mirror of
https://github.com/hiyouga/LLaMA-Factory.git
synced 2025-10-14 15:52:49 +08:00
218 lines
8.6 KiB
Python
218 lines
8.6 KiB
Python
from collections import defaultdict
|
|
from contextlib import nullcontext
|
|
from types import MethodType
|
|
from typing import TYPE_CHECKING, Dict, Optional, Tuple, Union
|
|
|
|
import torch
|
|
from transformers import Trainer
|
|
from trl import KTOTrainer
|
|
from trl.trainer.utils import disable_dropout_in_model
|
|
|
|
from ...extras.constants import IGNORE_INDEX
|
|
from ..utils import create_custom_optimzer, create_custom_scheduler
|
|
|
|
|
|
if TYPE_CHECKING:
|
|
from transformers import PreTrainedModel, ProcessorMixin
|
|
|
|
from ...hparams import FinetuningArguments
|
|
|
|
|
|
class CustomKTOTrainer(KTOTrainer):
|
|
def __init__(
|
|
self,
|
|
model: Union["PreTrainedModel", torch.nn.Module],
|
|
ref_model: Optional[Union["PreTrainedModel", torch.nn.Module]],
|
|
finetuning_args: "FinetuningArguments",
|
|
processor: Optional["ProcessorMixin"],
|
|
disable_dropout: bool = True,
|
|
**kwargs,
|
|
):
|
|
if disable_dropout:
|
|
disable_dropout_in_model(model)
|
|
if ref_model is not None:
|
|
disable_dropout_in_model(ref_model)
|
|
|
|
self.finetuning_args = finetuning_args
|
|
self.processor = processor
|
|
self.reference_free = False
|
|
self.use_dpo_data_collator = True # hack to avoid warning
|
|
self.generate_during_eval = False # disable at evaluation
|
|
self.label_pad_token_id = IGNORE_INDEX
|
|
self.padding_value = 0
|
|
self.is_encoder_decoder = model.config.is_encoder_decoder
|
|
self.precompute_ref_log_probs = False
|
|
self._precomputed_train_ref_log_probs = False
|
|
self._precomputed_eval_ref_log_probs = False
|
|
self._peft_has_been_casted_to_bf16 = False
|
|
|
|
self.ref_model = ref_model
|
|
self._stored_metrics = defaultdict(lambda: defaultdict(list))
|
|
|
|
# kto hyperparams
|
|
self.beta = finetuning_args.kto_beta
|
|
self.desirable_weight = finetuning_args.kto_chosen_weight
|
|
self.undesirable_weight = finetuning_args.kto_rejected_weight
|
|
self.ftx_gamma = finetuning_args.kto_ftx
|
|
|
|
Trainer.__init__(self, model=model, **kwargs)
|
|
if not hasattr(self, "accelerator"):
|
|
raise AttributeError("Please update `transformers`.")
|
|
|
|
if ref_model is not None:
|
|
if self.is_deepspeed_enabled:
|
|
if not (
|
|
getattr(ref_model, "is_loaded_in_8bit", False) or getattr(ref_model, "is_loaded_in_4bit", False)
|
|
): # quantized models are already set on the correct device
|
|
self.ref_model = self._prepare_deepspeed(self.ref_model)
|
|
else:
|
|
self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
|
|
|
|
if finetuning_args.use_badam:
|
|
from badam import clip_grad_norm_for_sparse_tensor
|
|
|
|
self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_for_sparse_tensor, self.accelerator)
|
|
|
|
def create_optimizer(self) -> "torch.optim.Optimizer":
|
|
if self.optimizer is None:
|
|
self.optimizer = create_custom_optimzer(self.model, self.args, self.finetuning_args)
|
|
return super().create_optimizer()
|
|
|
|
def create_scheduler(
|
|
self, num_training_steps: int, optimizer: Optional["torch.optim.Optimizer"] = None
|
|
) -> "torch.optim.lr_scheduler.LRScheduler":
|
|
create_custom_scheduler(self.args, num_training_steps, optimizer)
|
|
return super().create_scheduler(num_training_steps, optimizer)
|
|
|
|
def _save(self, output_dir: Optional[str] = None, state_dict: Optional[Dict[str, "torch.Tensor"]] = None) -> None:
|
|
super()._save(output_dir, state_dict)
|
|
if self.processor is not None:
|
|
output_dir = output_dir if output_dir is not None else self.args.output_dir
|
|
getattr(self.processor, "image_processor").save_pretrained(output_dir)
|
|
|
|
def sft_loss(self, chosen_logits: "torch.FloatTensor", chosen_labels: "torch.LongTensor") -> "torch.Tensor":
|
|
r"""
|
|
Computes supervised cross-entropy loss of given labels under the given logits.
|
|
|
|
Returns:
|
|
A tensor of shape (batch_size,) containing the cross-entropy loss of each samples.
