fix: resolve unexpected 'out of memory' issue in multi-GPU setup (#2373)

BienBoy

2025-02-01 c1e365fea09aafda387cac12fdff43d28c598979

 funasr/bin/train.py

@@ -1,115 +1,106 @@
#!/usr/bin/env python3
# -*- encoding: utf-8 -*-

import os
import sys
import torch
import torch.nn as nn
import hydra
import logging
import time
import argparse
from io import BytesIO

from contextlib import nullcontext
import torch.distributed as dist
from collections.abc import Sequence

from omegaconf import DictConfig, OmegaConf
from torch.cuda.amp import autocast, GradScaler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.algorithms.join import Join
from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler
from tensorboardX import SummaryWriter
from funasr.train_utils.average_nbest_models import average_checkpoints

from funasr.register import tables
from funasr.optimizers import optim_classes
from funasr.train_utils.trainer import Trainer
from funasr.schedulers import scheduler_classes
from funasr.train_utils.initialize import initialize
from funasr.download.download_from_hub import download_model
from funasr.download.download_model_from_hub import download_model
from funasr.models.lora.utils import mark_only_lora_as_trainable
from funasr.train_utils.set_all_random_seed import set_all_random_seed
from funasr.train_utils.load_pretrained_model import load_pretrained_model
# from funasr.tokenizer.build_tokenizer import build_tokenizer
# from funasr.tokenizer.token_id_converter import TokenIDConverter
# from funasr.tokenizer.funtoken import build_tokenizer
from funasr.utils.misc import prepare_model_dir
from funasr.train_utils.model_summary import model_summary
from funasr import AutoModel


@hydra.main(config_name=None, version_base=None)
def main_hydra(kwargs: DictConfig):
    if kwargs.get("debug", False):
        import pdb; pdb.set_trace()
        import pdb

        pdb.set_trace()

    assert "model" in kwargs
    if "model_conf" not in kwargs:
        logging.info("download models from model hub: {}".format(kwargs.get("model_hub", "ms")))
        logging.info("download models from model hub: {}".format(kwargs.get("hub", "ms")))
        kwargs = download_model(is_training=kwargs.get("is_training", True), **kwargs)
    

    main(**kwargs)


def main(**kwargs):
    

    # set random seed
    tables.print()
    set_all_random_seed(kwargs.get("seed", 0))
    torch.backends.cudnn.enabled = kwargs.get("cudnn_enabled", torch.backends.cudnn.enabled)
    torch.backends.cudnn.benchmark = kwargs.get("cudnn_benchmark", torch.backends.cudnn.benchmark)
    torch.backends.cudnn.deterministic = kwargs.get("cudnn_deterministic", True)
    
    local_rank = int(os.environ.get('LOCAL_RANK', 0))
    # open tf32
    torch.backends.cuda.matmul.allow_tf32 = kwargs.get("enable_tf32", True)

    local_rank = int(os.environ.get("LOCAL_RANK", 0))
    if local_rank == 0:
        tables.print()
    # Check if we are using DDP or FSDP
    use_ddp = 'WORLD_SIZE' in os.environ and int(os.environ["WORLD_SIZE"]) > 1
    use_fsdp = kwargs.get("use_fsdp", None)
    use_ddp = "WORLD_SIZE" in os.environ and int(os.environ["WORLD_SIZE"]) > 1
    use_fsdp = kwargs.get("use_fsdp", False)
    # use_ddp = False if use_fsdp else use_fsdp
    if use_ddp or use_fsdp:
        dist.init_process_group(backend=kwargs.get("backend", "nccl"), init_method='env://')
        dist.init_process_group(backend=kwargs.get("backend", "nccl"), init_method="env://")
        torch.cuda.set_device(local_rank)
    

    logging.info("Build model, frontend, tokenizer")
    device = kwargs.get("device", "cuda")
    kwargs["device"] = "cpu"
    model = AutoModel(**kwargs)

    # save config.yaml
    if (use_ddp or use_fsdp) and dist.get_rank() == 0 or not (use_ddp or use_fsdp) and local_rank == 0:
        os.makedirs(kwargs.get("output_dir", "./"), exist_ok=True)
        yaml_file = os.path.join(kwargs.get("output_dir", "./"), "config.yaml")
        OmegaConf.save(config=kwargs, f=yaml_file)
        logging.info("config.yaml is saved to: %s", yaml_file)
    if (
        (use_ddp or use_fsdp)
        and dist.get_rank() == 0
        or not (use_ddp or use_fsdp)
        and local_rank == 0
    ):
        prepare_model_dir(**kwargs)

    tokenizer = kwargs.get("tokenizer", None)
    if tokenizer is not None:
        tokenizer_class = tables.tokenizer_classes.get(tokenizer)
        tokenizer = tokenizer_class(**kwargs["tokenizer_conf"])
        kwargs["tokenizer"] = tokenizer
    
