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

BienBoy

2025-02-01 c1e365fea09aafda387cac12fdff43d28c598979

 funasr/bin/train.py

@@ -20,6 +20,7 @@
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
@@ -27,7 +28,7 @@
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
@@ -191,8 +192,6 @@
    tensorboard_dir = os.path.join(kwargs.get("output_dir"), "tensorboard")
    os.makedirs(tensorboard_dir, exist_ok=True)
    try:
        from tensorboardX import SummaryWriter

        writer = SummaryWriter(tensorboard_dir)  # if trainer.rank == 0 else None
    except:
        writer = None
@@ -222,14 +221,15 @@
            )
            trainer.start_step = 0

            torch.cuda.empty_cache()
            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)*time_escaped:.3f} hours"
                f"epoch: {((trainer.max_epoch - epoch - 1)*dataloader.data_split_num + dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours\n"
                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