Initial commit

LLaMA-Factory/src/llamafactory/train/dpo/trainer.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/trl/trainer/dpo_trainer.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from collections import defaultdict
+from contextlib import nullcontext
+from types import MethodType
+from typing import TYPE_CHECKING, Dict, List, Literal, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import Trainer
+from trl import DPOTrainer
+from trl.trainer import disable_dropout_in_model
+from typing_extensions import override
+
+from ...extras.constants import IGNORE_INDEX
+from ...extras.packages import is_transformers_version_equal_to_4_46
+from ..callbacks import SaveProcessorCallback
+from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, ProcessorMixin
+
+    from ...hparams import FinetuningArguments
+
+
+class CustomDPOTrainer(DPOTrainer):
+    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.f_divergence_type = "reverse_kl"
+        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))
+
+        # dpo hyperparams
+        self.beta = finetuning_args.pref_beta
+        self.loss_type = finetuning_args.pref_loss
+        self.ftx_gamma = finetuning_args.pref_ftx
+        self.label_smoothing = finetuning_args.dpo_label_smoothing
+        self.simpo_gamma = finetuning_args.simpo_gamma
+
+        Trainer.__init__(self, model=model, **kwargs)
+        if not hasattr(self, "accelerator"):
+            raise AttributeError("Please update `transformers`.")
+
+        warnings.simplefilter("ignore")  # remove gc warnings on ref model
+
+        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)
+                self.ref_model.eval()
+
+        if processor is not None:
+            self.add_callback(SaveProcessorCallback(processor))
+
+        if finetuning_args.use_badam:
+            from badam import BAdamCallback, clip_grad_norm_old_version  # type: ignore
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.add_callback(BAdamCallback)
+
+    @override
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimizer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    @override
+    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)
+
+    @override
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return super()._get_train_sampler()
+
+    @override
+    def get_batch_samples(self, epoch_iterator, num_batches):
+        r"""
+        Replaces the method of KTO Trainer with the one of the standard Trainer.
+        """
+        return Trainer.get_batch_samples(self, epoch_iterator, num_batches)
+
+    def odds_ratio_loss(self, chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor") -> "torch.Tensor":
+        r"""
+        Computes ORPO's odds ratio (OR) loss for batched log probabilities of the policy model.
+        """
+        log_odds = (chosen_logps - rejected_logps) - (
+            torch.log1p(-torch.exp(chosen_logps)) - torch.log1p(-torch.exp(rejected_logps))
+        )
+        sft_loss = -chosen_logps
+        odds_ratio_loss = -F.logsigmoid(log_odds)
+        orpo_loss = sft_loss + self.beta * odds_ratio_loss
+        return orpo_loss
+
+    def simpo_loss(self, chosen_logps: "torch.Tensor", rejected_logps: "torch.Tensor") -> "torch.Tensor":
+        r"""
+        Computes SimPO loss for batched log probabilities of the policy model.
+        """
+        pi_logratios = chosen_logps - rejected_logps
+        gamma_logratios = self.simpo_gamma / self.beta
+        logits = pi_logratios - gamma_logratios
+        simpo_loss = -F.logsigmoid(self.beta * logits)
+        return simpo_loss
+
+    def compute_preference_loss(
+        self,
+        policy_chosen_logps: "torch.Tensor",
+        policy_rejected_logps: "torch.Tensor",
+        reference_chosen_logps: Optional["torch.Tensor"],
+        reference_rejected_logps: Optional["torch.Tensor"],
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes loss for preference learning.
+        """
+        if not self.finetuning_args.use_ref_model:
+            if self.loss_type == "orpo":
+                losses = self.odds_ratio_loss(policy_chosen_logps, policy_rejected_logps)
+            elif self.loss_type == "simpo":
+                losses = self.simpo_loss(policy_chosen_logps, policy_rejected_logps)
+            else:
+                raise NotImplementedError(f"Unknown loss type: {self.loss_type}.")
+
+            chosen_rewards = self.beta * policy_chosen_logps.to(self.accelerator.device).detach()
+            rejected_rewards = self.beta * policy_rejected_logps.to(self.accelerator.device).detach()
+        else:
+            losses, chosen_rewards, rejected_rewards = self.dpo_loss(
+                policy_chosen_logps, policy_rejected_logps, reference_chosen_logps, reference_rejected_logps
+            )
+
+        return losses, chosen_rewards, rejected_rewards
+
+    @override
+    def concatenated_forward(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes the sum log probabilities of the labels under given logits if loss_type is not IPO, ORPO or SimPO.
+
+        Otherwise the average log probabilities.
+        """
+        if self.finetuning_args.use_ref_model:
+            batch = {k: v.detach().clone() for k, v in batch.items()}  # avoid error
+
+        all_logits: "torch.Tensor" = model(**batch, return_dict=True, use_cache=False).logits.to(torch.float32)
+        all_logps, valid_length = get_batch_logps(logits=all_logits, labels=batch["labels"])
+        if self.loss_type in ["ipo", "orpo", "simpo"]:
+            all_logps = all_logps / valid_length
+
+        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)
+        chosen_length, _ = valid_length.split(batch_size, dim=0)
+        return chosen_logps, rejected_logps, chosen_logits, rejected_logits, chosen_logps / chosen_length
+
+    @override
+    def compute_reference_log_probs(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple[Optional["torch.Tensor"], Optional["torch.Tensor"]]:
+        r"""
+        Computes log probabilities of the reference model.
+        """
+        if not self.finetuning_args.use_ref_model:
+            return None, None
+
+        if self.ref_model is None:
+            ref_model = model
+            ref_context = self.accelerator.unwrap_model(model).disable_adapter()
+        else:
+            ref_model = self.ref_model
+            ref_context = nullcontext()
+
+        with torch.no_grad(), ref_context:
+            reference_chosen_logps, reference_rejected_logps, *_ = self.concatenated_forward(ref_model, batch)
+
+        return reference_chosen_logps, reference_rejected_logps
+
+    @override
+    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 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_rejected_logits,
+            policy_chosen_logps_avg,
+        ) = self.concatenated_forward(model, batch)
+
+        reference_chosen_logps, reference_rejected_logps = self.compute_reference_log_probs(model, batch)
+        losses, chosen_rewards, rejected_rewards = self.compute_preference_loss(
+            policy_chosen_logps,
+            policy_rejected_logps,
+            reference_chosen_logps,
+            reference_rejected_logps,
+        )
+        sft_loss = -policy_chosen_logps_avg
+        if self.ftx_gamma > 1e-6:
+            losses += self.ftx_gamma * sft_loss
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics[f"{prefix}rewards/chosen"] = chosen_rewards.mean().item()
+        metrics[f"{prefix}rewards/rejected"] = rejected_rewards.mean().item()
+        metrics[f"{prefix}rewards/accuracies"] = (chosen_rewards > rejected_rewards).float().mean().item()
+        metrics[f"{prefix}rewards/margins"] = (chosen_rewards - rejected_rewards).mean().item()
+        metrics[f"{prefix}logps/chosen"] = policy_chosen_logps.mean().item()
+        metrics[f"{prefix}logps/rejected"] = policy_rejected_logps.mean().item()
+        metrics[f"{prefix}logits/chosen"] = policy_chosen_logits.mean().item()
+        metrics[f"{prefix}logits/rejected"] = policy_rejected_logits.mean().item()
+        if self.loss_type == "orpo":
+            metrics[f"{prefix}sft_loss"] = sft_loss.mean().item()
+            metrics[f"{prefix}odds_ratio_loss"] = ((losses - sft_loss) / self.beta).mean().item()
+
+        return losses.mean(), metrics
+
+    @override
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
+        r"""
+        Fixes the loss value for transformers 4.46.0.
+        https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605
+        """
+        loss = super().compute_loss(model, inputs, return_outputs)
+        if is_transformers_version_equal_to_4_46() and kwargs.pop("num_items_in_batch", False):
+            if return_outputs:
+                return (loss[0] / self.args.gradient_accumulation_steps, *loss[1:])
+            else:
+                return loss / self.args.gradient_accumulation_steps
+
+        return loss
+
+    @override
+    def log(self, logs: Dict[str, float]) -> None:
+        r"""
+        Log `logs` on the various objects watching training, including stored metrics.
+        """
+        # logs either has "loss" or "eval_loss"
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        key_list, metric_list = [], []
+        for key, metrics in self._stored_metrics[train_eval].items():
+            key_list.append(key)
+            metric_list.append(torch.tensor(metrics, dtype=torch.float).to(self.accelerator.device).mean().item())
+
+        del self._stored_metrics[train_eval]
+        if len(metric_list) < 10:  # pad to for all reduce
+            for i in range(10 - len(metric_list)):
+                key_list.append(f"dummy_{i}")
+                metric_list.append(0.0)
+
+        metric_list = torch.tensor(metric_list, dtype=torch.float).to(self.accelerator.device)
+        metric_list = self.accelerator.reduce(metric_list, "mean").tolist()
+        for key, metric in zip(key_list, metric_list):  # add remaining items
+            if not key.startswith("dummy_"):
+                logs[key] = metric
+
+        return Trainer.log(self, logs)

