conv-neural-network - Using Pre-trained ELMo to Generate Words embeddingE

Question

I'm trying to understand how ELMo is designed and how it works, and I have a couple of questions:

1. Is the ELMo architecture (visualized in the figure below) used for training the model, or for generating the context-dependent embeddings using the pre-trained model? Or is the same for both?

Source

2. Before passing the input to the Bi-LSTM layers, it is passed through a convolutional neural network (CNN) to convert the words into raw word vectors (character-based). How CNN does this? Any helpful references?

Thank you.