code/fine_tune_sft_dpo/unsloth_compiled_cache/UnslothPRMTrainer.py

"""
2026.2.1
2026.2.1
5.2.0
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.prm_trainer import (BaseImageProcessor, BaseTrainer, Callable, DataCollator, DataCollatorForTokenClassification, Dataset, EvalPrediction, FeatureExtractionMixin, Optional, PRMConfig, PRMTrainer, PartialState, Path, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, TrainerCallback, Union, chain, compute_accuracy, disable_dropout_in_model, features, nn, os, textwrap, torch, warnings, BaseImageProcessor, Callable, DataCollator, DataCollatorForTokenClassification, Dataset, EvalPrediction, FeatureExtractionMixin, Optional, PRMConfig, PartialState, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, TrainerCallback, Union, compute_accuracy, disable_dropout_in_model, features, nn, os, torch, warnings, PreTrainedModel, os, torch)


import os
from typing import *
from dataclasses import dataclass, field
from packaging.version import Version
import torch
import numpy as np
from contextlib import nullcontext
from torch.nn import functional as F
import inspect
from transformers import DataCollatorForSeq2Seq, DataCollatorForLanguageModeling as TransformersDataCollatorForLanguageModeling
from transformers.training_args import ParallelMode
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
@dataclass
class UnslothPRMConfig(PRMConfig):
    """
    
    Configuration class for the [`PRMTrainer`].

    This class includes only the parameters that are specific to PRM training. For a full list of training arguments,
    please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values in this class may
    differ from those in [`~transformers.TrainingArguments`].

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

    Parameters:
        max_length (`int` or `None`, *optional*, defaults to `1024`):
            Maximum length of the sequences (prompt + completion) used for truncation.
        max_prompt_length (`int` or `None`, *optional*, defaults to `512`):
            Maximum length of the prompt used for truncation.
        max_completion_length (`int`, *optional*):
            Maximum length of the completion used for truncation. The completion is the concatenation of the steps.
        disable_dropout (`bool`, *optional*, defaults to `True`):
            Whether to disable dropout in the model.
        step_separator (`str`, *optional*, defaults to `"\n"`):
            Separator used to separate each step of the reasoning process.
        train_on_last_step_only (`bool`, *optional*, defaults to `False`):
            Whether to train only on the last step.
        dataset_num_proc (`int`, *optional*):
            Number of processes to use for processing the dataset.
    
    """
    vllm_sampling_params: Optional[Any] = field(
        default = None,
        metadata = {'help': 'vLLM SamplingParams'},
    )
    unsloth_num_chunks : Optional[int] = field(
        default = -1,
        metadata = {'help': 'Chunk size to reduce memory usage. -1 is most efficient.'},
    )
    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.'},
    )
    max_seq_length : Optional[int] = field(
        default = None,
        metadata = {'help': 'Maximum sequence length to truncate to.'},
    )
    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,
        max_length = 1024,
        max_prompt_length = 512,
        max_completion_length = None,
        disable_dropout = True,
        step_separator = '\
',
        train_on_last_step_only = False,
        dataset_num_proc = None,
        vllm_sampling_params = None,
        unsloth_num_chunks = -1,
        unsloth_logit_chunk_multiplier = None, 
        unsloth_grpo_mini_batch = None, 
        max_seq_length = None,
        **kwargs,
    ):
        if learning_rate < 1e-7: print(f'Unsloth: Your learning rate of `{learning_rate}` is too small and less than 1e-7! Consider increasing it, otherwise gradient updates will be close to 0!')
        if learning_rate > 1: print(f'Unsloth: Your learning rate of `{learning_rate}` is way too larger > 1! Consider decreasing it to 1e-1, otherwise gradient updates will explode!')
        if 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))
        
        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,
            max_length = max_length,
            max_prompt_length = max_prompt_length,
            max_completion_length = max_completion_length,
            disable_dropout = disable_dropout,
            step_separator = step_separator,
            train_on_last_step_only = train_on_last_step_only,
            dataset_num_proc = dataset_num_proc,**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
        self.max_seq_length = max_seq_length

pass

class _UnslothPRMTrainer(BaseTrainer):
    """"""

