Prefill/decode disaggregation across heterogeneous silicon

A roofline cost model for chip × model routing

KV-cache economics on a tiered memory hierarchy

Batching under latency SLOs: the throughput floor