src/train/mmsft/trainer.py

# Copyright 2024 HuggingFace Inc., the LlamaFactory team, and the Llamole 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 warnings
import numpy as np
import torch
from transformers import Seq2SeqTrainer

from ...extras.constants import IGNORE_INDEX
from ...extras.logging import get_logger
from ..callbacks import PissaConvertCallback, SaveProcessorCallback
from ..trainer_utils import create_custom_optimzer, create_custom_scheduler


if TYPE_CHECKING:
    import optuna
    from torch.utils.data import Dataset
    from transformers import ProcessorMixin
    from transformers.trainer import PredictionOutput

    from ...hparams import FinetuningArguments

from transformers.trainer_utils import (
    enable_full_determinism,
    find_executable_batch_size,
    get_last_checkpoint,
    set_seed,
)

import huggingface_hub.utils as hf_hub_utils
from transformers.utils import is_sagemaker_mp_enabled
from transformers.trainer_callback import TrainerState

TRAINER_STATE_NAME = "trainer_state.json"

logger = 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:
        super().__init__(**kwargs)
        self.finetuning_args = finetuning_args

        if processor is not None:
            self.add_callback(SaveProcessorCallback(processor))

        if finetuning_args.pissa_convert:
            self.add_callback(PissaConvertCallback)


    def create_optimizer(self) -> "torch.optim.Optimizer":
        if self.optimizer is None:
            self.optimizer = create_custom_optimzer(self.model, self.args, self.finetuning_args)
        return super().create_optimizer()

    def create_scheduler(
        self, num_training_steps: int, optimizer: Optional["torch.optim.Optimizer"] = None
    ) -> "torch.optim.lr_scheduler.LRScheduler":
        create_custom_scheduler(self.args, num_training_steps, optimizer)
        return super().create_scheduler(num_training_steps, optimizer)

    def 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"].detach().clone() if "labels" in inputs else None  # backup labels
        if self.args.predict_with_generate:
            assert self.tokenizer.padding_side == "left", "This method only accepts left-padded tensor."
            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.tokenizer.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.tokenizer.pad_token_id is not None, "Pad token is required."
        padded_tensor = self.tokenizer.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") -> 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(f"Saving prediction results to {output_prediction_file}")

        labels = np.where(
            predict_results.label_ids != IGNORE_INDEX, predict_results.label_ids, self.tokenizer.pad_token_id
        )
        preds = np.where(
            predict_results.predictions != IGNORE_INDEX, predict_results.predictions, self.tokenizer.pad_token_id
        )

        for i in range(len(preds)):
            pad_len = np.nonzero(preds[i] != self.tokenizer.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.tokenizer.batch_decode(dataset["input_ids"], skip_special_tokens=True)
        decoded_labels = self.tokenizer.batch_decode(labels, skip_special_tokens=True)
        decoded_preds = self.tokenizer.batch_decode(preds, skip_special_tokens=True)

        with open(output_prediction_file, "w", encoding="utf-8") as writer:
            res: List[str] = []
            for text, label, pred in zip(decoded_inputs, decoded_labels, decoded_preds):
                res.append(json.dumps({"prompt": text, "label": label, "predict": pred}, ensure_ascii=False))

            writer.write("\n".join(res))

    def train(
        self,
        resume_from_checkpoint: Optional[Union[str, bool]] = None,
        trial: Union["optuna.Trial", Dict[str, Any]] = None,
        ignore_keys_for_eval: Optional[List[str]] = None,
        **kwargs,
    ):
        """
        Main training entry point.

        Args:
            resume_from_checkpoint (`str` or `bool`, *optional*):
                If a `str`, local path to a saved checkpoint as saved by a previous instance of [`Trainer`]. If a
                `bool` and equals `True`, load the last checkpoint in *args.output_dir* as saved by a previous instance
                of [`Trainer`]. If present, training will resume from the model/optimizer/scheduler states loaded here.
            trial (`optuna.Trial` or `Dict[str, Any]`, *optional*):
                The trial run or the hyperparameter dictionary for hyperparameter search.
            ignore_keys_for_eval (`List[str]`, *optional*)
                A list of keys in the output of your model (if it is a dictionary) that should be ignored when
                gathering predictions for evaluation during the training.
            kwargs (`Dict[str, Any]`, *optional*):
                Additional keyword arguments used to hide deprecated arguments
        """
        if resume_from_checkpoint is False:
            resume_from_checkpoint = None

        # memory metrics - must set up as early as possible
        self._memory_tracker.start()

        args = self.args

        self.is_in_train = True

        # Attach NEFTune hooks if necessary
        if self.neftune_noise_alpha is not None:
            self.model = self._activate_neftune(self.model)

        # do_train is not a reliable argument, as it might not be set and .train() still called, so
        # the following is a workaround:
        if (args.fp16_full_eval or args.bf16_full_eval) and not args.do_train:
            self._move_model_to_device(self.model, args.device)

        if "model_path" in kwargs:
            resume_from_checkpoint = kwargs.pop("model_path")
            warnings.warn(
                "`model_path` is deprecated and will be removed in a future version. Use `resume_from_checkpoint` "
                "instead.",
                FutureWarning,
            )
        if len(kwargs) > 0:
            raise TypeError(f"train() received got unexpected keyword arguments: {', '.join(list(kwargs.keys()))}.")
        # This might change the seed so needs to run first.
        self._hp_search_setup(trial)
        self._train_batch_size = self.args.train_batch_size

        # Model re-init
        model_reloaded = False
        if self.model_init is not None:
            # Seed must be set before instantiating the model when using model_init.
            enable_full_determinism(self.args.seed) if self.args.full_determinism else set_seed(self.args.seed)
            self.model = self.call_model_init(trial)
            model_reloaded = True
            # Reinitializes optimizer and scheduler
            self.optimizer, self.lr_scheduler = None, None

        # Load potential model checkpoint
        if isinstance(resume_from_checkpoint, bool) and resume_from_checkpoint:
            resume_from_checkpoint = get_last_checkpoint(args.output_dir)
            if resume_from_checkpoint is None:
                raise ValueError(f"No valid checkpoint found in output directory ({args.output_dir})")

        if resume_from_checkpoint is not None:
            if not is_sagemaker_mp_enabled() and not self.is_deepspeed_enabled and not self.is_fsdp_enabled:
                self._load_from_checkpoint(resume_from_checkpoint, self.model.language_model)
            # In case of repeating the find_executable_batch_size, set `self._train_batch_size` properly
            state = TrainerState.load_from_json(os.path.join(resume_from_checkpoint, TRAINER_STATE_NAME))
            if state.train_batch_size is not None:
                self._train_batch_size = state.train_batch_size

        # If model was re-initialized, put it on the right device and update self.model_wrapped
        if model_reloaded:
            if self.place_model_on_device:
                self._move_model_to_device(self.model, args.device)
            self.model_wrapped = self.model

        inner_training_loop = find_executable_batch_size(
            self._inner_training_loop, self._train_batch_size, args.auto_find_batch_size
        )

        return inner_training_loop(
            args=args,
            resume_from_checkpoint=resume_from_checkpoint,
            trial=trial,
            ignore_keys_for_eval=ignore_keys_for_eval,
        )