unsloth_compiled_cache/UnslothRLOOTrainer.py

"""
2025.11.3
2025.11.2
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.rloo_trainer import (Any, AutoConfig, AutoModelForSequenceClassification, AutoProcessor, AutoTokenizer, BaseTrainer, DataLoader, Dataset, FSDP, GenerationConfig, IterableDataset, Optional, Path, PeftConfig, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, RLOOConfig, RLOOTrainer, RepeatSampler, RewardFunc, Sampler, SyncRefModelCallback, TrainerCallback, Union, VLLMClient, apply_chat_template, broadcast_object_list, datasets, defaultdict, deque, disable_dropout_in_model, ensure_master_addr_port, entropy_from_logits, gather, gather_object, identity, inspect, is_conversational, is_datasets_available, is_flash_attn_2_available, is_peft_model, is_rich_available, is_vllm_available, logger, logging, maybe_apply_chat_template, nanmax, nanmin, nanstd, nn, nullcontext, os, pad, partial, prepare_deepspeed, prepare_fsdp, prepare_multimodal_messages, prepare_peft_model, print_prompt_completions_sample, profiling_context, profiling_decorator, seed_worker, selective_log_softmax, set_seed, shuffle_sequence_dict, split_pixel_values_by_grid, split_tensor_dict, textwrap, torch, transformers, unsplit_pixel_values_by_grid, unwrap_model_for_generation, wandb, warnings, FSDP, Optional, apply_chat_template, broadcast_object_list, gather, gather_object, is_flash_attn_2_available, maybe_apply_chat_template, nullcontext, os, pad, prepare_multimodal_messages, profiling_context, torch, transformers, unwrap_model_for_generation, FSDP, gather, is_peft_model, nn, nullcontext, os, profiling_decorator, Any, Union, profiling_decorator, shuffle_sequence_dict, split_pixel_values_by_grid, split_tensor_dict, torch, unsplit_pixel_values_by_grid, Optional, PreTrainedModel, logger, os, torch, FSDP, nn, os, FSDP, nn, 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
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()
        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
def vLLMSamplingParams(**kwargs):
    from vllm import SamplingParams

    sampling_params = SamplingParams(**kwargs)
    sampling_params._set_kwargs = kwargs
    return sampling_params
@dataclass
class UnslothRLOOConfig(RLOOConfig):
    """
    
    Configuration class for the [`RLOOTrainer`].

    This class includes only the parameters that are specific to RLOO 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:
        > Parameters that control the model and reference model

        model_init_kwargs (`str`, `dict[str, Any]`, *optional*):
            Keyword arguments for [`~transformers.AutoModelForCausalLM.from_pretrained`], used when the `model`
            argument of the [`RLOOTrainer`] is provided as a string.
        disable_dropout (`bool`, *optional*, defaults to `False`):
            Whether to disable dropout in the model. This is useful for training with a reference model, as it prevents
            the model from generating different logprobs for the same input.

        > Parameters that control the data preprocessing

        remove_unused_columns (`bool`, *optional*, defaults to `False`):
            Whether to only keep the column `"prompt"` in the dataset. If you use a custom reward function that
            requires any column other than `"prompts"` and `"completions"`, you should keep this to `False`.
        max_prompt_length (`int` or `None`, *optional*, defaults to `512`):
            Maximum length of the prompt. If the prompt is longer than this value, it will be truncated left.
        num_generations (`int` or `None`, *optional*, defaults to `2`):
            Number of generations per prompt to sample. The effective batch size (num_processes * per_device_batch_size
            * gradient_accumulation_steps) must be evenly divisible by this value.
        max_completion_length (`int` or `None`, *optional*, defaults to `256`):
            Maximum length of the generated completion.
        ds3_gather_for_generation (`bool`, *optional*, defaults to `True`):
            This setting applies to DeepSpeed ZeRO-3. If enabled, the policy model weights are gathered for generation,
            improving generation speed. However, disabling this option allows training models that exceed the VRAM
            capacity of a single GPU, albeit at the cost of slower generation. Disabling this option is not compatible
            with vLLM generation.
        shuffle_dataset (`bool`, *optional*, defaults to `True`):
            Whether to shuffle the training dataset.

        > Parameters that control generation

        generation_batch_size: (`int`, *optional*):
            Batch size to use for generation. If `None`, it defaults to the effective training batch size:
            `per_device_train_batch_size * num_processes * steps_per_generation`. In other words, there is one
            generation batch processed per optimization step. Mutually exclusive with `steps_per_generation`.
        steps_per_generation: (`int`, *optional*):
            Number of steps per generation. If `None`, it defaults to `gradient_accumulation_steps`. Mutually exclusive
            with `generation_batch_size`.
        temperature (`float`, defaults to `1.0`):
            Temperature for sampling. The higher the temperature, the more random the completions.
        top_p (`float`, *optional*, defaults to `1.0`):
            Float that controls the cumulative probability of the top tokens to consider. Must be in (0, 1]. Set to
            `1.0` to consider all tokens.
        top_k (`int`, *optional*):
            Number of highest probability vocabulary tokens to keep for top-k-filtering. If `None`, top-k-filtering is
            disabled and all tokens are considered.
        min_p (`float`, *optional*):
            Minimum token probability, which will be scaled by the probability of the most likely token. It must be a
            value between `0.0` and `1.0`. Typical values are in the `0.01-0.2` range.
        repetition_penalty (`float`, *optional*, defaults to `1.0`):
            Float that penalizes new tokens based on whether they appear in the prompt and the generated text so far.
            Values > `1.0` encourage the model to use new tokens, while values < `1.0` encourage the model to repeat
            tokens.
        use_transformers_paged (`bool`, *optional*, defaults to `False`):
            Whether to use the `transformers` paged implementation for generation. If set to `True`, the `transformers`
            paged implementation will be used for generation instead of the default padded implementation. This
            parameter is only effective when `use_vllm` is set to `False`.
        cache_implementation (`str`, *optional*):
            Implementation of the cache method for faster generation when `use_vllm` is set to `False`.
        generation_kwargs (`dict[str, Any]`, *optional*):
            Additional keyword arguments to pass to [`~transformers.GenerationConfig`] (if using transformers) or
            `SamplingParams` (if using vLLM) when sampling completions. This can be used to further customize the
            generation behavior, such as setting `suppress_tokens`, `num_beams`, etc. If it contains keys that conflict
            with the other generation parameters (like `min_p`, `top_p`, etc.), they will override them.

        > Parameters that control generation acceleration powered by vLLM

        use_vllm (`bool`, *optional*, defaults to `False`):
            Whether to use vLLM for generating completions. If set to `True`, the trainer will use vLLM for generation
            instead of the default model.generate(). Requires `vllm` to be installed.
        vllm_mode (`str`, *optional*, defaults to `"server"`):
            Mode to use for vLLM integration when `use_vllm` is set to `True`. Must be one of `"server"` or
            `"colocate"`.

            - `"server"`: The trainer will send generation requests to a separate vLLM server. Make sure a TRL vLLM
              server is running (start with `trl vllm-serve`).
            - `"colocate"`: vLLM will run in the same process and share the training GPUs. This avoids the need for a
              separate server but may cause resource contention with training.
        vllm_model_impl (`str`, *optional*, defaults to `"vllm"`):
            Model implementation to use for vLLM. Must be one of `"transformers"` or `"vllm"`. `"transformers"`: Use
            the `transformers` backend for model implementation. `"vllm"`: Use the `vllm` library for model
            implementation.
        vllm_guided_decoding_regex (`str`, *optional*):
            Regex for vLLM guided decoding. If `None` (default), guided decoding is disabled.

        > Parameters that control the vLLM server (only used when `vllm_mode` is `"server"`)

        vllm_server_base_url (`str`, *optional*):
            Base URL for the vLLM server (e.g., `"http://localhost:8000"`). If provided, `vllm_server_host` and
            `vllm_server_port` are ignored.
        vllm_server_host (`str`, *optional*, defaults to `"0.0.0.0"`):
            Host of the vLLM server to connect to. Ignored if `vllm_server_base_url` is provided.
        vllm_server_port (`int`, *optional*, defaults to `8000`):
            Port of the vLLM server to connect to. Ignored if `vllm_server_base_url` is provided.
        vllm_server_timeout (`float`, *optional*, defaults to `240.0`):
            Total timeout duration in seconds to wait for the vLLM server to be up. If the server is not up after the
            timeout, a `ConnectionError` is raised.

        > Parameters that control colocated vLLM execution (only used when `vllm_mode` is `"colocate"`)

        vllm_gpu_memory_utilization (`float`, *optional*, defaults to `0.3`):
            Control the GPU memory utilization for vLLM. This setting only applies when `vllm_mode` is set to
            `"colocate"`. If you are using `vllm_mode="server"`, this parameter must be passed separately when
            launching the vLLM server via the `--vllm_gpu_memory_utilization` flag.
        vllm_tensor_parallel_size (`int`, *optional*, defaults to `1`):
            Control the tensor parallel size for vLLM. This setting only applies when `vllm_mode` is set to
            `"colocate"`. If you are using `vllm_mode="server"`, this parameter must be passed separately when
            launching the vLLM server via the `--vllm_tensor_parallel_size` flag.
        vllm_enable_sleep_mode (`bool`, *optional*, defaults to `False`):
            Whether to enable sleep mode for vLLM. If `True`, vLLM will sleep during the optimization step and woken
            for weight sync and generation.

        > Parameters that control the training

        beta (`float`, *optional*, defaults to `0.05`):
            KL coefficient. If `0.0`, the reference model is not loaded, reducing memory usage and improving training
            speed.
        num_iterations (`int`, *optional*, defaults to `1`):
            Number of iterations per batch (denoted as μ in the algorithm).
        epsilon (`float`, *optional*, defaults to `0.2`):
            Epsilon value for clipping.
        epsilon_high (`float`, *optional*):
            Upper-bound epsilon value for clipping. If not specified, it defaults to the same value as the lower-bound
            specified in argument `epsilon`. Paper [DAPO](https://huggingface.co/papers/2503.14476) recommends `0.28`.
        reward_weights (`list[float]`, *optional*):
            Weights for each reward function. Must match the number of reward functions. If `None`, all rewards are
            weighted equally with weight `1.0`.
        normalize_advantages (`bool`, *optional*, defaults to `False`):
            Whether to normalize advantages. Normalization is done per generation batch to have mean `0.0` and standard
            deviation of `1.0`.
        reward_clip_range (`tuple[float, float]`, *optional*):
            Clip range for rewards as (min, max). If `None`, no clipping is applied.
        mask_truncated_completions (`bool`, *optional*, defaults to `False`):
            When enabled, truncated completions are excluded from the loss calculation, preventing them from being
            incorrectly penalized and introducing noise during training. According to the
            [DAPO](https://huggingface.co/papers/2503.14476) paper, this is a good practice for training stability.
        sync_ref_model (`bool`, *optional*, defaults to `False`):
            Whether to synchronize the reference model with the active model every `ref_model_sync_steps` steps, using
            the `ref_model_mixup_alpha` parameter. This synchronization originates from the
            [TR-DPO](https://huggingface.co/papers/2404.09656) paper.
        ref_model_mixup_alpha (`float`, *optional*, defaults to `0.6`):
            α parameter from the [TR-DPO](https://huggingface.co/papers/2404.09656) paper, which controls the mix
            between the current policy and the previous reference policy during updates. The reference policy is
            updated according to the equation: `π_ref = α * π_θ + (1 - α) * π_ref_prev`. To use this parameter, you
            must set `sync_ref_model=True`.
        ref_model_sync_steps (`int`, *optional*, defaults to `512`):
            τ parameter from the [TR-DPO](https://huggingface.co/papers/2404.09656) paper, which determines how
            frequently the current policy is synchronized with the reference policy. To use this parameter, you must
            set `sync_ref_model=True`.

