PaperlayerView exact revision on arXiv

Source-linked AI summary

MQSS-Selector: RL-Guided Pass Selection for an MLIR Compilation Pipeline

Andre Youssefi, Ercüment Kaya, Minh Chung, Jorge Echavarria, Laura B. Schulz, Martin Schulz

arXiv:2609.30104v1quant-phcs.DCcs.LGcs.PL

TL;DR

Quantum device selection and compiler-pass ordering are often handled without coordination, despite their interdependent design spaces. MQSS-Selector combines ML-based device selection with RL-based pass selection; experiments provide empirical evidence for the device-selection component and show that DAEP accelerates convergence and improves reward exploitation.

  • Problem

    Existing approaches often use heuristics or static policies without coordinating device selection and pass ordering, despite their interdependent, multi-objective design spaces.

    Source: §1 · p. 2

  • Method

    MQSS-Selector combines ML-based device selection with an RL-based, MLIR-native module for choosing and ordering quantum compilation passes, using DAEP as a training strategy.

    Source: §1 · p. 2, §6 · pp. 9–10

  • Results

    DAEP accelerates convergence and improves reward exploitation, while experiments provide empirical evidence that ML can address device selection.

    Source: §6 · pp. 9–10

  • Takeaways & Limitations

    The authors conclude that Pareto-optimal joint selection of a quantum device and compilation-pass sequence is theoretically feasible, with experiments demonstrating feasibility of both ML components.

    Source: §6 · p. 9

    RL without additional methods was not viable for pass-selection training under the resource constraints considered.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

A. Youssefi Leibniz Supercomputing Centre (LRZ) · E. Kaya Leibniz Supercomputing Centre · M. Chung Leibniz Supercomputing Centre (LRZ)

The supplied affiliation text places the Leibniz Supercomputing Centre in Garching bei München, Germany, and also references Munich (TUM).

M. Chung Leibniz Supercomputing Centre (LRZ)

J. Echavarria Munich Quantum Valley (MQV) Garching bei München, Germany · M. Schulz Leibniz Supercomputing Centre

The supplied passage lists Lemont, Illinois, United States, under the M. Schulz Leibniz Supercomputing Centre subsection.

M. Schulz Leibniz Supercomputing Centre

CCS Concepts

The paper is classified across quantum computation, just-in-time compilers, and sequential decision making.

CCS Concepts

  • The CCS concepts span quantum computation, just-in-time compilers, and sequential decision making.They are listed under hardware, software and its engineering, and theory of computation, respectively.Source: CCS Concepts · p. 1

1 Introduction

HPCQC execution requires coordinating quantum-device selection with runtime compilation, where NISQ device constraints make topology- and calibration-aware optimization important. MQSS-Selector unifies device selection and program transformation in an adaptive reinforcement-learning framework.

Introduction

  • NISQ devices’ error-proneness, qubit limits, and sensitivity to calibration and workload conditions require compilation tailored to device topology and current calibration data.Device selection and phase ordering are NP-hard, and prior approaches often use uncoordinated heuristics or static policies [9].Source: §1 · pp. 1–2
  • MQSS-Selector jointly addresses device selection and compiler-pass phase ordering, two interdependent multi-objective combinatorial problems, with decisions based on program characteristics, available passes, and quantum backends.The pipeline integrates device selection with phase ordering, or compiler pass-sequence optimization.Source: §1 · p. 2, Figure 1 · p. 2
  • MQSS-Selector is an RL-based pass selector that learns to choose and order quantum-circuit compilation passes, demonstrating data-driven, MLIR-native selection unlike prior work largely targeting Qiskit pipelines.MLIR represents programs across levels through extensible dialects and reusable, composable passes; quantum passes can be applied at runtime because critical inputs are not known statically.Source: §1 · pp. 1–2
  • The paper also contributes Dual-Annealed Exploration Priming for training efficiency, actor- and critic-network analysis metrics, and an expanded MLIR-based MQSS X Compilation Pass Suite.Source: §1 · p. 2

2 Problem Statement and Motivation

Device selection and compiler phase ordering are jointly NP-hard and determine Pareto-optimal compilation outcomes. Their interdependence motivates a unified selector that reasons over device compatibility and pass strategies together.

2 Problem Statement and Motivation

  • Device selection and phase ordering are NP-hard and jointly determine Pareto-optimal compilation results.The combined problem is NP-hard because either problem remains when the other is restricted to one choice.Source: §2 · pp. 2–3
  • Some pass combinations can degrade performance, so each program class needs a tailored pass sequence compatible with the target device architecture.The sequence must optimize compilation while producing a program compliant with the device architecture.Source: §2 · p. 3
  • The selector’s design space combines program features, device and queue conditions, available passes, and objectives such as compilation time, fidelity, and circuit depth.Relevant circuit features can include qubit count and program depth; backend information includes queue lengths, calibration data, and noise characteristics.Source: §2 · p. 3
  • Because exact solutions do not scale and device selection and pass ordering are interdependent, the paper motivates a unified RL-driven selector for multi-objective, changing execution conditions.The paper frames this mutual dependency as a bootstrapping dilemma and argues for Pareto-efficient configurations over sequential or isolated heuristics.Source: §2 · p. 3

3 Related Work

Prior work applies learning to quantum compilation, device selection, and scheduling, while MQT Predictor separates device selection from phase ordering. This work builds on Quake and proposes a unified selector that jointly considers compiler transformations, hardware constraints, and live system status.

