Open the offload file at dir (experts.bin + manifest.json) and route
expert weights through it. Call BEFORE the model is constructed.
Routing is process-global. The returned function restores the routing that
was active before this call, so an owner that activated a file for one
model can stop later loads from resolving against it. It never unmaps:
weights already created from the mapping keep borrowing its bytes for the
process lifetime (see expertOffloadArray).
Open the offload file at
dir(experts.bin + manifest.json) and route expert weights through it. Call BEFORE the model is constructed.Routing is process-global. The returned function restores the routing that was active before this call, so an owner that activated a file for one model can stop later loads from resolving against it. It never unmaps: weights already created from the mapping keep borrowing its bytes for the process lifetime (see
expertOffloadArray).