, context_length:]
+        ref_completion_ids, ref_completion_mask = truncate_right(
+            ref_completion_ids, self.processing_class.eos_token_id, self.processing_class.pad_token_id
+        )
+        ref_data = {
+            "input_ids": torch.cat((prompts["input_ids"], ref_completion_ids), dim=1),
+            "attention_mask": torch.cat((prompts["attention_mask"], ref_completion_mask), dim=1),
+            "raw": prompts["raw"],
+        }
+
+        return model_data, ref_data
+
+    def _compute_rewards(self, model_data, ref_data, context_length):
+        with torch.no_grad():
+            _, model_scores, _ = get_reward(
+                self.reward_funcs, model_data["input_ids"], self.processing_class.pad_token_id, context_length
+            )
+            _, ref_scores, _ = get_reward(
+                self.reward_funcs, ref_data["input_ids"], self.processing_class.pad_token_id, context_length
+            )
+
+        # Apply EOS penalty if needed
+        if self.args.missing_eos_penalty is not None:
+            model_contain_eos = torch.any(model_data["input_ids"] == self.processing_class.eos_token_id, dim=-1)
+            ref_contain_eos = torch.any(ref_data["input_ids"] == self.processing_class.eos_token_id, dim=-1)
+            model_scores[~model_contain_eos] -= self.args.missing_eos_penalty
+            ref_scores[~ref_contain_eos] -= self.args.missing_eos_penalty
+
+        return model_scores, ref_scores
+
+    def _compute_judge(self, model_data, ref_data, context_length):
+        prompts = model_data["raw"]
+        model_data_completions = self.processing_class.batch_decode(
+            model_data["input_ids"][:, context_length:], skip_special_tokens=True
+        )
+        model_data_completions = [completion.strip() for completion in model_data_completions]
+
+        ref_data_completions = self.processing_class.batch_decode(
+            ref_data["input_ids"][:, context_length:], skip_special_tokens=True
+        )
+        ref_data_completions = [completion.strip() for completion in ref_data_completions]
+
+        if is_conversational({"prompt": prompts[0]}):
+            model_data_completions = [
+                [{"role": "assistant", "content": completion}] for completion in model_data_completions
+            ]
+            environment = jinja2.Environment()
+            template = environment.from_string(SIMPLE_CHAT_TEMPLATE)
+            prompts = [template.render(messages=message) for message in prompts]
+            model_data_completions = [template.render(messages=completion) for completion in model_data_completions]
+
+            ref_data_completions = [
+                [{"role": "assistant", "content": completion}] for completion in ref_data_completions
+            ]
+            ref_data_completions = [template.render(messages=completion) for completion in ref_data_completions]
+
+        ranks_of_first_completion = self.judge.judge(
+            prompts,
+            list(zip(model_data_completions, ref_data_completions)),
+        )
+        # convert ranks to a True/False mask:
+        # when rank == 0, it means the first completion is the best
+        # when rank == 1, it means the second completion is the best
+        return torch.tensor([rank == 0 for rank in ranks_of_first_completion], device=model_data["input_ids"].device)
+
+    def _compute_logprobs(self, model, model_data, ref_data, context_length):
+        def compute_logprobs_for_data(m, data):
+            output = m(data["input_ids"], attention_mask=data["attention_mask"])
+            logits = output.logits[:, context_length - 1 : -1]
+            token_logprobs = selective_log_softmax(logits, data["input_ids"][:, context_length:])
+            return token_logprobs
+
+        # Compute logprobs for model completions
+        model_logprobs_model_data = compute_logprobs_for_data(model, model_data)
+        # Compute logprobs for model on reference completions (for XPO loss)
+        model_logprobs_ref_data = compute_logprobs_for_data(model, ref_data)
