Pith. sign in

REVIEW 3 major objections 5 minor 43 references

DOPPLER: Dual-Policy Learning for Device Assignment in Asynchronous Dataflow Graphs

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Dual-policy device assignment beats the critical-path heuristic by up to 78 percent and an RL baseline by up to 62.5 percent on asynchronous multi-GPU runtimes.

desk verdict DOPPLER is a real empirical contribution on device placement for asynchronous runtimes, with final-hardware results that mostly hold; the simulator-target mismatch and overbroad abstract claims are the issues to fix. read the letter →

arxiv 2505.23131 v2 pith:J4DX2DVD submitted 2025-05-29 cs.LG cs.DC

Xinyu Yao , Daniel Bourgeois , Abhinav Jain , Yuxin Tang , Jiawen Yao , Zhimin Ding , Arlei Silva , Chris Jermaine This is my paper
classification cs.LGcs.DC
keywords deviceassignmentmulti-GPUexecutionwork-conservingschedulerreinforcementlearningimitationgraphneuralnetworksdual-policydataflowgraphs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that device assignment in a work-conserving multi-GPU runtime—one that launches operations as soon as their inputs are available rather than synchronizing at layer barriers—can be learned as a two-step decision process. A selection policy chooses which operation in the dataflow graph to assign next, and a placement policy chooses which GPU should run it. The authors claim that this split, trained in three stages (imitating a critical-path heuristic, then reinforcement learning against a simulator, then reinforcement learning on the real system), yields assignments that run faster than the heuristic, an earlier RL method, and even a purpose-built enumerative optimizer. If the claim is right, asynchronous execution of large machine-learning workloads becomes more practical, and placement policies become much cheaper to train.

What carries the argument

The load-bearing mechanism is a dual-policy sequential assignment process: a SEL policy, fed a partially assigned dataflow graph, chooses the next vertex to assign, and a PLC policy, fed the graph plus the chosen vertex, assigns it to a device. Both policies are graph neural networks with message passing, and the key efficiency trick is to run message passing once per episode rather than once per assignment step, encoding the current placement in device features instead. Training proceeds through three stages: imitation of CRITICAL PATH, policy-gradient reinforcement learning against a stochastic simulator that implements Algorithm 1's work-conserving execution model, and fine-tuning on the real work-conserving runtime.

What would settle it

Collect a held-out set of device assignments, measure each in the simulator and on the real work-conserving runtime, and check whether the simulator's ranking disagrees with the hardware ranking on any clearly separated pair; frequent large inversions would mean simulation-pretrained policies are being optimized against the wrong objective.

Watch

Extended reading notes

Core claim

The paper's central claim is that DOPPLER's dual-policy sequential assignment produces the fastest real-system execution times among all methods tested on four dataflow graphs derived from neural-network workloads. In the reported experiments DOPPLER-SYS reduces execution time by up to 78.2% compared with CRITICAL PATH and up to 62.5% compared with PLACETO, and it beats the authors' ENUMERATIVEOPTIMIZER by up to 13.8%. The paper also claims that the imitation and simulation pretraining stages make the method sample-efficient—per-episode training time drops—and that a policy trained on small graphs transfers to larger transformer-style graphs with a few thousand fine-tuning episodes.

Load-bearing premise

The Stage II simulator must rank assignments roughly the way the real hardware does, because the pretraining rewards come from the simulator; the paper's own ablation reports only a Pearson correlation of 0.79 with real runtimes and says the simulator struggles with assignments of similar speed.