|
|
"""
|
|
all_logps = self.get_batch_logps(chosen_logits, chosen_labels, average_log_prob=True)
|
|
return -all_logps
|
|
|
|
def forward(
|
|
self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
|
|
) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
|
|
with torch.no_grad():
|
|
kl_logits = model(
|
|
input_ids=batch["kl_input_ids"],
|
|
attention_mask=batch["kl_attention_mask"],
|
|
return_dict=True,
|
|
use_cache=False,
|
|
).logits.to(torch.float32)
|
|
|
|
target_logits = model(
|
|
input_ids=batch["input_ids"],
|
|
attention_mask=batch["attention_mask"],
|
|
return_dict=True,
|
|
use_cache=False,
|
|
).logits.to(torch.float32)
|
|
|
|
target_logps = self.get_batch_logps(
|
|
logits=target_logits,
|
|
labels=batch["labels"],
|
|
average_log_prob=False,
|
|
is_encoder_decoder=self.is_encoder_decoder,
|
|
label_pad_token_id=self.label_pad_token_id,
|
|
)
|
|
|
|
kl_logps = self.get_batch_logps(
|
|
logits=kl_logits,
|
|
labels=batch["kl_labels"],
|
|
average_log_prob=False,
|
|
is_encoder_decoder=self.is_encoder_decoder,
|
|
label_pad_token_id=self.label_pad_token_id,
|
|
)
|
|
|
|
if len(target_logps) != len(batch["kto_tags"]):
|
|
raise ValueError("Mismatched shape of inputs and labels.")
|
|
|
|
chosen_idx = [i for i in range(len(target_logps)) if batch["kto_tags"][i]]
|
|
rejected_idx = [i for i in range(len(target_logps)) if not batch["kto_tags"][i]]
|
|
|
|
chosen_logps = target_logps[chosen_idx, ...]
|
|
rejected_logps = target_logps[rejected_idx, ...]
|
|
|
|
chosen_logits = target_logits[chosen_idx, ...]
|
|
rejected_logits = target_logits[rejected_idx, ...]
|
|
|
|
return chosen_logps, rejected_logps, chosen_logits, rejected_logits, kl_logps
|
|
|
|
def get_batch_loss_metrics(
|
|
self,
|
|
model: "PreTrainedModel",
|
|
batch: Dict[str, "torch.Tensor"],
|
|
) -> Tuple["torch.Tensor", Dict[str, "torch.Tensor"]]:
|
|
r"""
|
|
Computes the DPO loss and other metrics for the given batch of inputs for train or test.
|
|
"""
|
|
metrics = {}
|
|
(
|
|
policy_chosen_logps,
|
|
policy_rejected_logps,
|
|
policy_chosen_logits,
|
|
_,
|
|
policy_kl_logps,
|
|
) = self.forward(model, batch)
|
|
|
|
with torch.no_grad():
|
|
if self.ref_model is None:
|
|
ref_model = self.model
|
|
ref_context = self.accelerator.unwrap_model(self.model).disable_adapter()
|
|
else:
|
|
ref_model = self.ref_model
|
|
ref_context = nullcontext()
|
|
|
|
with ref_context:
|
|
(
|
|
reference_chosen_logps,
|
|
reference_rejected_logps,
|
|
_,
|
|
_,
|
|
reference_kl_logps,
|
|
) = self.forward(ref_model, batch)
|
|
|
|
losses, chosen_rewards, rejected_rewards, kl = self.kto_loss(
|
|
policy_chosen_logps,
|
|
policy_rejected_logps,
|
|
policy_kl_logps,
|
|
reference_chosen_logps,
|
|
reference_rejected_logps,
|
|
reference_kl_logps,
|
|
)
|
|
losses = losses.nanmean()
|
|
|
|
if self.ftx_gamma > 1e-6 and len(policy_chosen_logps) > 0: # remember to rescale
|
|
sft_loss = self.sft_loss(policy_chosen_logits, batch["labels"][batch["kto_tags"]])
|
|
losses += self.ftx_gamma * sft_loss.nanmean() / len(policy_chosen_logits) * len(batch["labels"])
|
|
|
|
num_chosen = torch.Tensor([len(chosen_rewards)]).to(self.accelerator.device)
|
|
num_rejected = torch.Tensor([len(rejected_rewards)]).to(self.accelerator.device)
|
|
|
|
all_num_chosen = self.accelerator.gather(num_chosen).sum().item()
|
|
all_num_rejected = self.accelerator.gather(num_rejected).sum().item()
|
|
|
|
if all_num_chosen > 0:
|
|
metrics["rewards/chosen_sum"] = self.accelerator.gather(chosen_rewards.nansum()).nansum().item()
|
|
metrics["logps/chosen_sum"] = self.accelerator.gather(policy_chosen_logps.nansum()).nansum().item()
|
|
metrics["count/chosen"] = all_num_chosen
|
|
|
|
if all_num_rejected > 0:
|
|
metrics["rewards/rejected_sum"] = self.accelerator.gather(rejected_rewards.nansum()).nansum().item()
|
|
metrics["logps/rejected_sum"] = self.accelerator.gather(policy_rejected_logps.nansum()).nansum().item()
|
|
metrics["count/rejected"] = all_num_rejected
|
|
|
|
metrics["kl"] = kl.item()
|
|
|
|
return losses, metrics
|