cciwon-code-review-cli 2.0.2 → 2.0.4

package/unsloth_compiled_cache/UnslothORPOTrainer.py

@@ -0,0 +1,1730 @@
+"""
+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.orpo_trainer import (Any, AutoModelForCausalLM, BaseImageProcessor, BaseTrainer, Callable, DPODataCollatorWithPadding, DataCollator, DataLoader, Dataset, EvalLoopOutput, F, FeatureExtractionMixin, Literal, ORPOConfig, ORPOTrainer, Optional, PartialState, Path, PeftModel, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, TrainerCallback, Union, add_bos_token_if_needed, add_eos_token_if_needed, autocast, defaultdict, disable_dropout_in_model, inspect, is_comet_available, is_peft_available, is_torch_fx_proxy, is_torch_xla_available, is_wandb_available, log_table_to_comet_experiment, logger, logging, maybe_apply_chat_template, maybe_extract_prompt, nn, np, nullcontext, os, pad_to_length, pd, peft_module_casting_to_bf16, prepare_model_for_kbit_training, random, selective_log_softmax, textwrap, torch, warnings, F, Optional, PeftModel, PreTrainedModel, is_peft_available, logger, os, torch)
+
+
+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 UnslothORPOConfig(ORPOConfig):
+    """
+    
+    Configuration class for the [`ORPOTrainer`].
+
+    This class includes only the parameters that are specific to ORPO training. For a full list of training arguments,
+    please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values in this class may
+    differ from those in [`~transformers.TrainingArguments`].
+
+    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. This argument is required if you want
+            to use the default data collator.
+        max_prompt_length (`int` or `None`, *optional*, defaults to `512`):
+            Maximum length of the prompt. This argument is required if you want to use the default data collator.
+        max_completion_length (`int`, *optional*):
+            Maximum length of the completion. This argument is required if you want to use the default data collator
+            and your model is an encoder-decoder.
+        beta (`float`, *optional*, defaults to `0.1`):
+            Parameter controlling the relative ratio loss weight in the ORPO loss. In the
+            [paper](https://huggingface.co/papers/2403.07691), it is denoted by λ. In the
+            [code](https://github.com/xfactlab/orpo), it is denoted by `alpha`.
+        disable_dropout (`bool`, *optional*, defaults to `True`):
+            Whether to disable dropout in the model.
+        label_pad_token_id (`int`, *optional*, defaults to `-100`):
+            Label pad token id. This argument is required if you want to use the default data collator.
+        padding_value (`int`, *optional*):
+            Padding value to use. If `None`, the padding value of the tokenizer is used.
+        truncation_mode (`str`, *optional*, defaults to `"keep_end"`):
+            Truncation mode to use when the prompt is too long. Possible values are `"keep_end"` or `"keep_start"`.
+            This argument is required if you want to use the default data collator.
+        generate_during_eval (`bool`, *optional*, defaults to `False`):
+            If `True`, generates and logs completions from the model to W&B or Comet during evaluation.
+        is_encoder_decoder (`bool`, *optional*):
+            When using the `model_init` argument (callable) to instantiate the model instead of the `model` argument,
+            you need to specify if the model returned by the callable is an encoder-decoder model.
+        model_init_kwargs (`dict[str, Any]`, *optional*):
+            Keyword arguments to pass to `AutoModelForCausalLM.from_pretrained` when instantiating the model from a
+            string.
+        dataset_num_proc (`int`, *optional*):
+            Number of processes to use for processing the dataset.
+    
+    """
+    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,
+        max_length = 1024,
+        max_prompt_length = 512,
+        max_completion_length = None,
+        beta = 0.1,
+        disable_dropout = True,
+        label_pad_token_id = -100,
+        padding_value = None,
+        truncation_mode = 'keep_end',
+        generate_during_eval = False,
+        is_encoder_decoder = None,
+        model_init_kwargs = None,
+        dataset_num_proc = 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)
+        
+        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,
+            max_length = max_length,
+            max_prompt_length = max_prompt_length,
+            max_completion_length = max_completion_length,
+            beta = beta,
+            disable_dropout = disable_dropout,
+            label_pad_token_id = label_pad_token_id,
+            padding_value = padding_value,
+            truncation_mode = truncation_mode,
+            generate_during_eval = generate_during_eval,
+            is_encoder_decoder = is_encoder_decoder,
+            model_init_kwargs = model_init_kwargs,
+            dataset_num_proc = dataset_num_proc,**kwargs)
+        self.vllm_sampling_params = vllm_sampling_params
+        self.unsloth_num_chunks = unsloth_num_chunks
+        self.max_seq_length = max_seq_length
+pass
+
+class _UnslothORPOTrainer(BaseTrainer):
+    r""""""
+
+    _tag_names = ["trl", "orpo"]
+    _name = "ORPO"
+    _paper = {
+        "title": "ORPO: Monolithic Preference Optimization without Reference Model",
+        "id": "2403.07691",
+        # docstyle-ignore
+        "citation": textwrap.dedent("""\
+            @article{hong2024orpo,
+                title        = {{ORPO: Monolithic Preference Optimization without Reference Model}},
+                author       = {Jiwoo Hong and Noah Lee and James Thorne},
+                year         = 2024,
+                eprint       = {arXiv:2403.07691}
+            }"""),
+    }
+
+    def __init__(
+        self,
+        model: Optional[Union[PreTrainedModel, nn.Module, str]] = None,
+        args: Optional[ORPOConfig] = 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,
+        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,
+        compute_metrics: Optional[Callable[[EvalLoopOutput], dict]] = None,
+    ):
+        if not os.environ.get("TRL_EXPERIMENTAL_SILENCE"):
+            warnings.warn(
+                "This trainer will soon be moved to trl.experimental and is a candidate for removal. If you rely on "
+                "it and want it to remain, please share your comments here: "
+                "https://github.com/huggingface/trl/issues/4223. Silence this warning by setting environment variable "
+                "TRL_EXPERIMENTAL_SILENCE=1."