    # build frontend if frontend is none None
    frontend = kwargs.get("frontend", None)
    if frontend is not None:
        frontend_class = tables.frontend_classes.get(frontend)
        frontend = frontend_class(**kwargs["frontend_conf"])
        kwargs["frontend"] = frontend
        kwargs["input_size"] = frontend.output_size()
    
    # import pdb;
    # pdb.set_trace()
    # build model
    model_class = tables.model_classes.get(kwargs["model"])
    model = model_class(**kwargs, **kwargs["model_conf"], vocab_size=len(tokenizer.token_list))



    # init_param
    init_param = kwargs.get("init_param", None)
    if init_param is not None:
        if not isinstance(init_param, (list, tuple)):
            init_param = (init_param,)
        logging.info("init_param is not None: %s", init_param)
        for p in init_param:
            logging.info(f"Loading pretrained params from {p}")
            load_pretrained_model(
                model=model,
                path=p,
                ignore_init_mismatch=kwargs.get("ignore_init_mismatch", True),
                oss_bucket=kwargs.get("oss_bucket", None),
                scope_map=kwargs.get("scope_map", None),
                excludes=kwargs.get("excludes", None),
            )
    else:
        initialize(model, kwargs.get("init", "kaiming_normal"))

    # parse kwargs
    kwargs = model.kwargs
    kwargs["device"] = device
    tokenizer = kwargs["tokenizer"]
    frontend = kwargs["frontend"]
    model = model.model
    del kwargs["model"]

    # freeze_param
    freeze_param = kwargs.get("freeze_param", None)
    if freeze_param is not None:
        freeze_param = eval(freeze_param)
        if isinstance(freeze_param, Sequence):
        if "," in freeze_param:
            freeze_param = eval(freeze_param)
        if not isinstance(freeze_param, (list, tuple)):
            freeze_param = (freeze_param,)
        logging.info("freeze_param is not None: %s", freeze_param)
        for t in freeze_param:
@@ -117,67 +108,156 @@
                if k.startswith(t + ".") or k == t:
                    logging.info(f"Setting {k}.requires_grad = False")
                    p.requires_grad = False
    
    if local_rank == 0:
        logging.info(f"{model_summary(model)}")

    if use_ddp:
        model = model.cuda(local_rank)
        model = DDP(model, device_ids=[local_rank],
                    find_unused_parameters=kwargs.get("train_conf", {}).get("find_unused_parameters", False))
        model = DDP(
            model,
            device_ids=[local_rank],
            find_unused_parameters=kwargs.get("train_conf", {}).get(
                "find_unused_parameters", False
            ),
        )
    elif use_fsdp:
        model = FSDP(model).cuda(local_rank)
        # model = FSDP(model).cuda(local_rank)

        def custom_auto_wrap_policy(
            module: nn.Module,
            recurse: bool,
            nonwrapped_numel: int,
            # Additional custom arguments
            min_num_params: int = int(1e8),
        ) -> bool:
            # 根据自定义逻辑决定是否包装模块
            is_large = unwrapped_params >= min_num_params
            requires_grad_uniform = len({p.requires_grad for p in module.parameters()}) == 1
            return is_large and requires_grad_uniform

        # Configure a custom `min_num_params`
        my_auto_wrap_policy = functools.partial(custom_auto_wrap_policy, min_num_params=int(1e5))
        torch.cuda.set_device(local_rank)
        model = FSDP(
            model,
            auto_wrap_policy=custom_auto_wrap_policy,
            mixed_precision=None,
            device_id=torch.cuda.current_device(),
        )
    else:
        model = model.to(device=kwargs.get("device", "cuda"))
        
        

    kwargs["device"] = next(model.parameters()).device

    # optim
    logging.info("Build optim")
    optim = kwargs.get("optim", "adam")
    assert optim in optim_classes
    optim_class = optim_classes.get(optim)
    optim = optim_class(model.parameters(), **kwargs.get("optim_conf"))
    

    # scheduler
    logging.info("Build scheduler")
    scheduler = kwargs.get("scheduler", "warmuplr")
    assert scheduler in scheduler_classes
    scheduler_class = scheduler_classes.get(scheduler)
    scheduler = scheduler_class(optim, **kwargs.get("scheduler_conf"))

    # import pdb;
    # pdb.set_trace()
    # dataset
    dataset_class = tables.dataset_classes.get(kwargs.get("dataset", "AudioDataset"))
    dataset_tr = dataset_class(kwargs.get("train_data_set_list"), frontend=frontend, tokenizer=tokenizer, **kwargs.get("dataset_conf"))