LLaMA-Factory/src/llamafactory/train/dpo/workflow.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/examples/scripts/dpo.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING, List, Optional
+
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, get_template_and_fix_tokenizer
+from ...extras.constants import IGNORE_INDEX
+from ...extras.misc import calculate_tps
+from ...extras.ploting import plot_loss
+from ...hparams import ModelArguments
+from ...model import load_model, load_tokenizer
+from ..trainer_utils import create_modelcard_and_push, create_ref_model
+from .trainer import CustomDPOTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments
+
+
+def run_dpo(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer_module = load_tokenizer(model_args)
+    tokenizer = tokenizer_module["tokenizer"]
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="rm", **tokenizer_module)
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
+
+    data_collator = PairwiseDataCollatorWithPadding(
+        template=template,
+        model=model,
+        pad_to_multiple_of=8,
+        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
+        **tokenizer_module,
+    )
+
+    # Create reference model
+    if finetuning_args.use_ref_model:
+        if finetuning_args.ref_model is None and (not training_args.do_train):  # use the model itself
+            ref_model = model
+        else:
+            ref_model = create_ref_model(model_args, finetuning_args)
+    else:
+        ref_model = None
+
+    # Update arguments
+    training_args.remove_unused_columns = False  # important for multimodal and pairwise dataset
+
+    # Initialize our Trainer
+    trainer = CustomDPOTrainer(
+        model=model,
+        ref_model=ref_model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        **dataset_module,
+        **tokenizer_module,
+    )
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        trainer.save_model()
+        if finetuning_args.include_effective_tokens_per_second:
+            train_result.metrics["effective_tokens_per_sec"] = calculate_tps(
+                dataset_module["train_dataset"], train_result.metrics, stage="rm"
+            )
+
+        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", "rewards/accuracies"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval")
+        if id(model) == id(ref_model):  # unable to compute rewards if reference model is the model itself
+            remove_keys = [key for key in metrics.keys() if "rewards" in key]
+            for key in remove_keys:
+                metrics.pop(key)
+        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)

LLaMA-Factory/src/llamafactory/train/kto/__init__.py

+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .workflow import run_kto
+
+
+__all__ = ["run_kto"]

LLaMA-Factory/src/llamafactory/train/kto/trainer.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/trl/trainer/kto_trainer.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from collections import defaultdict
+from contextlib import nullcontext
+from types import MethodType
+from typing import TYPE_CHECKING, Dict, List, Literal, Optional, Tuple, Union
+
+import torch
+from transformers import Trainer
+from trl import KTOTrainer
+from trl.trainer import disable_dropout_in_model
+from typing_extensions import override
+
+from ...extras.constants import IGNORE_INDEX
+from ...extras.packages import is_transformers_version_equal_to_4_46
+from ..callbacks import SaveProcessorCallback
+from ..trainer_utils import create_custom_optimizer, create_custom_scheduler, get_batch_logps
+
+
+if TYPE_CHECKING:
+    import torch.utils.data
+    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.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.pref_beta
+        self.desirable_weight = finetuning_args.kto_chosen_weight
+        self.undesirable_weight = finetuning_args.kto_rejected_weight
+        self.ftx_gamma = finetuning_args.pref_ftx
+
+        Trainer.__init__(self, model=model, **kwargs)
+        if not hasattr(self, "accelerator"):
+            raise AttributeError("Please update `transformers`.")
+
+        warnings.simplefilter("ignore")  # remove gc warnings on ref model
+
+        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)
+                self.ref_model.eval()
+
+        if processor is not None:
+            self.add_callback(SaveProcessorCallback(processor))
+
+        if finetuning_args.use_badam:
+            from badam import BAdamCallback, clip_grad_norm_old_version  # type: ignore
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.add_callback(BAdamCallback)
+
+    @override
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimizer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    @override
+    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)
+
+    @override
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        r"""
+        Replaces the sequential sampler of KTO Trainer created by trl with the random sampler.
+        """
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return Trainer._get_train_sampler(self)
+
+    @override
+    def get_batch_samples(self, epoch_iterator, num_batches):
+        r"""
+        Replaces the method of KTO Trainer with the one of the standard Trainer.
+        """
+        return Trainer.get_batch_samples(self, epoch_iterator, num_batches)
+
+    @override
+    def forward(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"], prefix: Literal["", "kl_"] = ""
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Runs forward pass and computes the log probabilities.
+        """
+        batch = {k: v.detach().clone() for k, v in batch.items()}  # avoid error
+        model_inputs = {
+            "input_ids": batch[f"{prefix}input_ids"],
+            "attention_mask": batch[f"{prefix}attention_mask"],
+        }
+        if f"{prefix}token_type_ids" in batch:
+            model_inputs["token_type_ids"] = batch[f"{prefix}token_type_ids"]
+
+        if "pixel_values" in batch:
+            model_inputs["pixel_values"] = batch["pixel_values"]
+
+        if "image_grid_thw" in batch:
+            model_inputs["image_grid_thw"] = batch["image_grid_thw"]
+
+        logits = model(**model_inputs, return_dict=True, use_cache=False).logits.to(torch.float32)
+        logps, valid_length = get_batch_logps(logits=logits, labels=batch[f"{prefix}labels"])
+        return logits, logps, logps / valid_length
+
+    @override
+    def concatenated_forward(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        target_logits, target_logps, target_logps_avg = self.forward(model, batch)
+        with torch.no_grad():
+            _, kl_logps, _ = self.forward(model, batch, prefix="kl_")
+
+        if len(target_logps) != len(batch["kto_tags"]):
+            raise ValueError("Mismatched shape of inputs and labels.")
+
+        chosen_logits = target_logits[batch["kto_tags"]]
+        chosen_logps = target_logps[batch["kto_tags"]]
+        rejected_logits = target_logits[~batch["kto_tags"]]
+        rejected_logps = target_logps[~batch["kto_tags"]]
+        chosen_logps_avg = target_logps_avg[batch["kto_tags"]]
+        return chosen_logps, rejected_logps, chosen_logits, rejected_logits, kl_logps, chosen_logps_avg
+
+    @override
+    def compute_reference_log_probs(
+        self, model: "PreTrainedModel", batch: Dict[str, "torch.Tensor"]
+    ) -> Tuple["torch.Tensor", "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Computes log probabilities of the reference model.
+        """
+        if self.ref_model is None:
+            ref_model = model
+            ref_context = self.accelerator.unwrap_model(model).disable_adapter()
+        else:
+            ref_model = self.ref_model
+            ref_context = nullcontext()
+
+        with torch.no_grad(), ref_context:
+            reference_chosen_logps, reference_rejected_logps, _, _, reference_kl_logps, _ = self.concatenated_forward(
+                ref_model, batch
+            )
+
+        return reference_chosen_logps, reference_rejected_logps, reference_kl_logps
+
+    @override
+    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_rejected_logits,
+            policy_kl_logps,
+            policy_chosen_logps_avg,
+        ) = self.concatenated_forward(model, batch)
+        reference_chosen_logps, reference_rejected_logps, reference_kl_logps = self.compute_reference_log_probs(
+            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 = -policy_chosen_logps_avg
+            losses += self.ftx_gamma * sft_loss.nanmean() / len(policy_chosen_logps) * len(batch["labels"])
+
+        num_chosen = len(chosen_rewards)
+        num_rejected = len(rejected_rewards)
+        if num_chosen > 0:
+            metrics["rewards/chosen_sum"] = chosen_rewards.nansum().item()
+            metrics["logps/chosen_sum"] = policy_chosen_logps.nansum().item()
+            metrics["logits/chosen_sum"] = policy_chosen_logits.nansum().item()
+            metrics["count/chosen"] = float(num_chosen)
+
+        if num_rejected > 0:
+            metrics["rewards/rejected_sum"] = rejected_rewards.nansum().item()
+            metrics["logps/rejected_sum"] = policy_rejected_logps.nansum().item()
+            metrics["logits/rejected_sum"] = policy_rejected_logits.nansum().item()
+            metrics["count/rejected"] = float(num_rejected)
+
+        metrics["kl"] = kl.item()
+        return losses, metrics
+
+    @override
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
+        r"""
+        Fixes the loss value for transformers 4.46.0.
+        https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605
+        """
+        loss = super().compute_loss(model, inputs, return_outputs)
+        if is_transformers_version_equal_to_4_46() and kwargs.pop("num_items_in_batch", False):
+            if return_outputs:
+                return (loss[0] / self.args.gradient_accumulation_steps, *loss[1:])
+            else:
+                return loss / self.args.gradient_accumulation_steps
+
+        return loss
+
+    @override
+    def log(self, logs: Dict[str, float]) -> None:
+        r"""
+        Log `logs` on the various objects watching training, including stored metrics.
+        """
+        # logs either has "loss" or "eval_loss"
+        train_eval = "train" if "loss" in logs else "eval"
+        prefix = "eval_" if train_eval == "eval" else ""
+        # Add averaged stored metrics to logs
+        key_list, metric_list = [], []
+        for key, metrics in self._stored_metrics[train_eval].items():
+            key_list.append(key)
+            metric_list.append(torch.tensor(metrics, dtype=torch.float).to(self.accelerator.device).sum().item())
+
+        del self._stored_metrics[train_eval]
+        if len(metric_list) < 9:  # pad to for all reduce
+            for i in range(9 - len(metric_list)):
+                key_list.append(f"dummy_{i}")
+                metric_list.append(0.0)
+
+        metric_list = torch.tensor(metric_list, dtype=torch.float).to(self.accelerator.device)
+        metric_list = self.accelerator.reduce(metric_list, "sum").tolist()
+        metric_dict: Dict[str, float] = dict(zip(key_list, metric_list))
+        for split in ["chosen", "rejected"]:  # accumulate average metrics from sums and lengths
+            if f"count/{split}" in metric_dict:
+                for key in ("rewards", "logps", "logits"):
+                    logs[f"{prefix}{key}/{split}"] = metric_dict[f"{key}/{split}_sum"] / metric_dict[f"count/{split}"]
+                    del metric_dict[f"{key}/{split}_sum"]
+                del metric_dict[f"count/{split}"]
+
+        if f"{prefix}rewards/chosen" in logs and f"{prefix}rewards/rejected" in logs:  # calculate reward margin
+            logs[f"{prefix}rewards/margins"] = logs[f"{prefix}rewards/chosen"] - logs[f"{prefix}rewards/rejected"]
+
+        for key, metric in metric_dict.items():  # add remaining items
+            if not key.startswith("dummy_"):
+                logs[key] = metric
+
+        return Trainer.log(self, logs)