+
+        # Compute logprobs for reference model completions
+        with torch.no_grad():
+            if self.ref_model is None:
+                with model.disable_adapter():
+                    ref_logprobs_model_data = compute_logprobs_for_data(model, model_data)
+                    ref_logprobs_ref_data = compute_logprobs_for_data(model, ref_data)
+            else:
+                ref_logprobs_model_data = compute_logprobs_for_data(self.ref_model, model_data)
+                ref_logprobs_ref_data = compute_logprobs_for_data(self.ref_model, ref_data)
+
+        # Mask padding tokens
+        model_padding_mask = model_data["attention_mask"][:, context_length:] == 0
+        ref_padding_mask = ref_data["attention_mask"][:, context_length:] == 0
+        model_logprobs_model_data = model_logprobs_model_data.masked_fill(model_padding_mask, 0.0)
+        model_logprobs_ref_data = model_logprobs_ref_data.masked_fill(ref_padding_mask, 0.0)
+        ref_logprobs_ref_data = ref_logprobs_ref_data.masked_fill(ref_padding_mask, 0.0)
+        ref_logprobs_model_data = ref_logprobs_model_data.masked_fill(model_padding_mask, 0.0)
+
+        return model_logprobs_model_data, model_logprobs_ref_data, ref_logprobs_ref_data, ref_logprobs_model_data
+
+    def _compute_losses(
+        self,
+        model_logprobs_model_data,
+        model_logprobs_ref_data,
+        ref_logprobs_ref_data,
+        ref_logprobs_model_data,
+        chosen_mask,
+    ):
+        # Compute log probs
+        model_logprobs_model_data_sum = model_logprobs_model_data.sum(1)
+        model_logprobs_ref_data_sum = model_logprobs_ref_data.sum(1)
+        ref_logprobs_ref_data_sum = ref_logprobs_ref_data.sum(1)
+        ref_logprobs_model_data_sum = ref_logprobs_model_data.sum(1)
+
+        chosen_model_logprobs = torch.where(chosen_mask, model_logprobs_model_data_sum, model_logprobs_ref_data_sum)
+        chosen_ref_logprobs = torch.where(chosen_mask, ref_logprobs_model_data_sum, ref_logprobs_ref_data_sum)
+        chosen_log_ratios = chosen_model_logprobs - chosen_ref_logprobs
+
+        rejected_model_logprobs = torch.where(~chosen_mask, model_logprobs_model_data_sum, model_logprobs_ref_data_sum)
+        rejected_ref_logprobs = torch.where(~chosen_mask, ref_logprobs_model_data_sum, ref_logprobs_ref_data_sum)
+        rejected_log_ratios = rejected_model_logprobs - rejected_ref_logprobs
+
+        # Compute logits as the difference between chosen and rejected log ratios
+        logits = chosen_log_ratios - rejected_log_ratios
+
+        if self.args.loss_type == "sigmoid":
+            dpo_losses = -F.logsigmoid(self.beta * logits)
+        elif self.args.loss_type == "ipo":
+            dpo_losses = (logits - 1 / (2 * self.beta)) ** 2
+        else:
+            raise NotImplementedError(f"invalid loss type {self.args.loss_type}")
+
+        # Compute XPO specific loss
+        xpo_losses = self.alpha * model_logprobs_ref_data_sum
+
+        # Total loss
+        loss = (dpo_losses + xpo_losses).mean()
+
+        return loss, dpo_losses, xpo_losses
+
+    def _log_statistics(
+        self,
+        model_data,
+        ref_data,
+        model_logprobs_model_data,
+        model_logprobs_ref_data,
+        ref_logprobs_ref_data,
+        ref_logprobs_model_data,
+        chosen_mask,
+        dpo_losses,
+        xpo_losses,
+        context_length,
+        model_scores=None,
+        ref_scores=None,
+    ):
+        # Helper function to gather and compute mean
+        def gather_mean(tensor):
+            return self.accelerator.gather_for_metrics(tensor).mean().item()
+
+        # Log losses
+        self.stats["loss/dpo"].append(gather_mean(dpo_losses))
+        self.stats["loss/xpo"].append(gather_mean(xpo_losses))
+
+        # Log scores
+        if self.reward_funcs is not None:
+            self.stats["objective/model_scores"].append(gather_mean(model_scores))
+            self.stats["objective/ref_scores"].append(gather_mean(ref_scores))
