用pnnx转换的deeplabv3-mobilevit-xx-small模型，使用默认的python推理就直接crash

[lanyuer]

error log | 日志或报错信息 | ログ

malloc(): unsorted double linked list corrupted
Aborted (core dumped)

context | 编译/运行环境 | バックグラウンド

WSL/Ubuntu 24.04
Google Colab均复现

how to reproduce | 复现步骤 | 再現方法

1. 用pnnx转换模型

`from transformers import MobileViTImageProcessor, MobileViTForSemanticSegmentation
import torch
import pnnx

model = MobileViTForSemanticSegmentation.from_pretrained("apple/deeplabv3-mobilevit-xx-small")
model.eval() # 设置为评估模式

class CustomMobileViT(MobileViTForSemanticSegmentation):
def forward(self, input):
output = super().forward(input)
return output.logits # 只返回 logits 部分

custom_model = CustomMobileViT.from_pretrained("apple/deeplabv3-mobilevit-xx-small")
custom_model.eval()

x = torch.rand(1, 3, 512, 512)

opt_model = pnnx.export(custom_model, "deeplabv3_mobilevit_xx_small.pt", x)

print(opt_model)`

2. 使用生成的测试代码直接运行模型

`import numpy as np
import ncnn
import torch

def test_inference():
torch.manual_seed(0)
in0 = torch.rand(1, 3, 512, 512, dtype=torch.float)
out = []
with ncnn.Net() as net:
net.load_param("deeplabv3_mobilevit_xx_small.ncnn.param")
net.load_model("deeplabv3_mobilevit_xx_small.ncnn.bin")

    with net.create_extractor() as ex:
        ex.input("in0", ncnn.Mat(in0.squeeze(0).numpy()).clone())

        _, out0 = ex.extract("out0")
        out.append(torch.from_numpy(np.array(out0)).unsqueeze(0))

if len(out) == 1:
    return out[0]
else:
    return tuple(out)

if name == "main":
print(test_inference())
`

more | 其他 | その他