AIConfigurator: Lightning-Fast Configuration Optimization for Multi-Framework LLM Serving

Optimizing Large Language Model (LLM) inference in production systems is increasingly difficult due to dynamic workloads, stringent latency/throughput targets, and a rapidly expanding configuration space. This complexity spans not only distributed parallelism strategies (tensor/pipeline/expert) but also intricate framework-specific runtime parameters such as those concerning the enablement of CUDA graphs, available KV-cache memory fractions, and maximum token capacity, which drastically impact performance. The diversity of modern inference frameworks (e.g., TRT-LLM, vLLM, SGLang), each employing distinct kernels and execution policies, makes manual tuning both framework-specific and computationally prohibitive. We present AIConfigurator, a unified performance-modeling system that enables rapid, framework-agnostic inference configuration search without requiring GPU-based profiling. AIConfigurator combines (1) a methodology that decomposes inference into analytically modelable primitives - GEMM, attention, communication, and memory operations while capturing framework-specific scheduling dynamics; (2) a calibrated kernel-level performance database for these primitives across a wide range of hardware platforms and popular open-weights models (GPT-OSS, Qwen, DeepSeek, LLama, Mistral); and (3) an abstraction layer that automatically resolves optimal launch parameters for the target backend, seamlessly integrating into production-grade orchestration systems. Evaluation on production LLM serving workloads demonstrates that AIConfigurator identifies superior serving configurations that improve performance by up to 40% for dense models (e.g., Qwen3-32B) and 50% for MoE architectures (e.g., DeepSeek-V3), while completing searches within 30 seconds on average. Enabling the rapid exploration of vast design spaces - from cluster topology down to engine specific flags.

Causal bandits with backdoor adjustment on unknown Gaussian DAGs

The causal bandit problem aims to sequentially learn the intervention that maximizes the expectation of a reward variable within a system governed by a causal graph. Most existing approaches assume prior knowledge of the graph structure, or impose unrealistically restrictive conditions on the graph. In this paper, we assume a Gaussian linear directed acyclic graph (DAG) over arms and the reward variable, and study the causal bandit problem when the graph structure is unknown. We identify backdoor adjustment sets for each arm using sequentially generated experimental and observational data during the decision process, which allows us to estimate causal effects and construct upper confidence bounds. By integrating estimates from both data sources, we develop a novel bandit algorithm, based on modified upper confidence bounds, to sequentially determine the optimal intervention. We establish both case-dependent and case-independent upper bounds on the cumulative regret for our algorithm, which improve upon the bounds of the standard multi-armed bandit algorithms. Our empirical study demonstrates its advantage over existing methods with respect to cumulative regret and computation time.