+            self.stats["objective/scores_margin"].append(gather_mean(model_scores - ref_scores))
+
+        # Log logprobs
+        model_logprobs_model_data_sum = model_logprobs_model_data.sum(1)
+        model_logprobs_ref_data_sum = model_logprobs_ref_data.sum(1)
+        ref_logprobs_ref_data_sum = ref_logprobs_ref_data.sum(1)
+        ref_logprobs_model_data_sum = ref_logprobs_model_data.sum(1)
+
+        chosen_model_logprobs = torch.where(chosen_mask, model_logprobs_model_data_sum, model_logprobs_ref_data_sum)
+        chosen_ref_logprobs = torch.where(chosen_mask, ref_logprobs_model_data_sum, ref_logprobs_ref_data_sum)
+        chosen_log_ratios = chosen_model_logprobs - chosen_ref_logprobs
+
+        rejected_model_logprobs = torch.where(~chosen_mask, model_logprobs_model_data_sum, model_logprobs_ref_data_sum)
+        rejected_ref_logprobs = torch.where(~chosen_mask, ref_logprobs_model_data_sum, ref_logprobs_ref_data_sum)
+        rejected_log_ratios = rejected_model_logprobs - rejected_ref_logprobs
+
+        self.stats["logps/chosen"].append(gather_mean(chosen_model_logprobs.mean() + chosen_ref_logprobs.mean()))
+        self.stats["logps/rejected"].append(gather_mean(rejected_model_logprobs.mean() + rejected_ref_logprobs.mean()))
+
+        # Log rewards
+        # Compute various statistics
+        chosen_rewards = chosen_log_ratios * self.beta
+        rejected_rewards = rejected_log_ratios * self.beta
+        self.stats["rewards/chosen"].append(gather_mean(chosen_rewards.mean()))
+        self.stats["rewards/rejected"].append(gather_mean(rejected_rewards.mean()))
+
+        # Calculate KL divergence for model and ref data
+        kl_model_data = model_logprobs_model_data - ref_logprobs_model_data
+        kl_ref_data = model_logprobs_ref_data - ref_logprobs_ref_data
+        mean_kl = (kl_model_data.sum(1) + kl_ref_data.sum(1)).mean() / 2
+        self.stats["objective/kl"].append(gather_mean(mean_kl))
+
+        # Calculate entropy for model and ref data
+        entropy_model_data = -model_logprobs_model_data.sum(1)
+        entropy_ref_data = -model_logprobs_ref_data.sum(1)
+        mean_entropy = (entropy_model_data.mean() + entropy_ref_data.mean()) / 2
+        self.stats["objective/entropy"].append(gather_mean(mean_entropy))
+
+        # Calculate margins
+        margin = chosen_rewards - rejected_rewards
+        self.stats["rewards/margins"].append(gather_mean(margin.mean()))
+
+        # Calculate accuracy
+        accuracy = (margin > 0).float()
+        self.stats["rewards/accuracies"].append(gather_mean(accuracy.mean()))
+
+        # Log EOS token statistics
+        model_eos = (model_data["input_ids"][:, context_length:] == self.processing_class.eos_token_id).any(dim=1)
+        ref_eos = (ref_data["input_ids"][:, context_length:] == self.processing_class.eos_token_id).any(dim=1)
+        self.stats["val/model_contain_eos_token"].append(gather_mean(model_eos.float()))
+        self.stats["val/ref_contain_eos_token"].append(gather_mean(ref_eos.float()))
+
+        # Log alpha and beta
+        self.stats["alpha"].append(self.alpha)
+        self.stats["beta"].append(self.beta)
+
+    def training_step(
+        self, model: nn.Module, inputs: dict[str, Union[torch.Tensor, Any]], num_items_in_batch: Optional[int] = None
+    ) -> torch.Tensor:
+        model.train()
+
+        # Apply chat template and tokenize the input
+        batch_size = len(next(iter(inputs.values())))
+        prompts = inputs["prompt"]
+        inputs = [{k: v[i] for k, v in inputs.items()} for i in range(batch_size)]
+        inputs = [maybe_apply_chat_template(x, self.processing_class) for x in inputs]
+        inputs = [self.tokenize_row(x, self.model.config.is_encoder_decoder, self.processing_class) for x in inputs]
+        inputs = self.data_collator(inputs)
+
+        # need the prompt_ only
+        inputs = self._prepare_inputs(inputs)
+        context_length = inputs["prompt_input_ids"].shape[1]
+        prompts = {
+            "input_ids": inputs["prompt_input_ids"],
