[0-size Tensor No.256] Add 0-size Tensor support for paddle.sgn API

[VVX94]

PR Category

Execute Infrastructure

PR Types

Improvements

Description

• 错误信息

2025-06-05 15:25:30.646645 GPU 0 64602 test begin: paddle.sgn(Tensor([0, 4],"complex128"), )
[paddle error] paddle.sgn(Tensor([0, 4],"complex128"), ) 
 (InvalidArgument) Expected the stride of last dimension of input(X) to be 1.But received 2. This means that the last dimension of theTensor(x) is not continuous and cannot be as_complex directly.You can call x.contiguous() to make the Tensor(x) contiguous first.
  [Hint: Expected x.strides()[x.strides().size() - 1] == 1, but received x.strides()[x.strides().size() - 1]:2 != 1:1.] (at ../paddle/phi/kernels/stride/as_complex_kernel.cc:35)

• paddle源代码
  • 发现问题在于paddle.as_real和paddle.as_complex，可能存在相关问题

def sgn(x, name=None):
    ...
    if paddle.is_complex(x):
        expand_x = paddle.as_real(x)
        ...
        return paddle.as_complex(output)
    else:
        return paddle.sign(x)

def as_real(x, name=None):
    if in_dynamic_or_pir_mode():
        return _C_ops.as_real(x)
    else:
       ...
def as_complex(x, name=None):
    if in_dynamic_or_pir_mode():
        return _C_ops.as_complex(x)
    else:
       ...

• 检查对应的infetmeta，没有发现问题
  • AsRealInferMeta
  • AsComplexInferMeta

void AsRealInferMeta(const MetaTensor& input, MetaTensor* output) {
  auto out_dims_v = common::vectorize(input.dims());
  out_dims_v.push_back(2);
  auto out_dims = common::make_ddim(out_dims_v);
  output->set_dims(out_dims);
  output->share_lod(input);
  output->set_dtype(dtype::ToReal(input.dtype()));
}

void AsComplexInferMeta(const MetaTensor& input, MetaTensor* output) {
  auto in_dims = input.dims();
  const int input_rank = in_dims.size();
  PADDLE_ENFORCE_GE(
      input_rank,
      1,
      common::errors::InvalidArgument(
          "The rank of input(X) is less than 1. "
          "Expected the rank of input(X) to be equal to or greater than 1."
          "But received rank of input(X) = %d",
          input_rank));
  const int last_dim_size = static_cast<int>(in_dims[input_rank - 1]);
  PADDLE_ENFORCE_EQ(
      last_dim_size,
      2,
      common::errors::InvalidArgument(
          "The size of the last dimension of input(X)"
          "does not equals 2."
          "Expected the size of last dimension of input(X) to be 2."
          "But received %d",
          last_dim_size));

  const phi::DDim out_dims(in_dims.Get(), input_rank - 1);
  output->set_dims(out_dims);
  output->share_lod(input);
  output->set_dtype(dtype::ToComplex(input.dtype()));
}

• 检查kernel函数，因调用了as_real和as_complex，对相应函数进行检查
  • 找到对应位置AsComplexStridedKernel，增加对0-size的检查

//as_complex_kernel.cc
void AsComplexStridedKernel(const Context& dev_ctx,
                            const DenseTensor& x,
                            DenseTensor* out) {
  if (!FLAGS_use_stride_kernel) {
    PADDLE_THROW(common::errors::Fatal(
        "FLAGS_use_stride_kernel is closed. Strided kernel "
        "be called, something wrong has happened!"));
  }

  PADDLE_ENFORCE_EQ(
      x.strides()[x.strides().size() - 1],
      1,
      common::errors::InvalidArgument(
          "Expected the stride of last dimension of input(X) to be 1."
          "But received %d. This means that the last dimension of the "
          "Tensor(x) is not continuous and cannot be as_complex directly."
          "You can call x.contiguous() to make the Tensor(x) contiguous first.",
          x.strides()[x.strides().size() - 1]));

    ...
}

• 经检测，反向传播不存在0-size问题
• 增加单测
• PaddleAPITest通过

[VVX94]

@luotao1 CI测试好像down了，而且无法重新构建，提示“无权限操作该流水线”

[luotao1]

最近Github不太稳定，请重新提交一个commit再次触发

[VVX94]

/re-run all-failed

[VVX94]

@luotao1 @DanielSun11 求review

[DanielSun11]

LGTM