When streaming experts from an SSD, if the expert weight blocks are large (a single expert may be several hundred MB), the read granularity can approach sequential, and the performance loss is relatively controllable.