单机多卡训练rt-detrv2-r101，loss反向传播报错ValueError: (InvalidArgument) Required tensor shall not be nullptr, but received nullptr.

[DoctorDream]

问题确认 Search before asking

• 我已经查询历史issue，没有发现相似的bug。I have searched the issues and found no similar bug report.

Bug组件 Bug Component

Training

Bug描述 Describe the Bug

当我使用下述指令训练rt-detr的时候：

python -m paddle.distributed.launch --gpus 0,1,2 tools/train.py -c configs/rtdetrv2/rtdetrv2_r101vd_6x_coco.yml --fleet --eval

会出现报错：

Traceback (most recent call last):
  File "/home/zqy/zqy/Codes/PaddleDetection/tools/train.py", line 209, in <module>
    main()
  File "/home/zqy/zqy/Codes/PaddleDetection/tools/train.py", line 205, in main
    run(FLAGS, cfg)
  File "/home/zqy/zqy/Codes/PaddleDetection/tools/train.py", line 158, in run
    trainer.train(FLAGS.eval)
  File "/home/zqy/zqy/Codes/PaddleDetection/ppdet/engine/trainer.py", line 614, in train
    loss.backward()
  File "/usr/local/lib/python3.10/dist-packages/decorator.py", line 232, in fun
    return caller(func, *(extras + args), **kw)
  File "/usr/local/lib/python3.10/dist-packages/paddle/base/wrapped_decorator.py", line 26, in __impl__
    return wrapped_func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/paddle/base/framework.py", line 593, in __impl__
    return func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/paddle/base/dygraph/tensor_patch_methods.py", line 342, in backward
    core.eager.run_backward([self], grad_tensor, retain_graph)
ValueError: (InvalidArgument) Required tensor shall not be nullptr, but received nullptr.
  [Hint: tensor should not be null.] (at ../paddle/phi/core/device_context.cc:142)

当我使用单卡训练的时候就不会报错了

复现环境 Environment

• OS：Linux
• PaddlePaddle: paddlepaddle-gpu 2.6.1.post117和 paddlepaddle-gpu 2.6.0.post117
• PaddleDetection: develop/2.7
• python: 3.10
• CUDA: 11.7

Bug描述确认 Bug description confirmation

• 我确认已经提供了Bug复现步骤、代码改动说明、以及环境信息，确认问题是可以复现的。I confirm that the bug replication steps, code change instructions, and environment information have been provided, and the problem can be reproduced.

是否愿意提交PR？ Are you willing to submit a PR?

• 我愿意提交PR！I'd like to help by submitting a PR!

[Sunting78]

您好，可以切换release/2.7.1试一下

[DoctorDream]

您好，可以切换release/2.7.1试一下

您好，我切换到PaddleDetection:release/2.7.1分支后，configs中并没有rtdetrv2文件夹，当我按照 release/2.7.0 分支的下述指令：

python -m paddle.distributed.launch --gpus 0,1,2 tools/train.py -c configs/rtdetrv2/rtdetrv2_r101vd_6x_coco.yml --fleet --eval

出现了如下报错：

Traceback (most recent call last):
  File "/home/zqy/zqy/Codes/PaddleDetection-release-2.7.1/tools/train.py", line 213, in <module>
    main()
  File "/home/zqy/zqy/Codes/PaddleDetection-release-2.7.1/tools/train.py", line 209, in main
    run(FLAGS, cfg)
  File "/home/zqy/zqy/Codes/PaddleDetection-release-2.7.1/tools/train.py", line 149, in run
    trainer = Trainer(cfg, mode='train')
  File "/home/zqy/zqy/Codes/PaddleDetection-release-2.7.1/ppdet/engine/trainer.py", line 116, in __init__
    self.model = create(cfg.architecture)
  File "/home/zqy/zqy/Codes/PaddleDetection-release-2.7.1/ppdet/core/workspace.py", line 255, in create
    cls_kwargs.update(cls.from_config(config, **kwargs))
  File "/home/zqy/zqy/Codes/PaddleDetection-release-2.7.1/ppdet/modeling/architectures/detr.py", line 63, in from_config
    transformer = create(cfg['transformer'], **kwargs)
  File "/home/zqy/zqy/Codes/PaddleDetection-release-2.7.1/ppdet/core/workspace.py", line 229, in create
    raise ValueError("The module {} is not registered".format(name))
ValueError: The module RTDETRTransformerv2 is not registered

[lyuwenyu]

先用v1跑一下同样的数据是不是有问题

[DoctorDream]

先用v1跑一下同样的数据是不是有问题

很抱歉，我之前填写运行环境时出现了错误，现在更正如下：

复现环境 Environment

• OS：Linux
• PaddlePaddle: paddlepaddle-gpu 2.6.1.post117和 paddlepaddle-gpu 2.6.0.post117
• PaddleDetection: develop （原本为2.7.0，现修正为develop，因为只有该版本有rtdetr v2）
• python: 3.10
• CUDA: 11.7

经过试验，在该环境下跑多卡跑rtdetr是没问题的，但是多卡跑rtdetr v2时会出现上述:

ValueError: (InvalidArgument) Required tensor shall not be nullptr, but received nullptr.
  [Hint: tensor should not be null.] (at ../paddle/phi/core/device_context.cc:142)

的报错

[DoctorDream]

先用v1跑一下同样的数据是不是有问题

请问这个问题短期内有解决方案吗？辛苦您了

[lyuwenyu]

收到 最近安排时间看下；其实v1 和 v2的第一阶段训练没啥区别

[yski]

先用v1跑一下同样的数据是不是有问题

大佬，导出问题看看吧，v2训练推理都没问题，但是导出报错，paddle3.0b1+paddledetection develop

[DoctorDream]

先用v1跑一下同样的数据是不是有问题

大佬，导出问题看看吧，v2训练推理都没问题，但是导出报错，paddle3.0b1+paddledetection develop

请问你是多卡训练也没问题吗

[zhang-prog]

@lyuwenyu 大佬看下呢

[yski]

先用v1跑一下同样的数据是不是有问题

大佬，导出问题看看吧，v2训练推理都没问题，但是导出报错，paddle3.0b1+paddledetection develop

请问你是多卡训练也没问题吗

没有测试，我一直用的windows单卡