unsloth_compiled_cache/UnslothPPOTrainer.py

"""
2026.3.2
2026.3.4
5.3.0
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 unsloth_zoo.temporary_patches.common import torch_compile
from typing import Any, List, Optional, Tuple, Union, Dict, Set, Callable
from trl.trainer.ppo_trainer import (Accelerator, BaseImageProcessor, BaseTrainer, CallbackHandler, DEFAULT_CALLBACKS, DEFAULT_PROGRESS_CALLBACK, DataCollatorWithPadding, DataLoader, Dataset, ExportableState, FeatureExtractionMixin, GenerationConfig, INVALID_LOGPROB, OnlineTrainerState, Optional, PPOConfig, PPOTrainer, Path, PeftConfig, PeftModel, PolicyAndValueWrapper, PreTrainedTokenizerBase, PrinterCallback, ProcessorMixin, TrainerCallback, TrainerControl, Union, batch_generation, broadcast, contextmanager, create_reference_model, defaultdict, disable_dropout_in_model, empty_cache, exact_div, first_true_indices, forward, gather_object, gc, get_peft_model, get_reporting_integration_callbacks, get_reward, is_peft_available, is_rich_available, log_table_to_comet_experiment, masked_mean, masked_whiten, math, nn, np, nullcontext, os, pd, peft_module_casting_to_bf16, prepare_deepspeed, print_rich_table, selective_log_softmax, textwrap, time, torch, truncate_response, unwrap_model_for_generation, warnings, Accelerator, BaseImageProcessor, CallbackHandler, DEFAULT_CALLBACKS, DEFAULT_PROGRESS_CALLBACK, DataCollatorWithPadding, DataLoader, Dataset, ExportableState, FeatureExtractionMixin, OnlineTrainerState, Optional, PPOConfig, PeftConfig, PeftModel, PolicyAndValueWrapper, PreTrainedTokenizerBase, PrinterCallback, ProcessorMixin, TrainerCallback, TrainerControl, Union, broadcast, create_reference_model, disable_dropout_in_model, exact_div, forward, get_peft_model, get_reporting_integration_callbacks, is_peft_available, math, nn, os, pd, peft_module_casting_to_bf16, prepare_deepspeed, time, torch, warnings, PeftModel, is_peft_available, os, torch)


import os
import math
import logging
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
from unsloth_zoo.device_type import DEVICE_TYPE, device_synchronize

# Wrap trainer with padding to right and enable training mode
# Also patches W&B since multiple runs must use wandb.finish()
import functools
from types import MethodType
try:
    from unsloth_zoo.gradient_checkpointing import reset_unsloth_gradient_checkpointing_buffers
except:
    def reset_unsloth_gradient_checkpointing_buffers(): pass
def prepare_for_training_mode(f):
    @functools.wraps(f)
    def wrapper(self, *args, **kwargs):
        # Enable training mode
        _was_training = None
        # Get gradient checkpointing setting from training arguments
        use_gc = getattr(self.args, 'gradient_checkpointing', True)
        if hasattr(self, 'model') and hasattr(self.model, "training"):
            _was_training = self.model.training
        if hasattr(self, 'model') and hasattr(self.model, "for_training"):
            self.model.for_training(use_gradient_checkpointing=use_gc)
        output = f(self, *args, **kwargs)
        # Restore previous mode when possible
        if hasattr(self, 'model') and hasattr(self.model, "for_inference"):
            if _was_training is False:
                self.model.for_inference()
            elif _was_training is True and hasattr(self.model, "for_training"):
                self.model.for_training(use_gradient_checkpointing=use_gc)
        # Reset gradient checkpointing buffers to free memory while staying ready for next run
        try:
            reset_unsloth_gradient_checkpointing_buffers()
        except:
            pass
        # Patch W&B to enable logging on future runs, otherwise it'll overwrite the first run
        try:
            import wandb
            wandb.finish()
        except:
            pass
        return output
    return wrapper
pass

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

@torch.compile(dynamic = True, fullgraph = True, options = torch_compile_options,)
def chunked_hidden_states_selective_log_softmax(
    hidden_states: torch.Tensor,
    lm_head: torch.Tensor,
    index: torch.Tensor,
    chunks: int = 4,
    logit_scale_multiply: float = 0.0,
    logit_scale_divide: float = 0.0,
    logit_softcapping: float = 0.0,
    temperature: float = 1.0,
) -> torch.Tensor:
    # All Unsloth Zoo code licensed under AGPL3
    flat_hidden_states = hidden_states.reshape(-1, hidden_states.shape[-1])
    flat_index = index.reshape(-1)

    chunked_hidden_states = torch.chunk(flat_hidden_states, chunks=chunks, dim=0)
    chunked_index = torch.chunk(flat_index, chunks=chunks, dim=0)

    all_per_token_logps = []

    for chunk_hidden_states, chunk_index in zip(chunked_hidden_states, chunked_index):
        chunk_logits = chunk_hidden_states.to(lm_head.dtype) @ lm_head.t()

        if logit_scale_multiply != 0.0:
            chunk_logits = chunk_logits * logit_scale_multiply
        if logit_scale_divide != 0.0:
            chunk_logits = chunk_logits / logit_scale_divide
        if logit_softcapping != 0.0:
            chunk_logits = chunk_logits * torch.tanh(chunk_logits / logit_softcapping)

        chunk_logits = chunk_logits.to(torch.float32)

        if temperature != 1.0:
            chunk_logits = chunk_logits / temperature

        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)

    all_per_token_logps = torch.concat(all_per_token_logps)

    all_per_token_logps = all_per_token_logps.reshape((hidden_states.shape[0], hidden_states.shape[1]))
    return all_per_token_logps

@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 autotune_batch_and_chunks(
    total_input_rows,
    seq_len,
    hidden_size,
    vocab_size,
    dtype_bytes=16,
    multiplier=None
):
    if multiplier is None:
        final_m = max(4, seq_len // 4096)
    else:
        final_m = multiplier

    if torch.cuda.is_available():
        free_bytes, _ = torch.cuda.mem_get_info()
        limit_gb = (free_bytes / (1024**3))*.80
    elif hasattr(torch, "xpu") and torch.xpu.is_available():
        # For XPU: estimate free memory from total - reserved
        total_mem = torch.xpu.get_device_properties(0).total_memory
        reserved_mem = torch.xpu.memory_reserved()
        free_bytes = total_mem - reserved_mem
        limit_gb = (free_bytes / (1024**3)) * 0.80
    else:
        # Fallback: assume 8GB available
        limit_gb = 8.0

    bytes_to_gb = 1024**3

    b_vals = torch.arange(total_input_rows, 0, -1, device='cpu', dtype=torch.float32)

    hidden_gb = (b_vals * seq_len * hidden_size * dtype_bytes) / bytes_to_gb

    base_logits = ((b_vals/total_input_rows) * b_vals * seq_len * vocab_size * dtype_bytes) / bytes_to_gb
    logits_gb = base_logits / final_m

    total_mem_gb = hidden_gb + logits_gb

    valid_mask = total_mem_gb <= limit_gb
    valid_indices = torch.nonzero(valid_mask, as_tuple=False)

    if valid_indices.shape[0] == 0:
        #This means your GPU will OOM
        return 4, final_m

    best_idx = valid_indices[0].item()
    final_b = int(b_vals[best_idx].item())

    return final_b, final_m

def sanitize_logprob(logprob):
    """Local port of trl.scripts.vllm_serve.sanitize_logprob.
    Filters NaN logprobs from vLLM outputs."""
    value = logprob.logprob
    if math.isnan(value):
        logging.getLogger(__name__).warning(
            f"Generated NaN logprob, token logprob '{logprob}' will be ignored"
        )
        return None
    return value
@dataclass
class UnslothPPOConfig(PPOConfig):
    """
    
    Configuration class for the [`PPOTrainer`].

