793 lines
37 KiB
Python
793 lines
37 KiB
Python
"""
|
|
2025.6.1
|
|
2025.6.2
|
|
4.52.4
|
|
0.18.2
|
|
__UNSLOTH_VERSIONING__
|
|
"""
|
|
from torch import Tensor
|
|
import torch
|
|
import torch.nn as nn
|
|
from torch.nn import functional as F
|
|
from trl.trainer.prm_trainer import (BaseImageProcessor, Callable, DataCollator, DataCollatorForTokenClassification, Dataset, EvalPrediction, FeatureExtractionMixin, Optional, PRMConfig, PRMTrainer, PartialState, PeftModel, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, Trainer, TrainerCallback, Union, chain, compute_accuracy, disable_dropout_in_model, features, generate_model_card, inspect, is_peft_available, is_wandb_available, nn, os, prepare_model_for_kbit_training, textwrap, torch, warnings)
|
|
|
|
|
|
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
|
|
from transformers import DataCollatorForSeq2Seq, DataCollatorForLanguageModeling as TransformersDataCollatorForLanguageModeling
|
|
|
|
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 selective_log_softmax(logits, index):
|
|
logits = logits.to(torch.float32)
|
|
selected_logits = torch.gather(logits, dim = -1, index = index.unsqueeze(-1)).squeeze(-1)
|
|
# loop to reduce peak mem consumption
|
|
# logsumexp_values = torch.stack([torch.logsumexp(lg, dim=-1) for lg in logits])
|
|
logsumexp_values = torch.logsumexp(logits, dim = -1)
|
|
per_token_logps = selected_logits - logsumexp_values # log_softmax(x_i) = x_i - logsumexp(x)
|
|
return per_token_logps
|
|
@dataclass
|
|
class UnslothPRMConfig(PRMConfig):
|
|
"""
|
|
|
|
Configuration class for the [`PRMTrainer`].
|
|
|
|
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:
|
|
learning_rate (`float`, *optional*, defaults to `1e-5`):
|
|
Initial learning rate for [`AdamW`] optimizer. The default value replaces that of
|
|
[`~transformers.TrainingArguments`].
|
|
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` or `None`, *optional*, defaults to `None`):
|
|
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*, defaults to `None`):
|
|
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.'},
|
|
)
|
|
def __init__(
|
|
self,
|
|
output_dir = None,
|
|
overwrite_output_dir = None,
|
|
do_train = False,
|
|
do_eval = False,
|
|
do_predict = False,
|
|
eval_strategy = 'no',
|
|
prediction_loss_only = False,
|
|
per_device_train_batch_size = 4,
|
|
per_device_eval_batch_size = 4,
|
|
per_gpu_train_batch_size = None,
|
|
per_gpu_eval_batch_size = None,
|
|
gradient_accumulation_steps = 2,
|
|
eval_accumulation_steps = 2,
|
|
eval_delay = 0,
|
|
torch_empty_cache_steps = 250,
|
|
learning_rate = 5e-05,
|
|
weight_decay = 0.01,
|
|
adam_beta1 = 0.9,
|
|
adam_beta2 = 0.999,
|
|
adam_epsilon = 1e-08,
|
|
max_grad_norm = 1.0,
|
|
num_train_epochs = 3.0,
|
|
max_steps = -1,
|
|
lr_scheduler_type = 'linear',
|
|
warmup_ratio = 0.1,
|
|
warmup_steps = 0,
|
|
log_level = 'passive',
|
|
log_level_replica = 'warning',
|
|
log_on_each_node = True,
|
|
logging_dir = None,
|
|
logging_strategy = 'steps',
|
|
logging_first_step = False,
|
|
logging_steps = 1,
|
|