3 Related Work

  • Quantum-related work spans RL-based circuit transformations and pass selection, supervised compilation-option prediction, and systems for scheduling quantum backends alongside classical nodes [18] [20] [21] [22].Other examples include MLGO [16] and AutoPhase [17] for classical LLVM optimization, and TuniQ [23], which separates phase ordering into mapping, decomposition, optimization, and later stages.Source: §3 · p. 3
  • MQT Predictor reports improved fidelity and reduced depth across diverse backends, but trains device selection and phase ordering as independent tasks and omits live backend conditions.It uses a supervised device-selection model and a separate RL phase-ordering agent; the cited limitations include queue length, device availability, and runtime calibration data.Source: §3 · pp. 3–4
  • The paper builds on the Quake dialect rather than existing quantum pass pipelines and motivates a unified selector for phase ordering, device selection, and live system status [6].The proposed multi-objective scope includes fidelity, execution latency, and compilation effort.Source: §3 · p. 4

4 A Software Approach Toward Unified Scheduling Framework

The section presents supervised device selection as a current building block, while framing joint device and pass selection as a future unified scheduler. It describes the pass suite, an actor–critic phase-ordering component, reward design, and two conceptual integration schemes.

4.1 Device-selection Component

  • The device selector uses Hellinger fidelity against noiseless simulated distributions and execution time to label runs, then predicts device scores with MLP, RF, KNN, or SVM to preserve rankings.The relative structure across devices matters more than exact fidelity values.Source: §4.1 · p. 4

4.2 MQSS Pass Suite

  • At data-generation time, the MQSS suite had 92 passes supporting necessary Quake-dialect operations and a mapping pass; its pybind11 selector enables Python action-selection policies independently of C++ pass execution.Source: §4.2 · pp. 4–5

4.3 Phase-ordering Component

  • The phase-ordering component uses an actor–critic model: the actor selects a pass or Finish from program and device context, while the critic estimates state value and GAE supports stable, sample-efficient advantage estimates.The state space spans devices, arbitrary programs, and pass-sequence stages.Source: §4.3 · p. 5

4.3.1 Reward Functions.

  • The tested potential reward functions performed best and encode preferences for executable states, shorter paths to equivalent executable results, and monotonic reductions in operation count and depth.Degradations in operation count or depth do not yield positive reward contributions.Source: §4.3.1 · p. 5
  • The proposed DAEP approach adds decaying guidance to actor–critic learning to constrain exploration when successful compilation rewards are difficult to discover.Guidance can overwrite the actor’s choice during rollouts; separate decaying coefficients control how often guidance is applied and how strongly it is reinforced.Source: §4.3.1 · pp. 6–7

4.4 Unified Scheduling Framework

  • The unified scheduler treats device and pass selection as joint RL decisions, with one scheme linking specialized modules and another combining both into a single action set.Its state can include program features, backend status, and intermediate compilation results, while rewards can reflect fidelity, compilation overhead, and latency.Source: §4.4 · p. 7

5 Experiments

Experiments assess device-selection models on benchmark and real-device workloads and evaluate pass selection on molecule programs. The reported results identify RF as the current MQT-Bench candidate and show that DAEP yields smaller average depth and operation count than no guidance.

5.1 Device selection

  • Device-selection experiments compared MLP, RF, SVM, and KNN using two datasets; weighted-loss and dual-headed MLP variants addressed infeasibility prediction, evaluated with MAE and R2.The datasets comprised 128 MQTBench programs on twelve mock devices and 2,880 random-benchmarking programs on two real devices; MAE alone does not distinguish many small errors from a few large deviations.Source: §5.1 · p. 7
  • RF is the current best candidate on the 128-program MQT-Bench set: MLP performed worse, while SVM and KNN may not generalize beyond the limited data.Figure 4 shows fidelity predictions for the 128 programs on twelve backends; the authors expect broader program diversity and depth variance to challenge SVM and KNN in deployment.Source: §5.1 · p. 8, Figure 4 · p. 8
  • On the real-device RB set, all models performed nearly equally in MAE and R2 because every program could run on every device, with little change among MLP variants.The passage attributes device differences to device properties and environmental influences.Source: §5.1 · p. 8, Figure 5 · p. 8

5.2 Pass selection

  • DAEP yields better compilation performance than No Guidance, with smaller average resulting depth and operation count.The comparison is reported in the chemistry-dataset validation summary; the passage provides no cell values.Source: Table 1 · p. 9, §5.2 · p. 9
  • The actor learned to finish when further passes were not expected to improve a program, but one rollout hit the 20-step horizon and seven samples used unfamiliar gates.The actor’s efficient termination leaves the critic with less training; the guided critic is conservative but performs better than the unguided critic.Source: §5.2 · p. 9

6 Conclusions

The paper presents a unified scheduler for jointly selecting quantum devices and compilation-pass sequences, with theoretical analysis and experiments supporting the feasibility of its ML components. The experiments distinguish effective ML-based device selection from resource-constrained pass-selection training, which required additional guidance.

6 Conclusions

  • The authors introduce a unified scheduler and theoretically demonstrate the feasibility of jointly selecting a Pareto-optimal quantum device and compilation-pass sequence for heterogeneous HPCQC stacks.Experiments demonstrate the feasibility of the ML components for device selection and pass selection.Source: §6 · p. 9
  • ML techniques effectively addressed device selection, whereas RL alone was not viable for pass-selection training under the considered resource constraints.The pass-selection experiments also indicate that more extensive training is needed.Source: §6 · pp. 9–10
  • Decaying action guidance in DAEP accelerated convergence and improved reward exploitation, while pass-selection training across devices showed the approach was not limited to specific native operation sets or physical constraints.The paper also states that defined properties ensure the actor learns a policy aligned with the intended objective.Source: §6 · p. 10
  • The unified scheduler adds complexity but could outperform independently optimized pipelines by exploiting device-selection and compilation interdependencies, and it supports online improvement from execution data.The paper identifies the potential to continuously improve predictions through online training on newly acquired data.Source: §6 · p. 10
Loading 2609.30104v1…