    This class includes only the parameters that are specific to PPO training. For a full list of training arguments,
    please refer to the [`~transformers.TrainingArguments`] and [`OnPolicyConfig`] 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:
        exp_name (`str`, *optional*, defaults to `os.path.basename(__file__)[:-3]`):
            Name of this experiment.
        reward_model_path (`str`, *optional*, defaults to `"EleutherAI/pythia-160m"`):
            Path to the reward model.
        model_adapter_name (`str`, *optional*):
            Name of the train target PEFT adapter, when using LoRA with multiple adapters.
        ref_adapter_name (`str`, *optional*):
            Name of the reference PEFT adapter, when using LoRA with multiple adapters.
        num_ppo_epochs (`int`, *optional*, defaults to `4`):
            Number of epochs to train.
        whiten_rewards (`bool`, *optional*, defaults to `False`):
            Whether to whiten the rewards.
        kl_coef (`float`, *optional*, defaults to `0.05`):
            KL coefficient.
        kl_estimator (`Literal["k1", "k3"]`, *optional*, defaults to `"k1"`):
            Which estimator for KL-Divergence to use from [Approximating KL
            Divergence](http://joschu.net/blog/kl-approx.html). Defaults to "k1", a straightforward, unbiased
            estimator. Can be set to "k3", an unbiased estimator with lower variance which "appears to be a strictly
            better estimator". Cannot be set to "k2", as it is used for logging purposes.
        cliprange (`float`, *optional*, defaults to `0.2`):
            Clip range.
        vf_coef (`float`, *optional*, defaults to `0.1`):
            Value function coefficient.
        cliprange_value (`float`, *optional*, defaults to `0.2`):
            Clip range for the value function.
        gamma (`float`, *optional*, defaults to `1.0`):
            Discount factor.
        lam (`float`, *optional*, defaults to `0.95`):
            Lambda value for GAE.
        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.
    
    """
    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.'},
    )
    unsloth_logit_chunk_multiplier : Optional[int] = field(
            default = None,
            metadata = {'help': 'Multiplier for chunked logit computations.'},
        )
    unsloth_grpo_mini_batch : Optional[int] = field(
        default = None,
        metadata = {'help': 'Mini batch size for GRPO hidden state accumulation. Default is None unless user defines it.'},
    )
    
    def __init__(
        self,
        output_dir = None,
        per_device_train_batch_size = 4,
        num_train_epochs = 3.0,
        max_steps = -1,
        learning_rate = 5e-05,
        lr_scheduler_type = 'linear',
        lr_scheduler_kwargs = None,
        warmup_steps = 0.1,
        optim = 'adamw_8bit',
        optim_args = None,
        weight_decay = 0.01,
        adam_beta1 = 0.9,
        adam_beta2 = 0.999,
        adam_epsilon = 1e-08,
        optim_target_modules = None,
        gradient_accumulation_steps = 2,
        average_tokens_across_devices = True,
        max_grad_norm = 1.0,
        label_smoothing_factor = 0.0,
        bf16 = False,
        fp16 = False,
        bf16_full_eval = False,
        fp16_full_eval = False,
        tf32 = None,
        gradient_checkpointing = True,
        gradient_checkpointing_kwargs = None,
        torch_compile = False,
        torch_compile_backend = None,
        torch_compile_mode = None,
        use_liger_kernel = False,
        liger_kernel_config = None,
        use_cache = False,
        neftune_noise_alpha = None,
        torch_empty_cache_steps = 250,
        auto_find_batch_size = False,
        logging_strategy = 'steps',
        logging_steps = 1,
        logging_first_step = False,
        log_on_each_node = True,
        logging_nan_inf_filter = False,
        include_num_input_tokens_seen = False,
        log_level = 'passive',
        log_level_replica = 'warning',
        disable_tqdm = None,
        report_to = 'none',
        run_name = None,
        project = 'huggingface',
        trackio_space_id = 'trackio',
        eval_strategy = 'no',
        eval_steps = None,
        eval_delay = 0,
        per_device_eval_batch_size = 4,
        prediction_loss_only = False,
        eval_on_start = False,
        eval_do_concat_batches = True,
        eval_use_gather_object = False,
        eval_accumulation_steps = 2,
        batch_eval_metrics = False,
        save_only_model = False,
        save_strategy = 'steps',
        save_steps = 500,
        save_on_each_node = False,
        save_total_limit = None,
        enable_jit_checkpoint = False,
        push_to_hub = False,
        hub_token = None,
        hub_private_repo = None,
        hub_model_id = None,
        hub_strategy = 'every_save',
        hub_always_push = False,
        hub_revision = None,
        load_best_model_at_end = False,
        metric_for_best_model = None,
        greater_is_better = None,
        ignore_data_skip = False,
        restore_callback_states_from_checkpoint = False,
        full_determinism = False,
        seed = 3407,
        data_seed = 3407,
        use_cpu = False,
        accelerator_config = None,
        parallelism_config = None,
        dataloader_drop_last = False,
        dataloader_num_workers = 0,
        dataloader_pin_memory = True,
        dataloader_persistent_workers = False,
        dataloader_prefetch_factor = None,
        remove_unused_columns = True,
        label_names = None,
        train_sampling_strategy = 'random',
        length_column_name = 'length',
        ddp_find_unused_parameters = None,
        ddp_bucket_cap_mb = None,
        ddp_broadcast_buffers = None,
        ddp_backend = None,
        ddp_timeout = 1800,
        fsdp = None,
        fsdp_config = None,
        deepspeed = None,
        debug = '',
        skip_memory_metrics = True,
        do_train = False,
        do_eval = False,
        do_predict = False,
        resume_from_checkpoint = None,
        warmup_ratio = None,
        logging_dir = None,
        local_rank = -1,
        dataset_num_proc = None,
        num_mini_batches = 1,
        total_episodes = None,
        local_rollout_forward_batch_size = 64,
        num_sample_generations = 10,
        response_length = 53,
        stop_token = None,
        stop_token_id = None,
        temperature = 0.7,
        missing_eos_penalty = None,
        sft_model_path = 'EleutherAI/pythia-160m',
        world_size = None,
        num_total_batches = None,
        micro_batch_size = None,
        local_batch_size = None,
        batch_size = None,
        local_mini_batch_size = None,
        mini_batch_size = None,
        exp_name = 'ppo_config',
        reward_model_path = 'EleutherAI/pythia-160m',
        model_adapter_name = None,
        ref_adapter_name = None,
        num_ppo_epochs = 4,
        whiten_rewards = False,
        kl_coef = 0.05,
        kl_estimator = 'k1',
        cliprange = 0.2,
        vf_coef = 0.1,
        cliprange_value = 0.2,
        gamma = 1.0,
        lam = 0.95,
        ds3_gather_for_generation = True,
        vllm_sampling_params = None,
        unsloth_num_chunks = -1,
        unsloth_logit_chunk_multiplier = None, 
        unsloth_grpo_mini_batch = None, 
        
        **kwargs,
    ):
        if learning_rate < 1e-7: print(f'Unsloth: Your learning rate of `{learning_rate}` is too small and less than 1e-7! Consider increasing it, otherwise gradient updates will be close to 0!')
        if learning_rate > 1: print(f'Unsloth: Your learning rate of `{learning_rate}` is way too larger > 1! Consider decreasing it to 1e-1, otherwise gradient updates will explode!')
        if num_train_epochs is None:
            num_train_epochs = 3.0  # Default to 3 epochs if None, max_steps will override
        if output_dir is None and save_strategy == 'steps' and save_steps == 500:
            output_dir = 'unsloth_training_checkpoints'
            save_strategy = 'no'
        import multiprocessing as _mp
        if _mp.get_start_method() != 'fork':
            dataset_num_proc = None
        elif dataset_num_proc is None:
            import psutil
            dataset_num_proc = min(max((psutil.cpu_count() or 1)+4, 2), 64)
            memory_gb_left = psutil.virtual_memory().available / (1024**3)
            if memory_gb_left <= 2: dataset_num_proc = 1
            else: dataset_num_proc = min(dataset_num_proc, int(memory_gb_left))
        if temperature <= 0:
            raise ValueError('Unsloth: Please set a positive non-zero temperature since your results will be wrong.')
        elif temperature >= 10:
            raise ValueError('Unsloth: Please set a positive non-zero temperature less than 10, since sampling will be quite erratic.')
        
