"""
2026.6.7
2026.6.9
5.5.0
1.7.0
__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 .
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.reward_trainer import (Any, AutoModelForSequenceClassification, AutoTokenizer, Callable, DataCollator, DataCollatorForPreference, Dataset, EvalPrediction, IterableDataset, PartialState, Path, PeftConfig, PeftModel, PreTrainedModel, PreTrainedTokenizerBase, RewardConfig, RewardTrainer, TrainerCallback, Version, _BaseTrainer, clone_chat_template, contextlib, create_model_from_path, dataclass, defaultdict, disable_dropout_in_model, get_act_offloading_ctx_manager, get_config_model_id, get_dataset_column_names, get_peft_model, is_conversational, is_peft_available, is_peft_model, json, logger, logging, nn, os, pad, peft, re, set_seed, suppress_seqcls_warning, torch, transformers, warnings, Any, AutoModelForSequenceClassification, AutoTokenizer, Callable, DataCollator, DataCollatorForPreference, Dataset, EvalPrediction, IterableDataset, PeftConfig, PeftModel, PreTrainedModel, PreTrainedTokenizerBase, RewardConfig, TrainerCallback, Version, clone_chat_template, contextlib, create_model_from_path, defaultdict, disable_dropout_in_model, get_act_offloading_ctx_manager, get_config_model_id, get_peft_model, is_peft_available, is_peft_model, logger, os, pad, peft, re, set_seed, suppress_seqcls_warning, torch, transformers, PeftModel, PreTrainedModel, is_peft_available, logger, os, peft, re, 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
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
# Canonical reset lives in unsloth.models._utils so the SFT auto-packing wrapper and the plain
# Trainer loop can import the same helper; fall back to a no-op only if it can't be imported.
try:
from unsloth.models._utils import _unsloth_reset_stray_compile_cache
except Exception:
def _unsloth_reset_stray_compile_cache(self): pass
def prepare_for_training_mode(f):
@functools.wraps(f)
def wrapper(self, *args, **kwargs):
# Drop any torch.compile graph cache poisoned by a stray pre-train forward.
try:
_unsloth_reset_stray_compile_cache(self)
except Exception:
pass
# Finish the previous W&B run if this is a subsequent train() call.
# We do this at the START of train() (not the end) so that
# evaluate() / log() still work after train() completes.
# HF's WandbCallback.setup() will call wandb.init() for the new run.
# See: https://github.com/unslothai/unsloth/issues/3954
if getattr(self, '_unsloth_training_completed', False):
try:
import wandb
if wandb.run is not None:
wandb.finish()
# Reset HF's WandbCallback so it calls wandb.init() for the new run
for cb in self.callback_handler.callbacks:
if type(cb).__name__ == 'WandbCallback':
cb._initialized = False
break
except:
pass
# 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
# Mark that training completed so the next train() call can
# finish this W&B run before starting a new one
self._unsloth_training_completed = True
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 = logit_softcapping * 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,
temperature: float = 1.0,
chunks: int = 4,
):
chunked_logits = torch.chunk(logits.reshape(-1, logits.shape[-1]), chunks = chunks, dim = 0)
chunked_index = torch.chunk(index.reshape(-1), chunks = chunks, dim = 0)
all_per_token_logps = []
# Per-chunk selective_log_softmax.
for chunk_logits, chunk_index in zip(chunked_logits, chunked_index):
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)
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:
"""Count left-padded tokens per sequence, e.g. [pad, pad, pad, cat] -> 3."""
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:
"""Build a completion mask that zeros leading prompt and trailing pad tokens.
For [p,p,p,c,c,c,pad,pad,pad] (p=sliced prompt, c=completion, pad=padding)
this returns [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:
"""Move all padding tokens in each sequence to the right."""
mask = (tensor != pad_id)
# stable=True since the binary mask 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:
"""Align 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
# Destination row indices, shape [batch_size, logprob_seq_len].
row_indices = torch.arange(batch_size, device=device).unsqueeze(1).expand_as(dest_indices)
# Keep only in-bounds destinations, then scatter via advanced indexing.
valid_mask = dest_indices < mask_seq_len
valid_rows = row_indices[valid_mask]
valid_cols = dest_indices[valid_mask]
valid_vals = logprob_tensor[valid_mask]
padded_logprobs[valid_rows, valid_cols] = valid_vals
return padded_logprobs
def align_completion_tool_mask(
tool_mask: torch.Tensor,
completion_mask: torch.Tensor,
) -> torch.Tensor:
"""Align a raw completion-length tool/env mask with Unsloth's repacked loss mask."""
