how to save tsai model and load in libtorch C++

[MichaelCarrik]

1. I am trainning model in python like this

           model= tsai.LSTM_FCNPlus(train_dsets.vars, train_dsets.c, train_dsets.len, se=8)
           learn = tsai.Learner(train_dsets, model, metrics=[tsai.mae, tsai.rmse])
           learn.fit(80, learnRate *0.9,cbs=[EarlyStoppingCallback(patience=7), ShowGraphCallback()]
                      )
          torch.jit.trace(model,).save("model.pt"）
          predict1 =  learn.get_X_preds(inputs)
   

2. I load model which step 1 save in libtorch c++ like this

   module = torch::jit::load("model.pt");
   predict2 =   module.forward(inputs).toTensor();
   

I can ensure they have same weights.
however predict1 is not same as predict2 . can anyone tell me correct way  how to save tsai model and load in c++, appreciatlly

[MichaelCarrik]

by research the reason is the inference learn.get_X_preds(inputs) != learn.model(inputs) 。 is different so save model as torchscript can not be loaded for inference. Anyone know how to solve it?