+            )
+        if args.model_init_kwargs is None:
+            model_init_kwargs = {}
+        elif not isinstance(model, str):
+            raise ValueError("You passed model_kwargs to the ORPOTrainer. But your model is already instantiated.")
+        else:
+            model_init_kwargs = args.model_init_kwargs
+            dtype = model_init_kwargs.get("dtype")
+            if dtype is not None:
+                # Convert to `torch.dtype` if an str is passed
+                if isinstance(dtype, str) and dtype != "auto":
+                    dtype = getattr(torch, dtype)
+                if dtype != "auto" and not isinstance(dtype, torch.dtype):
+                    raise ValueError(
+                        f"Invalid `dtype` passed to the ORPOConfig. Expected a string with either `torch.dtype` or 'auto', but got {dtype}."
+                    )
+                model_init_kwargs["dtype"] = dtype
+
+        if isinstance(model, str):
+            model = AutoModelForCausalLM.from_pretrained(model, **model_init_kwargs)
+
+        # Initialize this variable to False. This helps tracking the case when `peft_module_casting_to_bf16`
+        # has been called in order to properly call autocast if needed.
+        self._peft_has_been_casted_to_bf16 = False
+
+        if not is_peft_available() and peft_config is not None:
+            raise ValueError(
+                "PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it to use the PEFT models"
+            )
+        elif is_peft_available() and peft_config is not None:
+            # if model is a peft model and we have a peft_config, we merge and unload it first
+            if isinstance(model, PeftModel):
+                model = model.merge_and_unload()
+
+            if getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_loaded_in_4bit", False):
+                _support_gc_kwargs = hasattr(
+                    args, "gradient_checkpointing_kwargs"
+                ) and "gradient_checkpointing_kwargs" in list(
+                    inspect.signature(prepare_model_for_kbit_training).parameters
+                )
+
+                prepare_model_kwargs = {"use_gradient_checkpointing": args.gradient_checkpointing}
+
+                if _support_gc_kwargs:
+                    prepare_model_kwargs["gradient_checkpointing_kwargs"] = args.gradient_checkpointing_kwargs
+
+                model = prepare_model_for_kbit_training(model, **prepare_model_kwargs)
+            elif args.gradient_checkpointing:
+                # For backward compatibility with older versions of transformers
+                if hasattr(model, "enable_input_require_grads"):
+                    model.enable_input_require_grads()
+                else:
+
+                    def make_inputs_require_grad(module, input, output):
+                        output.requires_grad_(True)
+
+                    model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
+
+            # get peft model with the given config
+            model = model
+            if args.bf16 and getattr(model, "is_loaded_in_4bit", False):
+                peft_module_casting_to_bf16(model)
+                # If args.bf16 we need to explicitly call `generate` with torch amp autocast context manager
+                self._peft_has_been_casted_to_bf16 = True
+
+        # For models that use gradient_checkpointing, we need to attach a hook that enables input
+        # to explicitly have `requires_grad=True`, otherwise training will either silently
+        # fail or completely fail.
+        elif args.gradient_checkpointing:
+            # For backward compatibility with older versions of transformers
+            if hasattr(model, "enable_input_require_grads"):
+                model.enable_input_require_grads()
+            else:
+
+                def make_inputs_require_grad(module, input, output):
+                    output.requires_grad_(True)
+
+                model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
+
+        if args.generate_during_eval and not (is_wandb_available() or is_comet_available()):
+            raise ValueError(
+                "`generate_during_eval=True` requires Weights and Biases or Comet to be installed."
+                " Please install `wandb` or `comet-ml` to resolve."
+            )
+
+        if model is not None:
+            self.is_encoder_decoder = model.config.is_encoder_decoder
+        elif args.is_encoder_decoder is None:
+            raise ValueError("When no model is provided, you need to pass the parameter is_encoder_decoder.")
+        else:
+            self.is_encoder_decoder = args.is_encoder_decoder
+
+        if self.is_encoder_decoder:
+            self.decoder_start_token_id = model.config.decoder_start_token_id
+            self.pad_token_id = model.config.pad_token_id
+
+        if processing_class is None:
+            raise ValueError("processing_class must be specified to tokenize a ORPO dataset.")
+        if args.max_length is None:
+            logger.warning(
+                "`max_length` is not set in the ORPOConfig's init"
+                " it will default to `512` by default, but you should do it yourself in the future.",
+            )
+            max_length = 512
+        else:
+            max_length = args.max_length
+        if args.max_prompt_length is None:
+            logger.warning(
+                "`max_prompt_length` is not set in the ORPOConfig's init"
+                " it will default to `128` by default, but you should do it yourself in the future.",
+            )
+            max_prompt_length = 128
+        else:
+            max_prompt_length = args.max_prompt_length
+
+        if args.max_completion_length is None and self.is_encoder_decoder:
+            logger.warning(
+                "When using an encoder decoder architecture, you should set `max_completion_length` in the ORPOConfig's init"
+                " it will default to `128` by default, but you should do it yourself in the future.",
+            )
+            self.max_completion_length = 128
+        else:
+            self.max_completion_length = args.max_completion_length
+
+        if data_collator is None:
+            data_collator = DPODataCollatorWithPadding(
+                pad_token_id=processing_class.pad_token_id,
+                label_pad_token_id=args.label_pad_token_id,
+                is_encoder_decoder=self.is_encoder_decoder,
+            )
+
+            if args.remove_unused_columns:
+                args.remove_unused_columns = False
+                # warn users
+                logger.warning(
+                    "When using DPODataCollatorWithPadding, you should set `remove_unused_columns=False` in your TrainingArguments"
+                    " we have set it for you, but you should do it yourself in the future.",
+                )
+
+            self.use_dpo_data_collator = True
+        else:
+            self.use_dpo_data_collator = False
+
+        # Disable dropout in the model and reference model
+        if args.disable_dropout:
+            disable_dropout_in_model(model)
+
+        self.max_length = max_length
+        self.generate_during_eval = args.generate_during_eval
+        self.label_pad_token_id = args.label_pad_token_id
+        self.padding_value = args.padding_value if args.padding_value is not None else processing_class.pad_token_id
+        self.max_prompt_length = max_prompt_length
+        self.truncation_mode = args.truncation_mode
+        self.processing_class = processing_class
+
+        self.beta = args.beta
+        self.aux_loss_enabled = getattr(model.config, "output_router_logits", False)
+        self.aux_loss_coef = getattr(model.config, "router_aux_loss_coef", 0.0)
+        if self.aux_loss_enabled and self.aux_loss_coef == 0.0:
+            logger.warning(
+                "You set `output_router_logits` to `True` in the model config, but `router_aux_loss_coef` is set to "
+                "`0.0`, meaning the auxiliary loss will not be used. Either set `router_aux_loss_coef` to a value "
+                "greater than `0.0`, or set `output_router_logits` to `False` if you don't want to use the auxiliary "
+                "loss.",
+            )
+
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+
+        # 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 ORPO, the sampled data does not include the
+        # "input_ids" key. Instead, the available keys are "prompt_input_ids", "chosen_input_ids", and
+        # "rejected_input_ids". 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
+
+        # Compute that only on the main process for faster data processing.