if tool_mask is None:
return completion_mask
if tool_mask.shape[0] != completion_mask.shape[0]:
raise ValueError("tool_mask batch size must match completion_mask batch size.")
tool_mask = tool_mask.to(device=completion_mask.device)
if tool_mask.shape == completion_mask.shape:
aligned_tool_mask = tool_mask
else:
aligned_tool_mask = align_logprobs_with_mask(
tool_mask,
completion_mask,
pad_value=0,
)
return completion_mask * aligned_tool_mask.to(dtype=completion_mask.dtype)
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():
# XPU: estimate free memory as 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 UnslothRewardConfig(RewardConfig):
"""
Configuration class for the [`RewardTrainer`].
This class includes only the parameters that are specific to Reward training. For a full list of training
arguments, please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values in this
class may differ from those in [`~transformers.TrainingArguments`].
Using [`~transformers.HfArgumentParser`] we can turn this class into
[argparse](https://docs.python.org/3/library/argparse#module-argparse) arguments that can be specified on the
command line.
Parameters:
> Parameters that control the model
model_init_kwargs (`dict[str, Any]`, *optional*):
Keyword arguments for [`~transformers.AutoModelForCausalLM.from_pretrained`], used when the `model`
argument of the [`RewardTrainer`] is provided as a string.
trust_remote_code (`bool`, *optional*, defaults to `False`):
Whether to allow loading models and tokenizers that ship custom Python code from the Hub. Forwarded to
[`~transformers.AutoModelForSequenceClassification.from_pretrained`] and
[`~transformers.AutoTokenizer.from_pretrained`].
chat_template_path (`str`, *optional*):
If specified, sets the model's chat template. This can either be the path to a tokenizer (local directory
or Hugging Face Hub model) or a direct path to a Jinja template file. When using a Jinja file, you must
ensure that any special tokens referenced in the template are added to the tokenizer and that the model's
embedding layer is resized accordingly.
disable_dropout (`bool`, *optional*, defaults to `True`):
Whether to disable dropout in the model.
> Parameters that control the data preprocessing
dataset_num_proc (`int`, *optional*):
Number of processes to use for processing the dataset.
eos_token (`str`, *optional*):
Token used to indicate the end of a turn or sequence. If `None`, it defaults to
`processing_class.eos_token`.
max_length (`int` or `None`, *optional*, defaults to `1024`):
Maximum length of the tokenized sequence. Samples are filtered out if either chosen or rejected sequence
exceeds this value. If `None`, no filtering is applied.
pad_to_multiple_of (`int`, *optional*):
If set, the sequences will be padded to a multiple of this value.
> Parameters that control the training
center_rewards_coefficient (`float`, *optional*):
Coefficient to incentivize the reward model to output mean-zero rewards (proposed by
https://huggingface.co/papers/2312.09244, Eq. 2). Recommended value: `0.01`.
activation_offloading (`bool`, *optional*, defaults to `False`):
Whether to offload the activations to the CPU.
> Deprecated parameters
pad_token:
Parameter `pad_token` is deprecated and will be removed in version v2.0.0. Set `tokenizer.pad_token`
directly and pass it as `processing_class` to the trainer instead.
> [!NOTE]
> These parameters have default values different from [`~transformers.TrainingArguments`]:
> - `logging_steps`: Defaults to `10` instead of `500`.
> - `gradient_checkpointing`: Defaults to `True` instead of `False`.
> - `bf16`: Defaults to `True` if `fp16` is not set, instead of `False`.
> - `learning_rate`: Defaults to `1e-4` instead of `5e-5`.
