Regulating Branch Parallelism in LLM Serving

Collection of systems and HPC research on efficient execution of LLMs and scientific kernels. Key contributions include TAPER (per-step admission control to regulate branch parallelism in LLM serving, improving goodput up to 1.77× vs baselines while preserving SLOs), FATE (future-state-aware scheduler for heterogeneous multi-stage LLM workflows that reduces makespan and P95 latency by ~32% vs simple heuristics), RcLLM (distributed generative-recommendation inference with beyond-prefix KV caching reducing TTFT 1.31–9.51×), HEXISEQ (heterogeneous CP/HP for long-context LLM training improving avg