arXiv:2609.30104
MQSS-Selector: RL-Guided Pass Selection for an MLIR Compilation Pipeline
Andre Youssefi, Ercüment Kaya, Minh Chung, Jorge Echavarria, Laura B. Schulz, Martin Schulz
Abstract
High Performance Computing (HPC) and Quantum Computing (QC) systems are increasingly converging towards unified High Performance Computing-Quantum Computing (HPCQC) infrastructures, driven by a growing need to bridge classical and quantum workflows, which affects all levels of the system stack, from the hardware to compilers and runtimes, all the way to applications. However, today's QC devices are still in the Noisy Intermediate-Scale Quantum (NISQ) era, are error-prone and resource-limited, and therefore require specialized optimizations and topology mappings to achieve sufficient fidelity. This places special emphasis on proper compilation and optimization within the overall quantum software stack. Many existing stacks remain fragmented, with separate components responsible for device selection, compiler-pass optimization, and job queue scheduling. This paper proposes a unified, learning-based selector that integrates these disparate stages into a cohesive framework. Our proposed selector scheme leverages reinforcement learning and deep learning models that can be extended to simultaneously optimize multiple objectives -- such as fidelity, compilation time, and scheduling latency -- while dynamically adapting to circuit characteristics and device conditions.
AI Summary
The paper studies learned quantum-device selection and compiler-pass ordering for heterogeneous HPC–quantum workflows. It implements separate device-selection and MLIR-native pass-selection components; their integration into a jointly optimized scheduler remains conceptual. Device selection uses supervised models to predict output-distribution Hellinger fidelity across backends. Evaluation uses 128 MQTBench circuits on 12 mock IBM backends and 2,880 randomized-benchmarking circuits on two real devices. Random forests are identified as the preferred model on the mock-backend dataset; models perform similarly on the real-device dataset. Pass selection uses an actor–critic agent with 92 compilation passes, a device-dependent mapping action, and a Finish action. Rewards combine changes in depth, operation count, and unsupported operations with executable termination and shorter compilation sequences. Dual-Annealed Exploration Priming (DAEP) temporarily substitutes heuristic-guided actions and adds a behavior-cloning loss, with separately decaying intervention probability and imitation weight. On 154 validation circuits from a 1,548-circuit chemistry dataset, DAEP yields 95.5% executable outputs versus 42.1% without guidance. Finish is selected in 153 versus zero cases. Mean depth and operation-count changes remain positive: approximately $+115$ and $+325$ with DAEP, versus $+135$ and $+376$ without it. The paper notes limited dataset coverage, failures involving unfamiliar gates, a 20-step rollout horizon, and the need for further critic training. Evaluation compares guided and unguided agents rather than external compilers; the stated rationale is that compiler comparisons require matching pass sets. Dynamic, multi-objective joint scheduling is proposed but not experimentally demonstrated.
Central claim may not hold: The abstract presents an integrated selector, but the implemented device-selection and pass-selection components remain separate and their integration is explicitly future work. The experiments do not demonstrate joint device/pass optimization, job-queue scheduling, or adaptation to changing backend conditions. The hardness discussion neither establishes joint Pareto-optimal selection nor justifies efficient Pareto-optimal heuristics, and its phase-ordering argument states a reduction in the wrong direction. The pass-selection experiment establishes improved executability relative to unguided training but provides no executable baseline comparison establishing compilation quality or optimality. The weighted reward potential can decrease when depth decreases but operation count increases sufficiently, contradicting the stated structural-optimization property and undermining the claimed assurance about the learned objective.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 11 pages, 5 figures, 1 table