logging_nan_inf_filter = False,
|
|
save_strategy = 'steps',
|
|
save_steps = 500,
|
|
save_total_limit = None,
|
|
save_safetensors = True,
|
|
save_on_each_node = False,
|
|
save_only_model = False,
|
|
restore_callback_states_from_checkpoint = False,
|
|
no_cuda = False,
|
|
use_cpu = False,
|
|
use_mps_device = False,
|
|
seed = 3407,
|
|
data_seed = 3407,
|
|
jit_mode_eval = False,
|
|
use_ipex = False,
|
|
bf16 = False,
|
|
fp16 = False,
|
|
fp16_opt_level = 'O1',
|
|
half_precision_backend = 'auto',
|
|
bf16_full_eval = False,
|
|
fp16_full_eval = False,
|
|
tf32 = None,
|
|
local_rank = -1,
|
|
ddp_backend = None,
|
|
tpu_num_cores = None,
|
|
tpu_metrics_debug = False,
|
|
debug = '',
|
|
dataloader_drop_last = False,
|
|
eval_steps = None,
|
|
dataloader_num_workers = 0,
|
|
dataloader_prefetch_factor = None,
|
|
past_index = -1,
|
|
run_name = None,
|
|
disable_tqdm = None,
|
|
remove_unused_columns = True,
|
|
label_names = None,
|
|
load_best_model_at_end = False,
|
|
metric_for_best_model = None,
|
|
greater_is_better = None,
|
|
ignore_data_skip = False,
|
|
fsdp = '',
|
|
fsdp_min_num_params = 0,
|
|
fsdp_config = None,
|
|
fsdp_transformer_layer_cls_to_wrap = None,
|
|
accelerator_config = None,
|
|
deepspeed = None,
|
|
label_smoothing_factor = 0.0,
|
|
optim = 'adamw_8bit',
|
|
optim_args = None,
|
|
adafactor = False,
|
|
group_by_length = False,
|
|
length_column_name = 'length',
|
|
report_to = None,
|
|
ddp_find_unused_parameters = None,
|
|
ddp_bucket_cap_mb = None,
|
|
ddp_broadcast_buffers = None,
|
|
dataloader_pin_memory = True,
|
|
dataloader_persistent_workers = False,
|
|
skip_memory_metrics = True,
|
|
use_legacy_prediction_loop = False,
|
|
push_to_hub = False,
|
|
resume_from_checkpoint = None,
|
|
hub_model_id = None,
|
|
hub_strategy = 'every_save',
|
|
hub_token = None,
|
|
hub_private_repo = None,
|
|
hub_always_push = False,
|
|
gradient_checkpointing = False,
|
|
gradient_checkpointing_kwargs = None,
|
|
include_inputs_for_metrics = False,
|
|
eval_do_concat_batches = True,
|
|
fp16_backend = 'auto',
|
|
push_to_hub_model_id = None,
|
|
push_to_hub_organization = None,
|
|
push_to_hub_token = None,
|
|
mp_parameters = '',
|
|
auto_find_batch_size = False,
|
|
full_determinism = False,
|
|
torchdynamo = None,
|
|
ray_scope = 'last',
|
|
ddp_timeout = 1800,
|
|
torch_compile = False,
|
|
torch_compile_backend = None,
|
|
torch_compile_mode = None,
|
|
include_tokens_per_second = False,
|
|
include_num_input_tokens_seen = False,
|
|
neftune_noise_alpha = None,
|
|
optim_target_modules = None,
|
|
batch_eval_metrics = False,
|
|
eval_on_start = False,
|
|
use_liger_kernel = False,
|
|
eval_use_gather_object = False,
|
|
average_tokens_across_devices = False,
|
|
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,
|
|
**kwargs,
|
|
):
|
|
if learning_rate < 1e-7: raise FloatingPointError(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: raise OverflowError(f'Unsloth: Your learning rate of `{learning_rate}` is way too larger > 1! Consider decreasing it to 1e-1, otherwise gradient updates will explode!')
