How to convert a batch of logits of a (Embedding + Linear Layer) to a one-hot encoded batch ? (Pytorch)

Question

I am building a FlaubertForClassification model from scratch (I don't want to use the FlauBertForSequenceClassification model of HuggingFace). Here is the model class :

class FlauBertForClassification(nn.Module):
    def __init__(self, embedding_dim, num_intents, dropout):
        super(FlauBertForClassification, self).__init__()
        self.flaubert = FlaubertModel.from_pretrained("flaubert/flaubert_base_cased")
        self.dropout = nn.Dropout(dropout)
        self.linear = nn.Linear(embedding_dim, num_intents)

    def forward(self, utterances, masks):
        embeddings = self.flaubert(input_ids=utterances, attention_mask=masks)[0]  # (batch size x sequence length x embedding dim)
        dropout_output = self.dropout(embeddings) # (batch size x sequence length x embedding dim)
        logits = self.linear(dropout_output) 
        return logits

1. When applying the model, the logits are of shape (batch size x sequence length x number of classes). How can I get rid of the sequence length to be able to compute several metrics such as Accuracy with the true labels (of shape : batch size x number of classes).

2. I don't understand why it worked "correctly" when computing the loss (CrossEntropy) without any modification of the predicted labels.