Editorial extensions

If this is right

  • If the reported results hold, asynchronous work-conserving runtimes can gain up to 78.2% in execution time over the critical-path heuristic and up to 62.5% over the PLACETO RL baseline on the tested network graphs.
  • The three-stage recipe means most exploration can happen offline: the paper reports lower per-episode training time because the imitation and simulation stages reduce the amount of real-system interaction needed.
  • The select and place policies contribute separately, and the paper's ablation shows the combined dual policy is best on the more complex graphs, so future systems can tune each policy independently.
  • Trained policies transfer across graphs: training on FFNN and CHAINMM then fine-tuning for 4,000 episodes on LLAMA-style graphs matches assignments from full training, which reduces the cost of adapting to new workloads.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the select-then-place decomposition could be transferred to other sequential resource-allocation problems, such as circuit placement or cluster scheduling, wherever the order of decisions is itself part of the optimization.
  • Editorial inference: since the paper's simulator has only 0.79 Pearson correlation with real runtimes, improving simulator fidelity or adding an online correction step could directly close the gap between simulation-pretrained and real-system-tuned policies.
  • Editorial inference: the once-per-episode message-passing approximation suggests training cost grows mainly with the number of episodes and graph depth; combining it with the paper's repeated-structure idea could scale placement to much larger graphs.
  • Editorial inference: the evidence is limited to a few neural-network dataflow graphs on two GPU configurations, so a natural stress test is whether the same recipe beats reactive heuristics on irregular non-neural DAGs under the same work-conserving scheduler.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies device assignment for dataflow graphs executed on a work-conserving multi-GPU runtime. It proposes DOPPLER, a dual-policy framework in which one policy (SEL) selects the next vertex to assign and a second policy (PLC) places it on a device. Training proceeds in three stages: imitation learning from a critical-path teacher, reinforcement learning against a simulator of the work-conserving scheduler, and reinforcement learning against the real system. Experiments on four P100 GPUs and eight V100 GPUs compare DOPPLER with CRITICAL PATH, PLACETO, and a purpose-built ENUMERATIVEOPTIMIZER, reporting large runtime reductions on ChainMM, FFNN, Llama-block, and Llama-layer workloads, plus ablations for the two policies, the three training stages, the simulator, message passing, and transfer learning.

Significance. If the empirical claims hold, the paper makes a useful systems contribution: it demonstrates that a select-then-place decomposition of the device-assignment problem can be learned cost-effectively and can beat both heuristic and existing RL baselines on real hardware. The strengths are concrete: real-system runtimes averaged over 10 executions, ablations isolating the SEL and PLC policies, a stage-wise training ablation, a simulator fidelity study, a message-passing efficiency study, a seed study on ChainMM, and an additional hardware configuration with eight V100 GPUs. The main weakness is that the Stage II simulator is the load-bearing component for the training-recipe and sampling-efficiency claims, and the paper's own evidence for simulator fidelity is limited to aggregate correlations on imitation-learning trajectories. The final DOPPLER-SYS results could still be rescued by Stage III fine-tuning, but the pretraining and DOPPLER-SIM claims need stronger support.

major comments (3)
  1. [Section 5.1 / Appendix H.1 / Appendix D] The Stage II simulator is the load-bearing component of the training-recipe claims, but its fidelity is only weakly established. Section 5.1 defines the RL reward as r_H = R_{s_H} - \bar{R}_{s_H}, so policy-gradient updates depend on the simulator's relative ordering of assignments, not merely on aggregate agreement. Appendix H.1 reports Pearson 0.79 and Spearman 0.69 between simulated and real runtimes, measured on ChainMM assignments produced during imitation learning, and admits the simulator 'has some trouble differentiating between assignments with similar running times'; Appendix D states the communication factor was hand-tuned to 4 on the same hardware. A Spearman correlation of 0.69 permits many pairwise inversions, and with a baseline-subtracted reward even a few inversions among close assignments can misdirect gradients. Because the correlation is not measured on the assignments explored during Stage II RL, the claims that Stage II accelerates convergence, that DOPPLER-SIM is often second-best, and that the three-stage recipe improves sampling efficiency are not yet supported. Please add a rank-correlation/pairwise-inversion analysis on assignments sampled from the Stage II policy, or ablate Stage II by training from Stage I directly to Stage III.
  2. [Abstract / Section 7.2 / Table 9] The abstract claims DOPPLER 'outperforms all baseline methods across tasks,' but Table 9 shows that on the 8-V100 LLAMA-BLOCK configuration DOPPLER-SYS (109.7±3.0 ms) is slightly slower than ENUMERATIVEOPTIMIZER (109.6±4.2 ms). Section 7.2 is more careful ('outperforms the alternatives in most of the settings'), so the abstract and introduction overstate the result. Please soften the headline claim or report a statistical test showing that the difference is not significant; as written, the abstract is inconsistent with the paper's own data.
  3. [Section 7.1 / Appendix H.2] The main results are reported as a single training run per workload, with a seed study only for ChainMM. Because the comparisons to PLACETO and ENUMERATIVEOPTIMIZER are often within a few percent (e.g., Table 2 FFNN: 47.4 vs 50.2 ms; Table 9 8-V100 LLAMA-BLOCK: 109.7 vs 109.6 ms), the paper should either run multiple seeds on all four workloads or provide confidence intervals for the training process itself. Additionally, the paper should state explicitly whether the reported assignments are selected by best simulated reward, best real-system reward, or final-policy rollout; this selection rule affects the interpretation of all reported improvements.
minor comments (5)
  1. [Appendix H.1] The text says the comparison is 'on the top' and 'on the bottom' of Figure 18, but the figure is described as having a left line chart and a right scatter plot; please fix the cross-reference.
  2. [Table 2 caption] The caption says 'During training, both PLACETO and DOPPLER-SIM rely on a simulator to find good solutions, while others use a real system,' but DOPPLER-SYS also uses the simulator in Stage II; the caption should describe training stages rather than whole methods.
  3. [Algorithm 1] The readiness check uses rdy[v, A_v] in the while-loop condition while the initialization and updates use the two-argument form rdy[v, d]; please make the notation consistent.
  4. [Section 5.2 / Section 7.1] The epsilon-greedy exploration schedule is introduced in Section 5.2 but the specific values are deferred to Section 7.1; consider defining or cross-referencing epsilon in the MDP section.
  5. [Table 4 caption] The caption says '(Zero-shot, 3k-shot, 4k-shot)' while the table columns are labeled Zero-shot, 2K-shot, and 4K-shot; the caption contains a typo.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central results are real-system measurements, and the simulator fidelity issue is a validity risk, not a definitional reduction.

