as explained in another thread the BGP layer2 NLRI is more describing the "site" then providing info on the hosts in each site.
So in VPLS with auto discovery either LDP + MP BGP auto discovery either MP BGP + MP BGP auto discovery these configurations are manuallly configured in advance allocating labels to be used by remote PE.
Each PE has configured a site id that is a number.
When a PE wants to send an MPLS packet to a remote PE it will calculate the L2VPN label as:
base label value + <its own site-id>
This label is used as the inner label (more internal) and the frame is sent over the LSP with destination the remote PE loopback (this provides the external label also called IGP label).
This is true of VPLS using BGP both for autodiscovery and for creating pseudowires.