        > Parameters that control the logging

        log_completions (`bool`, *optional*, defaults to `False`):
            Whether to log a sample of (prompt, completion) pairs every `logging_steps` steps. If `rich` is installed,
            it prints the sample. If `wandb` logging is enabled, it logs it to `wandb`.
        num_completions_to_print (`int`, *optional*):
            Number of completions to print with `rich`. If `None`, all completions are logged.
        wandb_log_unique_prompts (`bool`, *optional*, defaults to `False`):
            Whether to log unique prompts in wandb. If `True`, only unique prompts are logged. If `False`, all prompts
            are logged.

        > Deprecated parameters

        rloo_k:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `num_generations` instead.

            </Deprecated>

        cliprange:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `epsilon` instead.

            </Deprecated>

        kl_coef:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `beta` instead.

            </Deprecated>

        exp_name:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `run_name` instead.

            </Deprecated>

        normalize_reward:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `normalize_advantages` instead.

            </Deprecated>

        num_ppo_epochs:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `num_iterations` instead.

            </Deprecated>

        num_mini_batches:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `steps_per_generation` instead.

            </Deprecated>

        total_episodes:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `max_steps` instead.

            </Deprecated>

        response_length:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `max_completion_length` instead.

            </Deprecated>

        token_level_kl:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. KL is now computed only at the sequence
            level.

            </Deprecated>

        dataset_num_proc:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. This parameter was unused, you can
            safely remove it from your scripts.

            </Deprecated>

        local_rollout_forward_batch_size:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Now it is automatically set to
            `per_device_train_batch_size` (or `per_device_eval_batch_size` during evaluation).

            </Deprecated>

        num_sample_generations:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `logging_steps` to control
            generation logging frequency.

            </Deprecated>

        stop_token:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0.

            </Deprecated>

        stop_token_id:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `processing_class.eos_token_id`
            instead.

            </Deprecated>

        missing_eos_penalty:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Replicate with a custom reward function
            checking if `eos_token_id` is in `completion_ids`.

            </Deprecated>
    
    """
    vllm_sampling_params: Optional[Any] = field(
        default = None,
        metadata = {'help': 'vLLM SamplingParams'},
    )
    unsloth_num_chunks : Optional[int] = field(
        default = -1,
        metadata = {'help': 'Chunk size to reduce memory usage. -1 is most efficient.'},
    )
    
    def __init__(
        self,
        output_dir = None,
        overwrite_output_dir = None,
        do_train = False,
        do_eval = False,
        do_predict = False,
        eval_strategy = 'no',
        prediction_loss_only = False,
        per_device_train_batch_size = 4,
        per_device_eval_batch_size = 4,
        per_gpu_train_batch_size = None,
        per_gpu_eval_batch_size = None,
        gradient_accumulation_steps = 2,
        eval_accumulation_steps = 2,
        eval_delay = 0,
        torch_empty_cache_steps = 250,
        learning_rate = 5e-05,
        weight_decay = 0.01,
        adam_beta1 = 0.9,
        adam_beta2 = 0.999,
        adam_epsilon = 1e-08,
        max_grad_norm = 1.0,
        num_train_epochs = 3.0,
        max_steps = -1,
        lr_scheduler_type = 'linear',
        warmup_ratio = 0.1,
        warmup_steps = 0,
        log_level = 'passive',
        log_level_replica = 'warning',
        log_on_each_node = True,
        logging_dir = None,
        logging_strategy = 'steps',
        logging_first_step = False,
        logging_steps = 1,
        logging_nan_inf_filter = False,
        save_strategy = 'steps',
        save_steps = 500,
        save_total_limit = None,
        save_safetensors = True,
        save_on_each_node = False,
        save_only_model = False,
        restore_callback_states_from_checkpoint = False,
        no_cuda = False,
        use_cpu = False,
        use_mps_device = False,
        seed = 3407,
        data_seed = 3407,
        jit_mode_eval = False,
        bf16 = False,
        fp16 = False,
        fp16_opt_level = 'O1',
        half_precision_backend = 'auto',
        bf16_full_eval = False,
        fp16_full_eval = False,
        tf32 = None,
        local_rank = -1,
        ddp_backend = None,
        tpu_num_cores = None,
        tpu_metrics_debug = False,
        debug = '',
        dataloader_drop_last = False,
        eval_steps = None,
        dataloader_num_workers = 0,
        dataloader_prefetch_factor = None,
        past_index = -1,
        run_name = None,
        disable_tqdm = None,
        remove_unused_columns = False,
        label_names = None,
        load_best_model_at_end = False,
        metric_for_best_model = None,
        greater_is_better = None,
        ignore_data_skip = False,
        fsdp = 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,
        model_init_kwargs = None,
        disable_dropout = False,
        max_prompt_length = 512,
        num_generations = 8,
        max_completion_length = 256,
        ds3_gather_for_generation = True,
        shuffle_dataset = True,
        generation_batch_size = None,
        steps_per_generation = None,
        temperature = 1.0,
        top_p = 1.0,
        top_k = None,
        min_p = None,
        generation_kwargs = {},
        repetition_penalty = 1.0,
        use_transformers_paged = False,
        cache_implementation = None,
        use_vllm = False,
        vllm_mode = 'colocate',
        vllm_model_impl = 'vllm',
        vllm_enable_sleep_mode = False,
        vllm_guided_decoding_regex = None,
        vllm_server_base_url = None,
        vllm_server_host = '0.0.0.0',
        vllm_server_port = 8000,
        vllm_server_timeout = 240.0,
        vllm_gpu_memory_utilization = 0.3,
        vllm_tensor_parallel_size = 1,
        beta = 0.05,
        num_iterations = 1,
        epsilon = 0.2,
        epsilon_high = None,
        reward_weights = None,
        normalize_advantages = False,
        reward_clip_range = None,
        mask_truncated_completions = False,
        sync_ref_model = False,
        ref_model_mixup_alpha = 0.6,
        ref_model_sync_steps = 512,
        log_completions = False,
        num_completions_to_print = None,
        wandb_log_unique_prompts = False,
        rloo_k = None,
        cliprange = None,
        kl_coef = None,
        exp_name = None,
        normalize_reward = None,
        num_ppo_epochs = None,
        num_mini_batches = None,
        total_episodes = None,
        response_length = None,
        token_level_kl = None,
        dataset_num_proc = None,
        local_rollout_forward_batch_size = None,
        num_sample_generations = None,
        stop_token = None,
        stop_token_id = None,
        missing_eos_penalty = None,
        vllm_sampling_params = None,
        unsloth_num_chunks = -1,
        
        **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 (per_device_train_batch_size // num_generations) * num_generations != per_device_train_batch_size:
            print('Unsloth: We now expect `per_device_train_batch_size` to be a multiple of `num_generations`.\nWe will change the batch size of ' + str(per_device_train_batch_size) + ' to the `num_generations` of ' + str(num_generations))
            per_device_train_batch_size = num_generations
        
        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,
            model_init_kwargs = model_init_kwargs,
            disable_dropout = disable_dropout,
            max_prompt_length = max_prompt_length,
            num_generations = num_generations,
            max_completion_length = max_completion_length,
            ds3_gather_for_generation = ds3_gather_for_generation,
            shuffle_dataset = shuffle_dataset,
            generation_batch_size = generation_batch_size,
            steps_per_generation = steps_per_generation,
            temperature = temperature,
            top_p = top_p,
            top_k = top_k,
            min_p = min_p,
            generation_kwargs = generation_kwargs,
            repetition_penalty = repetition_penalty,
            use_transformers_paged = use_transformers_paged,
            cache_implementation = cache_implementation,
            use_vllm = use_vllm,
            vllm_mode = vllm_mode,
            vllm_model_impl = vllm_model_impl,
            vllm_enable_sleep_mode = vllm_enable_sleep_mode,
            vllm_guided_decoding_regex = vllm_guided_decoding_regex,
            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_gpu_memory_utilization = vllm_gpu_memory_utilization,
            vllm_tensor_parallel_size = vllm_tensor_parallel_size,
            beta = beta,
            num_iterations = num_iterations,
            epsilon = epsilon,
            epsilon_high = epsilon_high,
            reward_weights = reward_weights,
            normalize_advantages = normalize_advantages,
            reward_clip_range = reward_clip_range,
            mask_truncated_completions = mask_truncated_completions,
            sync_ref_model = sync_ref_model,
            ref_model_mixup_alpha = ref_model_mixup_alpha,
            ref_model_sync_steps = ref_model_sync_steps,
            log_completions = log_completions,
            num_completions_to_print = num_completions_to_print,
            wandb_log_unique_prompts = wandb_log_unique_prompts,
            rloo_k = rloo_k,
            cliprange = cliprange,
            kl_coef = kl_coef,
            exp_name = exp_name,
            normalize_reward = normalize_reward,
            num_ppo_epochs = num_ppo_epochs,
            num_mini_batches = num_mini_batches,
            total_episodes = total_episodes,
            response_length = response_length,
            token_level_kl = token_level_kl,
            dataset_num_proc = dataset_num_proc,
            local_rollout_forward_batch_size = local_rollout_forward_batch_size,
            num_sample_generations = num_sample_generations,
            stop_token = stop_token,
            stop_token_id = stop_token_id,
            missing_eos_penalty = missing_eos_penalty,**kwargs)
        self.vllm_sampling_params = vllm_sampling_params
        self.unsloth_num_chunks = unsloth_num_chunks
        
pass

class _UnslothRLOOTrainer(BaseTrainer):
    """"""

    _tag_names = ["trl", "rloo"]
    _name = "RLOO"
    _paper = {
        "title": "Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs",
        "id": "2402.14740",
        # docstyle-ignore
        "citation": textwrap.dedent("""\
            @inproceedings{ahmadian2024back,
                title        = {{Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs}},
                author       = {Arash Ahmadian and Chris Cremer and Matthias Gall{\'{e}} and Marzieh Fadaee and Julia Kreutzer and Olivier Pietquin and Ahmet {\"{U}}st{\"{u}}n and Sara Hooker},
                year         = 2024,
                booktitle    = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), {ACL} 2024, Bangkok, Thailand, August 11-16, 2024},
                pages        = {12248--12267},
                publisher    = {Association for Computational Linguistics},
                editor       = {Lun{-}Wei Ku and Andre Martins and Vivek Srikumar},
            }"""),
    }

    def __init__(
        self,
        # Note for dev: we can remove the default None when we remove the deprecated model parameter in version 0.25.0
        model: Union[str, PreTrainedModel] = None,
        reward_funcs: Union[RewardFunc, list[RewardFunc]] = None,
        args: Optional[RLOOConfig] = None,
        train_dataset: Optional[Union[Dataset, IterableDataset]] = None,
        eval_dataset: Optional[Union[Dataset, IterableDataset, dict[str, Union[Dataset, IterableDataset]]]] = None,
        processing_class: Optional[Union[PreTrainedTokenizerBase, ProcessorMixin]] = None,
        reward_processing_classes: Optional[Union[PreTrainedTokenizerBase, list[PreTrainedTokenizerBase]]] = None,
        callbacks: Optional[list[TrainerCallback]] = None,
        optimizers: tuple[Optional[torch.optim.Optimizer], Optional[torch.optim.lr_scheduler.LambdaLR]] = (None, None),
        peft_config: Optional["PeftConfig"] = None,
        # Deprecated parameters
        config=None,
        reward_model=None,
        policy=None,
        ref_policy=None,
        data_collator=None,
    ):

        if hasattr(model, 'vllm_engine') and hasattr(args, 'use_vllm'):
            if (getattr(args, 'use_vllm', False) == False):
                args.use_vllm = True
        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."
            )
        # Handle deprecated parameters
        if config is not None:
            warnings.warn(
                "Parameter 'config' is deprecated and will be removed in version 0.25.0. Please use 'args' instead. "
                "We are setting args=config"
            )
            if args is None:
                args = config
            else:
                raise ValueError("Cannot specify both 'config' (deprecated) and 'args'. Please use 'args' only.")