+        # see: https://github.com/huggingface/trl/pull/1255
+        with PartialState().main_process_first():
+            # Extract the prompt if needed, and apply the chat template if needed
+            train_dataset = train_dataset.map(maybe_extract_prompt, num_proc=args.dataset_num_proc)
+            train_dataset = train_dataset.map(
+                maybe_apply_chat_template, fn_kwargs={"tokenizer": processing_class}, num_proc=args.dataset_num_proc
+            )
+            train_dataset = train_dataset.map(self.tokenize_row, num_proc=args.dataset_num_proc)
+            if eval_dataset is not None:
+                eval_dataset = eval_dataset.map(maybe_extract_prompt, num_proc=args.dataset_num_proc)
+                eval_dataset = eval_dataset.map(
+                    maybe_apply_chat_template,
+                    fn_kwargs={"tokenizer": processing_class},
+                    num_proc=args.dataset_num_proc,
+                )
+                eval_dataset = eval_dataset.map(self.tokenize_row, 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,
+        )
+
+        # Gradient accumulation requires scaled loss. Normally, loss scaling in the parent class depends on whether the
+        # model accepts loss-related kwargs. Since we compute our own loss, this check is irrelevant. We set
+        # self.model_accepts_loss_kwargs to False to enable scaling.
+        self.model_accepts_loss_kwargs = False
+
+        # 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)
+
+        if not hasattr(self, "accelerator"):
+            raise AttributeError(
+                "Your `Trainer` does not have an `accelerator` object. Consider upgrading `transformers`."
+            )
+
+    def build_tokenized_answer(self, prompt, answer):
+        """
+        Llama tokenizer does satisfy `enc(a + b) = enc(a) + enc(b)`. It does ensure `enc(a + b) = enc(a) + enc(a +
+        b)[len(enc(a)):]`. Reference:
+            https://github.com/EleutherAI/lm-evaluation-harness/pull/531#issuecomment-1595586257
+        """
+
+        full_tokenized = self.processing_class(prompt + answer, add_special_tokens=False)
+        prompt_input_ids = self.processing_class(prompt, add_special_tokens=False)["input_ids"]
+
+        answer_input_ids = full_tokenized["input_ids"][len(prompt_input_ids) :]
+        answer_attention_mask = full_tokenized["attention_mask"][len(prompt_input_ids) :]
+
+        # Concat tokens to form `enc(a) + enc(a + b)[len(enc(a)):]`
+        full_concat_input_ids = np.concatenate([prompt_input_ids, answer_input_ids])
+
+        # Prepare input tokens for token by token comparison
+        full_input_ids = np.array(full_tokenized["input_ids"])
+
+        if len(full_input_ids) != len(full_concat_input_ids):
+            raise ValueError("Prompt input ids and answer input ids should have the same length.")
+
+        # On some tokenizers, like Llama-2 tokenizer, there are occasions where tokens
+        # can be merged together when tokenizing prompt+answer. This could result
+        # on the last token from the prompt being different when tokenized on its own
+        # vs when done as prompt+answer.
+        response_token_ids_start_idx = len(prompt_input_ids)
+
+        # If tokenized prompt is different than both prompt+answer, then it means the
+        # last token has changed due to merging.
+        if prompt_input_ids != full_tokenized["input_ids"][:response_token_ids_start_idx]:
+            response_token_ids_start_idx -= 1
+
+        prompt_input_ids = full_tokenized["input_ids"][:response_token_ids_start_idx]
+        prompt_attention_mask = full_tokenized["attention_mask"][:response_token_ids_start_idx]
+
+        if len(prompt_input_ids) != len(prompt_attention_mask):
+            raise ValueError("Prompt input ids and attention mask should have the same length.")
+
+        answer_input_ids = full_tokenized["input_ids"][response_token_ids_start_idx:]
+        answer_attention_mask = full_tokenized["attention_mask"][response_token_ids_start_idx:]
+
+        return dict(
+            prompt_input_ids=prompt_input_ids,
+            prompt_attention_mask=prompt_attention_mask,
+            input_ids=answer_input_ids,
+            attention_mask=answer_attention_mask,
+        )
+
+    def tokenize_row(self, feature, model: Optional[Union[PreTrainedModel, nn.Module]] = None) -> dict:
+        """Tokenize a single row from a ORPO specific dataset.