full rationale

No significant circularity is present. DOPPLER is an empirical systems paper: the headline claim is that the learned dual-policy assignments reduce measured execution time on real GPU hardware (Table 2), and DOPPLER-SYS is trained and evaluated against the real work-conserving runtime, so the final result does not reduce to its training inputs by construction. Stage II uses a simulator as an RL reward, and Appendix D's communication factor (=4) is hand-tuned to the same P100 hardware; that is a fidelity and overfitting concern, not a circular derivation, because the paper does not present simulator runtimes as the predicted outcome. It reports real-system runtimes and even discloses in Appendix H.1 that the simulator has only moderate correlation with the real system (Pearson 0.79, Spearman 0.69) and 'has some trouble differentiating between assignments with similar running times.' The three-stage recipe's sampling-efficiency claim could be weakened if the simulator misranks high-quality assignments, but that is an empirical validity risk rather than a self-definitional or fitted-prediction circularity. The self-citations ([6,12]) describe the underlying C++ runtime used for the experiments; they are implementation references, not load-bearing mathematical premises or uniqueness theorems invoked to forbid alternatives. The dual-policy select-then-place design is explicitly framed as a learned list-scheduling heuristic in Section 8, so it is not a renamed known result presented as a derivation. No equation in the paper reduces a predicted quantity to a fitted parameter or to a self-citation by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper contributes an algorithmic recipe and introduces no new physical or ontological entities. The load-bearing external inputs are a calibrated simulator (with a communication factor fitted to the evaluation hardware), the authors' own asynchronous runtime from prior work [6,12], and standard RL and GNN machinery. The most consequential unverified premise is simulator fidelity: Stage II pretraining and the training-efficiency claims rest on a simulator that matches the real system only moderately (Pearson 0.79, Spearman 0.69, Appendix H.1).