        if reward_model is not None:
            warnings.warn(
                "Parameter 'reward_model' is deprecated and will be removed in version 0.25.0. Please use "
                "'reward_funcs' instead. We are setting reward_funcs=reward_model"
            )
            if reward_funcs is None:
                reward_funcs = reward_model
            else:
                raise ValueError(
                    "Cannot specify both 'reward_model' (deprecated) and 'reward_funcs'. Please use 'reward_funcs' "
                    "only."
                )
        if policy is not None:
            warnings.warn(
                "Parameter 'policy' is deprecated and will be removed in version 0.25.0. Please use 'model' instead. "
                "We are setting model=policy"
            )
            if model is None:
                model = policy
            else:
                raise ValueError("Cannot specify both 'policy' (deprecated) and 'model'. Please use 'model' only.")
        if ref_policy is not None:
            warnings.warn(
                "Parameter 'ref_policy' is deprecated and will be removed in version 0.25.0. To use the initial model "
                "as the reference model, simply omit this parameter. The parameter is ignored."
            )
        if data_collator is not None:
            warnings.warn(
                "Parameter 'data_collator' is deprecated and will be removed in version 0.25.0. The RLOOTrainer does "
                "not use a data collator, so this parameter is ignored."
            )
        if "input_ids" in train_dataset.column_names:
            warnings.warn(
                "The training dataset contains a column named 'input_ids', indicating that it is pre-tokenized. "
                "Support for pre-tokenized datasets is deprecated and will be removed in version 0.25. Please provide "
                "the raw dataset (conversational or standard) with a 'prompt' column instead."
            )

            def decode(example, tokenizer):
                return {"prompt": tokenizer.decode(example["input_ids"])}

            train_dataset = train_dataset.map(decode, fn_kwargs={"tokenizer": processing_class})
        if eval_dataset is not None and "input_ids" in eval_dataset.column_names:
            warnings.warn(
                "The evaluation dataset contains a column named 'input_ids', indicating that it is pre-tokenized. "
                "Support for pre-tokenized datasets is deprecated and will be removed in version 0.25. Please provide "
                "the raw dataset (conversational or standard) with a 'prompt' column instead."
            )

            def decode(example, tokenizer):
                return {"prompt": tokenizer.decode(example["input_ids"])}

            eval_dataset = eval_dataset.map(decode, fn_kwargs={"tokenizer": processing_class})

        # Args
        if args is None:
            model_name = model if isinstance(model, str) else model.config._name_or_path
            model_name = model_name.split("/")[-1]
            args = RLOOConfig(f"{model_name}-RLOO")

        # Models
        # Trained model
        model_init_kwargs = args.model_init_kwargs or {}
        if isinstance(model, str):
            model_id = model
            dtype = model_init_kwargs.get("dtype")
            if isinstance(dtype, torch.dtype) or dtype == "auto" or dtype is None:
                pass  # dtype is already a torch.dtype or "auto" or None
            elif isinstance(dtype, str):  # it's a str, but not "auto"
                dtype = getattr(torch, dtype)
                model_init_kwargs["dtype"] = dtype
            else:
                raise ValueError(
                    "Invalid `dtype` passed to `RLOOConfig`. Expected either 'auto' or a string representing "
                    f"a `torch.dtype` (e.g., 'float32'), but got {dtype}."
                )
            # Disable caching if gradient checkpointing is enabled [not supported]
            config = AutoConfig.from_pretrained(model_id)
            architecture = getattr(transformers, config.architectures[0])
            model = architecture.from_pretrained(model_id, **model_init_kwargs)
        else:
            model_id = model.config._name_or_path
            if args.model_init_kwargs is not None:
                logger.warning(
                    "You passed `model_init_kwargs` to the `RLOOConfig`, but your model is already instantiated. "
                    "The `model_init_kwargs` will be ignored."
                )

        # Some models [SmolVLM/Idefics3] don't support `logits_to_keep` argument and error out if we pass it
        # Inspect the forward method before we wrap the model with PEFT
        self.model_kwarg_keys = (
            inspect.signature(model.forward).parameters.keys()
            if not hasattr(model, "get_base_model")
            else inspect.signature(model.get_base_model().forward).parameters.keys()
        )

        if False:
            model = prepare_peft_model(model, peft_config, args)

        # Processing class
        if processing_class is None:
            processing_class = AutoProcessor.from_pretrained(model.config._name_or_path, truncation_side="left")

        # Handle pad token for processors or tokenizers
        if isinstance(processing_class, ProcessorMixin):
            tokenizer = processing_class.tokenizer
        elif isinstance(processing_class, PreTrainedTokenizerBase):
            tokenizer = processing_class
        else:
            raise TypeError("The `processing_class` must be either a `PreTrainedTokenizerBase` or a `ProcessorMixin`")

        if tokenizer.pad_token is None:
            tokenizer.pad_token = tokenizer.eos_token

        self.pad_token = tokenizer.pad_token
        self.pad_token_id = tokenizer.pad_token_id
        self.eos_token_id = tokenizer.eos_token_id

        # Reward functions
        if not isinstance(reward_funcs, list):
            reward_funcs = [reward_funcs]
        self.reward_func_names = []
        for i, reward_func in enumerate(reward_funcs):
            if isinstance(reward_func, str):
                reward_funcs[i] = AutoModelForSequenceClassification.from_pretrained(
                    reward_func, num_labels=1, **model_init_kwargs
                )
            if isinstance(reward_funcs[i], nn.Module):  # Use Module over PretrainedModel for compat w/ compiled models
                self.reward_func_names.append(reward_funcs[i].config._name_or_path.split("/")[-1])
            else:
                self.reward_func_names.append(reward_funcs[i].__name__)
        self.reward_funcs = reward_funcs

        # Reward weights
        if args.reward_weights is not None:
            if len(args.reward_weights) != len(reward_funcs):
                raise ValueError(
                    f"Number of reward weights ({len(args.reward_weights)}) must match number of reward "
                    f"functions ({len(reward_funcs)})"
                )
            self.reward_weights = torch.tensor(args.reward_weights, dtype=torch.float32)
        else:
            self.reward_weights = torch.ones(len(reward_funcs), dtype=torch.float32)

        # Reward processing class
        if reward_processing_classes is None:
            reward_processing_classes = [None] * len(reward_funcs)
        elif not isinstance(reward_processing_classes, list):
            reward_processing_classes = [reward_processing_classes]
        if len(reward_processing_classes) != len(reward_funcs):
            raise ValueError(
                f"The number of reward processing classes ({len(reward_processing_classes)}) must match the number of "
                f"reward functions ({len(reward_funcs)})."
            )

        for i, (reward_processing_class, reward_func) in enumerate(zip(reward_processing_classes, reward_funcs)):
            if isinstance(reward_func, PreTrainedModel):
                if reward_processing_class is None:
                    reward_processing_class = AutoTokenizer.from_pretrained(reward_func.config._name_or_path)
                if reward_processing_class.pad_token_id is None:
                    reward_processing_class.pad_token = reward_processing_class.eos_token
                # The reward model computes the reward for the latest non-padded token in the input sequence.
                # So it's important to set the pad token ID to the padding token ID of the processing class.
                reward_func.config.pad_token_id = reward_processing_class.pad_token_id
                reward_processing_classes[i] = reward_processing_class

        self.reward_processing_classes = reward_processing_classes

        # Training arguments
        self.max_prompt_length = args.max_prompt_length
        self.max_completion_length = args.max_completion_length
        self.num_generations = args.num_generations
        self.temperature = args.temperature
        self.top_p = args.top_p
        self.top_k = args.top_k
        self.min_p = args.min_p
        self.repetition_penalty = args.repetition_penalty
        self.use_transformers_paged = args.use_transformers_paged
        self.use_vllm = args.use_vllm
        self.vllm_mode = args.vllm_mode
        self.vllm_gpu_memory_utilization = args.vllm_gpu_memory_utilization  # only applies to colocation mode
        self.vllm_tensor_parallel_size = args.vllm_tensor_parallel_size  # only applies to colocation mode
        self.normalize_advantages = args.normalize_advantages
        self.mask_truncated_completions = args.mask_truncated_completions
        self.reward_clip_range = args.reward_clip_range

        # Datasets
        self.shuffle_dataset = args.shuffle_dataset

        if (
            isinstance(train_dataset, IterableDataset)
            or isinstance(eval_dataset, IterableDataset)
            or (
                isinstance(eval_dataset, dict) and any(isinstance(ds, IterableDataset) for ds in eval_dataset.values())
            )
        ):
            # See https://github.com/huggingface/trl/issues/3213
            raise NotImplementedError(
                "Iterable datasets are not yet supported in RLOOTrainer. Please use a standard dataset instead."
            )

        # Multi-step
        self.num_iterations = args.num_iterations
        self.epsilon_low = args.epsilon
        self.epsilon_high = args.epsilon_high if args.epsilon_high is not None else args.epsilon
        # Tracks the number of iterations [forward + backward passes], including those within a grad accum cycle
        self._step = 0
        # Buffer the batch to reuse generated outputs across multiple updates. For more details, see
        # `_get_train_sampler` and `_prepare_inputs`.
        self._buffered_inputs = None

        # 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 RLOO, the sampled data does not include the
        # "input_ids" key. Instead, the available keys is "prompt". 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

        super().__init__(
            model=model,
            args=args,
            data_collator=identity,  # No data collation is needed in RLOO
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            processing_class=processing_class,
            callbacks=callbacks,
            optimizers=optimizers,
        )

        # Reference model
        self.beta = args.beta
        if self.beta == 0.0:
            # If beta is 0.0, the reference model is not needed
            self.ref_model = None
        elif is_peft_model(model):
            # If PEFT is used, the reference model is not needed since the adapter can be disabled
            # to revert to the initial model.
            self.ref_model = None
        else:
            # For deepspeed, fsdp or non-distributed models, create a reference model from scratch
            config = AutoConfig.from_pretrained(model_id)
            architecture = getattr(transformers, config.architectures[0])
            self.ref_model = architecture.from_pretrained(model_id, **model_init_kwargs)

        # Disable dropout in the models
        if args.disable_dropout:
            disable_dropout_in_model(model)
            if self.ref_model is not None:
                disable_dropout_in_model(self.ref_model)