LLaMA-Factory/src/llamafactory/train/kto/workflow.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/examples/scripts/kto.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING, List, Optional
+
+from ...data import KTODataCollatorWithPadding, get_dataset, get_template_and_fix_tokenizer
+from ...extras.constants import IGNORE_INDEX
+from ...extras.ploting import plot_loss
+from ...hparams import ModelArguments
+from ...model import load_model, load_tokenizer
+from ..trainer_utils import create_modelcard_and_push, create_ref_model
+from .trainer import CustomKTOTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments
+
+
+def run_kto(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer_module = load_tokenizer(model_args)
+    tokenizer = tokenizer_module["tokenizer"]
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="kto", **tokenizer_module)
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
+
+    data_collator = KTODataCollatorWithPadding(
+        template=template,
+        model=model,
+        pad_to_multiple_of=8,
+        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
+        **tokenizer_module,
+    )
+
+    # Create reference model
+    if finetuning_args.ref_model is None and (not training_args.do_train):  # use the model itself
+        ref_model = model
+    else:
+        ref_model = create_ref_model(model_args, finetuning_args)
+
+    # Update arguments
+    training_args.remove_unused_columns = False  # important for multimodal and pairwise dataset
+
+    # Initialize our Trainer
+    trainer = CustomKTOTrainer(
+        model=model,
+        ref_model=ref_model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        **dataset_module,
+        **tokenizer_module,
+    )
+
+    # 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", "rewards/chosen"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval")
+        if id(model) == id(ref_model):  # unable to compute rewards without a reference model
+            remove_keys = [key for key in metrics.keys() if "rewards" in key]
+            for key in remove_keys:
+                metrics.pop(key)
+        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)

LLaMA-Factory/src/llamafactory/train/ppo/__init__.py

+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .workflow import run_ppo
+
+
+__all__ = ["run_ppo"]

LLaMA-Factory/src/llamafactory/train/ppo/ppo_utils.py

+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+from contextlib import nullcontext
+from typing import TYPE_CHECKING, Dict, List, Literal, Optional
+
+import torch
+from transformers.integrations import is_deepspeed_zero3_enabled
+
+from ...extras.packages import is_requests_available
+
+
+if is_requests_available():
+    import requests
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+    from trl import AutoModelForCausalLMWithValueHead
+
+
+def get_rewards_from_server(server_url: str, messages: List[str]) -> List["torch.Tensor"]:
+    r"""
+    Gets reward scores from the API server.
+    """
+    headers = {"Content-Type": "application/json"}
+    payload = {"model": "model", "messages": messages}
+    response = requests.post(server_url, json=payload, headers=headers)
+    rewards = json.loads(response.text)["scores"]
+    return torch.Tensor(rewards)
+
+
+def replace_model(model: "AutoModelForCausalLMWithValueHead", target: Literal["default", "reward"]) -> None:
+    r"""
+    Replaces the default/reward modules in the model. The model is already unwrapped.
+    """
+    v_head_layer = model.v_head.summary
+    if is_deepspeed_zero3_enabled():
+        import deepspeed  # type: ignore
+
+        params = [v_head_layer.weight, v_head_layer.bias]
+        context_maybe_zero3 = deepspeed.zero.GatheredParameters(params, modifier_rank=0)
+    else:
+        context_maybe_zero3 = nullcontext()
+
+    model.pretrained_model.set_adapter(target)  # set the LoRA adapter to be active
+    with context_maybe_zero3:
+        if target == "reward":  # save default head temporarily
+            setattr(model, "default_head_weight", v_head_layer.weight.data.detach().clone())
+            setattr(model, "default_head_bias", v_head_layer.bias.data.detach().clone())
+
+        device = v_head_layer.weight.device
+        v_head_layer.weight.data = model.get_buffer(f"{target}_head_weight").detach().clone().to(device)
+        v_head_layer.bias.data = model.get_buffer(f"{target}_head_bias").detach().clone().to(device)
+
+
+def dump_layernorm(model: "PreTrainedModel") -> Dict[str, "torch.Tensor"]:
+    r"""
+    Dumps the layernorm parameters in the model. The model is already unwrapped (and gathered).
+    """
+    layer_norm_params = {}
+    for name, param in model.named_parameters():
+        if param.data.dtype == torch.float32:
+            layer_norm_params[name] = param.data.detach().clone()
+            param.data = param.data.to(model.config.torch_dtype)
+
+    return layer_norm_params
+
+
+def restore_layernorm(model: "PreTrainedModel", layernorm_params: Optional[Dict[str, "torch.Tensor"]] = None) -> None:
+    r"""
+    Restores the layernorm parameters in the model. The model is already unwrapped (and gathered).
+    """
+    for name, param in model.named_parameters():
+        if name in layernorm_params:
+            param.data = layernorm_params[name]