|
|
if output_dir is None and save_strategy == 'steps' and save_steps == 500:
|
|
output_dir = 'unsloth_training_checkpoints'
|
|
save_strategy = 'no'
|
|
if dataset_num_proc is None:
|
|
from multiprocessing import cpu_count
|
|
dataset_num_proc = cpu_count()
|
|
|
|
super().__init__(
|
|
output_dir = output_dir,
|
|
overwrite_output_dir = overwrite_output_dir,
|
|
do_train = do_train,
|
|
do_eval = do_eval,
|
|
do_predict = do_predict,
|
|
eval_strategy = eval_strategy,
|
|
prediction_loss_only = prediction_loss_only,
|
|
per_device_train_batch_size = per_device_train_batch_size,
|
|
per_device_eval_batch_size = per_device_eval_batch_size,
|
|
per_gpu_train_batch_size = per_gpu_train_batch_size,
|
|
per_gpu_eval_batch_size = per_gpu_eval_batch_size,
|
|
gradient_accumulation_steps = gradient_accumulation_steps,
|
|
eval_accumulation_steps = eval_accumulation_steps,
|
|
eval_delay = eval_delay,
|
|
torch_empty_cache_steps = torch_empty_cache_steps,
|
|
learning_rate = learning_rate,
|
|
weight_decay = weight_decay,
|
|
adam_beta1 = adam_beta1,
|
|
adam_beta2 = adam_beta2,
|
|
adam_epsilon = adam_epsilon,
|
|
max_grad_norm = max_grad_norm,
|
|
num_train_epochs = num_train_epochs,
|
|
max_steps = max_steps,
|
|
lr_scheduler_type = lr_scheduler_type,
|
|
warmup_ratio = warmup_ratio,
|
|
warmup_steps = warmup_steps,
|
|
log_level = log_level,
|
|
log_level_replica = log_level_replica,
|
|
log_on_each_node = log_on_each_node,
|
|
logging_dir = logging_dir,
|
|
logging_strategy = logging_strategy,
|
|
logging_first_step = logging_first_step,
|
|
logging_steps = logging_steps,
|
|
logging_nan_inf_filter = logging_nan_inf_filter,
|
|
save_strategy = save_strategy,
|
|
save_steps = save_steps,
|
|
save_total_limit = save_total_limit,
|
|
save_safetensors = save_safetensors,
|
|
save_on_each_node = save_on_each_node,
|
|
save_only_model = save_only_model,
|
|
restore_callback_states_from_checkpoint = restore_callback_states_from_checkpoint,
|
|
no_cuda = no_cuda,
|
|
use_cpu = use_cpu,
|
|
use_mps_device = use_mps_device,
|
|
seed = seed,
|
|
data_seed = data_seed,
|
|
jit_mode_eval = jit_mode_eval,
|
|
use_ipex = use_ipex,
|
|
bf16 = bf16,
|
|
fp16 = fp16,
|
|
fp16_opt_level = fp16_opt_level,
|
|
half_precision_backend = half_precision_backend,
|
|
bf16_full_eval = bf16_full_eval,
|
|
fp16_full_eval = fp16_full_eval,
|
|
tf32 = tf32,
|
|
local_rank = local_rank,
|
|
ddp_backend = ddp_backend,
|
|
tpu_num_cores = tpu_num_cores,
|
|
tpu_metrics_debug = tpu_metrics_debug,
|
|
debug = debug,
|
|
dataloader_drop_last = dataloader_drop_last,
|
|
eval_steps = eval_steps,
|
|
dataloader_num_workers = dataloader_num_workers,
|
|
dataloader_prefetch_factor = dataloader_prefetch_factor,
|
|
past_index = past_index,
|
|
run_name = run_name,
|
|
disable_tqdm = disable_tqdm,
|
|
remove_unused_columns = remove_unused_columns,
|
|
label_names = label_names,
|
|
load_best_model_at_end = load_best_model_at_end,
|
|
metric_for_best_model = metric_for_best_model,
|
|
greater_is_better = greater_is_better,
|
|
ignore_data_skip = ignore_data_skip,
|
|
fsdp = fsdp,
|
|
fsdp_min_num_params = fsdp_min_num_params,
|
|
fsdp_config = fsdp_config,
|
|
fsdp_transformer_layer_cls_to_wrap = fsdp_transformer_layer_cls_to_wrap,
|
|
accelerator_config = accelerator_config,
|
|
deepspeed = deepspeed,
|
|