        # Initialize the metrics
        self._metrics = {"train": defaultdict(list), "eval": defaultdict(list)}
        self._total_train_tokens = 0
        self.log_completions = args.log_completions
        self.wandb_log_unique_prompts = args.wandb_log_unique_prompts
        self.num_completions_to_print = args.num_completions_to_print
        # Keep logs sized to the generation batch to record only outputs from the latest model update.
        self._logs = {
            "images": deque(maxlen=args.generation_batch_size),
            "prompt": deque(maxlen=args.generation_batch_size),
            "completion": deque(maxlen=args.generation_batch_size),
            "rewards": defaultdict(lambda: deque(maxlen=args.generation_batch_size)),
            "advantages": deque(maxlen=args.generation_batch_size),
        }

        # Ensure each process receives a unique seed to prevent duplicate completions when generating with
        # transformers if num_generations exceeds per_device_train_batch_size. We could skip it if we use vLLM, but
        # it's safer to set it in all cases.
        set_seed(args.seed, device_specific=True)

        if self.use_vllm:
            if not is_vllm_available():
                raise ImportError(
                    "vLLM is not available and `use_vllm` is set to True. Please install vLLM with "
                    "`pip install trl[vllm]` to use it."
                )

            if self.vllm_mode == "server":
                if self.accelerator.is_main_process:
                    if args.vllm_server_base_url is not None:
                        base_url = args.vllm_server_base_url
                    else:
                        base_url = f"http://{args.vllm_server_host}:{args.vllm_server_port}"
                    self.vllm_client = VLLMClient(base_url=base_url, connection_timeout=args.vllm_server_timeout)
                    self.vllm_client.init_communicator(device=torch.cuda.current_device())

            elif self.vllm_mode == "colocate":
                if not self.accelerator.num_processes % self.vllm_tensor_parallel_size == 0:
                    raise ValueError(
                        f"vllm_tensor_parallel_size ({self.vllm_tensor_parallel_size}) must divide world size "
                        f"({self.accelerator.num_processes}) evenly."
                    )

                if self.vllm_tensor_parallel_size > 1:
                    self.tp_group, _ = torch.distributed.new_subgroups_by_enumeration(
                        [
                            list(range(i * self.vllm_tensor_parallel_size, (i + 1) * self.vllm_tensor_parallel_size))
                            for i in range(self.accelerator.num_processes // self.vllm_tensor_parallel_size)
                        ]
                    )
                os.environ["RANK"] = str(self.accelerator.process_index)
                os.environ["LOCAL_RANK"] = str(self.accelerator.local_process_index)
                os.environ["WORLD_SIZE"] = str(self.accelerator.num_processes)
                ensure_master_addr_port()

                if self.max_prompt_length is not None and self.max_completion_length is not None:
                    max_model_len = self.max_prompt_length + self.max_completion_length
                else:
                    max_model_len = None
                self.llm = model.vllm_engine
                if self.args.vllm_enable_sleep_mode:
                    self.llm.sleep(level=1)
            else:
                raise ValueError(f"vllm_mode must be either 'server' or 'colocate', got '{self.vllm_mode}'.")
            self.guided_decoding_regex = args.vllm_guided_decoding_regex

            self._last_loaded_step = -1
            self.accelerator.wait_for_everyone()
        else:
            generation_kwargs = {
                "max_new_tokens": self.max_completion_length,
                "do_sample": True,
                "pad_token_id": tokenizer.pad_token_id,
                "bos_token_id": tokenizer.bos_token_id,
                "eos_token_id": tokenizer.eos_token_id,
                "temperature": self.temperature,
                "top_p": self.top_p,
                "top_k": self.top_k,
                "min_p": self.min_p,
                "repetition_penalty": self.repetition_penalty,
                "cache_implementation": args.cache_implementation,
            }
            if args.generation_kwargs is not None:
                generation_kwargs.update(args.generation_kwargs)
            self.generation_config = GenerationConfig(**generation_kwargs)

        # 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 to the model
        self.model.add_model_tags(self._tag_names)

        if self.ref_model is not None:
            if self.is_deepspeed_enabled:
                self.ref_model = prepare_deepspeed(self.ref_model, self.accelerator)
            elif self.is_fsdp_enabled:
                self.ref_model = prepare_fsdp(self.ref_model, self.accelerator)
            else:
                self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)

        if args.sync_ref_model:
            self.add_callback(SyncRefModelCallback(ref_model=self.ref_model, accelerator=self.accelerator))

        for i, reward_func in enumerate(self.reward_funcs):
            if isinstance(reward_func, PreTrainedModel):
                if self.is_deepspeed_enabled:
                    self.reward_funcs[i] = prepare_deepspeed(reward_func, self.accelerator)
                else:
                    # set device placement to True to make `prepare_model` move `reward_func` to device when using fsdp
                    self.reward_funcs[i] = self.accelerator.prepare_model(
                        reward_func, evaluation_mode=True, device_placement=True
                    )

    def _set_signature_columns_if_needed(self):
        # If `self.args.remove_unused_columns` is True, non-signature columns are removed.
        # By default, this method sets `self._signature_columns` to the model's expected inputs.
        # In RLOOTrainer, we preprocess data, so using the model's signature columns doesn't work.
        # Instead, we set them to the columns expected by the `training_step` method, hence the override.
        if self._signature_columns is None:
            self._signature_columns = ["prompt", "image", "images"]

    # This method overrides `Trainer.get_train_dataloader` to support our custom batching strategy.
    # Instead of returning a standard per-step batch (i.e., `per_device_batch_size), our dataloader loads an
    # *generation* batch (i.e., `per_device_batch_size × steps_per_generation`). This allows us to generate completions
    # once every steps_per_generation step—rather than once per accumulation step—which is significantly more
    # efficient. The only change from the original implementation is multiplying the batch size by
    # `steps_per_generation`. Thus, `_prepare_inputs` is called with this *generation* batch, and it handles the
    # splitting internally.
    # Maintenance note: This method is a copy-paste of the original `Trainer.get_train_dataloader` with only one line
    # modification. As a result, some parts of the method aren't relevant to RLOO, but we keep them to stay one line
    # apart from the super method, ensuring easier maintenance in the future.
    def get_train_dataloader(self):
        if self.train_dataset is None:
            raise ValueError("Trainer: training requires a train_dataset.")

        train_dataset = self.train_dataset
        data_collator = self.data_collator
        if is_datasets_available() and isinstance(train_dataset, datasets.Dataset):
            train_dataset = self._remove_unused_columns(train_dataset, description="training")
        else:
            data_collator = self._get_collator_with_removed_columns(data_collator, description="training")

        dataloader_params = {
            "batch_size": self._train_batch_size * self.args.steps_per_generation,  # < this is the change
            "collate_fn": data_collator,
            "num_workers": self.args.dataloader_num_workers,
            "pin_memory": self.args.dataloader_pin_memory,
            "persistent_workers": self.args.dataloader_persistent_workers,
        }

        if not isinstance(train_dataset, torch.utils.data.IterableDataset):
            dataloader_params["sampler"] = self._get_train_sampler()
            dataloader_params["drop_last"] = self.args.dataloader_drop_last
            dataloader_params["worker_init_fn"] = partial(
                seed_worker, num_workers=self.args.dataloader_num_workers, rank=self.args.process_index
            )

            dataloader_params["prefetch_factor"] = self.args.dataloader_prefetch_factor

        return self.accelerator.prepare(DataLoader(train_dataset, **dataloader_params))

    def _get_train_sampler(self, dataset: Optional[Dataset] = None) -> Sampler:
        # Returns a sampler that
        # 1. ensures each prompt is repeated across multiple processes. This guarantees that identical prompts are
        #    distributed to different GPUs, allowing rewards to be computed and normalized correctly within each prompt
        #    group. Using the same seed across processes ensures consistent prompt assignment, preventing discrepancies
        #    in group formation.
        # 2. repeats the batch multiple times to allow reusing generations across multiple updates. Refer to
        #    _prepare_inputs to see how the generations are stored and reused.

        # In the following figure, the values are the prompt indices. The first row shows the first sampled batch, the
        # second row shows the second sampled batch, and so on.
        #
        #                                      |   GPU 0  |   GPU 1  |
        #
        #                 global_step   step    <-───>  num_generations=2
        #                                       <-───────> per_device_train_batch_size=3
        #  grad_accum    ▲  ▲  0          0     0   0   1   1   2   2   <- Generate for the first `steps_per_generation` (prompts 0 to 11); store the completions; use the first slice to compute the loss
        #     =2         ▼  |  0          1     3   3   4   4   5   5   <- Take the stored generations and use the second slice to compute the loss
        #                   |
        #                   |  1          2     6   6   7   7   8   8   <- Take the stored generations and use the third slice to compute the loss
        #  steps_per_gen=4  ▼  1          3     9   9  10  10  11  11   <- Take the stored generations and use the fourth slice to compute the loss
        #
        #                      2          4    12  12  13  13  14  14   <- Generate for the second `steps_per_generation` (prompts 12 to 23); store the completions; use the first slice to compute the loss
        #                      2          5    15  15  16  16  17  17   <- Take the stored generations and use the second slice to compute the loss
        #                                          ...
        if dataset is None:
            dataset = self.train_dataset
        return RepeatSampler(
            data_source=dataset,
            mini_repeat_count=self.num_generations,
            batch_size=self.args.generation_batch_size // self.num_generations,
            repeat_count=self.num_iterations * self.args.steps_per_generation,
            shuffle=self.shuffle_dataset,
            seed=self.args.seed,
        )

    def _get_eval_sampler(self, eval_dataset) -> Sampler:
        # See _get_train_sampler for an explanation of the sampler.
        return RepeatSampler(
            data_source=eval_dataset,
            mini_repeat_count=self.num_generations,
            seed=self.args.seed,
        )

    @profiling_decorator
    def _get_per_token_logps_and_entropies(
        self,
        model,
        input_ids,
        attention_mask,
        logits_to_keep,
        batch_size=None,
        compute_entropy=False,
        pixel_values=None,
        image_grid_thw=None,
        num_images=None,
        pixel_attention_mask=None,
        image_sizes=None,
        token_type_ids=None,
    ) -> dict[str, Optional[torch.Tensor]]:
        """Compute log-probs and (optionally) entropies for each token."""
        batch_size = batch_size or input_ids.size(0)  # Chunk inputs into smaller batches to reduce memory peak
        all_logps = []
        all_entropies = []
        for start in range(0, input_ids.size(0), batch_size):
            input_ids_batch = input_ids[start : start + batch_size]
            attention_mask_batch = attention_mask[start : start + batch_size]

            # Build model inputs - check if the model supports logits_to_keep (some models and VLMs don't)
            model_inputs = {"input_ids": input_ids_batch, "attention_mask": attention_mask_batch}

            if image_grid_thw is not None and pixel_values is not None:
                rows_per_image = image_grid_thw.prod(dim=-1)
                rows_per_sample = torch.split(rows_per_image, num_images)
                rows_per_sample = torch.stack([s.sum() for s in rows_per_sample])
                cum_rows = torch.cat([torch.tensor([0], device=rows_per_sample.device), rows_per_sample.cumsum(0)])
                row_start, row_end = cum_rows[start].item(), cum_rows[start + batch_size].item()
                model_inputs["pixel_values"] = pixel_values[row_start:row_end]
                cum_imgs = torch.tensor([0] + num_images).cumsum(0)
                img_start, img_end = cum_imgs[start], cum_imgs[start + batch_size]
                model_inputs["image_grid_thw"] = image_grid_thw[img_start:img_end]
            elif pixel_values is not None:
                model_inputs["pixel_values"] = pixel_values[start : start + batch_size]
            if pixel_attention_mask is not None:
                model_inputs["pixel_attention_mask"] = pixel_attention_mask[start : start + batch_size]
            if image_sizes is not None:
                model_inputs["image_sizes"] = image_sizes[start : start + batch_size]
            if token_type_ids is not None:
                model_inputs["token_type_ids"] = token_type_ids[start : start + batch_size]