free parameters (6)
  • simulator communication factor = 4 (searched over 1..10)
    Appendix D: the factor scaling edge byte counts in the simulator's cost model is chosen to make simulator runtimes match the real engine; all Stage II training rewards inherit this fitted calibration.
  • simulator completion-time model P = not specified
    Algorithm 1 says the distribution over completion times is realized by a model that takes into account factors such as FLOPs and bytes, but the model form and its fitted constants are not given, so the simulation used for training is under-specified and partially fitted.
  • epsilon-greedy exploration schedule = 0.2 to 0.0 for DOPPLER; 0.5 to 0.0 for PLACETO
    Section 7.1 reports schedules chosen by experiment; reported assignment quality depends on these hand-picked schedules.
  • learning rate schedule = 1e-4 to 1e-7 for DOPPLER; 1e-3 to 1e-6 for PLACETO
    Section 7.1 states schedules were selected after trying initial values {1e-3, 1e-4, 1e-5}; this is hyperparameter tuning, not derivation.
  • entropy weight = 1e-2
    Section 7.1 applies a single entropy weight to all RL methods; it is chosen by hand and affects exploration quality.
  • GNN and FFNN architecture sizes = not reported
    Section 5.2 describes the network family, but the appendix does not give layer counts, hidden widths, or activation details, so the exact networks cannot be reconstructed.
assumptions (5)
  • domain assumption Algorithm 1 is a faithful proxy or digital twin for a real work-conserving scheduler
    Section 3 states the algorithm can serve as a reasonable proxy or digital twin for a real-life scheduler; all rewards in Stages II and III are defined through it, yet Appendix H.1 shows only moderate agreement with the real system.
  • domain assumption Completion times follow a stationary stochastic model P over FLOPs and byte counts
    Algorithm 1 models the next completed task by a fixed distribution given the schedule; real PCIe arbitration and kernel launch jitter are not guaranteed stationary, but the model assumes they are.
  • ad hoc to paper Sequential node-by-node assignment can express near-optimal assignments
    Section 5.1 restricts solutions to the class produced by Algorithm 3's single pass over |V| nodes; no argument shows this class contains the optimal assignment, only empirical evidence on four graphs.
  • standard math Policy gradient and the episodic MDP reduction are valid for this bandit problem
    Section 4 reformulates assignment as an episodic MDP following the combinatorial bandit result [32] and uses the policy gradient theorem [41]; these are standard results used without proof in the text.
  • domain assumption Sharded dataflow graphs from the authors' decomposition framework preserve the original computation's semantics
    The graphs come from the group's EinDecomp-style decomposition [6,12]; this paper does not argue correctness of the decomposition, it inherits it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DOPPLER: Dual-Policy Learning for Device Assignment in Asynchronous Dataflow Graphs." pith.science (2026). https://pith.science/paper/J4DX2DVD

@misc{pith2026250523131,
  author       = {Pith},
  title        = {Pith review of: DOPPLER: Dual-Policy Learning for Device Assignment in Asynchronous Dataflow Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J4DX2DVD}},
  note         = {Machine review of arXiv:2505.23131}
}
abstract

We study the problem of assigning operations in a dataflow graph to devices to minimize execution time in a work-conserving system, with emphasis on complex machine learning workloads. Prior learning-based methods often struggle due to three key limitations: (1) reliance on bulk-synchronous systems like TensorFlow, which under-utilize devices due to barrier synchronization; (2) lack of awareness of the scheduling mechanism of underlying systems when designing learning-based methods; and (3) exclusive dependence on reinforcement learning, ignoring the structure of effective heuristics designed by experts. In this paper, we propose Doppler, a three-stage framework for training dual-policy networks consisting of 1) a $\mathsf{SEL}$ policy for selecting operations and 2) a $\mathsf{PLC}$ policy for placing chosen operations on devices. Our experiments show that Doppler outperforms all baseline methods across tasks by reducing system execution time and additionally demonstrates sampling efficiency by reducing per-episode training time.

Figures

Figures reproduced from arXiv: 2505.23131 by the authors.

