Hi @charlestang06! Sounds like an interesting setting. We do not support this in full generality, but if I understand you correctly you can get a fairly good approximation by modelling it as follows.

Although you do not have a central depot, PyVRP requires at least one. Just add a dummy depot with default time windows, and set all edge distances and durations from and to that depot to 0.

I assume you have a list of $(O, D, q)_i$-triplets for each shipment $i$, where $O_i$ is an origin location, $D_i$ a destination, and $q_i > 2$ some demand. We are going to do a full-truckload approximation. Simply count $n_i = \lceil q_i / 2 \rceil$, the minimum number of shipments that are needed to move $q_i$ from $O_i$ to $D_i$.

Add $n_i$ dummy clients with zero demand to your model, each at the location $D_i$. These replace the $O_i$ and $D_i$ locations for shipment $i$. Arcs leaving any of the new clients are the same as those originally leaving $D_i$, but any arcs into the new clients need to have a distance and duration that takes into account that on the arc we first (implicitly) go to $O_i$ to pick up a full truckload of demand. The picked up demand is then immediately dropped of at $D_i$ through one of the dummy clients. This should hopefully get you a good approximate solution.