"""
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 = 0.001,
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,
model_init_kwargs = None,
trust_remote_code = False,
chat_template_path = None,
disable_dropout = True,
dataset_num_proc = None,
eos_token = None,
max_length = 1024,
pad_to_multiple_of = None,
center_rewards_coefficient = None,
activation_offloading = False,
pad_token = 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 dataset_num_proc is None:
if _mp.get_start_method() != 'fork':
dataset_num_proc = None
else:
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 os.environ.get('UNSLOTH_ENABLE_FLEX_ATTENTION', '0') == '1':
from unsloth_zoo.flex_attention import HAS_FLEX_ATTENTION
if HAS_FLEX_ATTENTION and pad_to_multiple_of is None:
from unsloth_zoo.flex_attention import FLEX_ATTENTION_BLOCK_SIZE
pad_to_multiple_of = FLEX_ATTENTION_BLOCK_SIZE
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,
model_init_kwargs = model_init_kwargs,
trust_remote_code = trust_remote_code,
chat_template_path = chat_template_path,
disable_dropout = disable_dropout,
dataset_num_proc = dataset_num_proc,
eos_token = eos_token,
max_length = max_length,
pad_to_multiple_of = pad_to_multiple_of,
center_rewards_coefficient = center_rewards_coefficient,
activation_offloading = activation_offloading,
pad_token = pad_token,**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
# Unsloth: Remove use_reentrant=False forced by TRL 0.27.0+
if getattr(self, 'gradient_checkpointing_kwargs', None) is not None:
if 'use_reentrant' in self.gradient_checkpointing_kwargs:
del self.gradient_checkpointing_kwargs['use_reentrant']
pass
class _UnslothRewardTrainer(_BaseTrainer):
""""""
_tag_names = ["trl", "reward-trainer"]
_name = "Reward"
_template_file = "rm_model_card.md"
def __init__(
self,
model: "str | PreTrainedModel | PeftModel",
args: RewardConfig | None = None,
data_collator: DataCollator | None = None,
train_dataset: Dataset | IterableDataset | None = None,
eval_dataset: Dataset | IterableDataset | dict[str, Dataset | IterableDataset] | None = None,
processing_class: PreTrainedTokenizerBase | None = None,
compute_metrics: Callable[[EvalPrediction], dict] | None = None,
callbacks: list[TrainerCallback] | None = None,
optimizers: tuple[torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None] = (None, None),
optimizer_cls_and_kwargs: tuple[type[torch.optim.Optimizer], dict[str, Any]] | None = None,
preprocess_logits_for_metrics: Callable[[torch.Tensor, torch.Tensor], torch.Tensor] | None = None,
peft_config: "PeftConfig | None" = None,
):
# Args
if args is None:
model_name = model if isinstance(model, str) else get_config_model_id(model.config)
model_name = model_name.split("/")[-1]
args = RewardConfig(f"{model_name}-Reward")
if train_dataset is None:
raise ValueError("`train_dataset` is required")
elif isinstance(train_dataset, IterableDataset):
# IterableDataset requires dispatch_batches=False because Accelerate's dispatch mode may try to concatenate
# batches from multiple processes, leading to mismatch errors.
if args.accelerator_config.dispatch_batches is True:
logger.warning(
"You are using an `IterableDataset` for training with `dispatch_batches=True`. `dispatch_batches` "
"is forced to `False` when using an `IterableDataset`. To remove this warning, unset "
"`dispatch_batches` in `RewardConfig` or set it to `False`."
)
args.accelerator_config.dispatch_batches = False
# Model
# As AutoModelForSequenceClassification.from_pretrained[] will add a random head for the model, set_seed must
# be done before loading the model to ensure reproducibility.
set_seed(args.seed)
if isinstance(model, str):
model_init_kwargs = args.model_init_kwargs or {}
# Distributed training requires device_map=None ["auto" fails]
if args.distributed_state.distributed_type in ["MULTI_GPU", "DEEPSPEED"]:
model_init_kwargs["device_map"] = None
model_init_kwargs["num_labels"] = 1 # the only output of the model is the reward score
model_init_kwargs.setdefault("trust_remote_code", args.trust_remote_code)
with suppress_seqcls_warning():
model = create_model_from_path(model, AutoModelForSequenceClassification, **model_init_kwargs)
else:
if args.model_init_kwargs is not None:
logger.warning(
"You passed `model_init_kwargs` to the `RewardConfig`, but your model is already instantiated. "
"The `model_init_kwargs` will be ignored."
)
# Validate that the model has num_labels = 1 [required for reward models]
if getattr(model.config, "num_labels", None) != 1:
raise ValueError(
f"The model has `num_labels={model.config.num_labels}`, but reward models require `num_labels=1` "
"to output a single scalar reward per sequence. Please instantiate your model with `num_labels=1` "
"or pass a model name as a string to have it configured automatically."