Figure 1
Figure 1. A decomposition of a matrix multiplication [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (Left) The ASSIGN algorithm, which sequentially produces an assignment A using SELθ policy and placed using PLCθ policy. (Right) A graphical depiction of the algorithm’s implementation. governs the “next completed task.” Given a schedule S and a current time tin, P is a joint distribution over the next task to complete and the time tout at which this task completes. Second, the function ChooseTask encapsulates the u… view at source ↗
Figure 3
Figure 3. Real engine execution times (in milliseconds) for [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Assignments for FFNN found by DOPPLER and PLACETO. Colors show the mapping of computations to GPUs. DOPPLER is more effective at load balancing across GPUs. Improving training us￾ing imitation learning, simulation-based RL, and real system RL (Q3) [PITH_FULL_IMAGE:fig…
Figure 5
Figure 5. Figure 5: Llama-block and Llama-layer architecture. Figure from [42] [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Assignment found by Doppler for ChainMM [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Assignment found by EnumerativeOptimizer for ChainMM [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Device and transfer utilization for DOPPLER, CHAINMM. 0 20 40 60 80 100 120 Time (ms) GPU 3 GPU 2 GPU 1 GPU 0 Memory Kernel Memory Kernel Memory Kernel Memory Kernel [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Device and transfer utilization for ENUMERATIVEOPTIMIZER, CHAINMM. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Device and transfer utilization for DOPPLER, FFNN. 0 20 40 60 80 100 120 Time (ms) GPU 3 GPU 2 GPU 1 GPU 0 Memory Kernel Memory Kernel Memory Kernel Memory Kernel [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Device and transfer utilization for PLACETO, FFNN. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Assignment found by DOPPLER for LLAMA-BLOCK The experiments demonstrate DOPPLER’s capability to efficiently achieve load balancing in a dis￾tributed computing environment, particularly when multiple GPUs are utilized for computation. Across varied workloads, we observ…
Figure 13
Figure 13. Figure 13: Assignment found by EnumerativeOptimizer for L [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: Assignment found by DOPPLER for LLAMA-LAYER 24 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: Assignment found by PLACETO for LLAMA-LAYER 25 [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]
Figure 16
Figure 16. Figure 16: Assignment found by CRITICAL PATH for LLAMA-LAYER F Synchronous system Configuration In [PITH_FULL_IMAGE:figures/full_fig_p026_16.png]
Figure 17
Figure 17. Figure 17: Kernel and memory operations across four GPUs under two scheduling strategies: [PITH_FULL_IMAGE:figures/full_fig_p027_17.png]
Figure 18
Figure 18. Figure 18: (left) A line chart showing a comparison between the simulator running time and the real [PITH_FULL_IMAGE:figures/full_fig_p028_18.png]
Figure 19
Figure 19. Figure 19: Three-stage framework for training DOOPLER cost-effectively by combining imitation learning, simulation-based RL, and real-system RL [PITH_FULL_IMAGE:figures/full_fig_p031_19.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 27 canonical work pages

  1. [1]

    {TensorFlow}: a system for {Large-Scale} machine learning

    Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. {TensorFlow}: a system for {Large-Scale} machine learning. In12th USENIX symposium on operating systems design and implementation (OSDI 16), pages 265–283, 2016

    2016

  2. [2]

    Hybrid symbiotic organisms search optimization algorithm for scheduling of tasks on cloud computing environment.PloS one, 11(6):e0158229, 2016

    Mohammed Abdullahi and Md Asri Ngadi. Hybrid symbiotic organisms search optimization algorithm for scheduling of tasks on cloud computing environment.PloS one, 11(6):e0158229, 2016

    work page 2016

  3. [3]

    Solving np-hard problems on graphs with extended alphago zero.arXiv preprint arXiv:1905.11623, 2019

    Kenshin Abe, Zijian Xu, Issei Sato, and Masashi Sugiyama. Solving np-hard problems on graphs with extended alphago zero.arXiv preprint arXiv:1905.11623, 2019

    arXiv 1905

  4. [4]

    Placeto: Learning generalizable device placement algorithms for distributed machine learning.arXiv preprint arXiv:1906.08879, 2019

    Ravichandra Addanki, Shaileshh Bojja Venkatakrishnan, Shreyan Gupta, Hongzi Mao, and Mo- hammad Alizadeh. Placeto: Learning generalizable device placement algorithms for distributed machine learning.arXiv preprint arXiv:1906.08879, 2019

    arXiv 1906

  5. [5]

    Neural combina- torial optimization with reinforcement learning.arXiv preprint arXiv:1611.09940, 2016

    Irwan Bello, Hieu Pham, Quoc V Le, Mohammad Norouzi, and Samy Bengio. Neural combina- torial optimization with reinforcement learning.arXiv preprint arXiv:1611.09940, 2016

    arXiv 2016

  6. [6]

    EinDecomp: Decomposition of declaratively- specified machine learning and numerical computations for parallel execution.Proceedings of the VLDB Endowment, 18(7):2240–2253, 2025