        
        super().__init__(
            output_dir = output_dir,
            per_device_train_batch_size = per_device_train_batch_size,
            num_train_epochs = num_train_epochs,
            max_steps = max_steps,
            learning_rate = learning_rate,
            lr_scheduler_type = lr_scheduler_type,
            lr_scheduler_kwargs = lr_scheduler_kwargs,
            warmup_steps = warmup_steps,
            optim = optim,
            optim_args = optim_args,
            weight_decay = weight_decay,
            adam_beta1 = adam_beta1,
            adam_beta2 = adam_beta2,
            adam_epsilon = adam_epsilon,
            optim_target_modules = optim_target_modules,
            gradient_accumulation_steps = gradient_accumulation_steps,
            average_tokens_across_devices = average_tokens_across_devices,
            max_grad_norm = max_grad_norm,
            label_smoothing_factor = label_smoothing_factor,
            bf16 = bf16,
            fp16 = fp16,
            bf16_full_eval = bf16_full_eval,
            fp16_full_eval = fp16_full_eval,
            tf32 = tf32,
            gradient_checkpointing = gradient_checkpointing,
            gradient_checkpointing_kwargs = gradient_checkpointing_kwargs,
            torch_compile = torch_compile,
            torch_compile_backend = torch_compile_backend,
            torch_compile_mode = torch_compile_mode,
            use_liger_kernel = use_liger_kernel,
            liger_kernel_config = liger_kernel_config,
            use_cache = use_cache,
            neftune_noise_alpha = neftune_noise_alpha,
            torch_empty_cache_steps = torch_empty_cache_steps,
            auto_find_batch_size = auto_find_batch_size,
            logging_strategy = logging_strategy,
            logging_steps = logging_steps,
            logging_first_step = logging_first_step,
            log_on_each_node = log_on_each_node,
            logging_nan_inf_filter = logging_nan_inf_filter,
            include_num_input_tokens_seen = include_num_input_tokens_seen,
            log_level = log_level,
            log_level_replica = log_level_replica,
            disable_tqdm = disable_tqdm,
            report_to = report_to,
            run_name = run_name,
            project = project,
            trackio_space_id = trackio_space_id,
            eval_strategy = eval_strategy,
            eval_steps = eval_steps,
            eval_delay = eval_delay,
            per_device_eval_batch_size = per_device_eval_batch_size,
            prediction_loss_only = prediction_loss_only,
            eval_on_start = eval_on_start,
            eval_do_concat_batches = eval_do_concat_batches,
            eval_use_gather_object = eval_use_gather_object,
            eval_accumulation_steps = eval_accumulation_steps,
            batch_eval_metrics = batch_eval_metrics,
            save_only_model = save_only_model,
            save_strategy = save_strategy,
            save_steps = save_steps,
            save_on_each_node = save_on_each_node,
            save_total_limit = save_total_limit,
            enable_jit_checkpoint = enable_jit_checkpoint,
            push_to_hub = push_to_hub,
            hub_token = hub_token,
            hub_private_repo = hub_private_repo,
            hub_model_id = hub_model_id,
            hub_strategy = hub_strategy,
            hub_always_push = hub_always_push,
            hub_revision = hub_revision,
            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,
            restore_callback_states_from_checkpoint = restore_callback_states_from_checkpoint,
            full_determinism = full_determinism,
            seed = seed,
            data_seed = data_seed,
            use_cpu = use_cpu,
            accelerator_config = accelerator_config,
            parallelism_config = parallelism_config,
            dataloader_drop_last = dataloader_drop_last,
            dataloader_num_workers = dataloader_num_workers,
            dataloader_pin_memory = dataloader_pin_memory,
            dataloader_persistent_workers = dataloader_persistent_workers,
            dataloader_prefetch_factor = dataloader_prefetch_factor,
            remove_unused_columns = remove_unused_columns,
            label_names = label_names,
            train_sampling_strategy = train_sampling_strategy,
            length_column_name = length_column_name,
            ddp_find_unused_parameters = ddp_find_unused_parameters,
            ddp_bucket_cap_mb = ddp_bucket_cap_mb,
            ddp_broadcast_buffers = ddp_broadcast_buffers,
            ddp_backend = ddp_backend,
            ddp_timeout = ddp_timeout,
            fsdp = fsdp,
            fsdp_config = fsdp_config,
            deepspeed = deepspeed,
            debug = debug,
            skip_memory_metrics = skip_memory_metrics,
            do_train = do_train,
            do_eval = do_eval,
            do_predict = do_predict,
            resume_from_checkpoint = resume_from_checkpoint,
            warmup_ratio = warmup_ratio,
            logging_dir = logging_dir,
            local_rank = local_rank,
            dataset_num_proc = dataset_num_proc,
            num_mini_batches = num_mini_batches,
            total_episodes = total_episodes,
            local_rollout_forward_batch_size = local_rollout_forward_batch_size,
            num_sample_generations = num_sample_generations,
            response_length = response_length,
            stop_token = stop_token,
            stop_token_id = stop_token_id,
            temperature = temperature,
            missing_eos_penalty = missing_eos_penalty,
            sft_model_path = sft_model_path,
            world_size = world_size,
            num_total_batches = num_total_batches,
            micro_batch_size = micro_batch_size,
            local_batch_size = local_batch_size,
            batch_size = batch_size,
            local_mini_batch_size = local_mini_batch_size,
            mini_batch_size = mini_batch_size,
            exp_name = exp_name,
            reward_model_path = reward_model_path,
            model_adapter_name = model_adapter_name,
            ref_adapter_name = ref_adapter_name,
            num_ppo_epochs = num_ppo_epochs,
            whiten_rewards = whiten_rewards,
            kl_coef = kl_coef,
            kl_estimator = kl_estimator,
            cliprange = cliprange,
            vf_coef = vf_coef,
            cliprange_value = cliprange_value,
            gamma = gamma,
            lam = lam,
            ds3_gather_for_generation = ds3_gather_for_generation,**kwargs)
        self.vllm_sampling_params = vllm_sampling_params
        self.unsloth_num_chunks = unsloth_num_chunks
        if unsloth_grpo_mini_batch is not None:
            if self.generation_batch_size >= unsloth_grpo_mini_batch:
                self.unsloth_grpo_mini_batch = unsloth_grpo_mini_batch
            else:
                raise ValueError(
                    f"Unsloth GRPO mini batch size needs to be less than or equal to the effective generation batch size, "
                    f"which is self.per_device_train_batch_size * gradient_accumulation_steps."
                )
        self.unsloth_logit_chunk_multiplier = unsloth_logit_chunk_multiplier
        

pass

class _UnslothPPOTrainer(BaseTrainer):
    """"""

