Inference result is SequenceClassifierOutput instance?

Hi, I am training the model with Trainer API, and in the forward method, the return is SequenceClassifierOutput.

def forward(**kwargs):
   return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

No problems happen in training. But when the model inferences, what I get is something like this

model = BertPrefixForSequenceClassification.from_pretrained(model_path)
model(**inputs)

The result is:

SequenceClassifierOutput(loss=None, logits=tensor([[-3.1394, -3.2234, -2.6458, -2.6055, -2.0099, -3.1522, -2.5696, -2.3495,
         -1.8810, -2.7378, -3.1791, -2.0395, -2.7319, -3.0667, -3.2499, -2.8271,
         -2.0556, -2.5394, -2.6604, -2.5847, -3.4659, -2.2318]],
       grad_fn=<AddmmBackward0>), hidden_states=None, attentions=None)

I am confused how to ONLY get the sigmoid value in inference but not the SequenceClassifierOutput.
It seems quite strange if I do this:

class NeuralNetwork(nn.Module):
    def __init__(self):
        super(NeuralNetwork, self).__init__()
        self.model = BertPrefixForSequenceClassification.from_pretrained(model_path)

    def forward(self, input_ids, token_type_ids, attention_mask):
        x = self.model(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask)
        logits = torch.nn.Sigmoid()(x)
        return logits

and if I did so, something wrong in torch.jit.save as well.