    Daniel Bourgeois, Zhimin Ding, Dimitrije Jankov, Jiehui Li, Mahmoud Sleem, Yuxin Tang, Jiawen Yao, Xinyu Yao, and Chris Jermaine. EinDecomp: Decomposition of declaratively- specified machine learning and numerical computations for parallel execution.Proceedings of the VLDB Endowment, 18(7):2240–2253, 2025

    work page 2025

  7. [7]

    Reinforcement learning driven heuristic optimization.arXiv preprint arXiv:1906.06639, 2019

    Qingpeng Cai, Will Hang, Azalia Mirhoseini, George Tucker, Jingtao Wang, and Wei Wei. Reinforcement learning driven heuristic optimization.arXiv preprint arXiv:1906.06639, 2019

    arXiv 1906

  8. [8]

    Improving optimization bounds using machine learning: Decision diagrams meet deep reinforcement learning

    Quentin Cappart, Emmanuel Goutierre, David Bergman, and Louis-Martin Rousseau. Improving optimization bounds using machine learning: Decision diagrams meet deep reinforcement learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 1443–1451, 2019

    work page 2019

Show all 43 references
  1. [9]

    Combinatorial bandits.Journal of Computer and System Sciences, 78(5):1404–1422, 2012

    Nicolo Cesa-Bianchi and Gábor Lugosi. Combinatorial bandits.Journal of Computer and System Sciences, 78(5):1404–1422, 2012

    2012

  2. [10]

    Combinatorial multi-armed bandit: General framework and applications

    Wei Chen, Yajun Wang, and Yang Yuan. Combinatorial multi-armed bandit: General framework and applications. InInternational conference on machine learning, pages 151–159. PMLR, 2013

    2013

  3. [11]

    Learning to perform local rewriting for combinatorial optimization.Advances in neural information processing systems, 32, 2019

    Xinyun Chen and Yuandong Tian. Learning to perform local rewriting for combinatorial optimization.Advances in neural information processing systems, 32, 2019

    2019

  4. [12]

    nondeterministic

    Zhimin Ding, Jiawen Yao, Brianna Barrow, Tania Lorido Botran, Christopher Jermaine, Yuxin Tang, Jiehui Li, Xinyu Yao, Sleem Mahmoud Abdelghafar, and Daniel Bourgeois. Turnip: A" nondeterministic" gpu runtime with cpu ram offload.arXiv preprint arXiv:2405.16283, 2024

    2024 arXiv

  5. [13]

    Learning to solve combinatorial optimization problems on real-world graphs in linear time

    Iddo Drori, Anant Kharkar, William R Sickinger, Brandon Kates, Qiang Ma, Suwen Ge, Eden Dolev, Brenda Dietrich, David P Williamson, and Madeleine Udell. Learning to solve combinatorial optimization problems on real-world graphs in linear time. In2020 19th IEEE International Co...

    2020

  6. [14]

    A structure-aware frame- work for learning device placements on computation graphs.arXiv preprint arXiv:2405.14185, 2024

    Shukai Duan, Heng Ping, Nikos Kanakaris, Xiongye Xiao, Peiyu Zhang, Panagiotis Kyriakis, Nesreen K Ahmed, Guixiang Ma, Mihai Capota, Shahin Nazarian, et al. A structure-aware frame- work for learning device placements on computation graphs.arXiv preprint arXiv:2405.14185, 2024

    2024 arXiv

  7. [15]

    Learning permutations with sinkhorn policy gradient.arXiv preprint arXiv:1805.07010, 2018

    Patrick Emami and Sanjay Ranka. Learning permutations with sinkhorn policy gradient.arXiv preprint arXiv:1805.07010, 2018

    2018 arXiv

  8. [16]

    A linear-time heuristic for improving network partitions

    Charles M Fiduccia and Robert M Mattheyses. A linear-time heuristic for improving network partitions. InPapers on Twenty-five years of electronic design automation, pages 241–247. 1988. 10

    1988

  9. [17]

    Compiling machine learning programs via high-level tracing.Systems for Machine Learning, 4(9), 2018

    Roy Frostig, Matthew James Johnson, and Chris Leary. Compiling machine learning programs via high-level tracing.Systems for Machine Learning, 4(9), 2018

    2018

  10. [18]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. InInternational conference on machine learning, pages 1263–1272. PMLR, 2017

    2017

  11. [19]