    _tag_names = ["trl", "ppo"]
    _name = "PPO"
    _paper = {
        "title": "Fine-Tuning Language Models from Human Preferences",
        "id": "1909.08593",
        # docstyle-ignore
        "citation": textwrap.dedent("""\
            @article{mziegler2019fine-tuning,
                title        = {{Fine-Tuning Language Models from Human Preferences}},
                author       = {Daniel M. Ziegler and Nisan Stiennon and Jeffrey Wu and Tom B. Brown and Alec Radford and Dario Amodei and Paul F. Christiano and Geoffrey Irving},
                year         = 2019,
                eprint       = {arXiv:1909.08593}
            }"""),
    }

    def __init__(
        self,
        args: PPOConfig,
        processing_class: Union[PreTrainedTokenizerBase, BaseImageProcessor, FeatureExtractionMixin, ProcessorMixin],
        model: nn.Module,
        ref_model: Optional[nn.Module],
        reward_model: nn.Module,
        train_dataset: Dataset,
        value_model: nn.Module,
        data_collator: Optional[DataCollatorWithPadding] = None,
        eval_dataset: Optional[Union[Dataset, dict[str, Dataset]]] = None,
        # less commonly used
        optimizers: tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (None, None),
        callbacks: Optional[list[TrainerCallback]] = None,
        peft_config: Optional["PeftConfig"] = None,
    ) -> None:
        if not os.environ.get("TRL_EXPERIMENTAL_SILENCE"):
            warnings.warn(
                "This trainer will soon be moved to trl.experimental and is a candidate for removal. If you rely on "
                "it and want it to remain, please share your comments here: "
                "https://github.com/huggingface/trl/issues/4223. Silence this warning by setting environment variable "
                "TRL_EXPERIMENTAL_SILENCE=1."
            )
        if ref_model is model:
            raise ValueError(
                "`model` and `ref_model` cannot be the same object. If you want `ref_model` to be the "
                "same as `model`, you must make a copy of it, or `None` if you use peft."
            )

        self.args = args
        self.processing_class = processing_class
        self.policy_model = model

        # Define the collator if not provided
        if data_collator is None:
            data_collator = DataCollatorWithPadding(self.processing_class)

        # Handle stop token settings: update policy model's generation_config to use provided stop token
        if args.stop_token and args.stop_token_id:
            raise ValueError("You cannot set both `stop_token` and `stop_token_id`.")
        elif args.stop_token:
            if args.stop_token == "eos":
                self.policy_model.generation_config.eos_token_id = self.stop_token_id = processing_class.eos_token_id
            else:
                raise ValueError(
                    f"Unknown `stop_token` {args.stop_token}. Allowed values are: `'eos'` and `None` (no stop token)."
                )
        else:
            self.policy_model.generation_config.eos_token_id = self.stop_token_id = args.stop_token_id  # None or int

        # Check that the kl estimator is valid
        if self.args.kl_estimator not in {"k1", "k3"}:
            raise ValueError(
                "kl_estimator must be either 'k1' (straightforward, unbiased) or 'k3' (lower variance, unbiased, "
                "appears to be a strictly better estimator). See "
                "[Approximating KL Divergence](http://joschu.net/blog/kl-approx.html) for details."
            )

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

            # get peft model with the given config
            self.policy_model = get_peft_model(self.policy_model, peft_config)
            if args.bf16 and getattr(self.policy_model, "is_loaded_in_4bit", False):
                peft_module_casting_to_bf16(self.policy_model)

        self.is_peft_model = is_peft_available() and isinstance(self.policy_model, PeftModel)
        self.model_adapter_name = args.model_adapter_name
        self.ref_adapter_name = args.ref_adapter_name

        if ref_model:
            self.ref_model = ref_model
        elif self.is_peft_model:
            self.ref_model = None
        else:
            self.ref_model = create_reference_model(self.policy_model)

        self.reward_model = reward_model
        self.train_dataset = train_dataset
        self.train_dataset_len = len(train_dataset)
        self.value_model = value_model
        self.data_collator = data_collator
        self.eval_dataset = eval_dataset
        self.optimizer, self.lr_scheduler = optimizers
        self.optimizer_cls_and_kwargs = None  # needed for transformers >= 4.47

        #########
        # calculate various batch sizes
        #########
        if args.total_episodes is None:  # allow the users to define episodes in terms of epochs.
            args.total_episodes = int(args.num_train_epochs * self.train_dataset_len)
        accelerator = Accelerator(gradient_accumulation_steps=args.gradient_accumulation_steps)
        self.accelerator = accelerator
        args.world_size = accelerator.num_processes
        args.local_batch_size = args.per_device_train_batch_size * args.gradient_accumulation_steps
        args.micro_batch_size = int(args.per_device_train_batch_size * args.world_size)
        args.batch_size = int(args.local_batch_size * args.world_size)
        args.mini_batch_size = exact_div(
            args.batch_size, args.num_mini_batches, "`batch_size` must be a multiple of `num_mini_batches`"
        )
        args.local_mini_batch_size = exact_div(
            args.local_batch_size, args.num_mini_batches, "`local_batch_size` must be a multiple of `num_mini_batches`"
        )
        if args.whiten_rewards:
            assert args.local_mini_batch_size >= 8, (
                f"Per-rank minibatch size {args.local_mini_batch_size} is insufficient for whitening"
            )
        # `per_rank_rollout_batch_size` is our `args.local_batch_size`
        # `per_rank_minibatch_size` is our `args.local_mini_batch_size`
        args.num_total_batches = math.ceil(
            args.total_episodes / args.batch_size
        )  # we may train for more than `total_episodes`
        time_tensor = torch.tensor(int(time.time()), device=accelerator.device)
        time_int = broadcast(time_tensor, 0).item()  # avoid different timestamps across processes
        args.run_name = f"{args.exp_name}__{args.seed}__{time_int}"
        self.local_seed = args.seed + accelerator.process_index * 100003  # Prime
        if args.num_sample_generations > 0:
            self.sample_generations_freq = max(1, args.num_total_batches // args.num_sample_generations)
        self.local_dataloader_batch_size = args.local_batch_size

        #########
        # setup model, optimizer, and others
        #########
        for module in [self.policy_model, self.ref_model, self.value_model, self.reward_model]:
            if module is not None:
                disable_dropout_in_model(module)
        self.model = PolicyAndValueWrapper(self.policy_model, self.value_model)
        self.model.config = self.policy_model.config  # needed for pushing to hub
        self.create_optimizer_and_scheduler(
            num_training_steps=args.num_total_batches
        )  # note that we are calling `self.lr_scheduler.step[]` manually only at the batch level

        #########
        # trainer specifics
        #########
        default_callbacks = DEFAULT_CALLBACKS + get_reporting_integration_callbacks(self.args.report_to)
        self.callbacks = default_callbacks if callbacks is None else default_callbacks + callbacks
        self.callback_handler = CallbackHandler(
            self.callbacks, self.model, self.processing_class, self.optimizer, self.lr_scheduler
        )
        self.add_callback(PrinterCallback if self.args.disable_tqdm else DEFAULT_PROGRESS_CALLBACK)
        self.control = TrainerControl()
        self.state = OnlineTrainerState(
            is_local_process_zero=self.is_local_process_zero(),
            is_world_process_zero=self.is_world_process_zero(),
            stateful_callbacks=[
                cb for cb in self.callback_handler.callbacks + [self.control] if isinstance(cb, ExportableState)
            ],
        )
        self.current_flos = 0
        self.hp_search_backend = None
        self.is_deepspeed_enabled = getattr(self.accelerator.state, "deepspeed_plugin", None) is not None
        self.is_fsdp_enabled = getattr(self.accelerator.state, "fsdp_plugin", None) is not None
        # Create distant repo and output directory if needed
        self.hub_model_id = None
        if self.args.push_to_hub:
            self.init_hf_repo()
        if self.args.should_save:
            os.makedirs(self.args.output_dir, exist_ok=True)