+
+        At this stage, we don't convert to PyTorch tensors yet; we just handle the truncation in case the prompt +
+        chosen or prompt + rejected responses is/are too long. First we truncate the prompt; if we're still too long,
+        we truncate the chosen/rejected.
+
+        We also create the labels for the chosen/rejected responses, which are of length equal to the sum of the length
+        of the prompt and the chosen/rejected response, with label_pad_token_id for the prompt tokens.
+        """
+        batch = {}
+        prompt = feature["prompt"]
+        chosen = feature["chosen"]
+        rejected = feature["rejected"]
+
+        if not self.is_encoder_decoder:
+            # Check issues below for more details
+            #  1. https://github.com/huggingface/trl/issues/907
+            #  2. https://github.com/EleutherAI/lm-evaluation-harness/pull/531#issuecomment-1595586257
+            #  3. https://github.com/LianjiaTech/BELLE/issues/337
+
+            if not isinstance(prompt, str):
+                raise ValueError(f"prompt should be an str but got {type(prompt)}")
+            prompt_tokens = self.processing_class(prompt, add_special_tokens=False)
+            prompt_tokens = {f"prompt_{k}": v for k, v in prompt_tokens.items()}
+
+            if not isinstance(chosen, str):
+                raise ValueError(f"chosen should be an str but got {type(chosen)}")
+            chosen_tokens = self.build_tokenized_answer(prompt, chosen)
+
+            if not isinstance(rejected, str):
+                raise ValueError(f"rejected should be an str but got {type(rejected)}")
+            rejected_tokens = self.build_tokenized_answer(prompt, rejected)
+
+            # Last prompt token might get merged by tokenizer and
+            # it should not be included for generation if that happens
+            prompt_len_input_ids = len(prompt_tokens["prompt_input_ids"])
+
+            chosen_prompt_len_input_ids = len(chosen_tokens["prompt_input_ids"])
+            rejected_prompt_len_input_ids = len(rejected_tokens["prompt_input_ids"])
+            prompt_len_input_ids = min(chosen_prompt_len_input_ids, rejected_prompt_len_input_ids)
+
+            for k, v in prompt_tokens.items():
+                prompt_tokens[k] = v[:prompt_len_input_ids]
+
+            # Make sure prompts only have one different token at most an
+            # and length only differs by 1 at most
+            num_diff_tokens = sum(
+                a != b for a, b in zip(chosen_tokens["prompt_input_ids"], rejected_tokens["prompt_input_ids"])
+            )
+            num_diff_len = abs(chosen_prompt_len_input_ids - rejected_prompt_len_input_ids)
+            if num_diff_tokens > 1 or num_diff_len > 1:
+                raise ValueError(
+                    "Chosen and rejected prompt_input_ids might only differ on the "
+                    "last token due to tokenizer merge ops."
+                )
+
+            # add BOS token to head of prompt. Avoid adding if it's already there
+            prompt_tokens, chosen_tokens, rejected_tokens = add_bos_token_if_needed(
+                self.processing_class.bos_token_id,
+                prompt_len_input_ids,
+                prompt_tokens,
+                chosen_prompt_len_input_ids,
+                chosen_tokens,
+                rejected_prompt_len_input_ids,
+                rejected_tokens,
+            )
+
+            # add EOS token to end of answer. Avoid adding if it's already there
+            chosen_tokens, rejected_tokens = add_eos_token_if_needed(
+                self.processing_class.eos_token_id, chosen_tokens, rejected_tokens
+            )
+
+            longer_response_length = max(len(chosen_tokens["input_ids"]), len(rejected_tokens["input_ids"]))
+
+            # if combined sequence is too long, truncate the prompt
+            for answer_tokens in [chosen_tokens, rejected_tokens, prompt_tokens]:
+                if len(answer_tokens["prompt_input_ids"]) + longer_response_length > self.max_length:
+                    if self.truncation_mode == "keep_start":
+                        for k in ["prompt_input_ids", "prompt_attention_mask"]:
+                            answer_tokens[k] = answer_tokens[k][: self.max_prompt_length]
+                    elif self.truncation_mode == "keep_end":
+                        for k in ["prompt_input_ids", "prompt_attention_mask"]:
+                            answer_tokens[k] = answer_tokens[k][-self.max_prompt_length :]
+                    else:
+                        raise ValueError(f"Unknown truncation mode: {self.truncation_mode}")
+
+            # if that's still too long, truncate the response
+            for answer_tokens in [chosen_tokens, rejected_tokens]:
+                if len(answer_tokens["prompt_input_ids"]) + longer_response_length > self.max_length:
+                    for k in ["input_ids", "attention_mask"]:
+                        answer_tokens[k] = answer_tokens[k][: self.max_length - self.max_prompt_length]
+
+            # Create labels
+            chosen_sequence_tokens = {
+                k: chosen_tokens[f"prompt_{k}"] + chosen_tokens[k] for k in ["input_ids", "attention_mask"]
+            }
+            rejected_sequence_tokens = {
+                k: rejected_tokens[f"prompt_{k}"] + rejected_tokens[k] for k in ["input_ids", "attention_mask"]
+            }
+            chosen_sequence_tokens["labels"] = chosen_sequence_tokens["input_ids"][:]
+            chosen_sequence_tokens["labels"][: len(chosen_tokens["prompt_input_ids"])] = [
+                self.label_pad_token_id
+            ] * len(chosen_tokens["prompt_input_ids"])
+            rejected_sequence_tokens["labels"] = rejected_sequence_tokens["input_ids"][:]
+            rejected_sequence_tokens["labels"][: len(rejected_tokens["prompt_input_ids"])] = [
+                self.label_pad_token_id
+            ] * len(rejected_tokens["prompt_input_ids"])
+
+            for k, toks in {
+                "chosen_": chosen_sequence_tokens,
+                "rejected_": rejected_sequence_tokens,
+                "": prompt_tokens,
+            }.items():
+                for type_key, tokens in toks.items():