    _tag_names = ["trl", "prm"]
    _name = "PRM"
    _paper = {
        "title": "Solving math word problems with process-and outcome-based feedback",
        "id": "2211.14275",
        # docstyle-ignore
        "citation": textwrap.dedent("""\
            @article{uesato2022solving,
                title        = {{Solving Math Word Problems With Process- and Outcome-Based Feedback}},
                author       = {Uesato, Jonathan and Kushman, Nate and Kumar, Ramana and Song, Francis and Siegel, Noah and Wang, Lisa and Creswell, Antonia and Irving, Geoffrey and Higgins, Irina},
                year         = 2022,
                journal      = {arXiv preprint arXiv:2211.14275}
            }"""),
    }

    def __init__(
        self,
        model: Optional[Union[PreTrainedModel, nn.Module]] = None,
        args: Optional[PRMConfig] = None,
        data_collator: Optional[DataCollator] = None,
        train_dataset: Optional[Dataset] = None,
        eval_dataset: Optional[Union[Dataset, dict[str, Dataset]]] = None,
        processing_class: Optional[
            Union[PreTrainedTokenizerBase, BaseImageProcessor, FeatureExtractionMixin, ProcessorMixin]
        ] = None,
        model_init: Optional[Callable[[], PreTrainedModel]] = None,
        compute_metrics: Optional[Callable[[EvalPrediction], dict]] = None,
        callbacks: Optional[list[TrainerCallback]] = None,
        optimizers: tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR] = (
            None,
            None,
        ),
        preprocess_logits_for_metrics: Optional[Callable[[torch.Tensor, torch.Tensor], torch.Tensor]] = None,
        peft_config: Optional[dict] = None,
    ):
        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 False:
            pass

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

        if compute_metrics is None:
            compute_metrics = compute_accuracy

        if data_collator is None:
            if processing_class is None:
                raise ValueError(
                    "A processing_class must be specified when using the default DataCollatorForTokenClassification"
                )
            data_collator = DataCollatorForTokenClassification(processing_class, max_length=args.max_length)

        if "input_ids" not in train_dataset.column_names:
            with PartialState().main_process_first():
                fn_kwargs = {
                    "tokenizer": processing_class,
                    "step_separator": args.step_separator,
                    "max_length": args.max_length,
                    "max_prompt_length": args.max_prompt_length,
                    "max_completion_length": args.max_completion_length,
                    "train_on_last_step_only": args.train_on_last_step_only,
                }
                train_fn_kwargs = {**fn_kwargs, "is_eval": False}
                train_dataset = train_dataset.map(
                    self.tokenize_row,
                    fn_kwargs=train_fn_kwargs,
                    num_proc=args.dataset_num_proc,
                    remove_columns=train_dataset.features,
                    desc="Tokenizing train dataset",
                    features=features.Features(  # needed to avoid map to cast labels to bool
                        {
                            "labels": features.Sequence(features.Value("int64")),
                            "input_ids": features.Sequence(features.Value("int64")),
                        }
                    ),
                )

                eval_fn_kwargs = {**fn_kwargs, "is_eval": True}
                if eval_dataset is not None:
                    eval_dataset = eval_dataset.map(
                        self.tokenize_row,
                        fn_kwargs=eval_fn_kwargs,
                        num_proc=args.dataset_num_proc,
                        remove_columns=eval_dataset.features,
                        desc="Tokenizing eval dataset",
                        features=features.Features(  # needed to avoid map to cast labels to bool
                            {
                                "labels": features.Sequence(features.Value("int64")),
                                "input_ids": features.Sequence(features.Value("int64")),
                            }
                        ),
                    )

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

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

    @staticmethod
    def tokenize_row(
        features,
        tokenizer,
        step_separator,
        max_length,
        max_prompt_length,
        max_completion_length,
        train_on_last_step_only,
        is_eval,
    ):
        r"""
        Tokenize a row of the dataset.

        Args:
            features (`dict[str, str]`):
                Row of the dataset, should contain the keys `"prompt"`, `"completions"`, and `"labels"`.
            tokenizer ([`~transformers.PreTrainedTokenizerBase`]):
                Tokenizer used to process the data.
            step_separator (`str`):
                Separator between steps in the completion.
            max_length (`int` or `None`):
               Maximum length of the sequences (prompt + completion). If `None`, the sequences are not truncated.
            max_prompt_length (`int` or `None`):
                Maximum length of the prompt. If `None`, the prompt is not truncated.
            max_completion_length (`int` or `None`):
                Maximum length of the completion sequences. If `None`, the completion sequences are not truncated.
            train_on_last_step_only (`bool`):
                Whether to train only on the last step. If `True`, the labels are `-100` for all tokens except the last
                token of the completion.
            is_eval (`bool`):
                Whether the function is used to tokenize samples from a training or an evaluation dataset. Used only if
                `train_on_last_step_only` is set to `True`.

        Returns:
            `dict[str, list[int]]`:
                Tokenized sequences with the keys `"input_ids"`, and `"labels".

