Research

Simulate, understand, and optimize LLM serving. ServingStudio compares serving configurations using performance models grounded in real GPU measurements, then lets an agent build and validate the strongest candidates in real serving frameworks.

Real-world coding agent traces collected from Claude Code and Codex, capturing multi-step agent sessions, LLM requests, tool calls, cache behavior, latency distributions, and workload characteristics for studying agentic AI systems. Integrated into NVIDIA AIPerf for trace replay.

A throughput-oriented LLM serving framework that exploits intra-device parallelism, overlapping compute, memory, and network operations within a single GPU (OSDI 2025). Its asynchronous scheduling is adopted in SGLang, and similar overlapping designs now appear in most mainstream serving engines.