LLaMA-Factory/src/llamafactory/train/ppo/trainer.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/trl/trainer/ppo_trainer.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import os
+import sys
+import warnings
+from types import MethodType
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
+
+import torch
+from accelerate.utils import DistributedDataParallelKwargs
+from tqdm import tqdm
+from transformers import GenerationConfig, Trainer, TrainerControl, TrainerState
+from transformers.optimization import get_scheduler
+from transformers.trainer import DEFAULT_CALLBACKS
+from transformers.trainer_callback import CallbackHandler
+from transformers.trainer_pt_utils import remove_dummy_checkpoint
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+from transformers.utils import SAFE_WEIGHTS_NAME, WEIGHTS_NAME
+from trl import PPOConfig, PPOTrainer
+from trl.core import PPODecorators, logprobs_from_logits
+from trl.models.utils import unwrap_model_for_generation
+from typing_extensions import override
+
+from ...extras import logging
+from ...extras.misc import AverageMeter, count_parameters, get_current_device, get_logits_processor
+from ..callbacks import FixValueHeadModelCallback, SaveProcessorCallback
+from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
+from .ppo_utils import dump_layernorm, get_rewards_from_server, replace_model, restore_layernorm
+
+
+if TYPE_CHECKING:
+    from datasets import Dataset
+    from transformers import (
+        DataCollatorWithPadding,
+        PreTrainedTokenizer,
+        ProcessorMixin,
+        Seq2SeqTrainingArguments,
+        TrainerCallback,
+    )
+    from trl import AutoModelForCausalLMWithValueHead
+
+    from ...hparams import FinetuningArguments, GeneratingArguments, ModelArguments
+
+
+logger = logging.get_logger(__name__)
+
+
+class CustomPPOTrainer(PPOTrainer, Trainer):
+    r"""
+    Inherits PPOTrainer.
+    """
+
+    def __init__(
+        self,
+        model_args: "ModelArguments",
+        training_args: "Seq2SeqTrainingArguments",
+        finetuning_args: "FinetuningArguments",
+        generating_args: "GeneratingArguments",
+        callbacks: Optional[List["TrainerCallback"]],
+        model: "AutoModelForCausalLMWithValueHead",
+        reward_model: Optional["AutoModelForCausalLMWithValueHead"],
+        ref_model: Optional["AutoModelForCausalLMWithValueHead"],
+        tokenizer: "PreTrainedTokenizer",
+        processor: Optional["ProcessorMixin"],
+        data_collator: "DataCollatorWithPadding",
+        train_dataset: Optional["Dataset"] = None,
+        eval_dataset: Optional["Dataset"] = None,
+    ) -> None:
+        if eval_dataset is not None:
+            raise NotImplementedError("PPOTrainer does not support eval dataset yet.")
+
+        backward_batch_size = training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps
+        ppo_config = PPOConfig(
+            model_name=model_args.model_name_or_path,
+            learning_rate=training_args.learning_rate,
+            mini_batch_size=training_args.per_device_train_batch_size,
+            batch_size=backward_batch_size * finetuning_args.ppo_buffer_size,
+            gradient_accumulation_steps=training_args.gradient_accumulation_steps,
+            ppo_epochs=finetuning_args.ppo_epochs,
+            max_grad_norm=training_args.max_grad_norm,
+            seed=training_args.seed,
+            optimize_device_cache=True,
+            target=finetuning_args.ppo_target,
+            use_score_scaling=finetuning_args.ppo_score_norm,
+            use_score_norm=finetuning_args.ppo_score_norm,
+            whiten_rewards=finetuning_args.ppo_whiten_rewards,
+            accelerator_kwargs={"step_scheduler_with_optimizer": False},
+            log_with=training_args.report_to[0] if training_args.report_to else None,
+            project_kwargs={"logging_dir": training_args.logging_dir},
+        )
+
+        # Add deepspeed config
+        if training_args.deepspeed_plugin is not None:
+            ppo_config.accelerator_kwargs["kwargs_handlers"] = [
+                DistributedDataParallelKwargs(find_unused_parameters=training_args.ddp_find_unused_parameters)
+            ]
+            ppo_config.accelerator_kwargs["deepspeed_plugin"] = training_args.deepspeed_plugin
+            if ppo_config.log_with is not None:
+                logger.warning_rank0("PPOTrainer cannot use external logger when DeepSpeed is enabled.")
+                ppo_config.log_with = None
+
+        # Create optimizer and scheduler
+        if training_args.max_steps > 0:
+            num_training_steps = training_args.max_steps
+        else:
+            total_train_batch_size = backward_batch_size * finetuning_args.ppo_buffer_size * training_args.world_size
+            num_training_steps = training_args.num_train_epochs * math.ceil(
+                len(train_dataset) / total_train_batch_size
+            )
+
+        optimizer = self.create_optimizer(model, training_args, finetuning_args)
+        scheduler = self.create_scheduler(training_args, num_training_steps, optimizer)
+
+        PPOTrainer.__init__(
+            self,
+            config=ppo_config,
+            model=model,
+            ref_model=ref_model,
+            tokenizer=tokenizer,
+            dataset=train_dataset,
+            optimizer=optimizer,
+            data_collator=data_collator,
+            lr_scheduler=scheduler,
+        )
+
+        self.args = training_args
+        self.model_args = model_args
+        self.finetuning_args = finetuning_args
+        self.reward_model = reward_model
+        self.current_device = get_current_device()  # patch for deepspeed training
+
+        self.generation_config = GenerationConfig(
+            pad_token_id=self.tokenizer.pad_token_id,
+            eos_token_id=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
+            **generating_args.to_dict(),
+        )
+
+        self.state = TrainerState()
+        self.control = TrainerControl()
+        self.is_deepspeed_enabled = getattr(self.accelerator.state, "deepspeed_plugin", None) is not None
+        self.is_fsdp_enabled = getattr(self.accelerator.state, "fsdp_plugin", None) is not None
+        callbacks = DEFAULT_CALLBACKS if callbacks is None else DEFAULT_CALLBACKS + callbacks
+        self.callback_handler = CallbackHandler(
+            callbacks, self.accelerator.unwrap_model(self.model), self.tokenizer, self.optimizer, self.lr_scheduler
+        )
+        if self.args.max_steps > 0:
+            logger.info_rank0("max_steps is given, it will override any value given in num_train_epochs")
+
+        self.amp_context = torch.autocast(self.current_device.type)
+        warnings.simplefilter("ignore")  # remove gc warnings on ref model
+
+        if finetuning_args.reward_model_type == "full":
+            if self.is_deepspeed_enabled:
+                if not (
+                    getattr(reward_model.pretrained_model, "is_loaded_in_8bit", False)
+                    or getattr(reward_model.pretrained_model, "is_loaded_in_4bit", False)
+                ):  # quantized models are already set on the correct device
+                    self.reward_model = self._prepare_deepspeed(self.reward_model)
+            else:
+                self.reward_model = self.accelerator.prepare_model(self.reward_model, evaluation_mode=True)
+
+        self.add_callback(FixValueHeadModelCallback)
+
+        if processor is not None:
+            self.add_callback(SaveProcessorCallback(processor))
+
+        if finetuning_args.use_badam:
+            from badam import BAdamCallback, clip_grad_norm_old_version  # type: ignore
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.add_callback(BAdamCallback)
+
+    def ppo_train(self, resume_from_checkpoint: Optional[str] = None) -> None:
+        r"""
+        Implements training loop for the PPO stage, like _inner_training_loop() in Huggingface's Trainer.
+        """
+        if resume_from_checkpoint is not None:
+            raise ValueError("`resume_from_checkpoint` will be supported in the future version.")
+
+        total_train_batch_size = (
+            self.args.per_device_train_batch_size
+            * self.args.gradient_accumulation_steps
+            * self.finetuning_args.ppo_buffer_size
+            * self.args.world_size
+        )
+        if self.args.max_steps > 0:
+            num_examples = total_train_batch_size * self.args.max_steps
+            num_train_epochs = sys.maxsize
+            max_steps = self.args.max_steps
+            steps_in_epoch = self.args.max_steps
+        else:
+            len_dataloader = len(self.dataloader)
+            num_examples = len(self.dataset)
+            num_train_epochs = self.args.num_train_epochs
+            max_steps = math.ceil(num_train_epochs * len_dataloader)
+            steps_in_epoch = len_dataloader
+
+        self.state.max_steps = max_steps
+        self.state.num_train_epochs = num_train_epochs
+        self.state.is_local_process_zero = self.is_local_process_zero()
+        self.state.is_world_process_zero = self.is_world_process_zero()
+
+        logger.info_rank0("***** Running training *****")
+        logger.info_rank0(f"  Num examples = {num_examples:,}")
+        logger.info_rank0(f"  Num Epochs = {num_train_epochs:,}")
+        logger.info_rank0(f"  Instantaneous batch size per device = {self.args.per_device_train_batch_size:,}")
+        logger.info_rank0(
+            "  Total train batch size (w. parallel, buffer, distributed & accumulation) = {:,}".format(
+                total_train_batch_size
+            )
+        )
+        logger.info_rank0(f"  Gradient Accumulation steps = {self.args.gradient_accumulation_steps:,}")
+        logger.info_rank0(f"  Num optimization epochs per batch = {self.finetuning_args.ppo_epochs:,}")
+        logger.info_rank0(f"  Total training steps = {max_steps:,}")
+        logger.info_rank0(f"  Number of trainable parameters = {count_parameters(self.model)[0]:,}")
+
+        dataiter = iter(self.dataloader)
+        loss_meter = AverageMeter()
+        reward_meter = AverageMeter()
+        self.callback_handler.on_train_begin(self.args, self.state, self.control)
+
+        for step in tqdm(range(max_steps), disable=not self.is_local_process_zero()):
+            try:
+                batch = next(dataiter)
+            except StopIteration:
+                dataiter = iter(self.dataloader)
+                batch = next(dataiter)
+
+            # Get inputs
+            self.model.eval()
+            self.tokenizer.padding_side = "right"  # change padding side
+            queries, responses, rewards = [], [], []
+            for idx in range(0, self.config.batch_size, self.config.mini_batch_size):
+                mini_batch_queries, mini_batch_responses = self.get_inputs(
+                    batch[idx : idx + self.config.mini_batch_size]
+                )
+                mini_batch_rewards = self.get_rewards(mini_batch_queries, mini_batch_responses)
+                queries.extend(mini_batch_queries)
+                responses.extend(mini_batch_responses)
+                rewards.extend(mini_batch_rewards)
+
+            # Run PPO step
+            self.model.train()
+            stats = self.step(queries, responses, rewards)
+            self.tokenizer.padding_side = "left"  # restore padding side
+            loss_meter.update(float(stats["ppo/loss/total"]), n=len(rewards))
+            reward_meter.update(torch.stack(rewards).mean().item(), n=len(rewards))
+
+            if self.config.log_with is not None:
+                try:
+                    batch["query"] = self.tokenizer.batch_decode(queries, skip_special_tokens=True)
+                    batch["response"] = self.tokenizer.batch_decode(responses, skip_special_tokens=True)
+                    self.log_stats(stats, batch, rewards)
+                except Exception:
+                    logger.warning_rank0("Failed to save stats due to unknown errors.")
+
+            self.state.global_step += 1
+            self.callback_handler.on_step_end(self.args, self.state, self.control)
+
+            if self.is_local_process_zero() and (step + 1) % self.args.logging_steps == 0:
+                logs = dict(
+                    loss=round(loss_meter.avg, 4),
+                    reward=round(reward_meter.avg, 4),
+                    learning_rate=stats["ppo/learning_rate"],
+                    epoch=round(step / steps_in_epoch, 2),
+                )
+                tqdm.write(str(logs))
+                logs["step"] = step
+                self.state.log_history.append(logs)
+                self.callback_handler.on_log(self.args, self.state, self.control, logs)
+                loss_meter.reset()
+                reward_meter.reset()
+
+            if (step + 1) % self.args.save_steps == 0:  # save checkpoint
+                self.save_model(
+                    os.path.join(self.args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}")
+                )
+                self.callback_handler.on_save(self.args, self.state, self.control)
+
+            if self.control.should_epoch_stop or self.control.should_training_stop:
+                break
+
+        self.callback_handler.on_train_end(self.args, self.state, self.control)
+
+    @override
+    def create_optimizer(
+        self,
+        model: "AutoModelForCausalLMWithValueHead",
+        training_args: "Seq2SeqTrainingArguments",
+        finetuning_args: "FinetuningArguments",
+    ) -> "torch.optim.Optimizer":
+        optimizer = create_custom_optimizer(model, training_args, finetuning_args)
+        if optimizer is None:
+            decay_params, nodecay_params = [], []
+            decay_param_names = self.get_decay_parameter_names(model)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    if name in decay_param_names:
+                        decay_params.append(param)
+                    else:
+                        nodecay_params.append(param)
+
+            optim_class, optim_kwargs = Trainer.get_optimizer_cls_and_kwargs(training_args)
+            param_groups = [
+                dict(params=nodecay_params),
+                dict(params=decay_params, weight_decay=training_args.weight_decay),
+            ]
+            optimizer = optim_class(param_groups, **optim_kwargs)
+
+        return optimizer
+
+    @override
+    def create_scheduler(
+        self, training_args: "Seq2SeqTrainingArguments", num_training_steps: int, optimizer: "torch.optim.Optimizer"
+    ) -> "torch.optim.lr_scheduler.LRScheduler":
+        create_custom_scheduler(training_args, num_training_steps, optimizer)
+        lr_scheduler = get_scheduler(
+            training_args.lr_scheduler_type,
+            optimizer=optimizer,
+            num_warmup_steps=training_args.get_warmup_steps(num_training_steps),
+            num_training_steps=num_training_steps,
+        )
+        return lr_scheduler
+
+    @torch.no_grad()
+    def get_inputs(self, batch: Dict[str, "torch.Tensor"]) -> Tuple[List["torch.Tensor"], List["torch.Tensor"]]:
+        r"""
+        Generates model's responses given queries.
+        """
+        if batch["input_ids"].size(0) == 1:  # handle llama2 ppo with gradient accumulation > 1
+            start_index = (batch["input_ids"][0] != self.tokenizer.pad_token_id).nonzero()[0].item()
+            for k, v in batch.items():
+                batch[k] = v[:, start_index:]
+
+        with unwrap_model_for_generation(self.model, self.accelerator) as unwrapped_model:
+            unwrapped_model: "AutoModelForCausalLMWithValueHead" = self.accelerator.unwrap_model(self.model)
+            if self.model_args.upcast_layernorm:
+                layernorm_params = dump_layernorm(unwrapped_model)
+
+            generate_output: "torch.Tensor" = unwrapped_model.generate(
+                generation_config=self.generation_config, logits_processor=get_logits_processor(), **batch
+            )
+            if self.model_args.upcast_layernorm:
+                restore_layernorm(unwrapped_model, layernorm_params)
+
+        query = batch["input_ids"].detach().cpu()
+        response = generate_output[:, batch["input_ids"].size(-1) :].detach().cpu()
+        queries, responses = [], []
+        for i in range(len(query)):
+            query_start_index = (query[i] != self.tokenizer.pad_token_id).nonzero()[0].item()
+            response_indexes = (response[i] != self.tokenizer.pad_token_id).nonzero()
+
+            if len(response_indexes) == 0:  # allow empty response
+                response_length = 1
+            elif self.tokenizer.eos_token_id == self.tokenizer.pad_token_id:  # include eos token
+                response_length = response_indexes[-1].item() + 2
+            else:
+                response_length = response_indexes[-1].item() + 1
+
+            queries.append(query[i, query_start_index:])  # remove padding from left
+            responses.append(response[i, :response_length])  # remove padding from right
+
+        return queries, responses
+
+    @torch.no_grad()
+    def get_rewards(
+        self,
+        queries: List["torch.Tensor"],
+        responses: List["torch.Tensor"],
+    ) -> List["torch.Tensor"]:
+        r"""
+        Computes scores using given reward model.
+
+        Both inputs and outputs are put on CPU.
+        """
+        if self.finetuning_args.reward_model_type == "api":
+            token_ids = [torch.cat((q, r), dim=-1).tolist() for q, r in zip(queries, responses)]
+            messages = self.tokenizer.batch_decode(token_ids, skip_special_tokens=False)
+            return get_rewards_from_server(self.reward_model, messages)
+
+        batch: Dict[str, "torch.Tensor"] = self.prepare_model_inputs(queries, responses)
+        unwrapped_model: "AutoModelForCausalLMWithValueHead" = self.accelerator.unwrap_model(self.model)
+
+        if self.finetuning_args.reward_model_type == "lora":
+            replace_model(unwrapped_model, target="reward")
+            reward_model = self.model
+        else:
+            reward_model = self.reward_model
+
+        with unwrap_model_for_generation(reward_model, self.accelerator), self.amp_context:  # support bf16
+            values: "torch.Tensor" = reward_model(**batch, return_dict=True, use_cache=False)[-1]
+
+        if self.finetuning_args.reward_model_type == "lora":
+            replace_model(unwrapped_model, target="default")
+
+        rewards = values.gather(dim=-1, index=(batch["attention_mask"].sum(dim=-1, keepdim=True) - 1))
+        return rewards.float().detach()  # use fp32 type
+
+    @override
+    @PPODecorators.empty_device_cache()
+    def batched_forward_pass(
+        self,
+        model: "AutoModelForCausalLMWithValueHead",
+        queries: "torch.Tensor",
+        responses: "torch.Tensor",
+        model_inputs: Dict[str, Any],
+        return_logits: bool = False,
+        response_masks: Optional["torch.Tensor"] = None,
+    ) -> Tuple["torch.Tensor", Optional["torch.Tensor"], "torch.Tensor", "torch.Tensor"]:
+        r"""
+        Calculates model outputs in multiple batches.
+
+        Subclass and override to inject custom behavior.
+        """
+        bs = len(queries)
+        fbs = self.config.mini_batch_size
+        all_logprobs = []
+        all_logits = []
+        all_masks = []
+        all_values = []
+
+        for i in range(math.ceil(bs / fbs)):
+            input_kwargs = {key: value[i * fbs : (i + 1) * fbs] for key, value in model_inputs.items()}
+            query_batch = queries[i * fbs : (i + 1) * fbs]
+            response_batch = responses[i * fbs : (i + 1) * fbs]
+            if response_masks is not None:
+                response_masks_batch = response_masks[i * fbs : (i + 1) * fbs]
+            input_ids = input_kwargs["input_ids"]
+            attention_mask = input_kwargs["attention_mask"]
+
+            with self.amp_context:  # support bf16
+                logits, _, values = model(**input_kwargs, return_dict=True, use_cache=False)
+
+            logprobs = logprobs_from_logits(logits[:, :-1, :], input_ids[:, 1:])
+            masks = torch.zeros_like(attention_mask)
+            masks[:, :-1] = attention_mask[:, 1:]
+
+            for j in range(len(query_batch)):
+                start = len(query_batch[j]) - 1
+                if attention_mask[j, 0] == 0:  # offset left padding
+                    start += attention_mask[j, :].nonzero()[0].item()
+                end = start + len(response_batch[j])
+
+                if response_masks is not None:
+                    response_masks_batch = torch.cat((torch.zeros_like(query_batch[j]), response_masks_batch[j]))[1:]
+
+                masks[j, :start] = 0
+                masks[j, end:] = 0
+                if response_masks is not None:
+                    masks[j, start:end] = masks[j, start:end] * response_masks_batch[j][start:end]
+
+            if return_logits:
+                all_logits.append(logits)
+            else:
+                del logits
+
+            all_values.append(values)
+            all_logprobs.append(logprobs)
+            all_masks.append(masks)
+
+        return (
+            torch.cat(all_logprobs),
+            torch.cat(all_logits)[:, :-1] if return_logits else None,
+            torch.cat(all_values)[:, :-1],
+            torch.cat(all_masks)[:, :-1],
+        )
+
+    @override
+    def save_model(self, output_dir: Optional[str] = None) -> None:
+        r"""
+        Saves model checkpoint.
+
+        Subclass and override to inject custom behavior.
+        """
+        if output_dir is None:
+            output_dir = self.args.output_dir
+
+        if self.is_fsdp_enabled or self.is_deepspeed_enabled:
+            try:
+                state_dict = self.accelerator.get_state_dict(self.model)  # must be called at all ranks
+                if self.args.should_save:
+                    self._save(output_dir, state_dict=state_dict)
+            except ValueError:
+                logger.warning_rank0(
+                    " stage3_gather_16bit_weights_on_model_save=false. Saving the full checkpoint instead,"
+                    " use zero_to_fp32.py to recover weights"
+                )
+                if self.args.should_save:
+                    self._save(output_dir, state_dict={})
+                # remove the dummy state_dict
+                remove_dummy_checkpoint(self.args.should_save, output_dir, [WEIGHTS_NAME, SAFE_WEIGHTS_NAME])
+                self.model.save_checkpoint(output_dir)
+
+        elif self.args.should_save:
+            unwrapped_model: "AutoModelForCausalLMWithValueHead" = self.accelerator.unwrap_model(self.model)
+            self._save(output_dir, state_dict=unwrapped_model.state_dict())