)
# Non-quantized models do not have the `is_loaded_in_{8,4}bit` attributes, whereas quantized models do
_is_quantized_model = getattr(model, "is_loaded_in_4bit", False) or getattr(model, "is_loaded_in_8bit", False)
# Processing class
if processing_class is None:
processing_class = AutoTokenizer.from_pretrained(
get_config_model_id(model.config), trust_remote_code=args.trust_remote_code
)
# Handle pad token for processors or tokenizers
if args.eos_token is not None:
if args.eos_token not in processing_class.get_vocab():
raise ValueError(
f"The specified `eos_token` ('{args.eos_token}') is not found in the vocabulary of the given "
f"`processing_class` ({processing_class.__class__.__name__}). Ensure that the `eos_token` exists "
"in the vocabulary before using it as an EOS token."
)
processing_class.eos_token = args.eos_token
if args.chat_template_path is not None:
if os.path.isfile(args.chat_template_path) and args.chat_template_path.endswith((".jinja", ".j2")):
with open(args.chat_template_path, encoding="utf-8") as chat_template_file:
processing_class.chat_template = chat_template_file.read()
added_tokens = []
else:
model, processing_class, added_tokens = clone_chat_template(
model, processing_class, args.chat_template_path
)
else:
added_tokens = []
# PEFT
if False:
if not is_peft_available():
raise ImportError(
"You passed `peft_config` but the `peft` library is not installed. "
"Install it with `pip install trl[peft]`."
)
if not isinstance(peft_config, PeftConfig):
raise TypeError(
f"`peft_config` must be a `peft.PeftConfig` instance (e.g. `peft.LoraConfig`), "
f"got {type(peft_config).__name__}."
)
if is_peft_model(model):
raise ValueError(
"You passed a `PeftModel` instance together with a `peft_config` to the trainer. Please first merge "
"and unload the existing adapter, save the resulting base model, and then pass that base model along "
"with the new `peft_config` to the trainer."
)
if added_tokens:
# Ensure that the added tokens are trainable
if peft_config.trainable_token_indices is None:
peft_config.trainable_token_indices = {"embed_tokens": added_tokens}
elif "embed_tokens" not in peft_config.trainable_token_indices:
peft_config.trainable_token_indices["embed_tokens"] = added_tokens
else:
peft_config.trainable_token_indices["embed_tokens"].extend(added_tokens)
# Ensure that the lm_head is trainable
if peft_config.modules_to_save is None or "lm_head" not in peft_config.modules_to_save:
logger.warning(
"Cloning chat template added new tokens to the tokenizer, but 'lm_head' is not in PEFT's "
"`modules_to_save`. As a result, the model may not learn to generate outputs with these new "
"tokens, leading to degraded generation quality. To fix this, add "
"`modules_to_save=['lm_head']` to your PEFT configuration."
)
if peft_config.modules_to_save is None:
peft_config.modules_to_save = ["lm_head"]
else:
peft_config.modules_to_save.append("lm_head")
