Incorrect accuracy pytorch(1.0.1) and libtorch because of nn.LSTM

[wuzuowuyou]

pytorch version1.0.1
libtorch version 1.0.1

...
self.lstm_1 = nn.LSTM(512 * 8, 100, 1, bidirectional=False)
...

traced_script_module = torch.jit.trace(net, img)
traced_script_module.save("./lstm20211124.pt")

and in c++ libtorch

torch::Tensor output = m_model->forward({tensor_input}).toTensor();

The output here is inconsistent with that of pytorch.
I have tested that the output results of the upper layer of LSTM are consistent

[ejguan]

Do you want to update PyTorch to 1.10 and check if the problem persists? 1.0.1 is released two years ago.

[wuzuowuyou]

谢谢你的回答！
我尝试了用pytorch1.1.0,libtorch1.1.0,cuda10.0，用同样的代码完全没有问题。
所以我现在确认肯定是pytorch1.0,libtorch1.0,cuda8有问题。 lstm算子转libtorch有问题
因为我们工程还是基于cuda8，所以很难升级最新版本，
中国有句古话叫“牵一发而动全身”，不知道你了解过这句话吗？

[wuzuowuyou]

Thank you for your answer!
I tried pytorch1.1.0，cuda10.0，libtorch1.1.0, and there was no problem with the same code.

So now I'm sure there must be a problem with pytorch1.0，cuda8.0，libtorch1.0. （about lstm）

Because our project is still based on cuda8, it is difficult to upgrade the latest version,
There is an old Chinese saying that "one hair touches the whole body". Don't you know this sentence?

[ejguan]

Maybe you can try to build libtorch from source code with cuda8.

[wuzuowuyou]

That's what I did

我本来就是这么做的