LLaMA-Factory/src/llamafactory/train/ppo/workflow.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's TRL library.
+# https://github.com/huggingface/trl/blob/v0.8.0/examples/scripts/ppo.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING, List, Optional
+
+from ...data import MultiModalDataCollatorForSeq2Seq, get_dataset, get_template_and_fix_tokenizer
+from ...extras.ploting import plot_loss
+from ...model import load_model, load_tokenizer
+from ..callbacks import fix_valuehead_checkpoint
+from ..trainer_utils import create_ref_model, create_reward_model
+from .trainer import CustomPPOTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
+
+
+def run_ppo(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    generating_args: "GeneratingArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer_module = load_tokenizer(model_args)
+    tokenizer = tokenizer_module["tokenizer"]
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="ppo", **tokenizer_module)
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train, add_valuehead=True)
+
+    tokenizer.padding_side = "left"  # use left-padding in generation while using right-padding in training
+    data_collator = MultiModalDataCollatorForSeq2Seq(template=template, model=model, **tokenizer_module)
+
+    # Create reference model and reward model
+    ref_model = create_ref_model(model_args, finetuning_args, add_valuehead=True)
+    reward_model = create_reward_model(model, model_args, finetuning_args)
+
+    # Initialize our Trainer
+    ppo_trainer: "CustomPPOTrainer" = CustomPPOTrainer(
+        model_args=model_args,
+        training_args=training_args,
+        finetuning_args=finetuning_args,
+        generating_args=generating_args,
+        callbacks=callbacks,
+        model=model,
+        reward_model=reward_model,
+        ref_model=ref_model,
+        data_collator=data_collator,
+        **dataset_module,
+        **tokenizer_module,
+    )
+
+    # Training
+    if training_args.do_train:
+        ppo_trainer.ppo_train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        ppo_trainer.save_model()
+        if training_args.should_save:
+            fix_valuehead_checkpoint(model, training_args.output_dir, training_args.save_safetensors)
+
+        ppo_trainer.save_state()  # must be called after save_model to have a folder
+        if ppo_trainer.is_world_process_zero() and finetuning_args.plot_loss:
+            plot_loss(training_args.output_dir, keys=["loss", "reward"])