        Example:
        ```python
        >>> from transformers import AutoTokenizer

        >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B")
        >>> features = {
        ...     "prompt": "Which number is larger, 9.8 or 9.11?",
        ...     "completions": ["11 is greater than 8.", "Hence, 9.11 > 9.8."],
        ...     "labels": [True, False],
        ... }
        >>> PRMTrainer.tokenize_row(
        ...     features, tokenizer, "\n", max_completion_length=None, train_on_last_step_only=False, is_eval=False
        ... )
        {'input_ids': [23085, 1372, 374, 8131, 11, 220, 24, 13, 23, 476, 220, 24, 13, 16, 16, 30, 16, 16, 374, 7046, 1091, 220, 23, 13, 198, 39, 763, 11, 220, 24, 13, 16, 16, 861, 220, 24, 13, 23, 13, 198],
         'labels': [-100, -100, -100, -100, -100, -100, -100, -100, 1, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, 0]}
        ```
        """
        # Tokenize the prompt and completions
        prompt_ids = tokenizer(features["prompt"], add_special_tokens=False)["input_ids"]
        completions_ids = [
            tokenizer(completion, add_special_tokens=False)["input_ids"] for completion in features["completions"]
        ]
        if train_on_last_step_only and not is_eval:
            labels = [-100] * (len(features["labels"]) - 1) + [int(features["labels"][-1])]
        else:
            labels = [int(label) for label in features["labels"]]

        # Get the ID of the separator token and add it to the completions
        separator_ids = tokenizer.encode(step_separator, add_special_tokens=False)
        completions_ids = [completion + separator_ids for completion in completions_ids]

        # Create the label
        labels = [[-100] * (len(completion) - 1) + [label] for completion, label in zip(completions_ids, labels)]

        # Join the completions and labels steps
        completion_ids = list(chain(*completions_ids))
        labels = list(chain(*labels))

        if tokenizer.bos_token_id is not None:
            prompt_ids = [tokenizer.bos_token_id] + prompt_ids

        # Truncate prompt and completion sequences
        if max_prompt_length is not None:
            prompt_ids = prompt_ids[-max_prompt_length:]
        if max_completion_length is not None:
            completion_ids = completion_ids[:max_completion_length]
            labels = labels[:max_completion_length]

        input_ids = prompt_ids + completion_ids
        labels = [-100] * len(prompt_ids) + labels

        if max_length is not None:
            input_ids = input_ids[:max_length]
            labels = labels[:max_length]

        return {"input_ids": input_ids, "labels": labels}

    # 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 UnslothPRMTrainer(_UnslothPRMTrainer):
    """
    
    Initialize PRMTrainer.

    Args:
        model ([`~transformers.PreTrainedModel`]):
            The model to train, preferably an `AutoModelForTokenClassification`.
        args ([`PRMConfig`]):
            The arguments to use for training.
        data_collator ([`~transformers.DataCollator`]):
            The data collator to use for training. If None is specified, the default data collator
            ([`~transformers.DataCollatorForTokenClassification`]) will be used which will pad the sequences to the
            maximum length of the sequences in the batch, given a dataset of paired sequences.
        train_dataset ([`~datasets.Dataset`]):
            The dataset to use for training.
        eval_dataset ([`~datasets.Dataset`]):
            The dataset to use for evaluation.
        processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.BaseImageProcessor`], [`~transformers.FeatureExtractionMixin`] or [`~transformers.ProcessorMixin`], *optional*):
            Processing class used to process the data. If provided, will be used to automatically process the inputs
            for the model, and it will be saved along the model to make it easier to rerun an interrupted training or
            reuse the fine-tuned model.
        model_init (`Callable[[], transformers.PreTrainedModel]`):
            The model initializer to use for training. If None is specified, the default model initializer will be
            used.
        compute_metrics (`Callable[[transformers.EvalPrediction], dict]`, *optional* defaults to `compute_accuracy`):
            The metrics to use for evaluation. If no metrics are specified, the default metric (`compute_accuracy`)
            will be used.
        callbacks (`list[transformers.TrainerCallback]`):
            The callbacks to use for training.
        optimizers (`tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`):
            The optimizer and scheduler to use for training.
        preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`):
            The function to use to preprocess the logits before computing the metrics.
        peft_config (`dict`, defaults to `None`):
            The PEFT configuration to use for training. If you pass a PEFT configuration, the model will be wrapped in
            a PEFT model.
    