+                    if type_key == "token_type_ids":
+                        continue
+                    batch[f"{k}{type_key}"] = tokens
+
+        else:
+            chosen_tokens = self.processing_class(
+                chosen, truncation=True, max_length=self.max_completion_length, add_special_tokens=True
+            )
+            rejected_tokens = self.processing_class(
+                rejected, truncation=True, max_length=self.max_completion_length, add_special_tokens=True
+            )
+            prompt_tokens = self.processing_class(
+                prompt, truncation=True, max_length=self.max_prompt_length, add_special_tokens=True
+            )
+
+            batch["chosen_labels"] = chosen_tokens["input_ids"]
+            batch["rejected_labels"] = rejected_tokens["input_ids"]
+            batch["prompt_input_ids"] = prompt_tokens["input_ids"]
+            batch["prompt_attention_mask"] = prompt_tokens["attention_mask"]
+
+            if model is not None and hasattr(model, "prepare_decoder_input_ids_from_labels"):
+                batch["rejected_decoder_input_ids"] = model.prepare_decoder_input_ids_from_labels(
+                    labels=torch.tensor(batch["rejected_labels"])
+                )
+                batch["chosen_decoder_input_ids"] = model.prepare_decoder_input_ids_from_labels(
+                    labels=torch.tensor(batch["chosen_labels"])
+                )
+
+        if is_torch_xla_available():
+            # Pad the sequences to global max_length to avoid TorchXLA recompilation
+            for k in batch:
+                if "labels" in k or self.is_encoder_decoder:
+                    pad_value = self.label_pad_token_id
+                elif k.endswith("_input_ids"):
+                    pad_value = self.padding_value
+                elif k.endswith("_attention_mask"):
+                    pad_value = 0
+                batch[k] = batch[k] + [pad_value] * (self.max_length - len(batch[k]))
+        return batch
+
+    @staticmethod
+    def concatenated_inputs(
+        batch: dict[str, Union[list, torch.LongTensor]],
+        is_encoder_decoder: bool = False,
+        label_pad_token_id: int = -100,
+        padding_value: int = 0,
+        device: Optional[torch.device] = None,
+    ) -> dict[str, torch.LongTensor]:
+        """Concatenate the chosen and rejected inputs into a single tensor.
+
+        Args:
+            batch:
+                A batch of data. Must contain the keys 'chosen_input_ids' and 'rejected_input_ids', which are tensors
+                of shape (batch_size, sequence_length).
+            is_encoder_decoder:
+                Whether the model is an encoder-decoder model.
+            label_pad_token_id:
+                The label pad token id.
+            padding_value:
+                The padding value to use for the concatenated inputs_ids.
+            device:
+                The device for the concatenated inputs.
+
+        Returns:
+            A dictionary containing the concatenated inputs under the key 'concatenated_input_ids'.
+        """
+        concatenated_batch = {}
+
+        if is_encoder_decoder:
+            max_length = max(batch["chosen_labels"].shape[1], batch["rejected_labels"].shape[1])
+        else:
+            max_length = max(batch["chosen_input_ids"].shape[1], batch["rejected_input_ids"].shape[1])
+
+        for k in batch:
+            if k.startswith("chosen") and isinstance(batch[k], torch.Tensor):
+                if "labels" in k or is_encoder_decoder:
+                    pad_value = label_pad_token_id
+                elif k.endswith("_input_ids"):
+                    pad_value = padding_value
+                elif k.endswith("_attention_mask"):
+                    pad_value = 0
+                concatenated_key = k.replace("chosen", "concatenated")
+                concatenated_batch[concatenated_key] = pad_to_length(batch[k], max_length, pad_value=pad_value)
+        for k in batch:
+            if k.startswith("rejected") and isinstance(batch[k], torch.Tensor):
+                if "labels" in k or is_encoder_decoder:
+                    pad_value = label_pad_token_id
+                elif k.endswith("_input_ids"):
+                    pad_value = padding_value
+                elif k.endswith("_attention_mask"):
+                    pad_value = 0
+                concatenated_key = k.replace("rejected", "concatenated")
+                concatenated_batch[concatenated_key] = torch.cat(
+                    (
+                        concatenated_batch[concatenated_key],
+                        pad_to_length(batch[k], max_length, pad_value=pad_value),
+                    ),
+                    dim=0,
+                ).to(device=device)
+
+        if is_encoder_decoder:
+            concatenated_batch["concatenated_input_ids"] = batch["prompt_input_ids"].repeat(2, 1).to(device=device)
+            concatenated_batch["concatenated_attention_mask"] = (
+                batch["prompt_attention_mask"].repeat(2, 1).to(device=device)
+            )
+
+        return concatenated_batch
+
+    def odds_ratio_loss(
+        self,
+        policy_chosen_logps: torch.FloatTensor,
+        policy_rejected_logps: torch.FloatTensor,
+    ) -> tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
+        """Compute ORPO's odds ratio (OR) loss for a batch of policy and reference model log probabilities.
+
+        Args:
+            policy_chosen_logps:
+                Log probabilities of the policy model for the chosen responses. Shape: (batch_size,)
+            policy_rejected_logps:
+                Log probabilities of the policy model for the rejected responses. Shape: (batch_size,)
+
+        Returns:
+            A tuple of three tensors: (losses, chosen_rewards, rejected_rewards). The losses tensor contains the ORPO
+            loss for each example in the batch. The chosen_rewards and rejected_rewards tensors contain the rewards for
+            the chosen and rejected responses, respectively. The log odds ratio of the chosen responses over the
+            rejected responses ratio for logging purposes. The `log(sigmoid(log_odds_chosen))` for logging purposes.