    A deep learning algorithm for the max-cut problem based on pointer network structure with supervised learning and reinforcement learning strategies.Mathematics, 8(2):298, 2020

    Shenshen Gu and Yue Yang. A deep learning algorithm for the max-cut problem based on pointer network structure with supervised learning and reinforcement learning strategies.Mathematics, 8(2):298, 2020

    2020

  12. [20]

    New spectral methods for ratio cut partitioning and clustering

    Lars Hagen and Andrew B Kahng. New spectral methods for ratio cut partitioning and clustering. IEEE transactions on computer-aided design of integrated circuits and systems, 11(9):1074– 1085, 1992

    1992

  13. [21]

    A novel device placement approach based on position-aware subgraph neural networks.Neurocomputing, 582:127501, 2024

    Meng Han, Yan Zeng, Jilin Zhang, Yongjian Ren, Meiting Xue, and Mingyao Zhou. A novel device placement approach based on position-aware subgraph neural networks.Neurocomputing, 582:127501, 2024

    2024

  14. [22]

    Optimization by simulated annealing: An experimental evaluation; part i, graph partitioning.Operations research, 37(6):865–892, 1989

    David S Johnson, Cecilia R Aragon, Lyle A McGeoch, and Catherine Schevon. Optimization by simulated annealing: An experimental evaluation; part i, graph partitioning.Operations research, 37(6):865–892, 1989

    1989

  15. [23]

    Metis: Unstructured graph partitioning and sparse matrix ordering system

    George Karypis. Metis: Unstructured graph partitioning and sparse matrix ordering system. Technical report, 1997

    1997

  16. [24]

    An efficient heuristic procedure for partitioning graphs.The Bell system technical journal, 49(2):291–307, 1970

    Brian W Kernighan and Shen Lin. An efficient heuristic procedure for partitioning graphs.The Bell system technical journal, 49(2):291–307, 1970

    1970

  17. [25]

    Learning combinatorial optimization algorithms over graphs.Advances in neural information processing systems, 30, 2017

    Elias Khalil, Hanjun Dai, Yuyu Zhang, Bistra Dilkina, and Le Song. Learning combinatorial optimization algorithms over graphs.Advances in neural information processing systems, 30, 2017

    2017

  18. [26]

    Optimization by simulated annealing

    Scott Kirkpatrick, C Daniel Gelatt Jr, and Mario P Vecchi. Optimization by simulated annealing. science, 220(4598):671–680, 1983

    1983

  19. [27]

    A conservation law for a wide class of queueing disciplines.Naval Research Logistics Quarterly, 12(2):181–192, 1965

    Leonard Kleinrock. A conservation law for a wide class of queueing disciplines.Naval Research Logistics Quarterly, 12(2):181–192, 1965

    1965

  20. [28]

    Attention, learn to solve routing problems! arXiv preprint arXiv:1803.08475, 2018

    Wouter Kool, Herke Van Hoof, and Max Welling. Attention, learn to solve routing problems! arXiv preprint arXiv:1803.08475, 2018

    2018 arXiv

  21. [29]

    Static scheduling algorithms for allocating directed task graphs to multiprocessors.ACM Computing Surveys (CSUR), 31(4):406–471, 1999

    Yu-Kwong Kwok and Ishfaq Ahmad. Static scheduling algorithms for allocating directed task graphs to multiprocessors.ACM Computing Surveys (CSUR), 31(4):406–471, 1999

    1999

  22. [30]

    Ranked reward: Enabling self-play reinforcement learning for combinatorial optimization.arXiv preprint arXiv:1807.01672, 2018

    Alexandre Laterre, Yunguan Fu, Mohamed Khalil Jabri, Alain-Sam Cohen, David Kas, Karl Hajjar, Torbjorn S Dahl, Amine Kerkeni, and Karim Beguir. Ranked reward: Enabling self-play reinforcement learning for combinatorial optimization.arXiv preprint arXiv:1807.01672, 2018

    2018 arXiv

  23. [31]

    Solving packing problems by conditional query learning

    Dongda Li, Changwei Ren, Zhaoquan Gu, Yuexuan Wang, and Francis Lau. Solving packing problems by conditional query learning. 2020

    2020

  24. [32]