# Create PEFT model
# ZeRO-3 + PEFT for non-quantized models:
# - PEFT's default autocast_adapter_dtype=True upcasts LoRA adapter params to fp32 even when the base model is bf16.
# - ZeRO-3's _allgather_params_coalesced allocates output buffers using the dtype of the first persistent parameter,
# so mixed-dtype persistent_parameters [bf16 base + fp32 LoRA] cause a TypeError on the first optimizer step.
# - Passing autocast_adapter_dtype=False keeps adapter params in the base model dtype [bf16], fixing the mismatch.
# - This is safe: the fp32 upcast is a QLoRA-specific concern [low-bit quantized base models], not needed for
# non-quantized bf16 training.
# - See:
# - TRL issue: https://github.com/huggingface/trl/issues/6089
# - Upstream issue: https://github.com/deepspeedai/DeepSpeed/issues/8072
# - autocast_adapter_dtype was introduced in PEFT 0.12.0; before, no upcast existed: no need to pass the kwarg
get_peft_model_kwargs = {}
if (
args.deepspeed_plugin is not None
and args.deepspeed_plugin.zero_stage == 3
and not _is_quantized_model
and Version(peft.__version__) >= Version("0.12.0")
):
get_peft_model_kwargs["autocast_adapter_dtype"] = False
model = get_peft_model(model, peft_config, **get_peft_model_kwargs)
# When using gradient checkpointing with PEFT, we need to enable input gradients. transformers.Trainer normally
# handles this, but a bug currently prevents it; see https://github.com/huggingface/transformers/issues/42489
if is_peft_model(model) and args.gradient_checkpointing:
model.enable_input_require_grads()
# When using QLoRA, the PEFT adapter weights are converted to bf16 to follow the recommendations from the
# original paper [see https://huggingface.co/papers/2305.14314, paragraph 3]. Normally, this can be done by
# passing `autocast_adapter_dtype=False` to `get_peft_model`, but this option is not yet supported for
# quantized models. See: https://github.com/huggingface/peft/issues/2889
if _is_quantized_model:
for param in model.parameters():
if param.requires_grad:
param.data = param.data.to(torch.bfloat16)
# Disable dropout in the model
if args.disable_dropout:
disable_dropout_in_model(model)
# Pad token [needed for SequenceClassification models]
# If not provided, use the one from the processing class or the eos token if the processing class does not have
# a pad token.
pad_token = args.pad_token or processing_class.pad_token or processing_class.eos_token
if pad_token not in processing_class.get_vocab():
raise ValueError(
f"The specified `pad_token` ('{pad_token}') is not found in the vocabulary of the given "
f"`processing_class` ({processing_class.__class__.__name__}). Ensure that the `pad_token` exists "
"in the vocabulary before using it as a padding token."
)
processing_class.pad_token = pad_token
# SequenceClassification models need `config.pad_token_id` to locate the last non-pad token.
model.config.pad_token_id = processing_class.pad_token_id
# Data collator
if data_collator is None:
data_collator = DataCollatorForPreference(
pad_token_id=processing_class.pad_token_id,
pad_to_multiple_of=args.pad_to_multiple_of,
)
# Dataset
train_dataset = self._prepare_dataset(train_dataset, processing_class, args, "train")
if eval_dataset is not None:
if isinstance(eval_dataset, dict):
eval_dataset = {
key: self._prepare_dataset(dataset, processing_class, args, key)
for key, dataset in eval_dataset.items()
}
else:
eval_dataset = self._prepare_dataset(eval_dataset, processing_class, args, "eval")
# Transformers explicitly set use_reentrant=True in the past to silence a PyTorch warning, but the default was
# never updated once PyTorch switched to recommending use_reentrant=False. Until that change lands upstream
# [see https://github.com/huggingface/transformers/pull/43203] and is released [most likely in 5.0.0], we
# default to the recommended non-reentrant behavior here, while preserving any user-provided value.
if args.gradient_checkpointing and Version(transformers.__version__) < Version("5.0.0"):
args.gradient_checkpointing_kwargs = args.gradient_checkpointing_kwargs or {}
args.gradient_checkpointing_kwargs.setdefault("use_reentrant", False)
super().__init__(
model=model,
args=args,
data_collator=data_collator,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
processing_class=processing_class,
compute_metrics=compute_metrics,
callbacks=callbacks,
optimizers=optimizers,
optimizer_cls_and_kwargs=optimizer_cls_and_kwargs,
preprocess_logits_for_metrics=preprocess_logits_for_metrics,
)
# During evaluation, Trainer calls compute_loss[] only if can_return_loss is True and label_names is empty.
self.can_return_loss = True
self.label_names = []
# Initialize activation offloading context
if self.args.activation_offloading:
self.maybe_activation_offload_context = get_act_offloading_ctx_manager(model=self.model)
else:
self.maybe_activation_offload_context = contextlib.nullcontext()
# Initialize the metrics
self._metrics = {"train": defaultdict(list), "eval": defaultdict(list)}
self._total_train_tokens = 0
# Gradient accumulation requires scaled loss. Normally, loss scaling in the parent class depends on whether the
# model accepts loss-related kwargs. Since we compute our own loss, this check is irrelevant. We set
# self.model_accepts_loss_kwargs to False to enable scaling.
self.model_accepts_loss_kwargs = False
# Add tags to the model
self.model.add_model_tags(self._tag_names)
def _prepare_dataset(
self,
dataset: Dataset | IterableDataset,
processing_class: PreTrainedTokenizerBase,
args: RewardConfig,
dataset_name: str,
) -> Dataset | IterableDataset:
# If the dataset is already preprocessed (tokenized), skip the processing steps.
column_names = get_dataset_column_names(dataset)
is_processed = "chosen_ids" in column_names and "rejected_ids" in column_names
has_legacy_processed_columns = "chosen_input_ids" in column_names and "rejected_input_ids" in column_names
if has_legacy_processed_columns and not is_processed:
warnings.warn(
"Detected legacy dataset columns `chosen_input_ids`/`rejected_input_ids`; they are deprecated and "
"will not be supported in v1. Please migrate to `chosen_ids`/`rejected_ids`.",
FutureWarning,
stacklevel=2,
)
dataset = dataset.rename_column("chosen_input_ids", "chosen_ids")
dataset = dataset.rename_column("rejected_input_ids", "rejected_ids")
is_processed = True
# Build the kwargs for the `map` function
map_kwargs = {}
if isinstance(dataset, Dataset): # IterableDataset does not support num_proc
map_kwargs["num_proc"] = args.dataset_num_proc
with PartialState().main_process_first():
if not is_processed:
# Add EOS token if needed: non-conversational only
first_example = next(iter(dataset))
if not is_conversational(first_example):
if isinstance(dataset, Dataset): # `IterableDataset.map` does not support `desc`
map_kwargs["desc"] = f"Adding EOS to {dataset_name} dataset"
def add_eos(example, eos_token):
if not example["chosen"].endswith(eos_token):
example["chosen"] = example["chosen"] + eos_token
if "rejected" in example and not example["rejected"].endswith(eos_token):
example["rejected"] = example["rejected"] + eos_token
return example
dataset = dataset.map(
add_eos,
fn_kwargs={"eos_token": processing_class.eos_token},
**map_kwargs,
)
# Tokenize the dataset
if isinstance(dataset, Dataset): # `IterableDataset.map` does not support `desc`
map_kwargs["desc"] = f"Tokenizing {dataset_name} dataset"
def tokenize_fn(example, processing_class):
tools = example.get("tools")
tools = json.loads(tools) if isinstance(tools, str) else tools
if "prompt" in example: # explicit prompt case
example["chosen"] = example["prompt"] + example["chosen"]
example["rejected"] = example["prompt"] + example["rejected"]
if is_conversational(example):
chosen_ids = processing_class.apply_chat_template(
example["chosen"],
tools=tools,
return_dict=True,
**example.get("chat_template_kwargs", {}),
)["input_ids"]
rejected_ids = processing_class.apply_chat_template(
example["rejected"],
tools=tools,
return_dict=True,
**example.get("chat_template_kwargs", {}),
)["input_ids"]
output = {"chosen_ids": chosen_ids, "rejected_ids": rejected_ids}
else:
output = {
"chosen_ids": processing_class(text=example["chosen"])["input_ids"],
"rejected_ids": processing_class(text=example["rejected"])["input_ids"],
}
return output
dataset = dataset.map(tokenize_fn, fn_kwargs={"processing_class": processing_class}, **map_kwargs)
# Filter samples that are longer than `max_length`
if args.max_length is not None:
if isinstance(dataset, Dataset): # `IterableDataset.map` does not support `desc`
map_kwargs["desc"] = f"Filtering {dataset_name} >{args.max_length} tokens"
dataset = dataset.filter(
lambda example: len(example["chosen_ids"]) <= args.max_length
and len(example["rejected_ids"]) <= args.max_length,
**map_kwargs,
)
return dataset
def _set_signature_columns_if_needed(self):
# If `self.args.remove_unused_columns` is True, non-signature columns are removed.
# By default, this method sets `self._signature_columns` to the model's expected inputs (usually, "input_ids"
# and "attention_mask").
if self._signature_columns is None:
self._signature_columns = ["chosen_ids", "rejected_ids", "margin"]
def evaluate(
self,
eval_dataset: Dataset | dict[str, Dataset] | None = None,
ignore_keys: list[str] | None = None,
metric_key_prefix: str = "eval",
) -> dict[str, float]:
