ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

support ORPO

Browse Source 
Former-commit-id: f44a4c27e2461cdaa1b16865f597a31033c0e6d9
This commit is contained in:
hiyouga
2024-03-31 18:29:50 +08:00
parent 526111a303
commit d764cd8736
22 changed files with 395 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
src/llmtuner/train/dpo/trainer.py
View File

@@ -74,7 +74,7 @@ class CustomDPOTrainer(DPOTrainer):
create_custom_scheduler(self.args, num_training_steps, optimizer)
return super().create_scheduler(num_training_steps, optimizer)
def sft_loss(self, chosen_logits: torch.FloatTensor, chosen_labels: torch.LongTensor) -> torch.Tensor:
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.
@@ -85,8 +85,8 @@ class CustomDPOTrainer(DPOTrainer):
return -all_logps
def concatenated_forward(
self, model: "PreTrainedModel", batch: Dict[str, torch.Tensor]
) -> Tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
batch_copied = BatchEncoding({k: v.detach().clone() for k, v in batch.items()}) # avoid error
all_logits = model(
@@ -107,9 +107,9 @@ class CustomDPOTrainer(DPOTrainer):
def get_batch_loss_metrics(
self,
model: "PreTrainedModel",
batch: Dict[str, torch.Tensor],
batch: Dict[str, "torch.Tensor"],
train_eval: Literal["train", "eval"] = "train",
) -> Tuple[torch.Tensor, 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.
"""
@@ -142,21 +142,22 @@ class CustomDPOTrainer(DPOTrainer):
reference_chosen_logps,
reference_rejected_logps,
)
batch_loss = losses.mean()
if self.ftx_gamma > 1e-6:
batch_size = batch["input_ids"].size(0) // 2
chosen_labels, _ = batch["labels"].split(batch_size, dim=0)
losses += self.ftx_gamma * self.sft_loss(policy_chosen_logits, chosen_labels)
batch_loss += self.ftx_gamma * self.sft_loss(policy_chosen_logits, chosen_labels).mean()
reward_accuracies = (chosen_rewards > rejected_rewards).float()
prefix = "eval_" if train_eval == "eval" else ""
metrics[f"{prefix}rewards/chosen"] = chosen_rewards.cpu().mean()
metrics[f"{prefix}rewards/rejected"] = rejected_rewards.cpu().mean()
metrics[f"{prefix}rewards/accuracies"] = reward_accuracies.cpu().mean()
metrics[f"{prefix}rewards/margins"] = (chosen_rewards - rejected_rewards).cpu().mean()
metrics[f"{prefix}logps/rejected"] = policy_rejected_logps.detach().cpu().mean()
metrics[f"{prefix}logps/chosen"] = policy_chosen_logps.detach().cpu().mean()
metrics[f"{prefix}logits/rejected"] = policy_rejected_logits.detach().cpu().mean()
metrics[f"{prefix}logits/chosen"] = policy_chosen_logits.detach().cpu().mean()
metrics["{}rewards/chosen".format(prefix)] = chosen_rewards.cpu().mean()
metrics["{}rewards/rejected".format(prefix)] = rejected_rewards.cpu().mean()
metrics["{}rewards/accuracies".format(prefix)] = reward_accuracies.cpu().mean()
metrics["{}rewards/margins".format(prefix)] = (chosen_rewards - rejected_rewards).cpu().mean()
metrics["{}logps/rejected".format(prefix)] = policy_rejected_logps.detach().cpu().mean()
metrics["{}logps/chosen".format(prefix)] = policy_chosen_logps.detach().cpu().mean()
metrics["{}logits/rejected".format(prefix)] = policy_rejected_logits.detach().cpu().mean()
metrics["{}logits/chosen".format(prefix)] = policy_chosen_logits.detach().cpu().mean()
return losses.mean(), metrics
return batch_loss, metrics

7
src/llmtuner/train/dpo/workflow.py
View File

@@ -2,13 +2,12 @@
from typing import TYPE_CHECKING, List, Optional
from ...data import get_dataset, split_dataset
from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
from ...extras.constants import IGNORE_INDEX
from ...extras.ploting import plot_loss
from ...hparams import ModelArguments
from ...model import load_model, load_tokenizer
from ..utils import create_modelcard_and_push, create_ref_model
from .collator import DPODataCollatorWithPadding
from .trainer import CustomDPOTrainer
@@ -29,7 +28,7 @@ def run_dpo(
dataset = get_dataset(tokenizer, model_args, data_args, training_args, stage="rm")
model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
data_collator = DPODataCollatorWithPadding(
data_collator = PairwiseDataCollatorWithPadding(
tokenizer=tokenizer,
pad_to_multiple_of=8,
label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
@@ -64,7 +63,7 @@ def run_dpo(
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
if trainer.is_world_process_zero() and finetuning_args.plot_loss:
plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "accuracy"])
# Evaluation
if training_args.do_eval:

4
src/llmtuner/train/orpo/__init__.py Normal file
View File

@@ -0,0 +1,4 @@
from .workflow import run_orpo
__all__ = ["run_orpo"]

150
src/llmtuner/train/orpo/trainer.py Normal file
View File

@@ -0,0 +1,150 @@
from collections import defaultdict
from typing import TYPE_CHECKING, Dict, Literal, Optional, Tuple, Union
import torch
import torch.nn.functional as F
from transformers import Trainer
from trl import DPOTrainer
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
from ...hparams import FinetuningArguments
class CustomORPOTrainer(DPOTrainer):
def __init__(
self,
model: Union["PreTrainedModel", "torch.nn.Module"],
finetuning_args: "FinetuningArguments",
disable_dropout: bool = True,
**kwargs,
):
if disable_dropout:
disable_dropout_in_model(model)
self.finetuning_args = finetuning_args
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.beta = finetuning_args.orpo_beta
self._stored_metrics = defaultdict(lambda: defaultdict(list))
Trainer.__init__(self, model=model, **kwargs)
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 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
# Borrowed from:
# https://github.com/huggingface/trl/blob/0ee349dcd43b0f4b3169449f16751c38ac4a609f/trl/trainer/orpo_trainer.py#L592
def odds_ratio_loss(
self, chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor"
) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
r"""
Computes ORPO's odds ratio (OR) loss.
Args:
policy_chosen_logps: Log probabilities of the policy model for the chosen responses. Shape: (batch_size,)
policy_rejected_logps: Log probabilities of the policy model for the rejected responses. Shape: (batch_size,)
Returns:
A tuple of five tensors: (losses, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen).
"""
# Derived from Eqs. (4) and (7) from https://arxiv.org/abs/2403.07691 by using log identities and exp(log(P(y|x)) = P(y|x)
log_odds = (chosen_logps - rejected_logps) - (
torch.log(1 - torch.exp(chosen_logps)) - torch.log(1 - torch.exp(rejected_logps))
)
ratio = F.logsigmoid(log_odds)
losses = self.beta * ratio
chosen_rewards = self.beta * chosen_logps.detach()
rejected_rewards = self.beta * rejected_logps.detach()
return losses, chosen_rewards, rejected_rewards, ratio, log_odds
def concatenated_forward(
self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
all_logits = model(
input_ids=batch["input_ids"], attention_mask=batch["attention_mask"], return_dict=True
).logits.to(torch.float32)
all_logps = self.get_batch_logps(
all_logits,
batch["labels"],
average_log_prob=False,
label_pad_token_id=self.label_pad_token_id,
)
batch_size = batch["input_ids"].size(0) // 2
chosen_logps, rejected_logps = all_logps.split(batch_size, dim=0)
chosen_logits, rejected_logits = all_logits.split(batch_size, dim=0)
return chosen_logps, rejected_logps, chosen_logits, rejected_logits
def get_batch_loss_metrics(
self,
model: "PreTrainedModel",
batch: Dict[str, "torch.Tensor"],
train_eval: Literal["train", "eval"] = "train",
) -> Tuple["torch.Tensor", Dict[str, "torch.Tensor"]]:
r"""
Computes the ORPO loss and other metrics for the given batch of inputs for train or test.
"""
metrics = {}
chosen_logps, rejected_logps, chosen_logits, rejected_logits = self.concatenated_forward(model, batch)
losses, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen = self.odds_ratio_loss(
chosen_logps, rejected_logps
)
batch_size = batch["input_ids"].size(0) // 2
chosen_labels, _ = batch["labels"].split(batch_size, dim=0)
sft_loss = self.sft_loss(chosen_logits, chosen_labels)
batch_loss = (sft_loss - losses).mean()
reward_accuracies = (chosen_rewards > rejected_rewards).float()
prefix = "eval_" if train_eval == "eval" else ""
metrics["{}rewards/chosen".format(prefix)] = chosen_rewards.cpu().mean()
metrics["{}rewards/rejected".format(prefix)] = rejected_rewards.cpu().mean()
metrics["{}rewards/accuracies".format(prefix)] = reward_accuracies.cpu().mean()
metrics["{}rewards/margins".format(prefix)] = (chosen_rewards - rejected_rewards).cpu().mean()
metrics["{}logps/rejected".format(prefix)] = rejected_logps.detach().cpu().mean()
metrics["{}logps/chosen".format(prefix)] = chosen_logps.detach().cpu().mean()
metrics["{}logits/rejected".format(prefix)] = rejected_logits.detach().cpu().mean()
metrics["{}logits/chosen".format(prefix)] = chosen_logits.detach().cpu().mean()
metrics["{}sft_loss".format(prefix)] = sft_loss.detach().cpu().mean()
metrics["{}log_odds_ratio".format(prefix)] = log_odds_ratio.detach().cpu().mean()
metrics["{}log_odds_chosen".format(prefix)] = log_odds_chosen.detach().cpu().mean()
return batch_loss, metrics