+            "attention_mask": inputs["prompt_attention_mask"],
+            "raw": prompts,
+        }
+        del inputs
+
+        # Sample completions from both the model and the reference model
+        model_output, ref_output = self._generate_completions(prompts, model)
+
+        # Process model completions
+        model_data, ref_data = self._process_completions(model_output, ref_output, prompts)
+
+        # Compute rewards
+        if self.reward_funcs is not None:
+            model_scores, ref_scores = self._compute_rewards(model_data, ref_data, context_length)
+            chosen_mask = model_scores >= ref_scores
+        else:
+            model_scores, ref_scores = None, None
+            chosen_mask = self._compute_judge(model_data, ref_data, context_length)
+
+        # Compute logprobs
+        model_logprobs_model_data, model_logprobs_ref_data, ref_logprobs_ref_data, ref_logprobs_model_data = (
+            self._compute_logprobs(model, model_data, ref_data, context_length)
+        )
+
+        # Compute loss
+        loss, dpo_losses, xpo_losses = self._compute_losses(
+            model_logprobs_model_data,
+            model_logprobs_ref_data,
+            ref_logprobs_ref_data,
+            ref_logprobs_model_data,
+            chosen_mask,
+        )
+
+        # Log everything
+        self._log_statistics(
+            model_data,
+            ref_data,
+            model_logprobs_model_data.detach(),
+            model_logprobs_ref_data.detach(),
+            ref_logprobs_ref_data,
+            ref_logprobs_model_data,
+            chosen_mask,
+            dpo_losses.detach(),
+            xpo_losses.detach(),
+            context_length,
+            model_scores,
+            ref_scores,
+        )
+
+        if (
+            self.args.torch_empty_cache_steps is not None
+            and self.state.global_step % self.args.torch_empty_cache_steps == 0
+        ):
+            empty_cache()
+
+        kwargs = {}
+        # For LOMO optimizers you need to explicitly use the learning rate
+        if self.args.optim in [OptimizerNames.LOMO, OptimizerNames.ADALOMO]:
+            kwargs["learning_rate"] = self._get_learning_rate()
+
+        if self.args.n_gpu > 1:
+            loss = loss.mean()  # mean() to average on multi-gpu parallel training
+
+        self.accelerator.backward(loss, **kwargs)
+
+        return loss.detach() / self.args.gradient_accumulation_steps
+class UnslothXPOTrainer(_UnslothXPOTrainer):
+    """
+    
+    Trainer for Exploratory Preference Optimization (XPO).
+
+    It is implemented as a subclass of [`OnlineDPOTrainer`].
+
+    Args:
+        model ([`~transformers.PreTrainedModel`]):
+            The model to train, preferably an `AutoModelForCausalLM`.
+        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.
+        reward_funcs ([`~transformers.PreTrainedModel`]):
+            The reward model to score completions with, preferably an
+            [`~transformers.AutoModelForSequenceClassification`].
+        judge ([`BasePairwiseJudge`]):
+            The judge to use for pairwise comparison of model completions.
+        args ([`XPOConfig`]):
+            The XPO config 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.
+        train_dataset ([`~datasets.Dataset`]):
+            The dataset to use for training.
+        eval_dataset ([`~datasets.Dataset`]):
+            The dataset to use for evaluation.
+        processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.BaseImageProcessor`], [`~transformers.FeatureExtractionMixin`] or [`~transformers.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.
+        peft_config (`dict`):
+            The peft config to use for training.
+        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.
+        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.
+
+        reward_model:
+
+            <Deprecated version="0.22.0">
+
+            This parameter is deprecated and will be removed in version 0.25.0. Use `reward_funcs` instead.