    # dataloader
    batch_sampler = kwargs["dataset_conf"].get("batch_sampler", "DynamicBatchLocalShuffleSampler")
    batch_sampler_class = tables.batch_sampler_classes.get(batch_sampler)
    if batch_sampler is not None:
        batch_sampler = batch_sampler_class(dataset_tr, **kwargs.get("dataset_conf"))
    dataloader_tr = torch.utils.data.DataLoader(dataset_tr,
                                                collate_fn=dataset_tr.collator,
                                                batch_sampler=batch_sampler,
                                                num_workers=kwargs.get("dataset_conf").get("num_workers", 4),
                                                pin_memory=True)
    

    logging.info("Build dataloader")
    dataloader_class = tables.dataloader_classes.get(
        kwargs["dataset_conf"].get("dataloader", "DataloaderMapStyle")
    )
    dataloader = dataloader_class(**kwargs)
    # dataloader_tr, dataloader_val = dataloader_class(**kwargs)
    trainer = Trainer(
        model=model,
        optim=optim,
        scheduler=scheduler,
        dataloader_train=dataloader_tr,
        dataloader_val=None,
        local_rank=local_rank,
        use_ddp=use_ddp,
        use_fsdp=use_fsdp,
        device=kwargs["device"],
        output_dir=kwargs.get("output_dir", "./exp"),
        resume=kwargs.get("resume", True),
        **kwargs.get("train_conf"),
    )
    trainer.run()
    
    if use_ddp or use_fsdp:
        torch.distributed.destroy_process_group()

    
    scaler = GradScaler(enabled=trainer.use_fp16) if trainer.use_fp16 else None
    scaler = ShardedGradScaler(enabled=trainer.use_fp16) if trainer.use_fsdp else scaler

    trainer.resume_checkpoint(
        model=model,
        optim=optim,
        scheduler=scheduler,
        scaler=scaler,
    )

    tensorboard_dir = os.path.join(kwargs.get("output_dir"), "tensorboard")
    os.makedirs(tensorboard_dir, exist_ok=True)
    try:
        writer = SummaryWriter(tensorboard_dir)  # if trainer.rank == 0 else None
    except:
        writer = None

    dataloader_tr, dataloader_val = None, None
    for epoch in range(trainer.start_epoch, trainer.max_epoch):
        time1 = time.perf_counter()

        for data_split_i in range(trainer.start_data_split_i, dataloader.data_split_num):
            time_slice_i = time.perf_counter()
            dataloader_tr, dataloader_val = dataloader.build_iter(
                epoch, data_split_i=data_split_i, start_step=trainer.start_step
            )

            trainer.train_epoch(
                model=model,
                optim=optim,
                scheduler=scheduler,
                scaler=scaler,
                dataloader_train=dataloader_tr,
                dataloader_val=dataloader_val,
                epoch=epoch,
                writer=writer,
                data_split_i=data_split_i,
                data_split_num=dataloader.data_split_num,
                start_step=trainer.start_step,
            )
            trainer.start_step = 0

            with torch.cuda.device(kwargs["device"]):
                torch.cuda.empty_cache()

            time_escaped = (time.perf_counter() - time_slice_i) / 3600.0
            logging.info(
                f"rank: {local_rank}, "
                f"time_escaped_epoch: {time_escaped:.3f} hours, "
                f"estimated to finish {dataloader.data_split_num} data_slices, remaining: {dataloader.data_split_num-data_split_i} slices, {(dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours, "
                f"epoch: {trainer.max_epoch - epoch} epochs, {((trainer.max_epoch - epoch - 1)*dataloader.data_split_num + dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours\n"
            )

        trainer.start_data_split_i = 0
        trainer.validate_epoch(
            model=model, dataloader_val=dataloader_val, epoch=epoch + 1, writer=writer
        )
        scheduler.step()
        trainer.step_in_epoch = 0
        trainer.save_checkpoint(
            epoch + 1, model=model, optim=optim, scheduler=scheduler, scaler=scaler
        )

        time2 = time.perf_counter()
        time_escaped = (time2 - time1) / 3600.0
        logging.info(
            f"rank: {local_rank}, "
            f"time_escaped_epoch: {time_escaped:.3f} hours, "
            f"estimated to finish {trainer.max_epoch} "
            f"epoch: {(trainer.max_epoch - epoch) * time_escaped:.3f} hours\n"
        )
        trainer.train_acc_avg = 0.0
        trainer.train_loss_avg = 0.0

    if trainer.rank == 0:
        average_checkpoints(trainer.output_dir, trainer.avg_nbest_model)

    trainer.close()


if __name__ == "__main__":
    main_hydra()
    main_hydra()