+        """
+
+        # Derived from Eqs. (4) and (7) from https://huggingface.co/papers/2403.07691 by using log identities and exp(log(P(y|x)) = P(y|x)
+        log_odds = (policy_chosen_logps - policy_rejected_logps) - (
+            torch.log1p(-torch.exp(policy_chosen_logps)) - torch.log1p(-torch.exp(policy_rejected_logps))
+        )
+        ratio = F.logsigmoid(log_odds)
+        losses = self.beta * ratio
+
+        chosen_rewards = self.beta * (policy_chosen_logps.to(self.accelerator.device)).detach()
+        rejected_rewards = self.beta * (policy_rejected_logps.to(self.accelerator.device)).detach()
+
+        return losses, chosen_rewards, rejected_rewards, torch.mean(ratio), torch.mean(log_odds)
+
+    @staticmethod
+    def get_batch_logps(
+        logits: torch.FloatTensor,
+        labels: torch.LongTensor,
+        average_log_prob: bool = False,
+        label_pad_token_id: int = -100,
+        is_encoder_decoder: bool = False,
+    ) -> torch.FloatTensor:
+        """Compute the log probabilities of the given labels under the given logits.
+
+        Args:
+            logits: Logits of the model (unnormalized). Shape: (batch_size, sequence_length, vocab_size)
+            labels:
+                Labels for which to compute the log probabilities. Label tokens with a value of label_pad_token_id are
+                ignored. Shape: (batch_size, sequence_length)
+            average_log_prob:
+                If True, return the average log probability per (non-masked) token. Otherwise, return the sum of the
+                log probabilities of the (non-masked) tokens.
+            label_pad_token_id: The label pad token id.
+            is_encoder_decoder: Whether the model is an encoder-decoder model.
+
+        Returns:
+            A tensor of shape (batch_size,) containing the average/sum log probabilities of the given labels under the
+            given logits.
+        """
+        if logits.shape[:-1] != labels.shape:
+            raise ValueError("Logits (batch and sequence length dim) and labels must have the same shape.")
+
+        if not is_encoder_decoder:
+            labels = labels[:, 1:].clone()
+            logits = logits[:, :-1, :]
+        loss_mask = labels != label_pad_token_id
+
+        # dummy token; we'll ignore the losses on these tokens later
+        labels = torch.where(labels == label_pad_token_id, 0, labels)
+
+        per_token_logps = selective_log_softmax(logits, labels)
+
+        if average_log_prob:
+            return (per_token_logps * loss_mask).sum(-1) / loss_mask.sum(-1)
+        else:
+            return (per_token_logps * loss_mask).sum(-1)
+
+    def concatenated_forward(
+        self, model: nn.Module, batch: dict[str, Union[list, torch.LongTensor]]
+    ) -> tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
+        """Run the given model on the given batch of inputs, concatenating the chosen and rejected inputs together.
+
+        We do this to avoid doing two forward passes, because it's faster for FSDP.
+        """
+        concatenated_batch = self.concatenated_inputs(
+            batch,
+            is_encoder_decoder=self.is_encoder_decoder,
+            label_pad_token_id=self.label_pad_token_id,
+            padding_value=self.padding_value,
+            device=self.accelerator.device,
+        )
+        len_chosen = batch["chosen_labels"].shape[0]
+
+        model_kwargs = (
+            {
+                "decoder_input_ids": self._shift_right(concatenated_batch["concatenated_labels"]),
+            }
+            if self.is_encoder_decoder
+            else {}
+        )
+
+        if self.aux_loss_enabled:
+            model_kwargs["output_router_logits"] = True
+
+        outputs = model(
+            concatenated_batch["concatenated_input_ids"],
+            attention_mask=concatenated_batch["concatenated_attention_mask"],
+            use_cache=False,
+            **model_kwargs,
+        )
+        all_logits = outputs.logits
+
+        def cross_entropy_loss(logits, labels):
+            if not self.is_encoder_decoder:
+                # Shift so that tokens < n predict n
+                logits = logits[..., :-1, :].contiguous()
+                labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = nn.CrossEntropyLoss()
+            logits = logits.view(-1, logits.shape[-1])
+            labels = labels.view(-1)
+            # Enable model parallelism
+            labels = labels.to(logits.device)
+            loss = loss_fct(logits, labels)
+            return loss
+
+        if self.is_encoder_decoder:
+            labels = concatenated_batch["concatenated_labels"].clone()
+        else:
+            labels = concatenated_batch["concatenated_input_ids"].clone()
+            attention_mask = concatenated_batch["concatenated_attention_mask"]
+            labels = torch.where(attention_mask == 1, labels, self.label_pad_token_id)
+        # orpo chosen nll loss is computed over the full prompt and response
+        chosen_nll_loss = cross_entropy_loss(all_logits[:len_chosen], labels[:len_chosen])
+
+        all_logps = self.get_batch_logps(
+            all_logits,
+            concatenated_batch["concatenated_labels"],
+            average_log_prob=True,
+            is_encoder_decoder=self.is_encoder_decoder,
+            label_pad_token_id=self.label_pad_token_id,
+        )
+
+        chosen_logps = all_logps[:len_chosen]
+        rejected_logps = all_logps[len_chosen:]
+
+        if not self.is_encoder_decoder:
+            chosen_logits = all_logits[:len_chosen, :-1, :]
+            rejected_logits = all_logits[len_chosen:, :-1, :]
+        else:
+            chosen_logits = all_logits[:len_chosen]
+            rejected_logits = all_logits[len_chosen:]
+
+        if self.aux_loss_enabled:
+            return (chosen_logps, rejected_logps, chosen_logits, rejected_logits, chosen_nll_loss, outputs.aux_loss)
+
+        return (chosen_logps, rejected_logps, chosen_logits, rejected_logits, chosen_nll_loss)
+
+    def get_batch_loss_metrics(
+        self,
+        model,
+        batch: dict[str, Union[list, torch.LongTensor]],
+        train_eval: Literal["train", "eval"] = "train",
+    ):
+        """Compute the ORPO loss and other metrics for the given batch of inputs for train or test."""
