多卡训练模型如何转为静态图？

[chengyi192]

对于单卡训练，我使用以下代码将其转为静态图。运行正常。

   input_spec = [
        InputSpec(shape=[None,None,None,None], dtype='float32'),
    ]

    model = paddle.jit.to_static(
            model,
            input_spec=input_spec,
            full_graph=True,
        )

但对于多卡训练，python3 -m paddle.distributed.launch --gpus '0,1,2,3,4,5,6,7' train-test8.py -c  det_r50_vd_db.yml

    model = paddle.DataParallel(model) # 第 3 处改动，增加 paddle.DataParallel 封装

    model = paddle.jit.to_static(
            model,
            input_spec=input_spec,
            full_graph=True,
        )

转静态图会报错。不转静态图的话，动态图运行没有问题：

  preds = model(images)
  File "/usr/local/lib/python3.10/dist-packages/paddle/nn/layer/layers.py", line 1530, in __call__
    return self.forward(*inputs, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/paddle/jit/dy2static/program_translator.py", line 558, in __call__
    return self._perform_call(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/paddle/jit/dy2static/program_translator.py", line 860, in _perform_call
    args, kwargs = self._function_spec.unified_args_and_kwargs(args, kwargs)
  File "/usr/local/lib/python3.10/dist-packages/paddle/jit/dy2static/function_spec.py", line 97, in unified_args_and_kwargs
    raise ValueError(error_msg)
ValueError: The decorated function `forward` requires 0 arguments: [], but received 1 with (Tensor(shape=[16, 3, 640, 640], dtype=float32, place=Place(gpu_pinned), stop_gradient=True,

请问有人知道如何将多卡训练如何转为静态图吗？