src/llamafactory/train/hyper_parallel/workflow.py

# Copyright 2025 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 math
from typing import TYPE_CHECKING, Optional

from transformers import DataCollatorForLanguageModeling

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
from ...extras.packages import is_hyper_parallel_available, is_transformers_version_greater_than
from ...extras.ploting import plot_loss
from ...model import load_model, load_tokenizer
from ..sft.metric import ComputeAccuracy, ComputeSimilarity, eval_logit_processor
from ..trainer_utils import create_modelcard_and_push, create_ref_model
from .trainer import HyperParallelTrainer


if TYPE_CHECKING:
    from transformers import Seq2SeqTrainingArguments, TrainerCallback

    from ...hparams import DataArguments, FinetuningArguments, GeneratingArguments, ModelArguments


logger = get_logger(__name__)


def _prepare_hp_args(finetuning_args: "FinetuningArguments", model_args: "ModelArguments"):
    r"""Load HyperParallel arguments and apply LlamaFactory-side overrides.

    When activation optimization is enabled, skip native gradient checkpointing
    so HP can install its own via ``setup_activation_optimization``.
    """
    if not is_hyper_parallel_available():
        raise ImportError("hyper_parallel is not installed. Please install it with `pip install hyper_parallel`.")

    from hyper_parallel.integration.llamafactory import HyperParallelArguments  # pylint: disable=C0415

    hp_args = HyperParallelArguments.from_finetuning_args(finetuning_args)

    if getattr(hp_args, "cp_size", None) != finetuning_args.hyper_parallel_cp_size:
        setattr(hp_args, "cp_size", finetuning_args.hyper_parallel_cp_size)

    if hp_args.activation_mode != "none":
        model_args.disable_gradient_checkpointing = True
    return hp_args


def run_pt(
    model_args: "ModelArguments",
    data_args: "DataArguments",
    training_args: "Seq2SeqTrainingArguments",
    finetuning_args: "FinetuningArguments",
    callbacks: Optional[list["TrainerCallback"]] = None,
):
    hp_args = _prepare_hp_args(finetuning_args, model_args)

    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)

    trainer = HyperParallelTrainer(
        hp_args=hp_args,
        model=model,
        args=training_args,
        finetuning_args=finetuning_args,
        data_collator=data_collator,
        callbacks=callbacks,
        **dataset_module,
        **tokenizer_module,
    )

    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:
            keys = ["loss"]
            if isinstance(dataset_module.get("eval_dataset"), dict):
                keys += [f"eval_{key}_loss" for key in dataset_module["eval_dataset"].keys()]
            else:
                keys += ["eval_loss"]

            plot_loss(training_args.output_dir, keys=keys)

    if training_args.do_eval:
        metrics = trainer.evaluate(metric_key_prefix="eval")

        if isinstance(dataset_module.get("eval_dataset"), dict):
            for key in dataset_module["eval_dataset"].keys():
                try:
                    perplexity = math.exp(metrics[f"eval_{key}_loss"])
                except OverflowError:
                    perplexity = float("inf")

                metrics[f"eval_{key}_perplexity"] = perplexity
        else:
            try:
                perplexity = math.exp(metrics["eval_loss"])
            except OverflowError:
                perplexity = float("inf")

            metrics["eval_perplexity"] = perplexity

        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    create_modelcard_and_push(trainer, model_args, data_args, training_args, finetuning_args)


def run_sft(
    model_args: "ModelArguments",
    data_args: "DataArguments",
    training_args: "Seq2SeqTrainingArguments",
    finetuning_args: "FinetuningArguments",
    generating_args: "GeneratingArguments",
    callbacks: Optional[list["TrainerCallback"]] = None,
):
    hp_args = _prepare_hp_args(finetuning_args, model_args)

    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)

    ref_model = None
    if finetuning_args.use_asft_loss:
        ref_model = create_ref_model(model_args, finetuning_args)

    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,
        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,
    )

    # 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

    # Keyword arguments for `model.generate`
    gen_kwargs = generating_args.to_dict(obey_generation_config=True)
    if is_transformers_version_greater_than("4.58.0"):
        extra_ids = getattr(tokenizer, "additional_special_tokens_ids", None)
        if not isinstance(extra_ids, list):
            extra_special_tokens = getattr(tokenizer, "_extra_special_tokens", [])
            string_tokens = [str(t) for t in extra_special_tokens]
            extra_ids = tokenizer.convert_tokens_to_ids(string_tokens)
        all_eos_ids = [tokenizer.eos_token_id] + [i for i in extra_ids if i != -1]
        gen_kwargs["eos_token_id"] = list(dict.fromkeys(all_eos_ids))
    else:
        gen_kwargs["eos_token_id"] = [tokenizer.eos_token_id] + tokenizer.additional_special_tokens_ids
    gen_kwargs["pad_token_id"] = tokenizer.pad_token_id

    trainer = HyperParallelTrainer(
        hp_args=hp_args,
        model=model,
        args=training_args,
        finetuning_args=finetuning_args,
        data_collator=data_collator,
        callbacks=callbacks,
        gen_kwargs=gen_kwargs,
        ref_model=ref_model,
        **dataset_module,
        **tokenizer_module,
        **metric_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="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:
            keys = ["loss"]
            if isinstance(dataset_module.get("eval_dataset"), dict):
                keys += sum(
                    [[f"eval_{key}_loss", f"eval_{key}_accuracy"] for key in dataset_module["eval_dataset"].keys()],
                    [],
                )
            else:
                keys += ["eval_loss", "eval_accuracy"]

            plot_loss(training_args.output_dir, keys=keys)

    if training_args.predict_with_generate:
        tokenizer.padding_side = "left"

    # Evaluation
    if training_args.do_eval:
        metrics = trainer.evaluate(metric_key_prefix="eval", **gen_kwargs)
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    # Predict
    if training_args.do_predict:
        logger.warning_rank0_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)
        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)