    Combinatorial multivariant multi- armed bandits with applications to episodic reinforcement learning and beyond.arXiv preprint arXiv:2406.01386, 2024

    Xutong Liu, Siwei Wang, Jinhang Zuo, Han Zhong, Xuchuang Wang, Zhiyong Wang, Shuai Li, Mohammad Hajiesmaili, John Lui, and Wei Chen. Combinatorial multivariant multi- armed bandits with applications to episodic reinforcement learning and beyond.arXiv preprint arXiv:2406.01386, 2024

    2024 arXiv

  25. [33]

    A learning-based iterative method for solving vehicle routing problems

    Hao Lu, Xingwen Zhang, and Shuang Yang. A learning-based iterative method for solving vehicle routing problems. InInternational conference on learning representations, 2019

    2019

  26. [34]

    Learning heuristics over large graphs via deep reinforcement learning.arXiv preprint arXiv:1903.03332, 2019

    Sahil Manchanda, Akash Mittal, Anuj Dhawan, Sourav Medya, Sayan Ranu, and Ambuj Singh. Learning heuristics over large graphs via deep reinforcement learning.arXiv preprint arXiv:1903.03332, 2019. 11

    1903 arXiv

  27. [35]

    Reinforcement learning for combinatorial optimization: A survey.Computers & Operations Research, 134:105400, 2021

    Nina Mazyavkina, Sergey Sviridov, Sergei Ivanov, and Evgeny Burnaev. Reinforcement learning for combinatorial optimization: A survey.Computers & Operations Research, 134:105400, 2021

    2021

  28. [36]

    Device placement opti- mization with reinforcement learning

    Azalia Mirhoseini, Hieu Pham, Quoc V Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen Kumar, Mohammad Norouzi, Samy Bengio, and Jeff Dean. Device placement opti- mization with reinforcement learning. InInternational conference on machine learning, pages 2430–2439. PMLR, 2017

    2017

  29. [37]

    Reinforcement learning for solving the vehicle routing problem.Advances in neural information processing systems, 31, 2018

    Mohammadreza Nazari, Afshin Oroojlooy, Lawrence Snyder, and Martin Takác. Reinforcement learning for solving the vehicle routing problem.Advances in neural information processing systems, 31, 2018

    2018

  30. [38]

    Reinforced genetic algorithm learning for optimizing computation graphs.arXiv preprint arXiv:1905.02494, 2019

    Aditya Paliwal, Felix Gimeno, Vinod Nair, Yujia Li, Miles Lubin, Pushmeet Kohli, and Oriol Vinyals. Reinforced genetic algorithm learning for optimizing computation graphs.arXiv preprint arXiv:1905.02494, 2019

    1905 arXiv

  31. [39]

    Pytorch: An imperative style, high-performance deep learning library.Advances in neural information processing systems, 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library.Advances in neural information processing systems, 32, 2019

    2019

  32. [40]

    A parallelisable multi-level banded diffusion scheme for computing balanced partitions with smooth boundaries

    François Pellegrini. A parallelisable multi-level banded diffusion scheme for computing balanced partitions with smooth boundaries. InEuro-Par 2007 Parallel Processing: 13th International Euro-Par Conference, Rennes, France, August 28-31, 2007. Proceedings 13, pages 195–204. S...

    2007

  33. [41]

    Policy gradient meth- ods for reinforcement learning with function approximation.Advances in neural information processing systems, 12, 1999

    Richard S Sutton, David McAllester, Satinder Singh, and Yishay Mansour. Policy gradient meth- ods for reinforcement learning with function approximation.Advances in neural information processing systems, 12, 1999

    1999

  34. [42]

    Exploring and building the llama 3 architecture: A deep dive into components, coding, and inference techniques, 2024

    Vignesh Yaadav. Exploring and building the llama 3 architecture: A deep dive into components, coding, and inference techniques, 2024

    2024

  35. [43]

    meta-ops

    Lianmin Zheng, Zhuohan Li, Hao Zhang, Yonghao Zhuang, Zhifeng Chen, Yanping Huang, Yida Wang, Yuanzhong Xu, Danyang Zhuo, Eric P Xing, et al. Alpa: Automating inter-and {Intra-Operator} parallelism for distributed deep learning. In16th USENIX Symposium on Operating Systems Des...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.