Whenever you're ready, walk me through how you'd approach this.
How does multi-head attention operate in parallel, and how are the outputs from individual heads combined before passing to the feed-forward network?