# When a dataset is passed directly to `evaluate` (e.g. a held-out test set), preprocess it the same way
# `__init__` does, so that `evaluate` accepts the same dataset types as the trainer. `_prepare_dataset` is
# idempotent: it skips datasets that are already tokenized. A `str` selects a dataset that was already prepared
# at init time, so it's left untouched.
if eval_dataset is not None and not isinstance(eval_dataset, str):
if isinstance(eval_dataset, dict):
eval_dataset = {
key: self._prepare_dataset(dataset, self.processing_class, self.args, key)
for key, dataset in eval_dataset.items()
}
else:
eval_dataset = self._prepare_dataset(eval_dataset, self.processing_class, self.args, "eval")
return super().evaluate(
eval_dataset=eval_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix
)
def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
mode = "train" if self.model.training else "eval"
# If not set, defaults from model config and may warn since cache isn't compatible with gradient checkpointing
inputs["use_cache"] = False
outputs = model(**inputs)
# Split the rewards into chosen and rejected
rewards_chosen, rewards_rejected = torch.chunk(outputs.logits.squeeze(-1), chunks=2)
# Calculate loss, optionally modulate with margin
if "margin" in inputs:
loss = -nn.functional.logsigmoid(rewards_chosen - rewards_rejected - inputs["margin"]).mean()
else:
loss = -nn.functional.logsigmoid(rewards_chosen - rewards_rejected).mean()
if self.args.center_rewards_coefficient is not None:
loss += self.args.center_rewards_coefficient * torch.mean((rewards_chosen + rewards_rejected) ** 2)
if mode == "train":
num_tokens_in_batch = self.accelerator.gather_for_metrics(inputs["attention_mask"].sum()).sum().item()
self._total_train_tokens += num_tokens_in_batch
self._metrics[mode]["num_tokens"] = [self._total_train_tokens]
# Compute min, mean, max, accuracy and margin
with torch.no_grad():
all_rewards = self.accelerator.gather(outputs.logits)
self._metrics[mode]["min_reward"].append(all_rewards.min().item())
self._metrics[mode]["mean_reward"].append(all_rewards.mean().item())
self._metrics[mode]["max_reward"].append(all_rewards.max().item())
mean_accuracy = (rewards_chosen > rewards_rejected).float().mean()
mean_accuracy = self.accelerator.gather_for_metrics(mean_accuracy).mean().item()
self._metrics[mode]["accuracy"].append(mean_accuracy)
mean_margin = (rewards_chosen - rewards_rejected).mean()
mean_margin = self.accelerator.gather_for_metrics(mean_margin).mean()
self._metrics[mode]["margin"].append(mean_margin.item())
return (loss, outputs) if return_outputs else loss
# Override training step to add activation offloading context.
def training_step(self, *args, **kwargs):
with self.maybe_activation_offload_context:
return super().training_step(*args, **kwargs)
def log(self, logs: dict[str, float], start_time: float | None = None) -> None:
mode = "train" if self.model.training else "eval"
metrics = {key: sum(val) / len(val) for key, val in self._metrics[mode].items()} # average the metrics
# This method can be called both in training and evaluation. When called in evaluation, the keys in `logs`
# start with "eval_". We need to add the prefix "eval_" to the keys in `metrics` to match the format.
if mode == "eval":
metrics = {f"eval_{key}": val for key, val in metrics.items()}
logs.update(metrics)
super().log(logs, start_time)
self._metrics[mode].clear()
# 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 UnslothRewardTrainer(_UnslothRewardTrainer):
"""
Trainer for Outcome-supervised Reward Models (ORM).
This class is a wrapper around the [`~transformers.Trainer`] class and inherits all of its attributes and methods.
Example:
```python
>>> from trl import RewardTrainer
>>> from datasets import load_dataset
>>> dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")
>>> trainer = RewardTrainer(
... model="Qwen/Qwen2.5-0.5B-Instruct",
... train_dataset=dataset,
... )
>>> trainer.train()
```
Args:
model (`str` or [`~transformers.PreTrainedModel`] or [`~peft.PeftModel`]):
Model to be trained. Can be either:
- A string, being the *model id* of a pretrained model hosted inside a model repo on huggingface.co, or a
path to a *directory* containing model weights saved using
[`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded
using `AutoModelForSequenceClassification.from_pretrained` with the keyword arguments in
`args.model_init_kwargs`.
- A sequence classification [`~transformers.PreTrainedModel`] object.
- A sequence classification [`~peft.PeftModel`] object.
args ([`RewardConfig`], *optional*):
Configuration for this trainer. If `None`, a default configuration is used.
data_collator ([`~transformers.DataCollator`], *optional*):
Function to use to form a batch from a list of elements of the processed `train_dataset` or `eval_dataset`.