68
src/llmtuner/train/orpo/workflow.py Normal file
View File

@@ -0,0 +1,68 @@
# Inspired by: https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama_2/scripts/dpo_llama2.py
from typing import TYPE_CHECKING, List, Optional
from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
from ...extras.constants import IGNORE_INDEX
from ...extras.ploting import plot_loss
from ...hparams import ModelArguments
from ...model import load_model, load_tokenizer
from ..utils import create_modelcard_and_push
from .trainer import CustomORPOTrainer
if TYPE_CHECKING:
from transformers import Seq2SeqTrainingArguments, TrainerCallback
from ...hparams import DataArguments, FinetuningArguments
def run_orpo(
model_args: "ModelArguments",
data_args: "DataArguments",
training_args: "Seq2SeqTrainingArguments",
finetuning_args: "FinetuningArguments",
callbacks: Optional[List["TrainerCallback"]] = None,
):
tokenizer = load_tokenizer(model_args)
dataset = get_dataset(tokenizer, model_args, data_args, training_args, stage="rm")
model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
data_collator = PairwiseDataCollatorWithPadding(
tokenizer=tokenizer,
pad_to_multiple_of=8,
label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
)
# Update arguments
training_args.remove_unused_columns = False # important for pairwise dataset
# Initialize our Trainer
trainer = CustomORPOTrainer(
model=model,
args=training_args,
finetuning_args=finetuning_args,
tokenizer=tokenizer,
data_collator=data_collator,
callbacks=callbacks,
**split_dataset(dataset, data_args, training_args),
)
# Training
if training_args.do_train:
train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
trainer.save_model()
trainer.log_metrics("train", train_result.metrics)
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
if trainer.is_world_process_zero() and finetuning_args.plot_loss:
plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "accuracy"])
# Evaluation
if training_args.do_eval:
metrics = trainer.evaluate(metric_key_prefix="eval")
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Create model card
create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)

3
src/llmtuner/train/rm/workflow.py
View File

@@ -2,13 +2,12 @@
from typing import TYPE_CHECKING, List, Optional
from ...data import get_dataset, split_dataset
from ...data import PairwiseDataCollatorWithPadding, get_dataset, split_dataset
from ...extras.callbacks import FixValueHeadModelCallback
from ...extras.misc import fix_valuehead_checkpoint
from ...extras.ploting import plot_loss
from ...model import load_model, load_tokenizer
from ..utils import create_modelcard_and_push
from .collator import PairwiseDataCollatorWithPadding
from .metric import compute_accuracy
from .trainer import PairwiseTrainer

3
src/llmtuner/train/tuner.py
View File

@@ -9,6 +9,7 @@ from ..extras.logging import get_logger
from ..hparams import get_infer_args, get_train_args
from ..model import load_model_and_tokenizer
from .dpo import run_dpo
from .orpo import run_orpo
from .ppo import run_ppo
from .pt import run_pt
from .rm import run_rm
@@ -36,6 +37,8 @@ def run_exp(args: Optional[Dict[str, Any]] = None, callbacks: Optional[List["Tra
run_ppo(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
elif finetuning_args.stage == "dpo":
run_dpo(model_args, data_args, training_args, finetuning_args, callbacks)
elif finetuning_args.stage == "orpo":
run_orpo(model_args, data_args, training_args, finetuning_args, callbacks)
else:
raise ValueError("Unknown task.")