            # Only add logits_to_keep if the model supports it
            if "logits_to_keep" in self.model_kwarg_keys:
                # We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded
                model_inputs["logits_to_keep"] = logits_to_keep + 1

            model_inputs["use_cache"] = False  # only used in generation; set False to suppress warnings

            logits = model(**model_inputs).logits
            # Exclude the last value: it corresponds to the next token pred
            logits = logits[:, :-1, :]  # (B, L-1, H)
            # Only keep the last logits_to_keep. For model that support logits_to_keep, this is a no-op.
            logits = logits[:, -logits_to_keep:, :]  # (B, logits_to_keep, H)
            # Divide logits by sampling temperature.
            # See https://huggingface.co/blog/the_n_implementation_details_of_rlhf_with_ppo#policy-training-implementation-details
            logits = logits / self.temperature

            completion_ids = input_ids_batch[:, -logits_to_keep:]
            logps = selective_log_softmax(logits, completion_ids)  # compute logprobs
            all_logps.append(logps)

            if compute_entropy:
                with torch.no_grad():
                    entropies = entropy_from_logits(logits)
                all_entropies.append(entropies)

        logps = torch.cat(all_logps, dim=0)
        entropies = torch.cat(all_entropies, dim=0) if compute_entropy else None
        return logps, entropies

    def _fix_param_name_to_vllm(self, name, extra_prefixes: Optional[list[str]] = None):
        extra_prefixes = extra_prefixes or []
        prefixes = ["_checkpoint_wrapped_module."] + extra_prefixes
        for prefix in prefixes:
            name = name.replace(prefix, "")
        return name

    def _sync_fsdp1_params_to_vllm(self, module: nn.Module, prefix: str = "", visited=None):
        """Memory-efficient post-order traversal of FSDP modules to extract full parameters and sync with vLLM."""
        # For FSDP1, we need to recurse into children and also use summon_full_params
        if visited is None:
            visited = set()
        for child_name, child_module in module.named_children():
            child_prefix = f"{prefix}.{child_name}" if prefix else child_name
            self._sync_fsdp1_params_to_vllm(
                child_module, prefix=child_prefix, visited=visited
            )  # recurse into the child

        if isinstance(module, FSDP):
            with FSDP.summon_full_params(module, recurse=False, writeback=False):
                for param_name, param in module.named_parameters():
                    full_name = f"{prefix}.{param_name}" if prefix else param_name
                    full_name = self._fix_param_name_to_vllm(full_name, extra_prefixes=["_fsdp_wrapped_module."])

                    if full_name in visited:
                        continue  # skip FSDP subtrees already traversed
                    visited.add(full_name)

                    if self.vllm_mode == "server" and self.accelerator.is_main_process:
                        self.vllm_client.update_named_param(full_name, param.data)
                    elif self.vllm_mode == "colocate":

                        pass

                        pass

    def _sync_fsdp2_params_to_vllm(self, module: nn.Module):
        # For FSDP2, module already covers all parameters, so no need for recursion
        for name, param in module.items():
            if param.is_cpu:
                param = param.to(torch.device("cuda"))
            param = param.full_tensor()

            if self.vllm_mode == "server" and self.accelerator.is_main_process:
                self.vllm_client.update_named_param(name, param)
            elif self.vllm_mode == "colocate":

                pass

                pass

    @profiling_decorator
    def _move_model_to_vllm(self):
        # For DeepSpeed ZeRO-3 and FSDP, we need to gather all parameters before operations
        deepspeed_plugin = self.accelerator.state.deepspeed_plugin
        zero_stage_3 = deepspeed_plugin is not None and deepspeed_plugin.zero_stage == 3
        if zero_stage_3:
            import deepspeed

            gather_if_zero3 = deepspeed.zero.GatheredParameters
        else:
            gather_if_zero3 = nullcontext

        if is_peft_model(self.model):
            # With PEFT and FSDP/DeepSpeed ZeRO Stage 3, we must gather the full model at once before merging, as
            # merging adapters in a sharded manner is not supported.
            # TODO: does this work with FSDP?
            with gather_if_zero3(list(self.model.parameters())):
                self.model.merge_adapter()

                # Update vLLM weights while parameters are gathered
                if self.is_fsdp_enabled:  # note if using FSDP, gather_if_zero3 is nullcontext
                    # Update vLLM weights while parameters are gathered
                    # For PEFT with FSDP we need to use the memory efficient post-order traversal
                    fsdp_plugin = getattr(self.accelerator.state, "fsdp_plugin", None)
                    fsdp_version = getattr(fsdp_plugin, "fsdp_version", 1) if fsdp_plugin else 1
                    if fsdp_version == 1:
                        self._sync_fsdp1_params_to_vllm(
                            self.model
                        )  # use memory-efficient post-order traversal for FSDP
                    elif fsdp_version == 2:
                        self._sync_fsdp2_params_to_vllm(self.model)
                else:
                    # DeepSpeed ZeRO-3 with PEFT
                    for name, param in self.model.named_parameters():
                        # When using PEFT, we need to recover the original parameter name and discard some parameters
                        name = name.removeprefix("base_model.model.").replace(".base_layer", "")
                        if self.model.prefix in name:
                            continue
                        # When module to save, remove its prefix and discard the original module
                        if "original_module" in name:
                            continue
                        name = self._fix_param_name_to_vllm(name, extra_prefixes=["modules_to_save.default."])

                        if self.vllm_mode == "server" and self.accelerator.is_main_process:
                            self.vllm_client.update_named_param(name, param.data)
                        elif self.vllm_mode == "colocate":

                            pass

                            pass
                # Unmerge adapters while parameters are still gathered
                self.model.unmerge_adapter()
                # Parameters will automatically be repartitioned when exiting the context
        else:
            # For non-PEFT models, simply gather (if needed) and update each parameter individually.
            if self.is_fsdp_enabled:
                fsdp_plugin = getattr(self.accelerator.state, "fsdp_plugin", None)
                fsdp_version = getattr(fsdp_plugin, "fsdp_version", 1) if fsdp_plugin else 1
                if fsdp_version == 1:
                    self._sync_fsdp1_params_to_vllm(self.model)  # use memory-efficient post-order traversal for FSDP
                elif fsdp_version == 2:
                    self._sync_fsdp2_params_to_vllm(self.model)
            else:
                for name, param in self.model.named_parameters():
                    name = self._fix_param_name_to_vllm(name)
                    with gather_if_zero3([param]):
                        if self.vllm_mode == "server" and self.accelerator.is_main_process:
                            self.vllm_client.update_named_param(name, param.data)
                        elif self.vllm_mode == "colocate":

                            pass

                            pass

        # Reset cache on vLLM
        if self.vllm_mode == "server" and self.accelerator.is_main_process:
            self.vllm_client.reset_prefix_cache()
        elif self.vllm_mode == "colocate":
            self.llm.reset_prefix_cache()

    @profiling_decorator
    def _prepare_inputs(
        self, generation_batch: dict[str, Union[torch.Tensor, Any]]
    ) -> dict[str, Union[torch.Tensor, Any]]:
        # Prepares inputs for model training/evaluation by managing completion generation and batch handling.
        # During training:
        #   - Receives the local generation batch (Per-GPU batch size × steps per generation)
        #     from the modified training dataloader instead of the standard local batch
        #   - Generates completions once for the entire generation batch and splits it into batches of size
        #     `per_device_train_batch_size`
        #   - Buffers these completions and returns the appropriate slice for the current accumulation step
        #   - Optimizes by regenerating completions only periodically (every steps_per_generation * num_iterations)
        # During evaluation:
        #   - The input is treated as a standard local batch (no accumulation, no multiple iterations)
        #   - Completions are generated for each batch without buffering or reuse
        # Returns a single local batch in both cases.

        mode = "train" if self.model.training else "eval"
        if mode == "train":
            generate_every = self.args.steps_per_generation * self.num_iterations
            if self._step % generate_every == 0 or self._buffered_inputs is None:
                # self._buffered_inputs=None can occur when resuming from a checkpoint
                generation_batch = self._generate_and_score_completions(generation_batch)
                generation_batch = split_pixel_values_by_grid(generation_batch)

                try: generation_batch = shuffle_sequence_dict(generation_batch)

                except: pass
                generation_batches = split_tensor_dict(generation_batch, self.args.steps_per_generation)
                self._buffered_inputs = [unsplit_pixel_values_by_grid(batch) for batch in generation_batches]
            inputs = self._buffered_inputs[self._step % self.args.steps_per_generation]
            self._step += 1
        else:
            # In evaluation, there is neither batch grouping for generation, nor multiple iterations, hence
            # local generation batch == local eval batch
            inputs = self._generate_and_score_completions(generation_batch)
        return inputs

    @profiling_decorator
    def _calculate_rewards(self, inputs, prompts, completions, completion_ids_list):
        device = self.accelerator.device
        rewards_per_func = torch.zeros(len(prompts), len(self.reward_funcs), device=device)

        # Repeat all input columns (but "prompt", "completion", and "completion_ids") to match the num of generations
        keys = [key for key in inputs[0] if key not in ["prompt", "completion", "completion_ids"]]
        reward_kwargs = {key: [example[key] for example in inputs] for key in keys}

        # This allows for dynamic reward shaping based on training progress.
        reward_kwargs["trainer_state"] = self.state

        for i, (reward_func, reward_processing_class, reward_func_name) in enumerate(
            zip(self.reward_funcs, self.reward_processing_classes, self.reward_func_names)
        ):
            with profiling_context(self, reward_func_name):
                if isinstance(reward_func, nn.Module):  # Module (no PretrainedModel) for compat with compiled models
                    if is_conversational(inputs[0]):
                        messages = [{"messages": p + c} for p, c in zip(prompts, completions)]
                        texts = [apply_chat_template(x, reward_processing_class)["text"] for x in messages]
                    else:
                        texts = [p + c for p, c in zip(prompts, completions)]
                    reward_inputs = reward_processing_class(
                        text=texts, return_tensors="pt", padding=True, padding_side="right", add_special_tokens=False
                    )
                    reward_inputs = super()._prepare_inputs(reward_inputs)
                    with torch.inference_mode():
                        rewards_per_func[:, i] = reward_func(**reward_inputs).logits[:, 0]  # Shape (B*G,)
                else:
                    output_reward_func = reward_func(
                        prompts=prompts, completions=completions, completion_ids=completion_ids_list, **reward_kwargs
                    )
                    # Convert None values to NaN
                    output_reward_func = [reward if reward is not None else torch.nan for reward in output_reward_func]

                    rewards_per_func[:, i] = torch.tensor(output_reward_func, dtype=torch.float32, device=device)