LLaMA-Factory/src/llamafactory/train/pt/__init__.py

+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .workflow import run_pt
+
+
+__all__ = ["run_pt"]

LLaMA-Factory/src/llamafactory/train/pt/trainer.py

+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from types import MethodType
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import Trainer
+from typing_extensions import override
+
+from ...extras.packages import is_transformers_version_equal_to_4_46, is_transformers_version_greater_than
+from ..callbacks import SaveProcessorCallback
+from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, ProcessorMixin
+
+    from ...hparams import FinetuningArguments
+
+
+class CustomTrainer(Trainer):
+    r"""
+    Inherits Trainer for custom optimizer.
+    """
+
+    def __init__(
+        self, finetuning_args: "FinetuningArguments", processor: Optional["ProcessorMixin"], **kwargs
+    ) -> None:
+        if is_transformers_version_greater_than("4.46"):
+            kwargs["processing_class"] = kwargs.pop("tokenizer")
+
+        super().__init__(**kwargs)
+        self.finetuning_args = finetuning_args
+
+        if processor is not None:
+            self.add_callback(SaveProcessorCallback(processor))
+
+        if finetuning_args.use_badam:
+            from badam import BAdamCallback, clip_grad_norm_old_version  # type: ignore
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.add_callback(BAdamCallback)
+
+    @override
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimizer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    @override
+    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)
+
+    @override
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return super()._get_train_sampler()
+
+    @override
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
+        r"""
+        Fixes the loss value for transformers 4.46.0.
+        https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605
+        """
+        loss = super().compute_loss(model, inputs, return_outputs, **kwargs)
+        if is_transformers_version_equal_to_4_46() and not getattr(self, "model_accepts_loss_kwargs", False):
+            # other model should not scale the loss
+            if return_outputs:
+                return (loss[0] / self.args.gradient_accumulation_steps, *loss[1:])
+            else:
+                return loss / self.args.gradient_accumulation_steps
+
+        return loss

LLaMA-Factory/src/llamafactory/train/pt/workflow.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/language-modeling/run_clm.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import TYPE_CHECKING, List, Optional
+
+from transformers import DataCollatorForLanguageModeling
+
+from ...data import get_dataset, get_template_and_fix_tokenizer
+from ...extras.ploting import plot_loss
+from ...model import load_model, load_tokenizer
+from ..trainer_utils import create_modelcard_and_push
+from .trainer import CustomTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments, ModelArguments
+
+
+def run_pt(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer_module = load_tokenizer(model_args)
+    tokenizer = tokenizer_module["tokenizer"]
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="pt", **tokenizer_module)
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
+    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+
+    # Initialize our Trainer
+    trainer = CustomTrainer(
+        model=model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        **dataset_module,
+        **tokenizer_module,
+    )
+
+    # 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"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval")
+        try:
+            perplexity = math.exp(metrics["eval_loss"])
+        except OverflowError:
+            perplexity = float("inf")
+
+        metrics["perplexity"] = perplexity
+        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)

LLaMA-Factory/src/llamafactory/train/rm/__init__.py

+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .workflow import run_rm
+
+
+__all__ = ["run_rm"]

LLaMA-Factory/src/llamafactory/train/rm/metric.py

+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Dict, Optional
+
+import numpy as np
+
+from ...extras.misc import numpify
+
+
+if TYPE_CHECKING:
+    from transformers import EvalPrediction
+
+
+@dataclass
+class ComputeAccuracy:
+    r"""
+    Computes reward accuracy and supports `batch_eval_metrics`.
+    """
+
+    def _dump(self) -> Optional[Dict[str, float]]:
+        result = None
+        if hasattr(self, "score_dict"):
+            result = {k: float(np.mean(v)) for k, v in self.score_dict.items()}
+
+        self.score_dict = {"accuracy": []}
+        return result
+
+    def __post_init__(self):
+        self._dump()
+
+    def __call__(self, eval_preds: "EvalPrediction", compute_result: bool = True) -> Optional[Dict[str, float]]:
+        chosen_scores, rejected_scores = numpify(eval_preds.predictions[0]), numpify(eval_preds.predictions[1])
+        if not chosen_scores.shape:
+            self.score_dict["accuracy"].append(chosen_scores > rejected_scores)
+        else:
+            for i in range(len(chosen_scores)):
+                self.score_dict["accuracy"].append(chosen_scores[i] > rejected_scores[i])
+
+        if compute_result:
+            return self._dump()

LLaMA-Factory/src/llamafactory/train/rm/trainer.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/trainer.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+from types import MethodType
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import Trainer
+from typing_extensions import override
+
+from ...extras import logging
+from ...extras.packages import is_transformers_version_equal_to_4_46, is_transformers_version_greater_than
+from ..callbacks import FixValueHeadModelCallback, SaveProcessorCallback
+from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel, ProcessorMixin
+    from transformers.trainer import PredictionOutput
+
+    from ...hparams import FinetuningArguments
+
+
+logger = logging.get_logger(__name__)
+
+
+class PairwiseTrainer(Trainer):
+    r"""
+    Inherits Trainer to compute pairwise loss.
+    """
+
+    def __init__(
+        self, finetuning_args: "FinetuningArguments", processor: Optional["ProcessorMixin"], **kwargs
+    ) -> None:
+        if is_transformers_version_greater_than("4.46"):
+            kwargs["processing_class"] = kwargs.pop("tokenizer")
+
+        super().__init__(**kwargs)
+        self.finetuning_args = finetuning_args
+        self.can_return_loss = True  # override property to return eval_loss
+        self.add_callback(FixValueHeadModelCallback)
+
+        if processor is not None:
+            self.add_callback(SaveProcessorCallback(processor))
+
+        if finetuning_args.use_badam:
+            from badam import BAdamCallback, clip_grad_norm_old_version  # type: ignore
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.add_callback(BAdamCallback)
+
+    @override
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimizer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    @override
+    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)
+
+    @override
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return super()._get_train_sampler()
+
+    @override
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
+        r"""
+        Computes pairwise loss. The first n examples are chosen and the last n examples are rejected.
+
+        Subclass and override to inject custom behavior.
+
+        Note that the first element will be removed from the output tuple.
+        See: https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/trainer.py#L3842
+        """
+        _, _, values = model(**inputs, output_hidden_states=True, return_dict=True, use_cache=False)
+        batch_size = inputs["input_ids"].size(0) // 2
+        chosen_masks, rejected_masks = torch.split(inputs["attention_mask"], batch_size, dim=0)
+        chosen_rewards, rejected_rewards = torch.split(values, batch_size, dim=0)
+        chosen_scores = chosen_rewards.gather(dim=-1, index=(chosen_masks.sum(dim=-1, keepdim=True) - 1))
+        rejected_scores = rejected_rewards.gather(dim=-1, index=(rejected_masks.sum(dim=-1, keepdim=True) - 1))
+        chosen_scores, rejected_scores = chosen_scores.squeeze(), rejected_scores.squeeze()
+
+        loss = -torch.nn.functional.logsigmoid(chosen_scores.float() - rejected_scores.float()).mean()
+
+        if is_transformers_version_equal_to_4_46() and kwargs.pop("num_items_in_batch", False):
+            loss /= self.args.gradient_accumulation_steps  # fixes the loss value for transformers 4.46.0
+
+        if return_outputs:
+            return loss, (loss, chosen_scores, rejected_scores)
+        else:
+            return loss
+
+    def save_predictions(self, predict_results: "PredictionOutput") -> None:
+        r"""
+        Saves model predictions to `output_dir`.
+
+        A custom behavior that not contained in Seq2SeqTrainer.
+        """
+        if not self.is_world_process_zero():
+            return
+
+        output_prediction_file = os.path.join(self.args.output_dir, "generated_predictions.jsonl")
+        logger.info_rank0(f"Saving prediction results to {output_prediction_file}")
+        chosen_scores, rejected_scores = predict_results.predictions
+
+        with open(output_prediction_file, "w", encoding="utf-8") as writer:
+            res: List[str] = []
+            for c_score, r_score in zip(chosen_scores, rejected_scores):
+                res.append(json.dumps({"chosen": round(float(c_score), 2), "rejected": round(float(r_score), 2)}))
+
+            writer.write("\n".join(res))