    """
    def __init__(
        self,
        model = None,
        args = None,
        data_collator = None,
        train_dataset = None,
        eval_dataset = None,
        processing_class = None,
        model_init = None,
        compute_metrics = None,
        callbacks = None,
        preprocess_logits_for_metrics = None,
        peft_config = None,
        **kwargs
    ):
        if args is None: args = UnslothPRMConfig()
        use_bf16 = getattr(args, 'bf16', False)
        if type(use_bf16) is not bool: use_bf16 = False
        use_fp16 = getattr(args, 'fp16', False)
        if type(use_fp16) is not bool: use_fp16 = False
        force_float32 = False
        full_finetuning = os.environ.get('UNSLOTH_ENABLE_FULL_FINETUNING', '0') == '1'
        if not full_finetuning and (os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1'):
            print('Unsloth: Switching to float32 training since model cannot work with float16')
            force_float32 = True
        mixed_precision_dtype = os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32')
        dtype = getattr(model.config, 'dtype', None) or getattr(model.config, 'torch_dtype', None)
        if dtype is None: dtype = model.get_input_embeddings().weight.dtype
        from unsloth_zoo.utils import _get_dtype
        dtype = _get_dtype(dtype)
        float16 = dtype == torch.float16
        if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')
        if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')
        if force_float32:
            # Forced float32 training
            args.fp16 = False
            args.bf16 = False
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
            # args.mixed_precision is a new argument which needs to be set now
        elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':
            # Mixed precision training
            args.fp16 = float16
            args.bf16 = not float16
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'
            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'fp16' if float16 else 'bf16'
            # args.mixed_precision is a new argument which needs to be set now
        elif mixed_precision_dtype == 'bfloat16':
            # Both False since bfloat16 full finetuning doesn't do any autocasting.
            args.fp16 = False
            args.bf16 = False
            os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
            if hasattr(args, 'mixed_precision'): args.mixed_precision = 'no'
            # args.mixed_precision is a new argument which needs to be set now
        
        if getattr(args, 'eval_dataset', None) is not None and getattr(args, 'eval_strategy', 'no') == 'no':
            args.eval_strategy = 'steps'
            if getattr(args, 'eval_steps', None) is None: args.eval_steps = 0.1
        ga_steps = getattr(args, 'gradient_accumulation_steps', None)
        if ga_steps is not None and ga_steps > 1:
            from transformers import __version__ as transformers_version
            if Version(transformers_version) <= Version('4.45.2'):
                print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!\n'
                      '`pip install --upgrade --no-cache-dir --force-reinstall --no-deps unsloth transformers trl unsloth_zoo`')
        if getattr(args, 'eval_strategy', 'no') != 'no':
            eval_bsz = getattr(args, 'per_device_eval_batch_size', 8)
            if eval_bsz == 8 and args.per_device_train_batch_size < eval_bsz: args.per_device_eval_batch_size = args.per_device_train_batch_size
            if getattr(args, 'eval_accumulation_steps', None) is None and ga_steps is not None: args.eval_accumulation_steps = ga_steps
        fp16_full_eval = getattr(args, 'fp16_full_eval', False)
        if type(fp16_full_eval) is not bool: fp16_full_eval = False
        bf16_full_eval = getattr(args, 'bf16_full_eval', False)
        if type(bf16_full_eval) is not bool: bf16_full_eval = False
        if args.fp16 and bf16_full_eval: args.bf16_full_eval = False; args.fp16_full_eval = True
        if args.bf16 and fp16_full_eval: args.bf16_full_eval = True; args.fp16_full_eval = False
        if force_float32:
            args.bf16_full_eval = False
            args.fp16_full_eval = False
        elif os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32') == 'bfloat16':
            args.bf16_full_eval = True
            args.fp16_full_eval = False
        elif not bf16_full_eval and not fp16_full_eval:
            args.bf16_full_eval = args.bf16
            args.fp16_full_eval = args.fp16
        _output_logits = False
        if locals().get('compute_metrics', None) is not None: _output_logits = True
        if locals().get('preprocess_logits_for_metrics', None) is not None: _output_logits = True
        if _output_logits:
            os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
        if 'max_seq_length' not in locals() and not hasattr(args, 'max_seq_length'):
            pass
        else:
            model_max_seq_length = getattr(model, 'max_seq_length', None)
            args_max_seq_length  = getattr(args,  'max_seq_length', None)
            if args_max_seq_length is None and model_max_seq_length is not None:
                max_seq_length = model.max_seq_length
                if hasattr(args, 'max_seq_length'): args.max_seq_length = max_seq_length
            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('prm_trainer', other_metrics)
        
        # [TODO] Fix up DataParallel multiplying batch sizes
        # [TODO] DDP works, but DP seems to not work? [TODO]
        if getattr(args, "parallel_mode", None) == ParallelMode.NOT_DISTRIBUTED and args.n_gpu > 1:
            if getattr(args, "_n_gpu", 1) != 1:
                args._n_gpu = 1
        if "model" in locals() and hasattr(model, "for_training"):
            model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
        super().__init__(
            model = model,
            args = args,
            data_collator = data_collator,
            train_dataset = train_dataset,
            eval_dataset = eval_dataset,
            processing_class = processing_class,
            model_init = model_init,
            compute_metrics = compute_metrics,
            callbacks = callbacks,
            preprocess_logits_for_metrics = preprocess_logits_for_metrics,
            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