        # If all reward functions return None for a given row, issue a detailed warning
        if torch.isnan(rewards_per_func).all(dim=1).any():
            nan_row_idx = torch.isnan(rewards_per_func).all(dim=1).nonzero(as_tuple=True)[0][0]
            row_reward_kwargs = {
                key: value[nan_row_idx] for key, value in reward_kwargs.items() if key != "trainer_state"
            }
            row_reward_kwargs["prompt"] = prompts[nan_row_idx]
            row_reward_kwargs["completion"] = completions[nan_row_idx]
            logger.warning(
                f"All reward functions returned None for the following kwargs:\n{row_reward_kwargs}\n"
                "Please ensure that at least one reward function returns a valid reward."
            )

        # Gather the reward per function: this part is crucial, because the rewards are normalized per group and the
        # completions may be distributed across processes
        rewards_per_func = gather(rewards_per_func)
        return rewards_per_func

    def _generate_single_turn(self, prompts: list[str], images: Optional[list]):
        device = self.accelerator.device

        # If the prompts are conversational and the inputs contain images, we need to convert the prompts from
        # [{"role": "user", "content": "What color is the sky?"}] to
        # [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "What color is the sky?"}]}]
        kwargs = {}
        if images is not None:
            kwargs = {"images": images}
            for prompt, image_list in zip(prompts, images):
                if isinstance(prompt, list):  # i.e., when using conversational data
                    prepare_multimodal_messages(prompt, num_images=len(image_list))

        prompts_text = [
            maybe_apply_chat_template({"prompt": prompt}, self.processing_class)["prompt"] for prompt in prompts
        ]

        if images is not None:
            prompt_inputs = self.processing_class(text=prompts_text, padding=True, return_tensors="pt", **kwargs)
            prompt_inputs = super()._prepare_inputs(prompt_inputs)
            forward_kwargs = {k: v for k, v in prompt_inputs.items() if k not in ["input_ids", "attention_mask"]}
        else:
            forward_kwargs = {}

        # Generate completions using either vLLM or regular generation
        if self.use_vllm:
            if self.vllm_mode == "colocate" and self.args.vllm_enable_sleep_mode:
                # wake up colocated vLLM instances if needed
                torch.cuda.empty_cache()  # required to avoid OOM in some cases
                self.llm.wake_up()

            # First, update the vLLM weights if needed
            if self.state.global_step != self._last_loaded_step:
                self._move_model_to_vllm()
                self._last_loaded_step = self.state.global_step

            # Generate completions using vLLM: gather all prompts and use them in a single call in the main process
            if self.vllm_mode == "server":
                all_prompts_text = gather_object(prompts_text)
                if images is not None:
                    all_images = gather_object(images)

                if self.accelerator.is_main_process:
                    # Since 'prompts' contains 'num_generations' duplicates, we first take unique prompts, and generate
                    # num_generations outputs for each one. This is faster than generating outputs for each duplicate
                    # prompt individually.
                    ordered_set_of_prompts = all_prompts_text[:: self.num_generations]

                    if images is not None:
                        ordered_set_of_images = all_images[:: self.num_generations]
                    else:
                        ordered_set_of_images = None

                    with profiling_context(self, "vLLM.generate"):
                        output = self.vllm_client.generate(
                            prompts=ordered_set_of_prompts,
                            images=ordered_set_of_images,
                            n=self.num_generations,
                            repetition_penalty=self.repetition_penalty,
                            temperature=self.temperature,
                            top_p=self.top_p,
                            top_k=-1 if self.top_k is None else self.top_k,
                            min_p=0.0 if self.min_p is None else self.min_p,
                            max_tokens=self.max_completion_length,
                            truncate_prompt_tokens=self.max_prompt_length,
                            guided_decoding_regex=self.guided_decoding_regex,
                            generation_kwargs=self.args.generation_kwargs,
                        )
                        payload = (output["prompt_ids"], output["completion_ids"], output["logprobs"])
                else:
                    payload = None

                # Broadcast the completions from the main process to all processes, ensuring each process receives its corresponding slice.
                obj_list = [payload]
                broadcast_object_list(obj_list, from_process=0)
                all_prompt_ids, all_completion_ids, _ = obj_list[0]

                # At this point, we only get 1 copy of each prompt, so we need to repeat them num_generations times
                all_prompt_ids = [ids for ids in all_prompt_ids for _ in range(self.num_generations)]

                process_slice = slice(
                    self.accelerator.process_index * len(prompts),
                    (self.accelerator.process_index + 1) * len(prompts),
                )
                prompt_ids = all_prompt_ids[process_slice]
                completion_ids = all_completion_ids[process_slice]

            # Generate completions using colocated vLLM instances: each device holds vLLM copy and work on their own batch of prompts
            elif self.vllm_mode == "colocate":
                if self.guided_decoding_regex:
                    guided_decoding = GuidedDecodingParams(regex=self.guided_decoding_regex)
                else:
                    guided_decoding = None

                generation_kwargs = {
                    "n": 1,  # vLLM on each GPU generates only 1 in colocate mode
                    "repetition_penalty": self.repetition_penalty,
                    "temperature": self.temperature,
                    "top_p": self.top_p,
                    "top_k": -1 if self.top_k is None else self.top_k,
                    "min_p": 0.0 if self.min_p is None else self.min_p,
                    "max_tokens": self.max_completion_length,
                    "truncate_prompt_tokens": self.max_prompt_length,
                    "guided_decoding": guided_decoding,
                }
                if self.args.generation_kwargs is not None:
                    generation_kwargs.update(self.args.generation_kwargs)
                sampling_params = SamplingParams(**grpo_update_SamplingParams(SamplingParams, generation_kwargs, getattr(self.args, 'vllm_sampling_params', None)))

                if self.vllm_tensor_parallel_size > 1:
                    # Gather prompts from all ranks in the TP group and flatten.
                    # Each rank starts with its own prompts; after gathering, all ranks see the full group set.
                    orig_size = len(prompts_text)
                    gathered_prompts = [None for _ in range(self.vllm_tensor_parallel_size)]
                    torch.distributed.all_gather_object(gathered_prompts, prompts_text, group=self.tp_group)
                    all_prompts_text = [p for sublist in gathered_prompts for p in sublist]

                    if images is not None:
                        gathered_images = [None for _ in range(self.vllm_tensor_parallel_size)]
                        torch.distributed.all_gather_object(gathered_images, images, group=self.tp_group)
                        all_images = [img for sublist in gathered_images for img in sublist]
                    else:
                        all_images = None
                else:
                    all_prompts_text = prompts_text
                    all_images = images

                if images is not None and all_images:
                    vllm_inputs = []
                    for prompt, image_list in zip(all_prompts_text, all_images):
                        vllm_inputs.append({"prompt": prompt, "multi_modal_data": {"image": image_list}})

                else:
                    vllm_inputs = all_prompts_text

                with profiling_context(self, "vLLM.generate"):
                    all_outputs = self.llm.generate(vllm_inputs, sampling_params=sampling_params, use_tqdm=False, lora_request = self.model.load_lora('rloo_trainer_lora_model', load_tensors = True))

                all_prompt_ids = [output.prompt_token_ids for output in all_outputs]
                all_completion_ids = [output.token_ids for outputs in all_outputs for output in outputs.outputs]

                if self.vllm_tensor_parallel_size > 1:
                    # Slice completions for this rank within its TP group.
                    # Each rank generates all outputs — we keep only our share.
                    local_rank_in_group = torch.distributed.get_rank(group=self.tp_group)
                    tp_slice = slice(local_rank_in_group * orig_size, (local_rank_in_group + 1) * orig_size)
                    prompt_ids = all_prompt_ids[tp_slice]
                    completion_ids = all_completion_ids[tp_slice]
                else:
                    prompt_ids = all_prompt_ids
                    completion_ids = all_completion_ids

                if self.args.vllm_enable_sleep_mode:
                    self.llm.sleep(level=1)

        elif self.use_transformers_paged:
            # Re-process inputs for paged generation if needed
            # Note: images are already validated and preprocessed above
            paged_prompt_inputs = self.processing_class(text=prompts_text, **kwargs)
            previous_attn = self.model_wrapped.config._attn_implementation

            if is_flash_attn_2_available():
                self.model_wrapped.config._attn_implementation = "paged_attention"
            else:
                self.model_wrapped.config._attn_implementation = "sdpa_paged"
            with (
                profiling_context(self, "transformers.generate_batch"),
                unwrap_model_for_generation(
                    self.model_wrapped, self.accelerator, gather_deepspeed3_params=self.args.ds3_gather_for_generation
                ) as unwrapped_model,
                torch.no_grad(),
                FSDP.summon_full_params(self.model_wrapped, recurse=False) if self.is_fsdp_enabled else nullcontext(),
            ):
                # Cast to the appropriate dtype based on training configuration
                if self.args.bf16:
                    unwrapped_model.to(torch.bfloat16)
                elif self.args.fp16:
                    unwrapped_model.to(torch.float16)
                with torch.inference_mode():
                    all_outputs = unwrapped_model.generate_batch(
                        paged_prompt_inputs.input_ids, generation_config=self.generation_config, progress_bar=False
                    )
                    unwrapped_model.train()  # restore training mode, as generate_batch forces eval mode
            completion_ids = [output.generated_tokens for output in all_outputs.values()]
            prompt_ids = paged_prompt_inputs.input_ids
            # Restore the original attention implementation, training mode
            self.model_wrapped.config._attn_implementation = previous_attn

        else:
            # Regular generation path
            generate_inputs = self.processing_class(
                text=prompts_text,
                return_tensors="pt",
                padding=True,
                padding_side="left",
                max_length=self.max_prompt_length,
                truncation=True,
                add_special_tokens=False,
                **kwargs,
            )
            generate_inputs = super()._prepare_inputs(generate_inputs)

            with (
                profiling_context(self, "transformers.generate"),
                unwrap_model_for_generation(
                    self.model_wrapped, self.accelerator, gather_deepspeed3_params=self.args.ds3_gather_for_generation
                ) as unwrapped_model,
                torch.no_grad(),
                FSDP.summon_full_params(self.model_wrapped, recurse=False) if self.is_fsdp_enabled else nullcontext(),
            ):
                prompt_completion_ids = unwrapped_model.generate(
                    **generate_inputs, generation_config=self.generation_config, disable_compile=True
                )
            # Compute prompt length and extract completion ids
            prompt_ids, prompt_mask = generate_inputs["input_ids"], generate_inputs["attention_mask"]
            prompt_length = prompt_ids.size(1)
            completion_ids = prompt_completion_ids[:, prompt_length:]

            # Mask everything after the first EOS token
            is_eos = completion_ids == self.eos_token_id
            eos_idx = torch.full((is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device)
            eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)]
            sequence_indices = torch.arange(is_eos.size(1), device=device).expand(is_eos.size(0), -1)
            completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int()
            prompt_ids = [p[m].tolist() for p, m in zip(prompt_ids, prompt_mask.bool())]
            completion_ids = [c[m].tolist() for c, m in zip(completion_ids, completion_mask.bool())]

        return prompt_ids, completion_ids, forward_kwargs

    def _generate(self, prompts: list[str], images: Optional[list]):
        device = self.accelerator.device
        mode = "train" if self.model.training else "eval"

        prompt_ids, completion_ids, forward_kwargs = self._generate_single_turn(prompts, images)

        # Get completion length per sequence, used for logging
        prompt_lengths = torch.tensor([len(ids) for ids in prompt_ids], device=device)
        completion_lengths = torch.tensor([len(ids) for ids in completion_ids], device=device)
        agg_prompt_lengths = self.accelerator.gather(prompt_lengths)
        agg_completion_lengths = self.accelerator.gather(completion_lengths)
        total_prompt_tokens = agg_prompt_lengths.sum()
        total_completion_tokens = agg_completion_lengths.sum()  # = num_items_in_batch, required for the DAPO loss