+        metrics = {}
+
+        forward_output = self.concatenated_forward(model, batch)
+        (
+            policy_chosen_logps,
+            policy_rejected_logps,
+            policy_chosen_logits,
+            policy_rejected_logits,
+            policy_nll_loss,
+        ) = forward_output[:5]
+        if self.aux_loss_enabled:
+            aux_loss = forward_output[5]
+
+        losses, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen = self.odds_ratio_loss(
+            policy_chosen_logps, policy_rejected_logps
+        )
+        # full ORPO loss
+        loss = policy_nll_loss - losses.mean()
+
+        reward_accuracies = (chosen_rewards > rejected_rewards).float()
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics[f"{prefix}rewards/chosen"] = self.accelerator.gather_for_metrics(chosen_rewards).mean()
+        metrics[f"{prefix}rewards/rejected"] = self.accelerator.gather_for_metrics(rejected_rewards).mean()
+        metrics[f"{prefix}rewards/accuracies"] = self.accelerator.gather_for_metrics(reward_accuracies).mean()
+        metrics[f"{prefix}rewards/margins"] = self.accelerator.gather_for_metrics(
+            chosen_rewards - rejected_rewards
+        ).mean()
+        metrics[f"{prefix}logps/rejected"] = self.accelerator.gather_for_metrics(policy_rejected_logps).detach().mean()
+        metrics[f"{prefix}logps/chosen"] = self.accelerator.gather_for_metrics(policy_chosen_logps).detach().mean()
+        metrics[f"{prefix}logits/rejected"] = self.accelerator.gather_for_metrics(
+            policy_rejected_logits.detach().mean()
+        ).mean()
+        metrics[f"{prefix}logits/chosen"] = self.accelerator.gather_for_metrics(
+            policy_chosen_logits.detach().mean()
+        ).mean()
+        metrics[f"{prefix}nll_loss"] = self.accelerator.gather_for_metrics(policy_nll_loss).detach().mean()
+        metrics[f"{prefix}log_odds_ratio"] = self.accelerator.gather_for_metrics(log_odds_ratio).detach().mean()
+        metrics[f"{prefix}log_odds_chosen"] = self.accelerator.gather_for_metrics(log_odds_chosen).detach().mean()
+        if is_torch_xla_available():
+            xm.mark_step()  # needed because .item() calls
+        for k, v in metrics.items():
+            metrics[k] = v.item()
+        if self.aux_loss_enabled:
+            loss += self.aux_loss_coef * aux_loss
+
+        return loss, metrics
+
+    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]]]:
+        compute_loss_context_manager = (
+            autocast(self.accelerator.device.type) if self._peft_has_been_casted_to_bf16 else nullcontext()
+        )
+
+        with compute_loss_context_manager:
+            loss, metrics = self.get_batch_loss_metrics(model, inputs, train_eval="train")
+
+        # Make sure to move the loss to the device the original accumulating loss is at back in the `Trainer` class:
+        loss = loss.to(self.args.device)
+
+        # force log the metrics
+        self.store_metrics(metrics, train_eval="train")
+
+        if return_outputs:
+            return (loss, metrics)
+        return loss
+
+    def generate_from_model(self, model, batch: dict[str, torch.LongTensor]) -> str:
+        """Generate samples from the model and reference model for the given batch of inputs."""
+
+        # If one uses `generate_during_eval` with peft + bf16, we need to explicitly call generate with
+        # the torch amp context manager as some hidden states are silently casted to full precision.
+        generate_context_manager = (
+            autocast(self.accelerator.device.type) if self._peft_has_been_casted_to_bf16 else nullcontext()
+        )
+
+        with generate_context_manager:
+            policy_output = model.generate(
+                input_ids=batch["prompt_input_ids"],
+                attention_mask=batch["prompt_attention_mask"],
+                max_length=self.max_length,
+                do_sample=True,
+                pad_token_id=self.processing_class.pad_token_id,
+            )
+
+        policy_output = pad_to_length(policy_output, self.max_length, self.processing_class.pad_token_id)
+        policy_output_decoded = self.processing_class.batch_decode(policy_output, skip_special_tokens=True)
+
+        return policy_output_decoded
+
+    def prediction_step(
+        self,
+        model: Union[PreTrainedModel, nn.Module],
+        inputs: dict[str, Union[torch.Tensor, Any]],
+        prediction_loss_only: bool,
+        ignore_keys: Optional[list[str]] = None,
+    ):
+        if not self.use_dpo_data_collator:
+            logger.warning(
+                "prediction_step is only implemented for DPODataCollatorWithPadding, and you passed a datacollator that is different than "
+                "DPODataCollatorWithPadding - you might see unexpected behavior. Alternatively, you can implement your own prediction_step method if you are using a custom data collator"
+            )
+        if ignore_keys is None:
+            if hasattr(model, "config"):
+                ignore_keys = getattr(model.config, "keys_to_ignore_at_inference", [])
+            else:
+                ignore_keys = []
+
+        prediction_context_manager = (
+            autocast(self.accelerator.device.type) if self._peft_has_been_casted_to_bf16 else nullcontext()
+        )
+
+        with torch.no_grad(), prediction_context_manager:
+            loss, metrics = self.get_batch_loss_metrics(model, inputs, train_eval="eval")
+
+        # force log the metrics
+        self.store_metrics(metrics, train_eval="eval")
+
+        if prediction_loss_only:
+            return (loss.detach(), None, None)
+
+        # logits for the chosen and rejected samples from model
+        logits_dict = {
+            "eval_logits/chosen": metrics["eval_logits/chosen"],
+            "eval_logits/rejected": metrics["eval_logits/rejected"],
+        }
+        logits = [v for k, v in logits_dict.items() if k not in ignore_keys]
+        logits = torch.tensor(logits, device=self.accelerator.device)
+        labels = torch.zeros(logits.shape[0], device=self.accelerator.device)
+
+        return (loss.detach(), logits, labels)
+
+    def store_metrics(self, metrics: dict[str, float], train_eval: Literal["train", "eval"] = "train") -> None:
+        for key, value in metrics.items():
+            self._stored_metrics[train_eval][key].append(value)
+
+    def evaluation_loop(
+        self,
+        dataloader: DataLoader,
+        description: str,
+        prediction_loss_only: Optional[bool] = None,
+        ignore_keys: Optional[list[str]] = None,
+        metric_key_prefix: str = "eval",
+    ) -> EvalLoopOutput:
+        """
+        Overriding built-in evaluation loop to store metrics for each batch. Prediction/evaluation loop, shared by
+        `Trainer.evaluate()` and `Trainer.predict()`.
