Resident is not active
Resident GB must fit a socket. Active GB/token drives each memory path. MoE makes those two quantities different.
Dual-socket MoE residency
NUMA shard placement lab
Fit is not locality. Assign every shard to one memory domain, minimize remote active reads, and expose the bandwidth path that controls the modeled ceiling.
Shard and hardware inputs
GB
GB/s
GB/s
Synthetic six-shard sample ready. Values are not a DeepSeek trace.
Placement proof
RAM capacity is only the first constraint.
Analyze the packed layout, then solve the binary affinity-aware assignment and compare every modeled bandwidth path.
Socket 0
Socket 1
Modeled path time
Canonical result
What changed
Remote cost follows affinity
Putting a shard on socket 0 makes socket 1's requester share remote, and vice versa. The binary objective sums those active reads.
The slowest path wins
The upper bound is one divided by the largest per-token time across socket 0 memory, socket 1 memory, and the link.
Optimize a valid placement to unlock the exact JSON artifact.