[CPU/GPU] paddle.minimum fails on broadcastable zero-sized tensors

[tingPetty]

bug描述 Describe the Bug

paddle.minimum raises PreconditionNotMet when the inputs are broadcastable and the resulting tensor has a zero-sized dimension. The same operation should be well-defined by broadcasting rules: [0, 5] and [1, 5] broadcast to [0, 5], and no element computation is required.

Minimal reproducing example:

import paddle

print("paddle:", paddle.__version__)

# This also fails on CPU if "gpu:0" is changed to "cpu".
paddle.device.set_device("gpu:0")

x = paddle.empty([0, 5], dtype="int64")
y = paddle.ones([1, 5], dtype="int64")

out = paddle.minimum(x, y)
print(out.shape, out.dtype, out.place)

Expected result:

[0, 5] paddle.int64 Place(gpu:0)

Actual result:

Traceback (most recent call last):
  File "<string>", line 8, in <module>
  File ".../site-packages/paddle/tensor/math.py", line 1311, in minimum
    return _C_ops.minimum(x, y)
RuntimeError: (PreconditionNotMet) The meta data must be valid when call the mutable data function.
  [Hint: Expected valid() == true, but received valid():0 != true:1.]
  (at ../paddle/phi/core/dense_tensor.cc:127)

For reference, PyTorch handles the same broadcasted zero-sized shape:

import torch

x = torch.empty((0, 5), device="cuda", dtype=torch.int64)
y = torch.ones((1, 5), device="cuda", dtype=torch.int64)
print(torch.minimum(x, y).shape)

torch.Size([0, 5])

其他补充信息 Additional Supplementary Information

Reproduced in a fresh Python process with CUDA_LAUNCH_BLOCKING=1.
The same error is observed on CPU and GPU.
Environment used for reproduction:

• Python: 3.10.20
• PaddlePaddle: 2.6.1
• GPU: NVIDIA GeForce RTX 3090
• NVIDIA driver: 595.58.03
• Paddle CUDA runtime: 11.7