经过代码的重重排查，终于找到了问题代码块，但是还是不知道具体原因

源代码：

def train(epoch, dataloader, model, criterion, optimizer, scaler: GradScaler, logger):
    losses = AverageMeter('Loss', ':.4f')
    top1 = AverageMeter('Acc@1', ':6.2f')
    top5 = AverageMeter('Acc@5', ':6.2f')
    model.train()

    end = time.time()
    loss_list = torch.zeros(len(dataloader.dataset))
    for i, (images, labels, index) in enumerate(dataloader):
        images = images.cuda(non_blocking=True)
        labels = labels.cuda(non_blocking=True)

        optimizer.zero_grad()
        # Forward + Backward + Optimize
        with torch.autocast(device_type="cuda"):
            _, _, logits = model(images)
            loss = criterion(logits, labels)

        for j in range(images.size(0)):
            loss_list[index[j]] = loss[j]
        loss = loss.mean()
        scaler.scale(loss).backward()
        scaler.step(optimizer)
        scaler.update()

        acc1, acc5 = accuracy(logits.cpu(), labels.cpu(), topk=(1, 5))
        batch_size = images.size(0)
        top1.update(acc1[0], batch_size)
        top5.update(acc5[0], batch_size)
        losses.update(loss.item(), batch_size)

        if (i + 1) % 50 == 0:
            logger.info(f"Epoch[{epoch}][{i}/{len(dataloader)}] "
                        f"lr: {optimizer.param_groups[-1]['lr']: .5f}\t"
                        f"Loss {losses.val:.3f} ({losses.avg:.3f})\t"
                        f"Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t"
                        f"Acc@5 {top5.val:.3f} ({top5.avg:.3f})\t")

    logger.info(f'Train1 Summary '
                f'Time cost {time.time() - end:.3f}s\t'
                f'Loss {losses.avg:.3f}\t'
                f'Acc@1 {top1.avg:.3f}\t'
                f'Acc@5 {top5.avg:.3f}\t')
    print(f'{epoch} end train')
    return loss_list

loss_list = torch.zeros改成loss_list = np.zeros后解决了，如果收集的loss不需要梯度可以进一步改进：

for j in range(batch_size):
  loss_list[index[j]] = loss[j].item()