1. for a local observable, each MPS is "gauged" just as in normal, finite serial DMRG, with a special "orthogonality center" tensor which you can adjust by calling .position(j). However, you have to use the MPS from node n to measure sites which are fully contained in block number n.

2. for observables which span beyond the block covered by a single node, it gets more technical. One reference is the parallel DMRG paper by myself and Steve White. Basically, if there are two blocks, block 1 which goes from sites 1 to 10 and block 2 which goes from sites 11 to 20, then there will be a "V" tensor connecting the MPS tensor on site 10 to the MPS tensor on site 11. So the "true" MPS is A1-A2-A3-...-A10-V-A11-A12-...-A20, where "A" just means an MPS tensor, nothing about the gauge. Then to measure a correlation function from site 5 to site 12, for example, you'd need to put the gauge center of the first MPS on site 5 and of the second MPS on site 12, then use the tensors A5-A6-...-A10-V-A11-A12 to do the measurements. The other tensors can be ignored / cancel due to the gauge adjustment.

It's kind of an open question how to do global measurements while taking advantage of the parallel structure by the way.