+
+            </Deprecated>
+    
+    """
+    def __init__(
+        self,
+        model = None,
+        ref_model = None,
+        reward_funcs = None,
+        judge = None,
+        args = None,
+        data_collator = None,
+        train_dataset = None,
+        eval_dataset = None,
+        processing_class = None,
+        reward_processing_classes = None,
+        peft_config = None,
+        compute_metrics = None,
+        callbacks = None,
+        preprocess_logits_for_metrics = None,
+        reward_model = None,
+        **kwargs
+    ):
+        if args is None: args = UnslothXPOConfig()
+        use_bf16 = getattr(args, 'bf16', False)
+        if type(use_bf16) is not bool: use_bf16 = False
+        use_fp16 = getattr(args, 'fp16', False)
+        if type(use_fp16) is not bool: use_fp16 = False
+        force_float32 = False
+        full_finetuning = os.environ.get('UNSLOTH_ENABLE_FULL_FINETUNING', '0') == '1'
+        if not full_finetuning and (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, 'dtype', None) or getattr(model.config, 'torch_dtype', None)
+        if dtype is None: dtype = model.get_input_embeddings().weight.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:
+            # Forced float32 training
+            args.fp16 = False
+            args.bf16 = False
+            os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
+            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
+            # args.mixed_precision is a new argument which needs to be set now
+        elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':
+            # Mixed precision training
+            args.fp16 = float16
+            args.bf16 = not float16
+            os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'
+            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16' if float16 else 'bf16'
+            # args.mixed_precision is a new argument which needs to be set now
+        elif mixed_precision_dtype == 'bfloat16':
+            # Both False since bfloat16 full finetuning doesn't do any autocasting.
+            args.fp16 = False
+            args.bf16 = False
+            os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
+            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
+            # args.mixed_precision is a new argument which needs to be set now
+        
+        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)
+        if type(fp16_full_eval) is not bool: fp16_full_eval = False
+        bf16_full_eval = getattr(args, 'bf16_full_eval', False)
+        if type(bf16_full_eval) is not bool: 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(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
+        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 = TransformersDataCollatorForLanguageModeling(
+                    __tokenizer,
+                    mlm = False,
+                    mlm_probability = 0.0,
+                    pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
+                )
+            elif isinstance(data_collator, TransformersDataCollatorForLanguageModeling) and 'labels' in train_dataset.column_names:
+                data_collator = DataCollatorForSeq2Seq(
+                    __tokenizer,
+                    pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
+                )
+        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,
+                        pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
+                    )
+                else:
+                    data_collator = TransformersDataCollatorForLanguageModeling(
+                        __tokenizer.tokenizer,
+                        mlm = False,
+                        mlm_probability = 0.0,
+                        pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
+                    )
+        other_metrics = []
+        
+        from unsloth_zoo.logging_utils import PatchRLStatistics
+        PatchRLStatistics('xpo_trainer', other_metrics)
+        
+        # [TODO] Fix up DataParallel multiplying batch sizes
+        # [TODO] DDP works, but DP seems to not work? [TODO]
+        if getattr(args, "parallel_mode", None) == ParallelMode.NOT_DISTRIBUTED and args.n_gpu > 1:
+            if getattr(args, "_n_gpu", 1) != 1:
+                args._n_gpu = 1
+        if "model" in locals() and hasattr(model, "for_training"):
+            model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
+        super().__init__(
+            model = model,
+            ref_model = ref_model,
+            reward_funcs = reward_funcs,
+            judge = judge,
+            args = args,
+            data_collator = data_collator,
+            train_dataset = train_dataset,
+            eval_dataset = eval_dataset,
+            processing_class = processing_class,
+            reward_processing_classes = reward_processing_classes,
+            peft_config = peft_config,
+            compute_metrics = compute_metrics,
+            callbacks = callbacks,
+            preprocess_logits_for_metrics = preprocess_logits_for_metrics,
+            reward_model = reward_model,**kwargs)
+        if "model" in locals() and hasattr(model, "for_inference"):
+            model.for_inference()
+        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
+        if hasattr(self, 'accelerator'):
+            scaler = self.accelerator.scaler
+            current_model = model
+            while hasattr(current_model, 'model'):
+                current_model.accelerator_scaler = scaler
+                current_model = current_model.model
+            current_model.accelerator_scaler = scaler
+        pass
+        if hasattr(self, 'train'):
+            self.train = MethodType(prepare_for_training_mode(self.__class__.train), self)
+        pass
+        
+pass