Segmentation fault when juliacall is imported before paddle

[AndPuQing]

bug描述 Describe the Bug

当paddle import 在 juliacall 之后将会发生core dumped, 下面是最小复现代码

# paddlepaddle-gpu              3.0.0b1
import pysr
from pysr import PySRRegressor
import numpy as np
import paddle # <---- 此处如果在 pysr 前面被import 将不会core dumped

X = np.random.randn(100, 10)
y = np.ones(X.shape[0])
model = PySRRegressor(
    progress=False,
    max_evals=10000,
    model_selection="accuracy",
    extra_sympy_mappings={},
    output_paddle_format=True,
)
model.fit(X, y) <----- core dumped

---

When paddle is imported after juliacall, a core dump occurs. Below is the minimal reproducible code:

# paddlepaddle-gpu 3.0.0b1
import pysr
from pysr import PySRRegressor
import numpy as np
import paddle  # <---- Importing this *before* pysr prevents the core dump

X = np.random.randn(100, 10)
y = np.ones(X.shape[0])
model = PySRRegressor(
    progress=False,
    max_evals=10000,
    model_selection="accuracy",
    extra_sympy_mappings={},
    output_paddle_format=True,
)
model.fit(X, y)  # <----- core dumped

该问题在其他框架也有类似的问题

---

This problem also has similar problems in other frameworks

Related Issues

• Certain import order triggers segmentation fault pytorch/pytorch#78829
• Using numba and julia (or any other LLVM loader?) at the same time leads to segfaults numba/numba#7857
• 【Hackathon 7th No.1】 Integrate PaddlePaddle as a New Backend MilesCranmer/PySR#704

其他补充信息 Additional Supplementary Information

No response