label_smoothing_factor = label_smoothing_factor,
|
|
optim = optim,
|
|
optim_args = optim_args,
|
|
adafactor = adafactor,
|
|
group_by_length = group_by_length,
|
|
length_column_name = length_column_name,
|
|
report_to = report_to,
|
|
ddp_find_unused_parameters = ddp_find_unused_parameters,
|
|
ddp_bucket_cap_mb = ddp_bucket_cap_mb,
|
|
ddp_broadcast_buffers = ddp_broadcast_buffers,
|
|
dataloader_pin_memory = dataloader_pin_memory,
|
|
dataloader_persistent_workers = dataloader_persistent_workers,
|
|
skip_memory_metrics = skip_memory_metrics,
|
|
use_legacy_prediction_loop = use_legacy_prediction_loop,
|
|
push_to_hub = push_to_hub,
|
|
resume_from_checkpoint = resume_from_checkpoint,
|
|
hub_model_id = hub_model_id,
|
|
hub_strategy = hub_strategy,
|
|
hub_token = hub_token,
|
|
hub_private_repo = hub_private_repo,
|
|
hub_always_push = hub_always_push,
|
|
gradient_checkpointing = gradient_checkpointing,
|
|
gradient_checkpointing_kwargs = gradient_checkpointing_kwargs,
|
|
include_inputs_for_metrics = include_inputs_for_metrics,
|
|
eval_do_concat_batches = eval_do_concat_batches,
|
|
fp16_backend = fp16_backend,
|
|
push_to_hub_model_id = push_to_hub_model_id,
|
|
push_to_hub_organization = push_to_hub_organization,
|
|
push_to_hub_token = push_to_hub_token,
|
|
mp_parameters = mp_parameters,
|
|
auto_find_batch_size = auto_find_batch_size,
|
|
full_determinism = full_determinism,
|
|
torchdynamo = torchdynamo,
|
|
ray_scope = ray_scope,
|
|
ddp_timeout = ddp_timeout,
|
|
torch_compile = torch_compile,
|
|
torch_compile_backend = torch_compile_backend,
|
|
torch_compile_mode = torch_compile_mode,
|
|
include_tokens_per_second = include_tokens_per_second,
|
|
include_num_input_tokens_seen = include_num_input_tokens_seen,
|
|
neftune_noise_alpha = neftune_noise_alpha,
|
|
optim_target_modules = optim_target_modules,
|
|
batch_eval_metrics = batch_eval_metrics,
|
|
eval_on_start = eval_on_start,
|
|
use_liger_kernel = use_liger_kernel,
|
|
eval_use_gather_object = eval_use_gather_object,
|
|
average_tokens_across_devices = average_tokens_across_devices,
|
|
max_length = max_length,
|
|
max_prompt_length = max_prompt_length,
|
|
max_completion_length = max_completion_length,
|
|
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
|
|
pass
|
|
|
|
class _UnslothPRMTrainer(Trainer):
|
|
""""""
|
|
|
|
_tag_names = ["trl", "prm"]
|
|
|
|
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 is_peft_available() and peft_config is not None:
|
|
raise ValueError(
|
|
"PEFT is not installed and you passed a `peft_config` in the trainer's kwargs, please install it to use the PEFT models"
|
|
)
|
|
elif is_peft_available() and peft_config is not None:
|
|
if not isinstance(model, PeftModel):
|
|
if getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_quantized", False):
|
|
_supports_gc_kwargs = "gradient_checkpointing_kwargs" in list(
|
|
inspect.signature(prepare_model_for_kbit_training).parameters
|
|
)
|
|
|
|
prepare_model_kwargs = {"use_gradient_checkpointing": args.gradient_checkpointing}
|
|
|
|
if not _supports_gc_kwargs and args.gradient_checkpointing_kwargs is not None:
|
|
warnings.warn(
|
|
"You passed `gradient_checkpointing_kwargs` in the trainer's kwargs, but your peft version does not support it. "
|
|
"please update to the latest version of peft to use `gradient_checkpointing_kwargs`."
