Disaggregated Serving

SourcesSurvey §6.1 frames specialization and fleet flexibility; named serving systems are retained from the original outline.

5.21.1Prefill–Decode Disaggregation and Per-Phase Resource Needs#

5.21.2KV Transfer, Connectors, and the Data Plane#

5.21.3Independent Scaling of Prefill and Decode#

5.21.4Heterogeneous Hardware Selection, Data Formats, and Migration Cost#

5.21.5Attention–FFN Disaggregation and Fine-Grained Splitting#

5.21.6Resource-Pool Imbalance, Queue Growth, and Idle Capacity#

5.21.7Splitwise, DistServe, and Mooncake Case Studies#