paddle.nn.functional.silu backward returns -Inf for finite float64 gradient at x=-709

bug描述 Describe the Bug

paddle.nn.functional.silu has an incorrect backward result for float64 input x = -709.0.

For SiLU, the function is:

silu(x) = x * sigmoid(x)

Its derivative is:

silu'(x) = sigmoid(x) + x * sigmoid(x) * (1 - sigmoid(x))

At x = -709.0, the correct derivative is approximately:

-8.614807714413836e-306

This is a finite, normal float64 value. However, Paddle autograd returns -Inf.

Environment

• PaddlePaddle version: 3.3.1
• NumPy version: 2.2.6
• dtype: float64
• Device: CPU

Minimal reproducible code

import paddle
import paddle.nn.functional as F

paddle.set_device("cpu")

x = paddle.to_tensor(
    -709.0,
    dtype="float64",
    stop_gradient=False,
)

y = F.silu(x)
y.backward()

wrong = float(x.grad)
expected = -8.614807714413836e-306

print("wrong:", wrong)
print("expected:", expected)

Actual behavior

wrong: -inf

Expected behavior

expected: -8.614807714413836e-306

paddle.nn.functional.silu backward should return a finite float64 gradient for x = -709.0.

Instead, Paddle returns -Inf, even though the correct gradient is finite and representable in float64.

其他补充信息 Additional Supplementary Information

No response