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

        #########
        # setup dataloader
        #########
        self.dataloader = DataLoader(
            self.train_dataset,
            batch_size=self.local_dataloader_batch_size,
            shuffle=True,
            collate_fn=self.data_collator,
            drop_last=True,  # needed; otherwise the last batch will be of ragged shape
        )
        # sync random states for DataLoader[shuffle=True] before `accelerator.prepare`
        # see https://gist.github.com/vwxyzjn/2581bff1e48e185e0b85b6dfe1def79c
        torch.manual_seed(args.seed)
        self.model, self.optimizer, self.dataloader = accelerator.prepare(self.model, self.optimizer, self.dataloader)
        torch.manual_seed(self.local_seed)  # reset the local seed again

        self.eval_dataloader = DataLoader(
            self.eval_dataset,
            batch_size=args.per_device_eval_batch_size,
            collate_fn=self.data_collator,
            drop_last=True,
        )  # no need to shuffle eval dataset
        self.eval_dataloader = accelerator.prepare(self.eval_dataloader)

        if self.is_deepspeed_enabled:
            self.reward_model = prepare_deepspeed(
                self.reward_model, args.per_device_train_batch_size, args.fp16, args.bf16
            )

            if self.ref_model is None:
                if not self.is_peft_model:
                    raise ValueError("No reference model and model is not a Peft model.")
            else:
                self.ref_model = prepare_deepspeed(
                    self.ref_model, args.per_device_train_batch_size, args.fp16, args.bf16
                )
        else:
            if self.ref_model is None:
                if not self.is_peft_model:
                    raise ValueError("No reference model and model is not a Peft model.")
            else:
                self.ref_model = self.ref_model.to(self.accelerator.device)
            self.reward_model = self.reward_model.to(self.accelerator.device)

    def get_train_dataloader(self) -> DataLoader:
        return self.dataloader

    def get_eval_dataloader(self) -> DataLoader:
        return self.eval_dataloader

    @contextmanager
    def null_ref_context(self):
        """Context manager for handling null reference model (that is, peft adapter manipulation)."""
        with (
            self.accelerator.unwrap_model(self.model.policy).disable_adapter()
            if self.is_peft_model and not self.ref_adapter_name
            else nullcontext()
        ):
            if self.ref_adapter_name:
                self.model.policy.set_adapter(self.ref_adapter_name)
            yield
            if self.ref_adapter_name:
                self.model.policy.set_adapter(self.model_adapter_name or "default")

    def save_model(self, output_dir: Optional[str] = None, _internal_call: bool = False):
        backup_model = self.model
        self.model = self.model.policy  # save only the policy

        if self.is_deepspeed_enabled:
            backup_deepspeed = self.deepspeed
            self.deepspeed = self.model

        super().save_model(output_dir, _internal_call)

        self.model = backup_model

        if self.is_deepspeed_enabled:
            self.deepspeed = backup_deepspeed

    def train(self):
        args = self.args
        accelerator = self.accelerator
        optimizer = self.optimizer
        model = self.model
        ref_policy = self.ref_model
        reward_model = self.reward_model
        processing_class = self.processing_class
        dataloader = self.dataloader
        device = accelerator.device

        def repeat_generator():
            while True:
                yield from dataloader

        iter_dataloader = iter(repeat_generator())
        generation_config = GenerationConfig(
            max_new_tokens=args.response_length,
            temperature=(args.temperature + 1e-7),
            top_k=0.0,
            top_p=1.0,
            do_sample=True,
        )

        accelerator.print("===training policy===")
        start_time = time.time()
        stats_shape = (args.num_ppo_epochs, args.num_mini_batches, args.gradient_accumulation_steps)
        approxkl_stats = torch.zeros(stats_shape, device=device)
        pg_clipfrac_stats = torch.zeros(stats_shape, device=device)
        pg_loss_stats = torch.zeros(stats_shape, device=device)
        vf_loss_stats = torch.zeros(stats_shape, device=device)
        vf_clipfrac_stats = torch.zeros(stats_shape, device=device)
        entropy_stats = torch.zeros(stats_shape, device=device)
        ratio_stats = torch.zeros(stats_shape, device=device)
        model.train()

        # trainer state initialization
        self.state.global_step = 0
        self.state.episode = 0
        self.state.max_steps = args.num_total_batches
        self.state.num_train_epochs = args.total_episodes / self.train_dataset_len
        # Compute absolute values for logging, eval, and save if given as ratio
        if args.logging_steps is not None:
            if args.logging_steps < 1:
                self.state.logging_steps = math.ceil(self.state.max_steps * args.logging_steps)
            else:
                self.state.logging_steps = args.logging_steps
        if args.eval_steps is not None:
            if args.eval_steps < 1:
                self.state.eval_steps = math.ceil(self.state.max_steps * args.eval_steps)
            else:
                self.state.eval_steps = args.eval_steps
        if args.save_steps is not None:
            if args.save_steps < 1:
                self.state.save_steps = math.ceil(self.state.max_steps * args.save_steps)
            else:
                self.state.save_steps = args.save_steps
        self.control = self.callback_handler.on_train_begin(args, self.state, self.control)

        # backward compatibility
        if self.is_deepspeed_enabled:
            self.deepspeed = self.model
            self.model_wrapped = self.model

        for update in range(1, args.num_total_batches + 1):
            self.state.episode += 1 * args.batch_size
            data = next(iter_dataloader)
            with torch.no_grad():
                queries = data["input_ids"].to(device)
                context_length = queries.shape[1]
                responses = []
                postprocessed_responses = []
                logprobs = []
                ref_logprobs = []
                scores = []
                sequence_lengths = []
                values = []
                with unwrap_model_for_generation(
                    self.model, self.accelerator, gather_deepspeed3_params=self.args.ds3_gather_for_generation
                ) as unwrapped_model:
                    query_responses, logitss = batch_generation(
                        unwrapped_model.policy,
                        queries,
                        args.local_rollout_forward_batch_size,
                        processing_class.pad_token_id,
                        generation_config,
                    )

                for i in range(0, queries.shape[0], args.local_rollout_forward_batch_size):
                    query = queries[i : i + args.local_rollout_forward_batch_size]
                    query_response = query_responses[i : i + args.local_rollout_forward_batch_size]
                    response = query_response[:, context_length:]
                    logits = logitss[i : i + args.local_rollout_forward_batch_size]
                    logprob = selective_log_softmax(logits, response)
                    del logits
                    empty_cache()

                    if ref_policy is None:
                        with self.null_ref_context():
                            ref_output = forward(model.policy, query_response, processing_class.pad_token_id)
                    else:
                        ref_output = forward(ref_policy, query_response, processing_class.pad_token_id)
                    ref_logits = ref_output.logits[:, context_length - 1 : -1]
                    ref_logits /= args.temperature + 1e-7
                    ref_logprob = selective_log_softmax(ref_logits, response)
                    del ref_output, ref_logits
                    empty_cache()

                    # Response Processing 1. truncate response after the first occurrence of `stop_token_id`
                    postprocessed_response = response
                    if self.stop_token_id is not None:  # handle the edge case when stop_token_id exists but is 0
                        postprocessed_response = truncate_response(
                            self.stop_token_id, processing_class.pad_token_id, response
                        )

                    # Response Processing 2. run reward model on the truncated responses
                    postprocessed_query_response = torch.cat((query, postprocessed_response), 1)
                    sequence_length = first_true_indices(postprocessed_response == processing_class.pad_token_id) - 1
                    unwrapped_value_model = accelerator.unwrap_model(model).value_model
                    full_value, _, _ = get_reward(
                        unwrapped_value_model, query_response, processing_class.pad_token_id, context_length
                    )
                    value = full_value[:, context_length - 1 : -1].squeeze(-1)
                    _, score, _ = get_reward(
                        reward_model, postprocessed_query_response, processing_class.pad_token_id, context_length
                    )

                    responses.append(response)
                    postprocessed_responses.append(postprocessed_response)
                    logprobs.append(logprob)
                    ref_logprobs.append(ref_logprob)
                    sequence_lengths.append(sequence_length)
                    scores.append(score)
                    values.append(value)
                responses = torch.cat(responses, 0)
                postprocessed_responses = torch.cat(postprocessed_responses, 0)
                logprobs = torch.cat(logprobs, 0)
                ref_logprobs = torch.cat(ref_logprobs, 0)
                sequence_lengths = torch.cat(sequence_lengths, 0)
                scores = torch.cat(scores, 0)
                values = torch.cat(values, 0)
                del (logprob, ref_logprob, full_value, value, score, unwrapped_model)
                empty_cache()
                gc.collect()