+
+        Works both with or without labels.
+        """
+
+        # Sample and save to game log if requested (for one batch to save time)
+        if self.generate_during_eval:
+            # Generate random indices within the range of the total number of samples
+            num_samples = len(dataloader.dataset)
+            random_indices = random.sample(range(num_samples), k=self.args.eval_batch_size)
+
+            # Use dataloader.dataset.select to get the random batch without iterating over the DataLoader
+            random_batch_dataset = dataloader.dataset.select(random_indices)
+            random_batch = self.data_collator(random_batch_dataset)
+            random_batch = self._prepare_inputs(random_batch)
+
+            policy_output_decoded = self.generate_from_model(self.model, random_batch)
+
+            table = pd.DataFrame(
+                columns=["Prompt", "Policy"],
+                data=[
+                    [prompt, pol[len(prompt) :]] for prompt, pol in zip(random_batch["prompt"], policy_output_decoded)
+                ],
+            )
+            if "wandb" in self.args.report_to:
+                wandb.log({"game_log": wandb.Table(data=table)})
+
+            if "comet_ml" in self.args.report_to:
+                log_table_to_comet_experiment(
+                    name="game_log.csv",
+                    table=table,
+                )
+
+        # Base evaluation
+        initial_output = super().evaluation_loop(
+            dataloader, description, prediction_loss_only, ignore_keys, metric_key_prefix
+        )
+
+        return initial_output
+
+    def log(self, logs: dict[str, float], start_time: Optional[float] = None) -> None:
+        """
+        Log `logs` on the various objects watching training, including stored metrics.
+
+        Args:
+            logs (`dict[str, float]`):
+                The values to log.
+            start_time (`float`, *optional*):
+                Start time of the training.
+        """
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[key] = torch.tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+        return super().log(logs, start_time)
+
+    def _shift_right(self, input_ids):
+        if self.decoder_start_token_id is None:
+            raise ValueError(
+                "model.config.decoder_start_token_id has to be defined. It is usually set to the pad_token_id."
+            )
+
+        # shift inputs to the right
+        if is_torch_fx_proxy(input_ids):
+            # Item assignment is not supported natively for proxies.
+            shifted_input_ids = torch.full(input_ids.shape[:-1] + (1,), self.decoder_start_token_id)
+            shifted_input_ids = torch.cat([shifted_input_ids, input_ids[..., :-1]], dim=-1)
+        else:
+            shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+            shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
+            shifted_input_ids[..., 0] = self.decoder_start_token_id
+
+        if self.pad_token_id is None:
+            raise ValueError("model.config.pad_token_id has to be defined.")
+        # replace possible -100 values in labels by `pad_token_id`
+        shifted_input_ids.masked_fill_(shifted_input_ids == -100, self.pad_token_id)
+
+        return shifted_input_ids
+
+    # Ensure the model card is saved along with the checkpoint
+    def _save_checkpoint(self, model, trial):
+        if self.args.hub_model_id is None:
+            model_name = Path(self.args.output_dir).name
+        else:
+            model_name = self.args.hub_model_id.split("/")[-1]
+        self.create_model_card(model_name=model_name)
+        super()._save_checkpoint(model, trial)
+class UnslothORPOTrainer(_UnslothORPOTrainer):
+    """
+    
+    Initialize ORPOTrainer.
+
+    Args:
+        model ([`~transformers.PreTrainedModel`]):
+            The model to train, preferably an [`~transformers.AutoModelForSequenceClassification`].
+        args ([`ORPOConfig`]):
+            The ORPO 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.
+        model_init (`Callable[[], transformers.PreTrainedModel]`):
+            The model initializer to use for training. If None is specified, the default model initializer will be
+            used.
+        callbacks (`list[transformers.TrainerCallback]`):
+            The callbacks to use for training.
+        optimizers (`tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`):
+            The optimizer and scheduler to use for training.
+        preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`):
+            The function to use to preprocess the logits before computing the metrics.
+        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.
+        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.
+    
+    """
+    def __init__(
+        self,
+        model = None,
+        args = None,
+        data_collator = None,
+        train_dataset = None,
+        eval_dataset = None,
+        processing_class = None,
+        model_init = None,
+        callbacks = None,
+        preprocess_logits_for_metrics = None,
+        peft_config = None,
+        compute_metrics = None,
+        **kwargs
+    ):
+        if args is None: args = UnslothORPOConfig()
+        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('orpo_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,
+            args = args,
+            data_collator = data_collator,
+            train_dataset = train_dataset,
+            eval_dataset = eval_dataset,
+            processing_class = processing_class,
+            model_init = model_init,
+            callbacks = callbacks,
+            preprocess_logits_for_metrics = preprocess_logits_for_metrics,
+            peft_config = peft_config,
+            compute_metrics = compute_metrics,**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
+
+
+if hasattr(logger, "addFilter"):
+    import logging
+    class HideLoggingMessage(logging.Filter):
+        def __init__(self, text): self.text = text
+        def filter(self, x): return not (self.text in x.getMessage())
+    pass
+    logger.addFilter(HideLoggingMessage("`use_cache=True`"))
+