LLaMA-Factory/src/llamafactory/train/rm/workflow.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/summarization/run_summarization.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING, List, Optional
+
+from ...data import PairwiseDataCollatorWithPadding, get_dataset, get_template_and_fix_tokenizer
+from ...extras.ploting import plot_loss
+from ...model import load_model, load_tokenizer
+from ..callbacks import fix_valuehead_checkpoint
+from ..trainer_utils import create_modelcard_and_push
+from .metric import ComputeAccuracy
+from .trainer import PairwiseTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments, ModelArguments
+
+
+def run_rm(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer_module = load_tokenizer(model_args)
+    tokenizer = tokenizer_module["tokenizer"]
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="rm", **tokenizer_module)
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train, add_valuehead=True)
+    data_collator = PairwiseDataCollatorWithPadding(
+        template=template, model=model, pad_to_multiple_of=8, **tokenizer_module
+    )
+
+    # Update arguments
+    training_args.remove_unused_columns = False  # important for multimodal and pairwise dataset
+
+    # Initialize our Trainer
+    trainer = PairwiseTrainer(
+        model=model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        compute_metrics=ComputeAccuracy(),
+        **dataset_module,
+        **tokenizer_module,
+    )
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        trainer.save_model()
+        if training_args.should_save:
+            fix_valuehead_checkpoint(model, training_args.output_dir, training_args.save_safetensors)
+
+        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", "eval_accuracy"])
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval")
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Predict
+    if training_args.do_predict:
+        predict_results = trainer.predict(dataset_module["eval_dataset"], metric_key_prefix="predict")
+        trainer.log_metrics("predict", predict_results.metrics)
+        trainer.save_metrics("predict", predict_results.metrics)
+        trainer.save_predictions(predict_results)
+
+    # Create model card
+    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)

LLaMA-Factory/src/llamafactory/train/sft/__init__.py

+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .workflow import run_sft
+
+
+__all__ = ["run_sft"]

LLaMA-Factory/src/llamafactory/train/sft/metric.py

+# Copyright 2024 HuggingFace Inc., THUDM, and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library and the THUDM's ChatGLM implementation.
+# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/summarization/run_summarization.py
+# https://github.com/THUDM/ChatGLM-6B/blob/main/ptuning/main.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Dict, Optional
+
+import numpy as np
+import torch
+from transformers.utils import is_jieba_available, is_nltk_available
+
+from ...extras.constants import IGNORE_INDEX
+from ...extras.misc import numpify
+from ...extras.packages import is_rouge_available
+
+
+if TYPE_CHECKING:
+    from transformers import EvalPrediction, PreTrainedTokenizer
+
+
+if is_jieba_available():
+    import jieba  # type: ignore
+
+
+if is_nltk_available():
+    from nltk.translate.bleu_score import SmoothingFunction, sentence_bleu
+
+
+if is_rouge_available():
+    from rouge_chinese import Rouge
+
+
+def eval_logit_processor(logits: "torch.Tensor", labels: "torch.Tensor") -> "torch.Tensor":
+    r"""
+    Computes the token with the largest likelihood to reduce memory footprint.
+    """
+    if isinstance(logits, (list, tuple)):
+        if logits[0].dim() == 3:  # (batch_size, seq_len, vocab_size)
+            logits = logits[0]
+        else:  # moe models have aux loss
+            logits = logits[1]
+
+    if logits.dim() != 3:
+        raise ValueError("Cannot process the logits.")
+
+    return torch.argmax(logits, dim=-1)
+
+
+@dataclass
+class ComputeAccuracy:
+    r"""
+    Computes accuracy and supports `batch_eval_metrics`.
+    """
+
+    def _dump(self) -> Optional[Dict[str, float]]:
+        result = None
+        if hasattr(self, "score_dict"):
+            result = {k: float(np.mean(v)) for k, v in self.score_dict.items()}
+
+        self.score_dict = {"accuracy": []}
+        return result
+
+    def __post_init__(self):
+        self._dump()
+
+    def __call__(self, eval_preds: "EvalPrediction", compute_result: bool = True) -> Optional[Dict[str, float]]:
+        preds, labels = numpify(eval_preds.predictions), numpify(eval_preds.label_ids)
+        for i in range(len(preds)):
+            pred, label = preds[i, :-1], labels[i, 1:]
+            label_mask = label != IGNORE_INDEX
+            self.score_dict["accuracy"].append(np.mean(pred[label_mask] == label[label_mask]))
+
+        if compute_result:
+            return self._dump()
+
+
+@dataclass
+class ComputeSimilarity:
+    r"""
+    Computes text similarity scores and supports `batch_eval_metrics`.
+
+    Wraps the tokenizer into metric functions, used in CustomSeq2SeqTrainer.
+    """
+
+    tokenizer: "PreTrainedTokenizer"
+
+    def _dump(self) -> Optional[Dict[str, float]]:
+        result = None
+        if hasattr(self, "score_dict"):
+            result = {k: float(np.mean(v)) for k, v in self.score_dict.items()}
+
+        self.score_dict = {"rouge-1": [], "rouge-2": [], "rouge-l": [], "bleu-4": []}
+        return result
+
+    def __post_init__(self):
+        self._dump()
+
+    def __call__(self, eval_preds: "EvalPrediction", compute_result: bool = True) -> Optional[Dict[str, float]]:
+        preds, labels = numpify(eval_preds.predictions), numpify(eval_preds.label_ids)
+
+        preds = np.where(preds != IGNORE_INDEX, preds, self.tokenizer.pad_token_id)
+        labels = np.where(labels != IGNORE_INDEX, labels, self.tokenizer.pad_token_id)
+
+        decoded_preds = self.tokenizer.batch_decode(preds, skip_special_tokens=True)
+        decoded_labels = self.tokenizer.batch_decode(labels, skip_special_tokens=True)
+
+        for pred, label in zip(decoded_preds, decoded_labels):
+            hypothesis = list(jieba.cut(pred))
+            reference = list(jieba.cut(label))
+
+            if len(" ".join(hypothesis).split()) == 0 or len(" ".join(reference).split()) == 0:
+                result = {"rouge-1": {"f": 0.0}, "rouge-2": {"f": 0.0}, "rouge-l": {"f": 0.0}}
+            else:
+                rouge = Rouge()
+                scores = rouge.get_scores(" ".join(hypothesis), " ".join(reference))
+                result = scores[0]
+
+            for k, v in result.items():
+                self.score_dict[k].append(round(v["f"] * 100, 4))
+
+            bleu_score = sentence_bleu([list(label)], list(pred), smoothing_function=SmoothingFunction().method3)
+            self.score_dict["bleu-4"].append(round(bleu_score * 100, 4))
+
+        if compute_result:
+            return self._dump()