        # Log the metrics
        if mode == "train":
            self.state.num_input_tokens_seen += (total_prompt_tokens + total_completion_tokens).item()
        self._metrics[mode]["num_tokens"] = [self.state.num_input_tokens_seen]

        # Log completion lengths, mean, min, max
        agg_completion_lengths = self.accelerator.gather(completion_lengths)
        self._metrics[mode]["completions/mean_length"].append(agg_completion_lengths.float().mean().item())
        self._metrics[mode]["completions/min_length"].append(agg_completion_lengths.float().min().item())
        self._metrics[mode]["completions/max_length"].append(agg_completion_lengths.float().max().item())

        # Identify sequences that terminated with EOS and log their lengths
        eos_and_pad = [self.eos_token_id, self.pad_token_id]
        is_truncated = torch.tensor([ids[-1] not in eos_and_pad for ids in completion_ids], device=device)
        agg_is_truncated = self.accelerator.gather(is_truncated)
        self._metrics[mode]["completions/clipped_ratio"].append(agg_is_truncated.float().mean().item())
        term_completion_lengths = agg_completion_lengths[~agg_is_truncated]
        if len(term_completion_lengths) == 0:  # edge case where no terminated sequences are found
            term_completion_lengths = torch.zeros(1, device=device)
        self._metrics[mode]["completions/mean_terminated_length"].append(term_completion_lengths.float().mean().item())
        self._metrics[mode]["completions/min_terminated_length"].append(term_completion_lengths.float().min().item())
        self._metrics[mode]["completions/max_terminated_length"].append(term_completion_lengths.float().max().item())

        return prompt_ids, completion_ids, forward_kwargs

    def _generate_and_score_completions(
        self, inputs: list[dict[str, Union[torch.Tensor, Any]]]
    ) -> dict[str, Union[torch.Tensor, Any]]:
        device = self.accelerator.device
        mode = "train" if self.model.training else "eval"

        prompts = [x["prompt"] for x in inputs]

        if "images" in inputs[0]:
            images = [example.get("images") for example in inputs]
        elif "image" in inputs[0]:
            images = [[example.get("image")] if example.get("image") is not None else None for example in inputs]
        else:
            images = None
        # Transformers requires at least one image in the batch, otherwise it throws an error
        if images is not None and all(img_list == [] for img_list in images):
            images = None

        prompt_ids_list, completion_ids_list, forward_kwargs = self._generate(prompts, images)

        # Convert lists of token IDs to padded tensors
        prompt_ids = [torch.tensor(ids, device=device) for ids in prompt_ids_list]
        prompt_mask = [torch.ones_like(ids, dtype=torch.long) for ids in prompt_ids]
        prompt_ids = pad(prompt_ids, padding_value=self.pad_token_id, padding_side="left")
        prompt_mask = pad(prompt_mask, padding_value=0, padding_side="left")
        completion_ids = [torch.tensor(ids, device=device) for ids in completion_ids_list]
        completion_mask = [torch.ones_like(ids, dtype=torch.long) for ids in completion_ids]
        completion_ids = pad(completion_ids, padding_value=self.pad_token_id, padding_side="right")
        completion_mask = pad(completion_mask, padding_value=0, padding_side="right")

        # If mask_truncated_completions is enabled, zero out truncated completions in completion_mask
        if self.mask_truncated_completions:
            eos_and_pad = [self.eos_token_id, self.pad_token_id]
            is_truncated = torch.tensor([ids[-1] not in eos_and_pad for ids in completion_ids_list], device=device)
            completion_mask = completion_mask * (~is_truncated).unsqueeze(1).int()

        # Concatenate prompt_mask with completion_mask for logit computation
        prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1)  # (B, P+C)
        attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)  # (B, P+C)
        # If token_type_ids are used, extend them with zeros for the completion part
        if "token_type_ids" in forward_kwargs:
            token_type_ids = forward_kwargs["token_type_ids"]
            forward_kwargs["token_type_ids"] = torch.cat(
                [token_type_ids, token_type_ids.new_zeros(completion_ids.shape)], dim=1
            )

        logits_to_keep = completion_ids.size(1)  # we only need to compute the logits for the completion tokens
        batch_size = self.args.per_device_train_batch_size if mode == "train" else self.args.per_device_eval_batch_size

        num_images = [len(img_list) for img_list in images] if images is not None else None

        with torch.no_grad():
            # Compute the per-token log probabilities for the current model
            old_per_token_logps, _ = self._get_per_token_logps_and_entropies(
                self.model,
                prompt_completion_ids,
                attention_mask,
                logits_to_keep,
                batch_size,
                num_images=num_images,
                **forward_kwargs,  # may contain pixel_values, image_grid_thw, pixel_attention_mask and image_sizes
            )
            old_logps = (old_per_token_logps * completion_mask).sum(1)  # mask out padding and tokens after EOS

            # Compute the per-token log probabilities for the reference model
            if self.beta != 0.0:
                if self.ref_model is not None:
                    ref_per_token_logps, _ = self._get_per_token_logps_and_entropies(
                        self.ref_model,
                        prompt_completion_ids,
                        attention_mask,
                        logits_to_keep,
                        batch_size=batch_size,
                        num_images=num_images,
                        **forward_kwargs,  # may contain pixel_values, image_grid_thw, pixel_attention_mask and image_sizes
                    )
                else:
                    with self.accelerator.unwrap_model(self.model).disable_adapter():
                        ref_per_token_logps, _ = self._get_per_token_logps_and_entropies(
                            self.model,
                            prompt_completion_ids,
                            attention_mask,
                            logits_to_keep,
                            batch_size=batch_size,
                            num_images=num_images,
                            **forward_kwargs,  # may contain pixel_values, image_grid_thw, pixel_attention_mask and image_sizes
                        )
            else:
                ref_per_token_logps = None

        # Decode
        prompts_text = self.processing_class.batch_decode(prompt_ids, skip_special_tokens=True)
        completions_text = self.processing_class.batch_decode(completion_ids, skip_special_tokens=True)
        if is_conversational(inputs[0]):
            completions = []
            for prompt, completion in zip(prompts, completions_text):
                bootstrap = prompt.pop()["content"] if prompt[-1]["role"] == "assistant" else ""
                completions.append([{"role": "assistant", "content": bootstrap + completion}])
        else:
            completions = completions_text

        # Calculate rewards for each reward function. rewards_per_func aggregates rewards across all processes. This is
        # important because rewards will be normalized per group, and completions are distributed. We will later slice
        # rewards_per_func to extract each process's subset.
        rewards_per_func = self._calculate_rewards(inputs, prompts, completions, completion_ids_list)

        # Apply weights to each reward function's output and sum
        rewards = (rewards_per_func * self.reward_weights.to(device).unsqueeze(0)).nansum(dim=1)

        # Apply reward clipping if specified
        if self.reward_clip_range:
            rewards = rewards.clamp(min=self.reward_clip_range[0], max=self.reward_clip_range[1])

        # Include the KL penalty in the reward
        if self.beta != 0.0:
            per_token_kl = old_per_token_logps - ref_per_token_logps
            # Apply sequence-level KL penalty to rewards (sum KL across tokens first, then apply to each sequence)
            kl = (per_token_kl * completion_mask).sum(-1)
            kl = gather(kl)  # rewards are gathered, so kl must be too
            rewards = rewards - self.beta * kl

        grouped_rewards = rewards.view(-1, self.num_generations)
        mean_grouped_rewards = grouped_rewards.mean(dim=1)
        std_rewards = grouped_rewards.std(dim=1)
        is_std_zero = torch.isclose(std_rewards, torch.zeros_like(std_rewards))

        # RLOO advantages computation
        grouped_sum = grouped_rewards.sum(dim=1, keepdim=True)  # (num_prompts, 1)
        baselines = (grouped_sum - grouped_rewards) / (self.num_generations - 1)  # (num_prompts, num_generations)
        baselines = baselines.view(-1)  # Flatten back to match rewards shape
        advantages = rewards - baselines

        # Normalize advantages
        if self.normalize_advantages:
            advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-4)

        # Slice to keep only the local part of the data
        process_slice = slice(
            self.accelerator.process_index * len(prompts),
            (self.accelerator.process_index + 1) * len(prompts),
        )
        all_process_advantages = advantages.clone()  # keep the aggregated advantages for logging
        advantages = advantages[process_slice]

        # Calculate and log the mean KL divergence between current and reference model
        if self.beta != 0.0:
            mean_kl = (per_token_kl * completion_mask).sum() / completion_mask.sum().clamp(min=1.0)
            self._metrics[mode]["kl"].append(self.accelerator.gather(mean_kl).nanmean().item())

        # Calculate mean reward per function, but only for samples where the function was applied (non-NaN values)
        for i, reward_func_name in enumerate(self.reward_func_names):
            mean_rewards = torch.nanmean(rewards_per_func[:, i]).item()
            self._metrics[mode][f"rewards/{reward_func_name}/mean"].append(mean_rewards)
            std_func_rewards = nanstd(rewards_per_func[:, i]).item()
            self._metrics[mode][f"rewards/{reward_func_name}/std"].append(std_func_rewards)
        self._metrics[mode]["reward"].append(mean_grouped_rewards.mean().item())
        self._metrics[mode]["reward_std"].append(std_rewards.mean().item())
        self._metrics[mode]["frac_reward_zero_std"].append(is_std_zero.float().mean().item())

        # Log prompt and completion texts
        self._logs["prompt"].extend(gather_object(prompts_text))
        self._logs["completion"].extend(gather_object(completions_text))
        for i, name in enumerate(self.reward_func_names):
            self._logs["rewards"][name].extend(rewards_per_func[:, i].tolist())
        self._logs["advantages"].extend(all_process_advantages.tolist())

        if images is not None:
            self._logs["images"].extend(gather_object(images))

        output = {
            "prompt_ids": prompt_ids,
            "prompt_mask": prompt_mask,
            "completion_ids": completion_ids,
            "completion_mask": completion_mask,
            "old_logps": old_logps,
            "advantages": advantages,
        }
        if "pixel_values" in forward_kwargs:
            output["pixel_values"] = forward_kwargs["pixel_values"]
        if "image_grid_thw" in forward_kwargs:
            output["image_grid_thw"] = forward_kwargs["image_grid_thw"]
        if "pixel_attention_mask" in forward_kwargs:
            output["pixel_attention_mask"] = forward_kwargs["pixel_attention_mask"]
        if "image_sizes" in forward_kwargs:
            output["image_sizes"] = forward_kwargs["image_sizes"]
        if "token_type_ids" in forward_kwargs:
            output["token_type_ids"] = forward_kwargs["token_type_ids"]
        if images is not None:
            output["num_images"] = num_images
        return output

    @profiling_decorator
    def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
        if return_outputs:
            raise ValueError("The RLOOTrainer does not support returning outputs")
        return self._compute_loss(model, inputs)

    def _compute_loss(self, model, inputs):
        # Compute the per-token log probabilities for the model
        prompt_ids, prompt_mask = inputs["prompt_ids"], inputs["prompt_mask"]
        completion_ids, completion_mask = inputs["completion_ids"], inputs["completion_mask"]
        input_ids = torch.cat([prompt_ids, completion_ids], dim=1)
        attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)
        logits_to_keep = completion_ids.size(1)  # we only need to compute the logits for the completion tokens

