train in multi gpus in ddp

[azuredsky]

我修改了tools/train.py 如下内容
strategy = "ddp"

devices=[0,1,2,3]

trainer = pl.Trainer(
    default_root_dir=cfg.save_dir,
    max_epochs=cfg.schedule.total_epochs,
    check_val_every_n_epoch=cfg.schedule.val_intervals,
    accelerator=accelerator,
    devices=devices,
    gpus=len(devices),
    log_every_n_steps=cfg.log.interval,
    num_sanity_val_steps=0,
    callbacks=[TQDMProgressBar(refresh_rate=0)],  # disable tqdm bar
    logger=logger,
    benchmark=cfg.get("cudnn_benchmark", True),
    gradient_clip_val=cfg.get("grad_clip", 0.0),
    strategy=strategy,
    precision=precision,
)

训练后只能够使用单GPU，不知道我哪里设置的不对？
环境如下：
pytorch-lightning 1.9.5
pytorch 1.13.1 py3.10_cuda11.6_cudnn8.3.2_0 pytorch