                # Response Processing 3. Filter completion. Ensure that the sample contains stop_token_id
                # Completions not passing that filter will receive a lower score.
                contain_eos_token = torch.any(postprocessed_responses == self.processing_class.eos_token_id, dim=-1)
                if self.args.missing_eos_penalty is not None:
                    scores[~contain_eos_token] -= self.args.missing_eos_penalty
                # accelerator.print(f"{scores=}, {(contain_eos_token.sum() / len(contain_eos_token))=}")

                # be very careful with `padding_mask_p1`; see https://excalidraw.com/#json=LWnzG4w2k5DjF_EOL_xPt,e2w3a-hFJ_gX5vOfeyXGTw
                response_idxs = torch.arange(responses.shape[1], device=responses.device).repeat(responses.shape[0], 1)
                padding_mask = response_idxs > sequence_lengths.unsqueeze(1)
                logprobs = torch.masked_fill(logprobs, padding_mask, INVALID_LOGPROB)
                ref_logprobs = torch.masked_fill(ref_logprobs, padding_mask, INVALID_LOGPROB)
                sequence_lengths_p1 = sequence_lengths + 1
                padding_mask_p1 = response_idxs > (sequence_lengths_p1.unsqueeze(1))
                values = torch.masked_fill(values, padding_mask_p1, 0)

                # 4. compute rewards
                # Formula used by http://joschu.net/blog/kl-approx.html for the k1 and k3 estimators
                logr = ref_logprobs - logprobs
                kl = -logr if args.kl_estimator == "k1" else (logr.exp() - 1) - logr  # Else statement is k3
                non_score_reward = -args.kl_coef * kl
                rewards = non_score_reward.clone()
                actual_start = torch.arange(rewards.size(0), device=rewards.device)
                actual_end = torch.where(sequence_lengths_p1 < rewards.size(1), sequence_lengths_p1, sequence_lengths)
                rewards[[actual_start, actual_end]] += scores

                # 5. whiten rewards
                if args.whiten_rewards:
                    rewards = masked_whiten(rewards, mask=~padding_mask_p1, shift_mean=False)
                    rewards = torch.masked_fill(rewards, padding_mask_p1, 0)

                # 6. compute advantages and returns
                lastgaelam = 0
                advantages_reversed = []
                gen_length = responses.shape[1]
                for t in reversed(range(gen_length)):
                    nextvalues = values[:, t + 1] if t < gen_length - 1 else 0.0
                    delta = rewards[:, t] + args.gamma * nextvalues - values[:, t]
                    lastgaelam = delta + args.gamma * args.lam * lastgaelam
                    advantages_reversed.append(lastgaelam)
                advantages = torch.stack(advantages_reversed[::-1], axis=1)
                returns = advantages + values
                advantages = masked_whiten(advantages, ~padding_mask)
                advantages = torch.masked_fill(advantages, padding_mask, 0)
                empty_cache()

            # Do multiple epochs of PPO training, with a fresh random shuffle in each epoch
            for ppo_epoch_idx in range(args.num_ppo_epochs):
                b_inds = np.random.permutation(args.local_batch_size)
                minibatch_idx = 0
                for mini_batch_start in range(0, args.local_batch_size, args.local_mini_batch_size):
                    mini_batch_end = mini_batch_start + args.local_mini_batch_size
                    mini_batch_inds = b_inds[mini_batch_start:mini_batch_end]
                    gradient_accumulation_idx = 0
                    for micro_batch_start in range(0, args.local_mini_batch_size, args.per_device_train_batch_size):
                        with accelerator.accumulate(model):
                            micro_batch_end = micro_batch_start + args.per_device_train_batch_size
                            micro_batch_inds = mini_batch_inds[micro_batch_start:micro_batch_end]
                            mb_advantage = advantages[micro_batch_inds]
                            mb_responses = responses[micro_batch_inds]
                            mb_query_responses = query_responses[micro_batch_inds]
                            mb_logprobs = logprobs[micro_batch_inds]
                            mb_return = returns[micro_batch_inds]
                            mb_values = values[micro_batch_inds]

                            output, vpred_temp = forward(model, mb_query_responses, processing_class.pad_token_id)
                            logits = output.logits[:, context_length - 1 : -1]
                            logits /= args.temperature + 1e-7
                            new_logprobs = selective_log_softmax(logits, mb_responses)
                            new_logprobs = torch.masked_fill(
                                new_logprobs, padding_mask[micro_batch_inds], INVALID_LOGPROB
                            )
                            vpred = vpred_temp[:, context_length - 1 : -1].squeeze(-1)
                            vpred = torch.masked_fill(vpred, padding_mask_p1[micro_batch_inds], 0)
                            vpredclipped = torch.clamp(
                                vpred,
                                mb_values - args.cliprange_value,
                                mb_values + args.cliprange_value,
                            )
                            vf_losses1 = torch.square(vpred - mb_return)
                            vf_losses2 = torch.square(vpredclipped - mb_return)
                            vf_loss_max = torch.max(vf_losses1, vf_losses2)
                            vf_loss = 0.5 * masked_mean(vf_loss_max, ~padding_mask_p1[micro_batch_inds])
                            vf_clipfrac = masked_mean(
                                (vf_losses2 > vf_losses1).float(), ~padding_mask_p1[micro_batch_inds]
                            )
                            logprobs_diff = new_logprobs - mb_logprobs
                            ratio = torch.exp(logprobs_diff)
                            pg_losses = -mb_advantage * ratio
                            pg_losses2 = -mb_advantage * torch.clamp(ratio, 1.0 - args.cliprange, 1.0 + args.cliprange)
                            pg_loss_max = torch.max(pg_losses, pg_losses2)
                            pg_loss = masked_mean(pg_loss_max, ~padding_mask[micro_batch_inds])
                            loss = pg_loss + args.vf_coef * vf_loss
                            accelerator.backward(loss)
                            optimizer.step()
                            optimizer.zero_grad()
                            with torch.no_grad():
                                pg_clipfrac = masked_mean(
                                    (pg_losses2 > pg_losses).float(), ~padding_mask[micro_batch_inds]
                                )
                                prob_dist = torch.nn.functional.softmax(logits, dim=-1, dtype = torch.float32).to(logits.dtype)
                                entropy = torch.logsumexp(logits, dim=-1) - torch.sum(prob_dist * logits, dim=-1)
                                approxkl = 0.5 * (logprobs_diff**2).mean()
                                approxkl_stats[ppo_epoch_idx, minibatch_idx, gradient_accumulation_idx] = approxkl
                                pg_clipfrac_stats[ppo_epoch_idx, minibatch_idx, gradient_accumulation_idx] = (
                                    pg_clipfrac
                                )
                                pg_loss_stats[ppo_epoch_idx, minibatch_idx, gradient_accumulation_idx] = pg_loss
                                vf_loss_stats[ppo_epoch_idx, minibatch_idx, gradient_accumulation_idx] = vf_loss
                                vf_clipfrac_stats[ppo_epoch_idx, minibatch_idx, gradient_accumulation_idx] = (
                                    vf_clipfrac
                                )
                                entropy_stats[ppo_epoch_idx, minibatch_idx, gradient_accumulation_idx] = entropy.mean()
                                ratio_stats[ppo_epoch_idx, minibatch_idx, gradient_accumulation_idx] = ratio.mean()
                        gradient_accumulation_idx += 1
                    minibatch_idx += 1
                    # del everything and empty cache
                    # fmt: off
                    del (
                        output, vpred_temp, logits, new_logprobs, vpred, vpredclipped,
                        vf_losses1, vf_losses2, vf_loss, vf_clipfrac, logprobs_diff, ratio, pg_losses, pg_losses2, pg_loss_max,
                        pg_loss, loss, pg_clipfrac, prob_dist, entropy, approxkl, mb_return,
                        mb_advantage, mb_values, mb_responses, mb_query_responses, mb_logprobs,
                    )
                    # fmt: on
                    empty_cache()
            with torch.no_grad():
                mean_kl = kl.sum(1).mean()
                mean_entropy = (-logprobs).sum(1).mean()
                mean_non_score_reward = non_score_reward.sum(1).mean()
                rlhf_reward = mean_non_score_reward + scores.mean()
                eps = int(self.state.episode / (time.time() - start_time))
                metrics = {}
                metrics["eps"] = eps
                metrics["objective/kl"] = self.accelerator.gather_for_metrics(mean_kl).mean().item()
                metrics["objective/entropy"] = self.accelerator.gather_for_metrics(mean_entropy).mean().item()
                metrics["objective/non_score_reward"] = (
                    self.accelerator.gather_for_metrics(mean_non_score_reward).mean().item()
                )
                metrics["objective/rlhf_reward"] = self.accelerator.gather_for_metrics(rlhf_reward).mean().item()
                metrics["objective/scores"] = self.accelerator.gather_for_metrics(scores.mean()).mean().item()
                metrics["policy/approxkl_avg"] = self.accelerator.gather_for_metrics(approxkl_stats).mean().item()
                metrics["policy/clipfrac_avg"] = self.accelerator.gather_for_metrics(pg_clipfrac_stats).mean().item()
                metrics["loss/policy_avg"] = self.accelerator.gather_for_metrics(pg_loss_stats).mean().item()
                metrics["loss/value_avg"] = self.accelerator.gather_for_metrics(vf_loss_stats).mean().item()
                metrics["val/clipfrac_avg"] = self.accelerator.gather_for_metrics(vf_clipfrac_stats).mean().item()
                metrics["policy/entropy_avg"] = self.accelerator.gather_for_metrics(entropy_stats).mean().item()
                metrics["val/ratio"] = self.accelerator.gather_for_metrics(ratio_stats).mean().item()
                metrics["val/ratio_var"] = self.accelerator.gather_for_metrics(ratio_stats).var().item()
                metrics["val/num_eos_tokens"] = (responses == processing_class.eos_token_id).sum().item()
                metrics["lr"] = self.lr_scheduler.get_last_lr()[0]
                metrics["episode"] = self.state.episode
                self.state.epoch = self.state.episode / self.train_dataset_len  # used by self.log
                self.state.global_step += 1
                self.log(metrics)