|
|
)
|
|
elif _supports_gc_kwargs and args.gradient_checkpointing_kwargs is not None:
|
|
prepare_model_kwargs["gradient_checkpointing_kwargs"] = args.gradient_checkpointing_kwargs
|
|
|
|
model = prepare_model_for_kbit_training(model, **prepare_model_kwargs)
|
|
|
|
model = model
|
|
|
|
# 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 (`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}
|
|
|
|
def create_model_card(
|
|
self,
|
|
model_name: Optional[str] = None,
|
|
dataset_name: Optional[str] = None,
|
|
tags: Union[str, list[str], None] = None,
|
|
):
|
|
"""
|
|
Creates a draft of a model card using the information available to the `Trainer`.
|
|
|
|
Args:
|
|
model_name (`str` or `None`, *optional*, defaults to `None`):
|
|
Name of the model.
|
|
dataset_name (`str` or `None`, *optional*, defaults to `None`):
|
|
Name of the dataset used for training.
|
|
tags (`str`, `list[str]` or `None`, *optional*, defaults to `None`):
|
|
Tags to be associated with the model card.
|
|
"""
|
|
if not self.is_world_process_zero():
|
|
return
|
|
|
|
if hasattr(self.model.config, "_name_or_path") and not os.path.isdir(self.model.config._name_or_path):
|
|
base_model = self.model.config._name_or_path
|
|
else:
|
|
base_model = None
|
|
|
|
tags = tags or []
|
|
if isinstance(tags, str):
|
|
tags = [tags]
|
|
|
|
if hasattr(self.model.config, "unsloth_version"):
|
|
tags.append("unsloth")
|
|
|
|
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}
|
|
}""")
|
|
|
|
model_card = generate_model_card(
|
|
base_model=base_model,
|
|
model_name=model_name,
|
|
hub_model_id=self.hub_model_id,
|
|
dataset_name=dataset_name,
|
|
tags=tags,
|
|
wandb_url=wandb.run.get_url() if is_wandb_available() and wandb.run is not None else None,
|
|
trainer_name="PRM",
|
|
trainer_citation=citation,
|
|
paper_title="Solving math word problems with process-and outcome-based feedback",
|
|
)
|
|
|
|
model_card.save(os.path.join(self.args.output_dir, "README.md"))
|
|
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 (`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 (`PreTrainedTokenizerBase` or `BaseImageProcessor` or `FeatureExtractionMixin` or `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)
|
|
use_fp16 = getattr(args, 'fp16', False)
|
|
force_float32 = False
|
|
if 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, 'torch_dtype', None)
|
|
if dtype is None: dtype = model.get_input_embeddings().dtype
|
|
from unsloth_zoo.utils import _get_dtype
|
|
dtype = _get_dtype(dtype)
|
|
float16 = dtype == torch.float16
|
|
if not force_float32 and (float16 and use_bf16): raise TypeError('Unsloth: Model is in float16 precision but you want to use bfloat16 precision. Set fp16 to `True` and bf16 to `False`')
|
|
if not force_float32 and (not float16 and use_fp16): raise TypeError('Unsloth: Model is in bfloat16 precision but you want to use float16 precision. Set fp16 to `False` and bf16 to `True`')
|
|
if force_float32:
|
|
args.fp16 = False
|
|
args.bf16 = False
|
|
os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'
|
|
elif (not use_bf16 and not use_fp16) and mixed_precision_dtype == 'float32':
|
|
args.fp16 = float16
|
|
args.bf16 = not float16
|
|
os.environ['ACCELERATE_MIXED_PRECISION'] = 'fp16' if float16 else 'bf16'
|
|
if getattr(args, 'eval_dataset', None) is not None and getattr(args, 'eval_strategy', 'no') == 'no':
|
|
args.eval_strategy = 'steps'
|
|
if getattr(args, 'eval_steps', None) is None: args.eval_steps = 0.1
|
|
ga_steps = getattr(args, 'gradient_accumulation_steps', None)
|
|
if ga_steps is not None and ga_steps > 1:
