GPU Resource Management and Inference Serving Optimization

Covers low-level GPU allocation and scheduling such as fractional partitioning, time-slicing, memory limits and scheduler constraints, plus inference-serving optimizations including prefill/decode disaggregation, KV-cache transfer and latency/throughput trade-offs. Boundary is system-level compute efficiency and serving performance, distinct from chip design, data-center buildout, corporate strategy or valuation.