            self.lr_scheduler.step()
            self.control = self.callback_handler.on_step_end(args, self.state, self.control)
            if self.control.should_save:
                self._save_checkpoint(model, trial=None)
                self.control = self.callback_handler.on_save(self.args, self.state, self.control)
            del kl, mean_kl, mean_entropy, mean_non_score_reward, scores, metrics, non_score_reward
            empty_cache()
            gc.collect()

            if args.num_sample_generations > 0 and (update - 1) % self.sample_generations_freq == 0:
                self.generate_completions(sampling=True)
                empty_cache()
            del (
                query_responses,
                responses,
                postprocessed_responses,
                logprobs,
                ref_logprobs,
                values,
                sequence_lengths,
                contain_eos_token,
                sequence_lengths_p1,
                response_idxs,
                padding_mask,
                padding_mask_p1,
                rewards,
                actual_start,
                actual_end,
                advantages,
                returns,
            )
            empty_cache()

        # HF trainer specifics
        self.control = self.callback_handler.on_train_end(args, self.state, self.control)
        if self.control.should_save:
            self._save_checkpoint(model, trial=None)
            self.control = self.callback_handler.on_save(self.args, self.state, self.control)

    def generate_completions(self, sampling: bool = False):
        args = self.args
        processing_class = self.processing_class
        generation_config = GenerationConfig(
            max_new_tokens=self.args.response_length,
            temperature=(0.01 + 1e-7),
            top_k=0.0,
            top_p=1.0,
            do_sample=True,
        )

        table = defaultdict(list)
        with unwrap_model_for_generation(
            self.model, self.accelerator, gather_deepspeed3_params=self.args.ds3_gather_for_generation
        ) as unwrapped_model:
            for batch in self.eval_dataloader:
                query = batch["input_ids"]
                with torch.no_grad():
                    context_length = query.shape[1]
                    query_response, _ = batch_generation(
                        unwrapped_model.policy,
                        query,
                        query.shape[0],
                        processing_class.pad_token_id,
                        generation_config,
                    )
                    response = query_response[:, context_length:]
                    postprocessed_response = response
                    if self.stop_token_id is not None:  # handle the edge case when stop_token_id exists but is 0
                        postprocessed_response = truncate_response(
                            self.stop_token_id, processing_class.pad_token_id, response
                        )
                    table["query"].extend(
                        gather_object(processing_class.batch_decode(query, skip_special_tokens=True))
                    )
                    table["model response"].extend(
                        gather_object(processing_class.batch_decode(postprocessed_response))
                    )

                    postprocessed_query_response = torch.cat((query, postprocessed_response), 1)
                    _, score, _ = get_reward(
                        self.reward_model, postprocessed_query_response, processing_class.pad_token_id, context_length
                    )
                    table["score"].extend(self.accelerator.gather_for_metrics(score).float().cpu().numpy())

                if sampling:
                    break
        df = pd.DataFrame(table)

        if self.accelerator.is_main_process:
            if is_rich_available():
                print_rich_table(df.iloc[0 : 0 + 5])
            if "wandb" in args.report_to:
                import wandb

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

            if "comet_ml" in args.report_to:
                log_table_to_comet_experiment(
                    name="completions.csv",
                    table=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 UnslothPPOTrainer(_UnslothPPOTrainer):
    """
    Trainer for Proximal Policy Optimization (PPO).

    For details on PPO, see the paper: [Proximal Policy Optimization
    Algorithms](https://huggingface.co/papers/1707.06347).

    Args:
        args ([`PPOConfig`]):
            Training arguments.
        processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.BaseImageProcessor`], [`~transformers.FeatureExtractionMixin`] or [`~transformers.ProcessorMixin`]):
            Class to process the data.
        model (`torch.nn.Module`):
            Model to be trained. This is the policy model.
        ref_model (`torch.nn.Module`, *optional*):
            Reference model used to compute the KL divergence. If `None`, a copy of the policy model is created.
        reward_model (`torch.nn.Module`):
            Reward model used to compute the rewards.
        train_dataset ([`~datasets.Dataset`]):
            Dataset for training.
        value_model (`torch.nn.Module`):
            Value model used to predict the value of a state.
        data_collator ([`~transformers.DataCollatorWithPadding`], *optional*):
            Data collator to batch and pad samples from the dataset. If `None`, a default data collator is created
            using the `processing_class`.
        eval_dataset ([`~datasets.Dataset`] or `dict` of [`~datasets.Dataset`], *optional*):
            Dataset for evaluation.
        optimizers (`tuple` of `torch.optim.Optimizer` and `torch.optim.lr_scheduler.LambdaLR`, *optional*, defaults to `(None, None)`):
            Tuple containing the optimizer and the learning rate scheduler to use for training. If `None`, the
            optimizer and the learning rate scheduler are created using the
            [`~transformers.Trainer.create_optimizer_and_scheduler`] method.
        callbacks (`list` of [`~transformers.TrainerCallback`], *optional*):
            Callbacks to use during training.
        peft_config ([`~peft.PeftConfig`], *optional*):
            PEFT configuration to use PEFT for training. If `None`, PEFT is not used. If provided, the policy `model`
            will be wrapped with the specified PEFT adapter.
    
