Libra: Flexible Request Partitioning and Scheduling for Serving Unbalanced and Dynamic LLM Workloads

LLM inference must meet strict latency SLOs while maximizing throughput. Yet, real-world variability in prompt and response lengths skews compute-intensive prefill and memory-bound decode phases, making both colocated (even with chunked prefill) and disaggregated deployments unable to simultaneously deliver low tail latency and high throughput. We introduce Libra, a high performance LLM serving system that maximizes goodput under SLO constraints even when handling imbalanced and dynamic workloads. At the core of Libra is a micro-request based flexible partitioning and scheduling (FPS) abstraction. The abstraction splits each request at any token boundary into multiple cooperating segments. Libra then designs a two-level scheduling framework that balances micro-request load across unified GPU instances. The framework consists of a global scheduler that selects per-request split points, and a local scheduler on each GPU instance to form SLO-aware batches. Finally, Libra uses chunked KV cache transfers to support cross-instance micro-request execution. On real-world traces, Libra improves goodput by up to 1.91× and 1.61×, increases serving capacity from 1.15× to 3.07×, and improves serving performance by up to 74.2% in a hybrid workload under strict SLOs and A100/H100 GPUs compared to state-of-the-art colocated and disaggregated baselines.