Will default to [`~trainer.reward_trainer.DataCollatorForPreference`].
train_dataset ([`~datasets.Dataset`] or [`~datasets.IterableDataset`]):
Dataset to use for training. This trainer supports [preference](#preference) type (both implicit and
explicit prompt). The format of the samples can be either:
- [Standard](dataset_formats#standard): Each sample contains plain text.
- [Conversational](dataset_formats#conversational): Each sample contains structured messages (e.g., role
and content).
The trainer also supports processed datasets (tokenized) as long as they contain `chosen_ids` and
`rejected_ids` fields.
eval_dataset ([`~datasets.Dataset`], [`~datasets.IterableDataset`] or `dict[str, Dataset | IterableDataset]`):
Dataset to use for evaluation. It must meet the same requirements as `train_dataset`.
processing_class ([`~transformers.PreTrainedTokenizerBase`], *optional*):
Tokenizer used to process the data. If `None`, the tokenizer is loaded from the model's name with
[`~transformers.AutoTokenizer.from_pretrained`]. A padding token, `processing_class.pad_token`, must be
set. If the processing class has not set a padding token, `processing_class.eos_token` will be used as the
default.
compute_metrics (`Callable[[EvalPrediction], dict]`, *optional*):
The function that will be used to compute metrics at evaluation. Must take a
[`~transformers.EvalPrediction`] and return a dictionary string to metric values. When passing
[`RewardConfig`] with `batch_eval_metrics` set to `True`, your `compute_metrics` function must take a
boolean `compute_result` argument. This will be triggered after the last eval batch to signal that the
function needs to calculate and return the global summary statistics rather than accumulating the
batch-level statistics.
callbacks (list of [`~transformers.TrainerCallback`], *optional*):
List of callbacks to customize the training loop. Will add those to the list of default callbacks detailed
in [here](https://huggingface.co/docs/transformers/main_classes/callback).
If you want to remove one of the default callbacks used, use the [`~transformers.Trainer.remove_callback`]
method.
optimizers (`tuple[torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None]`, *optional*, defaults to `(None, None)`):
A tuple containing the optimizer and the scheduler to use. Will default to an instance of `AdamW` on your
model and a scheduler given by [`~transformers.get_linear_schedule_with_warmup`] controlled by `args`.
optimizer_cls_and_kwargs (`tuple[Type[torch.optim.Optimizer], Dict[str, Any]]`, *optional*):
A tuple containing the optimizer class and keyword arguments to use. Overrides `optim` and `optim_args` in
`args`. Incompatible with the `optimizers` argument.
Unlike `optimizers`, this argument avoids the need to place model parameters on the correct devices before
initializing the Trainer.
preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`, *optional*):
A function that preprocess the logits right before caching them at each evaluation step. Must take two
tensors, the logits and the labels, and return the logits once processed as desired. The modifications made
by this function will be reflected in the predictions received by `compute_metrics`.
Note that the labels (second parameter) will be `None` if the dataset does not have them.
peft_config ([`~peft.PeftConfig`], *optional*):
PEFT configuration used to wrap the model. If `None`, the model is not wrapped. Note that if the loaded
model is a causal LM, it's highly recommended to set `modules_to_save=["score"]` in the PEFT configuration
to ensure that the reward head is properly trained.
"""
def __init__(
self,
model,
args = None,
data_collator = None,
train_dataset = None,
eval_dataset = None,
processing_class = None,
compute_metrics = None,
callbacks = None,
optimizer_cls_and_kwargs = None,
preprocess_logits_for_metrics = None,
peft_config = None,
**kwargs
):
if args is None: args = UnslothRewardConfig()
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),
)
elif isinstance(data_collator, TransformersDataCollatorForLanguageModeling):
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('reward_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,
compute_metrics = compute_metrics,
callbacks = callbacks,
optimizer_cls_and_kwargs = optimizer_cls_and_kwargs,
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
if hasattr(logger, "addFilter"):
import logging
class HideLoggingMessage(logging.Filter):
def __init__(self, text): self.text = text
def filter(self, x): return not (self.text in x.getMessage())
pass
logger.addFilter(HideLoggingMessage("`use_cache=True`"))