        # Compute the per_token_logps and the entropy at each position in the completion
        per_token_logps, entropies = self._get_per_token_logps_and_entropies(
            model,
            input_ids,
            attention_mask,
            logits_to_keep,
            compute_entropy=True,
            pixel_values=inputs.get("pixel_values"),
            image_grid_thw=inputs.get("image_grid_thw"),
            num_images=inputs.get("num_images"),
            pixel_attention_mask=inputs.get("pixel_attention_mask"),
            image_sizes=inputs.get("image_sizes"),
            token_type_ids=inputs.get("token_type_ids"),
        )

        logps = (per_token_logps * completion_mask).sum(1)  # mask out padding and tokens after EOS
        old_logps = inputs["old_logps"]
        log_ratio = logps - old_logps

        # Compute the loss
        advantages = inputs["advantages"]
        coef_1 = torch.exp(log_ratio)
        coef_2 = torch.clamp(coef_1, 1 - self.epsilon_low, 1 + self.epsilon_high)
        per_sequence_loss1 = coef_1 * advantages
        per_sequence_loss2 = coef_2 * advantages
        per_sequence_loss = -torch.min(per_sequence_loss1, per_sequence_loss2)
        loss = per_sequence_loss.mean()

        # Log the metrics
        mode = "train" if self.model.training else "eval"

        # Entropy
        mean_entropy = (entropies * completion_mask).sum() / completion_mask.sum().clamp(min=1.0)
        self._metrics[mode]["entropy"].append(self.accelerator.gather(mean_entropy).nanmean().item())

        # Compute the clipped probability ratios
        is_low_clipped = (coef_1 < 1 - self.epsilon_low) & (advantages < 0)
        is_high_clipped = (coef_1 > 1 + self.epsilon_high) & (advantages > 0)
        is_region_clipped = is_low_clipped | is_high_clipped
        gathered_low_clip = self.accelerator.gather(is_low_clipped.float().mean())
        self._metrics[mode]["clip_ratio/low_mean"].append(gathered_low_clip.nanmean().item())
        self._metrics[mode]["clip_ratio/low_min"].append(nanmin(gathered_low_clip).item())
        gathered_high_clip = self.accelerator.gather(is_high_clipped.float().mean())
        self._metrics[mode]["clip_ratio/high_mean"].append(gathered_high_clip.nanmean().item())
        self._metrics[mode]["clip_ratio/high_max"].append(nanmax(gathered_high_clip).item())
        gathered_clip_ratio = self.accelerator.gather(is_region_clipped.float().mean())
        self._metrics[mode]["clip_ratio/region_mean"].append(gathered_clip_ratio.nanmean().item())
        return loss

    def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys: Optional[list[str]] = None):
        inputs = self._prepare_inputs(inputs)
        with torch.no_grad():
            with self.compute_loss_context_manager():
                loss = self.compute_loss(model, inputs)
            loss = loss.mean().detach()
        return loss, None, None

    def log(self, logs: dict[str, float], start_time: Optional[float] = None) -> None:
        mode = "train" if self.model.training else "eval"
        metrics = {key: sum(val) / len(val) for key, val in self._metrics[mode].items()}  # average the metrics

        # This method can be called both in training and evaluation. When called in evaluation, the keys in `logs`
        # start with "eval_". We need to add the prefix "eval_" to the keys in `metrics` to match the format.
        if mode == "eval":
            metrics = {f"eval_{key}": val for key, val in metrics.items()}

        logs = {**logs, **metrics}
        super().log(logs, start_time)
        self._metrics[mode].clear()

        if self.accelerator.is_main_process and self.log_completions:
            if is_rich_available():
                print_prompt_completions_sample(
                    self._logs["prompt"],
                    self._logs["completion"],
                    self._logs["rewards"],
                    self._logs["advantages"],
                    self.state.global_step,
                    self.num_completions_to_print,
                )

            if self.args.report_to and "wandb" in self.args.report_to and wandb.run is not None:
                import pandas as pd

                table = {
                    "step": [str(self.state.global_step)] * len(self._logs["prompt"]),
                    "prompt": self._logs["prompt"],
                    "completion": self._logs["completion"],
                    **self._logs["rewards"],
                    "advantage": self._logs["advantages"],
                }

                if self._logs["images"]:
                    table["images"] = []
                    for image_list in self._logs["images"]:
                        # Convert images to wandb Image objects for proper visualization
                        table["images"].append([wandb.Image(image) for image in image_list])

                df = pd.DataFrame(table)
                if self.wandb_log_unique_prompts:
                    df = df.drop_duplicates(subset=["prompt"])
                wandb.log({"completions": wandb.Table(dataframe=df)})

    # 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 UnslothRLOOTrainer(_UnslothRLOOTrainer):
    """
    
    Trainer for the Reinforce Leave One Out (RLOO) method. This algorithm was initially proposed in the paper [Back to
    Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in
    LLMs](https://huggingface.co/papers/2402.14740).

    Example:

    ```python
    from datasets import load_dataset
    from trl import RLOOTrainer

    dataset = load_dataset("trl-lib/tldr", split="train")
    def reward_func(completions, **kwargs):
        # Dummy reward function that rewards completions with more unique letters.
        return [float(len(set(completion))) for completion in completions]
    trainer = RLOOTrainer(
        model="Qwen/Qwen2-0.5B-Instruct",
        reward_funcs=reward_func,
        train_dataset=dataset,
    )

    trainer.train()
    ```

    Args:
        model (`Union[str, PreTrainedModel]`):
            Model to be trained. Can be either:

            - A string, being the *model id* of a pretrained model hosted inside a model repo on huggingface.co, or a
              path to a *directory* containing model weights saved using
              [`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded
              using [`~transformers.AutoModelForCausalLM.from_pretrained`] with the keyword arguments in
              `args.model_init_kwargs`.
            - A [`~transformers.PreTrainedModel`] object. Only causal language models are supported.
        reward_funcs (`Union[RewardFunc, list[RewardFunc]]`):
            Reward functions to be used for computing the rewards. To compute the rewards, we call all the reward
            functions with the prompts and completions and sum the rewards. Can be either:

            - A single reward function, such as:
                - A string: The *model ID* of a pretrained model hosted inside a model repo on huggingface.co, or a
                path to a *directory* containing model weights saved using
                [`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded
                using [`~transformers.AutoModelForSequenceClassification.from_pretrained`] with `num_labels=1` and the
                keyword arguments in `args.model_init_kwargs`.
                - A [`~transformers.PreTrainedModel`] object: Only sequence classification models are supported.
                - A custom reward function: The function is provided with the prompts and the generated completions,
                  plus any additional columns in the dataset. It should return a list of rewards. Custom reward
                  functions can also return `None` when the reward is not applicable to those samples. This is useful
                  for multi-task training where different reward functions apply to different types of samples. When a
                  reward function returns `None` for a sample, that reward function is excluded from the reward
                  calculation for that sample. For more details, see [Using a custom reward
                  function](#using-a-custom-reward-function).

                  The trainer's state is also passed to the reward function. The trainer's state is an instance of
                  [`~transformers.TrainerState`] and can be accessed by accessing the `trainer_state` argument to the
                  reward function's signature.
            - A list of reward functions, where each item can independently be any of the above types. Mixing different
            types within the list (e.g., a string model ID and a custom reward function) is allowed.
        args ([`RLOOConfig`], *optional*):
            Configuration for this trainer. If `None`, a default configuration is used.
        train_dataset ([`~datasets.Dataset`] or [`~datasets.IterableDataset`]):
            Dataset to use for training. It must include a column `"prompt"`. Any additional columns in the dataset is
            ignored. The format of the samples can be either:

            - [Standard](dataset_formats#standard): Each sample contains plain text.
            - [Conversational](dataset_formats#conversational): Each sample contains structured messages (e.g., role
              and content).
        eval_dataset ([`~datasets.Dataset`], [`~datasets.IterableDataset`] or `dict[str, Union[Dataset, IterableDataset]]`):
            Dataset to use for evaluation. It must meet the same requirements as `train_dataset`.
        processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.ProcessorMixin`], *optional*):
            Processing class used to process the data. The padding side must be set to "left". If `None`, the
            processing class is loaded from the model's name with [`~transformers.AutoProcessor.from_pretrained`]. A
            padding token, `tokenizer.pad_token`, must be set. If the processing class has not set a padding token,
            `tokenizer.eos_token` will be used as the default.
        reward_processing_classes ([`~transformers.PreTrainedTokenizerBase`] or `list[PreTrainedTokenizerBase]`, *optional*):
            Processing classes corresponding to the reward functions specified in `reward_funcs`. Can be either:

            - A single processing class: Used when `reward_funcs` contains only one reward function.
            - A list of processing classes: Must match the order and length of the reward functions in `reward_funcs`.
            If set to `None`, or if an element of the list corresponding to a [`~transformers.PreTrainedModel`] is
            `None`, the tokenizer for the model is automatically loaded using
            [`~transformers.AutoTokenizer.from_pretrained`]. For elements in `reward_funcs` that are custom reward
            functions (not [`~transformers.PreTrainedModel`]), the corresponding entries in `reward_processing_classes`
            are ignored.
        callbacks (list of [`~transformers.TrainerCallback`], *optional*):
            List of callbacks to customize the training loop. Will add those to the list of default callbacks detailed
            in [here](https://huggingface.co/docs/transformers/main_classes/callback).

            If you want to remove one of the default callbacks used, use the [`~transformers.Trainer.remove_callback`]
            method.
        optimizers (`tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`, *optional*, defaults to `(None, None)`):
            A tuple containing the optimizer and the scheduler to use. Will default to an instance of [`AdamW`] on your
            model and a scheduler given by [`get_linear_schedule_with_warmup`] controlled by `args`.
        peft_config ([`~peft.PeftConfig`], *optional*):
            PEFT configuration used to wrap the model. If `None`, the model is not wrapped.

        config:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `args` instead.

            </Deprecated>

        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>

        policy:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. Use `model` instead.

            </Deprecated>

        ref_policy:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. To use the initial model as the
            reference model, simply omit this parameter. The parameter is ignored.

            </Deprecated>

        data_collator:

            <Deprecated version="0.22.0">

            This parameter is deprecated and will be removed in version 0.25.0. The RLOOTrainer does not use a data
            collator, so this parameter is ignored.

            </Deprecated>
    
    """
    def __init__(
        self,
        model = None,
        reward_funcs = None,
        args = None,
        train_dataset = None,
        eval_dataset = None,
        processing_class = None,
        reward_processing_classes = None,
        callbacks = None,
        peft_config = None,
        config = None,
        reward_model = None,
        policy = None,
        ref_policy = None,
        data_collator = None,
        **kwargs
    ):
        if args is None: args = UnslothRLOOConfig()
        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().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'
        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 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()
        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('rloo_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()
        super().__init__(
            model = model,
            reward_funcs = reward_funcs,
            args = args,
            train_dataset = train_dataset,
            eval_dataset = eval_dataset,
            processing_class = processing_class,
            reward_processing_classes = reward_processing_classes,
            callbacks = callbacks,
            peft_config = peft_config,
            config = config,
            reward_model = reward_model,
            policy = policy,
            ref_policy = ref_policy,
            data_collator = data_collator,**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`"))