    """
    def __init__(
        self,
        args,
        processing_class,
        model,
        ref_model,
        reward_model,
        train_dataset,
        value_model,
        data_collator = None,
        eval_dataset = None,
        callbacks = None,
        peft_config = None,
        **kwargs
    ):
        if args is None: args = UnslothPPOConfig()
        use_bf16 = getattr(args, 'bf16', False)
        if type(use_bf16) is not bool: use_bf16 = False
        use_fp16 = getattr(args, 'fp16', False)
        if type(use_fp16) is not bool: use_fp16 = False
        force_float32 = False
        full_finetuning = os.environ.get('UNSLOTH_ENABLE_FULL_FINETUNING', '0') == '1'
        if not full_finetuning and (os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1'):
            print('Unsloth: Switching to float32 training since model cannot work with float16')
            force_float32 = True
        mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')
        dtype = getattr(model.config, 'dtype', None) or getattr(model.config, 'torch_dtype', None)
        if dtype is None: dtype = model.get_input_embeddings().weight.dtype
        from unsloth_zoo.utils import _get_dtype
        dtype = _get_dtype(dtype)
        float16 = dtype == torch.float16
        if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')
        if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')
        if force_float32:
            # Forced float32 training
            args.fp16 = False
            args.bf16 = False
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
            # args.mixed_precision is a new argument which needs to be set now
        elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':
            # Mixed precision training
            args.fp16 = float16
            args.bf16 = not float16
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'
            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16' if float16 else 'bf16'
            # args.mixed_precision is a new argument which needs to be set now
        elif mixed_precision_dtype == 'bfloat16':
            # Both False since bfloat16 full finetuning doesn't do any autocasting.
            args.fp16 = False
            args.bf16 = False
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
            # args.mixed_precision is a new argument which needs to be set now
        
        if getattr(args, 'eval_dataset', None) is not None and getattr(args, 'eval_strategy', 'no') == 'no':
            args.eval_strategy = 'steps'
            if getattr(args, 'eval_steps', None) is None: args.eval_steps = 0.1
        ga_steps = getattr(args, 'gradient_accumulation_steps', None)
        if ga_steps is not None and ga_steps > 1:
            from transformers import __version__ as transformers_version
            if Version(transformers_version) <= Version('4.45.2'):
                print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!\n'
                      '`pip install --upgrade --no-cache-dir --force-reinstall --no-deps unsloth transformers trl unsloth_zoo`')
        if getattr(args, 'eval_strategy', 'no') != 'no':
            eval_bsz = getattr(args, 'per_device_eval_batch_size', 8)
            if eval_bsz == 8 and args.per_device_train_batch_size < eval_bsz: args.per_device_eval_batch_size = args.per_device_train_batch_size
            if getattr(args, 'eval_accumulation_steps', None) is None and ga_steps is not None: args.eval_accumulation_steps = ga_steps
        fp16_full_eval = getattr(args, 'fp16_full_eval', False)
        if type(fp16_full_eval) is not bool: fp16_full_eval = False
        bf16_full_eval = getattr(args, 'bf16_full_eval', False)
        if type(bf16_full_eval) is not bool: bf16_full_eval = False
        if args.fp16 and bf16_full_eval: args.bf16_full_eval = False; args.fp16_full_eval = True
        if args.bf16 and fp16_full_eval: args.bf16_full_eval = True; args.fp16_full_eval = False
        if force_float32:
            args.bf16_full_eval = False
            args.fp16_full_eval = False
        elif os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32') == 'bfloat16':
            args.bf16_full_eval = True
            args.fp16_full_eval = False
        elif not bf16_full_eval and not fp16_full_eval:
            args.bf16_full_eval = args.bf16
            args.fp16_full_eval = args.fp16
        _output_logits = False
        if locals().get('compute_metrics', None) is not None: _output_logits = True
        if locals().get('preprocess_logits_for_metrics', None) is not None: _output_logits = True
        if _output_logits:
            os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
        if model is not None:
            _warnings_issued = getattr(model, 'warnings_issued', None)
            if _warnings_issued is None:
                model.warnings_issued = {}
            elif not isinstance(_warnings_issued, dict):
                try:
                    model.warnings_issued = dict(_warnings_issued)
                except Exception:
                    model.warnings_issued = {}
        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
            elif args_max_seq_length is not None and model_max_seq_length is not None:
                if args_max_seq_length > model_max_seq_length:
                    print('Unsloth: You set `max_seq_length` as ' + str(args_max_seq_length) + ' but '
                           'the maximum the model supports is ' + str(model_max_seq_length) + '. We shall reduce it.')
                    args.max_seq_length = model_max_seq_length
        if model is not None and hasattr(model, 'for_training'):
            model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
        if 'tokenizer' in locals() and hasattr(tokenizer, 'padding_side'): tokenizer.padding_side = 'right'
        if 'processing_class' in locals():
            if hasattr(processing_class, 'padding_side'): processing_class.padding_side = 'right'
            if hasattr(processing_class, 'tokenizer') and hasattr(processing_class.tokenizer, 'padding_side'): processing_class.tokenizer.padding_side = 'right'
        __tokenizer = processing_class if 'processing_class' in locals() else tokenizer
        from unsloth_zoo.vision_utils import UnslothVisionDataCollator
        if not isinstance(data_collator, UnslothVisionDataCollator):
            if isinstance(data_collator, DataCollatorForSeq2Seq) and 'labels' not in train_dataset.column_names:
                data_collator = TransformersDataCollatorForLanguageModeling(
                    __tokenizer,
                    mlm = False,
                    mlm_probability = 0.0,
                    pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
                )
            elif isinstance(data_collator, TransformersDataCollatorForLanguageModeling) and 'labels' in train_dataset.column_names:
                data_collator = DataCollatorForSeq2Seq(
                    __tokenizer,
                    pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
                )
        else:
            if hasattr(args, 'remove_unused_columns'): args.remove_unused_columns = False
            if hasattr(args, 'dataset_text_field'): args.dataset_text_field = ''
            if hasattr(args, 'dataset_kwargs'): args.dataset_kwargs = {'skip_prepare_dataset': True}
        if not isinstance(data_collator, UnslothVisionDataCollator):
            if not hasattr(__tokenizer, 'pad') and hasattr(__tokenizer, 'tokenizer'):
                if isinstance(data_collator, DataCollatorForSeq2Seq):
                    data_collator = DataCollatorForSeq2Seq(
                        __tokenizer.tokenizer,
                        pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
                    )
                else:
                    data_collator = TransformersDataCollatorForLanguageModeling(
                        __tokenizer.tokenizer,
                        mlm = False,
                        mlm_probability = 0.0,
                        pad_to_multiple_of = getattr(args, 'pad_to_multiple_of', None),
                    )
        other_metrics = []
        
        from unsloth_zoo.logging_utils import PatchRLStatistics
        PatchRLStatistics('ppo_trainer', other_metrics)
        
        # [TODO] Fix up DataParallel multiplying batch sizes
        # [TODO] DDP works, but DP seems to not work? [TODO]
        if getattr(args, "parallel_mode", None) == ParallelMode.NOT_DISTRIBUTED and args.n_gpu > 1:
            if getattr(args, "_n_gpu", 1) != 1:
                args._n_gpu = 1
        if "model" in locals() and hasattr(model, "for_training"):
            model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
        super().__init__(
            args = args,
            processing_class = processing_class,
            model = model,
            ref_model = ref_model,
            reward_model = reward_model,
            train_dataset = train_dataset,
            value_model = value_model,
            data_collator = data_collator,
            eval_dataset = eval_dataset,
            callbacks = callbacks,
            peft_config = peft_config,**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
        if hasattr(self, 'llm') and self.llm is not None and hasattr(self.llm, 'get_tokenizer'):
            _vllm_tok = self.llm.get_tokenizer()
            _pc = getattr(self, 'processing_class', None) or getattr(self, 'tokenizer', None)
            if _vllm_tok is not None and _pc is not None and getattr(_pc, 'chat_template', None) is not None and getattr(_vllm_tok, 'chat_template', None) is None:
                _vllm_tok.chat_template = _pc.chat_template
        pass
        
pass