LLaMA-Factory/src/llamafactory/train/sft/trainer.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/trainer_seq2seq.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+from types import MethodType
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import Seq2SeqTrainer
+from typing_extensions import override
+
+from ...extras import logging
+from ...extras.constants import IGNORE_INDEX
+from ...extras.packages import is_transformers_version_equal_to_4_46, is_transformers_version_greater_than
+from ..callbacks import SaveProcessorCallback
+from ..trainer_utils import create_custom_optimizer, create_custom_scheduler
+
+
+if TYPE_CHECKING:
+    from torch.utils.data import Dataset
+    from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
+    from transformers.trainer import PredictionOutput
+
+    from ...hparams import FinetuningArguments
+
+
+logger = logging.get_logger(__name__)
+
+
+class CustomSeq2SeqTrainer(Seq2SeqTrainer):
+    r"""
+    Inherits Seq2SeqTrainer to compute generative metrics such as BLEU and ROUGE.
+    """
+
+    def __init__(
+        self, finetuning_args: "FinetuningArguments", processor: Optional["ProcessorMixin"], **kwargs
+    ) -> None:
+        if is_transformers_version_greater_than("4.46"):
+            kwargs["processing_class"] = kwargs.pop("tokenizer")
+        else:
+            self.processing_class: "PreTrainedTokenizer" = kwargs.get("tokenizer")
+
+        super().__init__(**kwargs)
+        self.finetuning_args = finetuning_args
+
+        if processor is not None:
+            self.add_callback(SaveProcessorCallback(processor))
+
+        if finetuning_args.use_badam:
+            from badam import BAdamCallback, clip_grad_norm_old_version  # type: ignore
+
+            self.accelerator.clip_grad_norm_ = MethodType(clip_grad_norm_old_version, self.accelerator)
+            self.add_callback(BAdamCallback)
+
+    @override
+    def create_optimizer(self) -> "torch.optim.Optimizer":
+        if self.optimizer is None:
+            self.optimizer = create_custom_optimizer(self.model, self.args, self.finetuning_args)
+        return super().create_optimizer()
+
+    @override
+    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)
+
+    @override
+    def _get_train_sampler(self) -> Optional["torch.utils.data.Sampler"]:
+        if self.finetuning_args.disable_shuffling:
+            return torch.utils.data.SequentialSampler(self.train_dataset)
+
+        return super()._get_train_sampler()
+
+    @override
+    def compute_loss(
+        self, model: "PreTrainedModel", inputs: Dict[str, "torch.Tensor"], return_outputs: bool = False, **kwargs
+    ) -> Union["torch.Tensor", Tuple["torch.Tensor", List["torch.Tensor"]]]:
+        r"""
+        Fixes the loss value for transformers 4.46.0.
+        https://github.com/huggingface/transformers/blob/v4.46.0/src/transformers/trainer.py#L3605
+        """
+        loss = super().compute_loss(model, inputs, return_outputs, **kwargs)
+        if is_transformers_version_equal_to_4_46() and not getattr(self, "model_accepts_loss_kwargs", False):
+            # other model should not scale the loss
+            if return_outputs:
+                return (loss[0] / self.args.gradient_accumulation_steps, *loss[1:])
+            else:
+                return loss / self.args.gradient_accumulation_steps
+
+        return loss
+
+    @override
+    def prediction_step(
+        self,
+        model: "torch.nn.Module",
+        inputs: Dict[str, Union["torch.Tensor", Any]],
+        prediction_loss_only: bool,
+        ignore_keys: Optional[List[str]] = None,
+    ) -> Tuple[Optional[float], Optional["torch.Tensor"], Optional["torch.Tensor"]]:
+        r"""
+        Removes the prompt part in the generated tokens.
+
+        Subclass and override to inject custom behavior.
+        """
+        labels = inputs["labels"] if "labels" in inputs else None
+        if self.args.predict_with_generate:
+            assert self.processing_class.padding_side == "left", "This method only accepts left-padded tensor."
+            labels = labels.detach().clone() if labels is not None else None  # backup labels
+            prompt_len, label_len = inputs["input_ids"].size(-1), inputs["labels"].size(-1)
+            if prompt_len > label_len:
+                inputs["labels"] = self._pad_tensors_to_target_len(inputs["labels"], inputs["input_ids"])
+            if label_len > prompt_len:  # truncate the labels instead of padding the inputs (llama2 fp16 compatibility)
+                inputs["labels"] = inputs["labels"][:, :prompt_len]
+
+        loss, generated_tokens, _ = super().prediction_step(  # ignore the returned labels (may be truncated)
+            model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys
+        )
+        if generated_tokens is not None and self.args.predict_with_generate:
+            generated_tokens[:, :prompt_len] = self.processing_class.pad_token_id
+            generated_tokens = generated_tokens.contiguous()
+
+        return loss, generated_tokens, labels
+
+    def _pad_tensors_to_target_len(self, src_tensor: "torch.Tensor", tgt_tensor: "torch.Tensor") -> "torch.Tensor":
+        r"""
+        Pads the tensor to the same length as the target tensor.
+        """
+        assert self.processing_class.pad_token_id is not None, "Pad token is required."
+        padded_tensor = self.processing_class.pad_token_id * torch.ones_like(tgt_tensor)
+        padded_tensor[:, -src_tensor.shape[-1] :] = src_tensor  # adopt left-padding
+        return padded_tensor.contiguous()  # in contiguous memory
+
+    def save_predictions(
+        self, dataset: "Dataset", predict_results: "PredictionOutput", skip_special_tokens: bool = True
+    ) -> None:
+        r"""
+        Saves model predictions to `output_dir`.
+
+        A custom behavior that not contained in Seq2SeqTrainer.
+        """
+        if not self.is_world_process_zero():
+            return
+
+        output_prediction_file = os.path.join(self.args.output_dir, "generated_predictions.jsonl")
+        logger.info_rank0(f"Saving prediction results to {output_prediction_file}")
+
+        labels = np.where(
+            predict_results.label_ids != IGNORE_INDEX, predict_results.label_ids, self.processing_class.pad_token_id
+        )
+        preds = np.where(
+            predict_results.predictions != IGNORE_INDEX,
+            predict_results.predictions,
+            self.processing_class.pad_token_id,
+        )
+
+        for i in range(len(preds)):
+            pad_len = np.nonzero(preds[i] != self.processing_class.pad_token_id)[0]
+            if len(pad_len):  # move pad token to last
+                preds[i] = np.concatenate((preds[i][pad_len[0] :], preds[i][: pad_len[0]]), axis=-1)
+
+        decoded_inputs = self.processing_class.batch_decode(dataset["input_ids"], skip_special_tokens=False)
+        decoded_preds = self.processing_class.batch_decode(preds, skip_special_tokens=skip_special_tokens)
+        decoded_labels = self.processing_class.batch_decode(labels, skip_special_tokens=skip_special_tokens)
+
+        with open(output_prediction_file, "w", encoding="utf-8") as f:
+            for text, pred, label in zip(decoded_inputs, decoded_preds, decoded_labels):
+                f.write(json.dumps({"prompt": text, "predict": pred, "label": label}, ensure_ascii=False) + "\n")

LLaMA-Factory/src/llamafactory/train/sft/workflow.py

+# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+#
+# This code is inspired by the HuggingFace's transformers library.
+# https://github.com/huggingface/transformers/blob/v4.40.0/examples/pytorch/summarization/run_summarization.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING, List, Optional
+
+from ...data import SFTDataCollatorWith4DAttentionMask, get_dataset, get_template_and_fix_tokenizer
+from ...extras.constants import IGNORE_INDEX
+from ...extras.logging import get_logger
+from ...extras.misc import calculate_tps, get_logits_processor
+from ...extras.ploting import plot_loss
+from ...model import load_model, load_tokenizer
+from ..trainer_utils import create_modelcard_and_push
+from .metric import ComputeAccuracy, ComputeSimilarity, eval_logit_processor
+from .trainer import CustomSeq2SeqTrainer
+
+
+if TYPE_CHECKING:
+    from transformers import Seq2SeqTrainingArguments, TrainerCallback
+
+    from ...hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments
+
+
+logger = get_logger(__name__)
+
+
+def run_sft(
+    model_args: "ModelArguments",
+    data_args: "DataArguments",
+    training_args: "Seq2SeqTrainingArguments",
+    finetuning_args: "FinetuningArguments",
+    generating_args: "GeneratingArguments",
+    callbacks: Optional[List["TrainerCallback"]] = None,
+):
+    tokenizer_module = load_tokenizer(model_args)
+    tokenizer = tokenizer_module["tokenizer"]
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="sft", **tokenizer_module)
+    model = load_model(tokenizer, model_args, finetuning_args, training_args.do_train)
+
+    if getattr(model, "is_quantized", False) and not training_args.do_train:
+        setattr(model, "_hf_peft_config_loaded", True)  # hack here: make model compatible with prediction
+
+    data_collator = SFTDataCollatorWith4DAttentionMask(
+        template=template,
+        model=model if not training_args.predict_with_generate else None,
+        pad_to_multiple_of=8 if training_args.do_train else None,  # for shift short attention
+        label_pad_token_id=IGNORE_INDEX if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id,
+        block_diag_attn=model_args.block_diag_attn,
+        attn_implementation=getattr(model.config, "_attn_implementation", None),
+        compute_dtype=model_args.compute_dtype,
+        **tokenizer_module,
+    )
+
+    # Override the decoding parameters of Seq2SeqTrainer
+    training_args.generation_max_length = training_args.generation_max_length or data_args.cutoff_len
+    training_args.generation_num_beams = data_args.eval_num_beams or training_args.generation_num_beams
+    training_args.remove_unused_columns = False  # important for multimodal dataset
+
+    # Metric utils
+    metric_module = {}
+    if training_args.predict_with_generate:
+        metric_module["compute_metrics"] = ComputeSimilarity(tokenizer=tokenizer)
+    elif finetuning_args.compute_accuracy:
+        metric_module["compute_metrics"] = ComputeAccuracy()
+        metric_module["preprocess_logits_for_metrics"] = eval_logit_processor
+
+    # Initialize our Trainer
+    trainer = CustomSeq2SeqTrainer(
+        model=model,
+        args=training_args,
+        finetuning_args=finetuning_args,
+        data_collator=data_collator,
+        callbacks=callbacks,
+        **dataset_module,
+        **tokenizer_module,
+        **metric_module,
+    )
+
+    # Keyword arguments for `model.generate`
+    gen_kwargs = generating_args.to_dict(obey_generation_config=True)
+    gen_kwargs["eos_token_id"] = [tokenizer.eos_token_id] + tokenizer.additional_special_tokens_ids
+    gen_kwargs["pad_token_id"] = tokenizer.pad_token_id
+    gen_kwargs["logits_processor"] = get_logits_processor()
+
+    # Training
+    if training_args.do_train:
+        train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
+        trainer.save_model()
+        if finetuning_args.include_effective_tokens_per_second:
+            train_result.metrics["effective_tokens_per_sec"] = calculate_tps(
+                dataset_module["train_dataset"], train_result.metrics, stage="sft"
+            )
+
+        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", "eval_accuracy"])
+
+    if training_args.predict_with_generate:
+        tokenizer.padding_side = "left"  # use left-padding in generation
+
+    # Evaluation
+    if training_args.do_eval:
+        metrics = trainer.evaluate(metric_key_prefix="eval", **gen_kwargs)
+        if training_args.predict_with_generate:  # eval_loss will be wrong if predict_with_generate is enabled
+            metrics.pop("eval_loss", None)
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Predict
+    if training_args.do_predict:
+        logger.warning_once("Batch generation can be very slow. Consider using `scripts/vllm_infer.py` instead.")
+        predict_results = trainer.predict(dataset_module["eval_dataset"], metric_key_prefix="predict", **gen_kwargs)
+        if training_args.predict_with_generate:  # predict_loss will be wrong if predict_with_generate is enabled
+            predict_results.metrics.pop("predict_loss", None)
+        trainer.log_metrics("predict", predict_results.metrics)
+        trainer.save_metrics("predict", predict_results.metrics)
+        trainer.save_predictions(dataset_module["eval_dataset"], predict_results, generating_args.skip_special_tokens)
+
+    # Create model card
+    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)