|
|
from transformers import __version__ as transformers_version
|
|
if Version(transformers_version) <= Version('4.45.2'):
|
|
print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!\n'
|
|
'`pip install --upgrade --no-cache-dir --force-reinstall --no-deps unsloth transformers trl unsloth_zoo`')
|
|
if getattr(args, 'eval_strategy', 'no') != 'no':
|
|
eval_bsz = getattr(args, 'per_device_eval_batch_size', 8)
|
|
if eval_bsz == 8 and args.per_device_train_batch_size < eval_bsz: args.per_device_eval_batch_size = args.per_device_train_batch_size
|
|
if getattr(args, 'eval_accumulation_steps', None) is None and ga_steps is not None: args.eval_accumulation_steps = ga_steps
|
|
fp16_full_eval = getattr(args, 'fp16_full_eval', False)
|
|
bf16_full_eval = getattr(args, 'bf16_full_eval', False)
|
|
if args.fp16 and bf16_full_eval: args.bf16_full_eval = False; args.fp16_full_eval = True
|
|
if args.bf16 and fp16_full_eval: args.bf16_full_eval = True; args.fp16_full_eval = False
|
|
if force_float32:
|
|
args.bf16_full_eval = False
|
|
args.fp16_full_eval = False
|
|
elif os.environ.get('UNSLOTH_MIXED_PRECISION', 'float32') == 'bfloat16':
|
|
args.bf16_full_eval = True
|
|
args.fp16_full_eval = False
|
|
elif not bf16_full_eval and not fp16_full_eval:
|
|
args.bf16_full_eval = args.bf16
|
|
args.fp16_full_eval = args.fp16
|
|
_output_logits = False
|
|
if locals().get('compute_metrics', None) is not None: _output_logits = True
|
|
if locals().get('preprocess_logits_for_metrics', None) is not None: _output_logits = True
|
|
if _output_logits:
|
|
os.environ['UNSLOTH_RETURN_LOGITS'] = '1'
|
|
if 'max_seq_length' not in locals() and not hasattr(args, 'max_seq_length'):
|
|
pass
|
|
else:
|
|
model_max_seq_length = getattr(model, 'max_seq_length', None)
|
|
args_max_seq_length = getattr(args, 'max_seq_length', None)
|
|
if args_max_seq_length is None and model_max_seq_length is not None:
|
|
max_seq_length = model.max_seq_length
|
|
if hasattr(args, 'max_seq_length'): args.max_seq_length = max_seq_length
|
|
if model is not None and hasattr(model, 'for_training'):
|
|
model.for_training()
|
|
if 'tokenizer' in locals() and hasattr(tokenizer, 'padding_side'): tokenizer.padding_side = 'right'
|
|
if 'processing_class' in locals():
|
|
if hasattr(processing_class, 'padding_side'): processing_class.padding_side = 'right'
|
|
if hasattr(processing_class, 'tokenizer') and hasattr(processing_class.tokenizer, 'padding_side'): processing_class.tokenizer.padding_side = 'right'
|
|
__tokenizer = processing_class if 'processing_class' in locals() else tokenizer
|
|
from unsloth_zoo.vision_utils import UnslothVisionDataCollator
|
|
if not isinstance(data_collator, UnslothVisionDataCollator):
|
|
if isinstance(data_collator, DataCollatorForSeq2Seq) and 'labels' not in train_dataset.column_names:
|
|
data_collator = TransformersDataCollatorForLanguageModeling(__tokenizer, mlm = False, mlm_probability = 0.0)
|
|
elif isinstance(data_collator, TransformersDataCollatorForLanguageModeling) and 'labels' in train_dataset.column_names:
|
|
data_collator = DataCollatorForSeq2Seq(__tokenizer)
|
|
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)
|
|
else:
|
|
data_collator = TransformersDataCollatorForLanguageModeling(__tokenizer.tokenizer, mlm = False, mlm_probability = 0.0)
|
|
other_metrics = []
|
|
|
|
from unsloth_zoo.logging_utils import PatchRLStatistics
|
|
PatchRLStatistics('prm_trainer', other_metrics)
|
|
|
|
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 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
|
|
|
|
pass
|