Pith. sign in

REVIEW 3 major objections 2 minor 61 references

REACH: Reinforcement Learning for Efficient Allocation in Community and Heterogeneous Networks

T0 review · 3 major / 2 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read REACH is a Transformer-based reinforcement learning scheduler that claims to improve task completion rates on community GPU platforms by up to 17%, more than double the success rate for high-priority tasks, and reduce bandwidth penalties…

desk verdict This arXiv listing's abstract describes REACH, a GPU-scheduling RL system, but the full text is an unrelated empathetic-response paper—so there is no supporting content for any of REACH's claims. read the letter →

arxiv 2508.12857 v1 pith:R6L6YJY4 submitted 2025-08-18 cs.NI

Zhiwei Yu , Chengze Du , Heng Xu , Ying Zhou , Bo Liu , Jialong Li This is my paper · ORCID
classification cs.NI
keywords reinforcementlearningtaskschedulingcommunityGPUheterogeneousnetworksTransformersequencescoringbandwidthpenaltyresourcereliability
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

Community GPU platforms pool idle consumer GPUs from diverse, volatile, and network-constrained environments, but traditional schedulers struggle with such heterogeneity. REACH addresses this by recasting task scheduling as a sequence scoring problem and learning a Transformer-based reinforcement learning policy. The paper claims that, in extensive simulations, REACH outperforms state-of-the-art baselines by up to 17% in task completion rate, more than doubles the success rate for high-priority tasks, and reduces bandwidth penalties by over 80%. A sympathetic reader would care because an effective scheduler could make community GPU platforms a practical, cost-efficient alternative to centralized clusters for AI workloads.

What carries the argument

The central object is the Transformer-based reinforcement learning policy that scores candidate task-to-GPU allocation sequences. The scheduling problem is cast as sequence scoring: given the current global GPU states, task requirements, and network topology, the model produces a score for every feasible assignment sequence and selects the one with the highest score. This single learned decision mechanism incorporates data-computation co-location, priority-aware task handling, and mitigation of unreliable resources, replacing the multi-objective hand-tuning of traditional schedulers.

What would settle it

Deploy REACH and a strong baseline scheduler on a live community GPU testbed with dozens of heterogeneous consumer GPUs across multiple network domains, run the same workload mix for a week, and compare task completion rates and high-priority success rates; if REACH fails to improve upon the baseline or to more than double high-priority success, the central claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that scheduling on community GPU platforms can be effectively solved by a Transformer-based reinforcement learning agent that views task allocation as a sequence scoring problem. Instead of assigning tasks with hand-crafted heuristics, REACH learns a policy that maps the global GPU state, task requirements, and network conditions to a score for each candidate allocation sequence, then selects the highest-scoring sequence. This learned policy balances performance, reliability, cost, and network efficiency, and it adaptively co-locates computation with data, prioritizes critical jobs, and avoids unreliable resources. In simulation, REACH achieves up to 17% higher task completion rates, more than doubles the success rate for high-priority tasks, and cuts bandwidth penalties by over 80% relative to baseline schedulers, with stress tests showing resilience to GPU churn and network congestion and scalability experiments confirming gains in large-scale, high-contention settings.

Load-bearing premise

The simulation environment faithfully represents the real-world diversity, volatility, and network conditions of community GPU platforms; if real platforms behave differently, the reported gains may not reproduce in deployment.

Editorial extensions

If this is right

  • Existing schedulers for geo-distributed or volunteer GPU platforms could be replaced or augmented with learned sequence-scoring policies that adapt to heterogeneity and volatility.
  • Reliable scheduling would make community GPU platforms attractive for cost-sensitive AI training and inference, broadening access to computational resources.
  • The formulation could extend to other resource-constrained distributed systems such as edge computing, federated learning, or volunteer computing, where data location and resource reliability vary.
  • Reducing bandwidth penalties through learned co-location could lower the communication bottleneck in distributed training across wide-area networks.
  • The learned policy could be retrained or fine-tuned on new platform profiles without redesigning the scheduler, easing deployment across different community networks.

Reading between the lines

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

  • Because the reported gains come from simulation, a real-world deployment on a live community GPU testbed would be the natural next test; the policy's ability to transfer to unseen hardware and network behavior remains open.
  • REACH's sequence scoring could also be used as a re-ranking layer on top of existing heuristic schedulers, reducing adoption friction while still yielding improvements.
  • The paper's balancing of performance, reliability, cost, and network efficiency suggests a tunable scheduler could let users specify preferences (e.g., cheapest-first vs. fastest-first), an extension not explicitly explored.
  • Ablations that isolate the effects of data co-location versus reliability avoidance would clarify which component drives the measured gains and help generalize the approach to other platforms.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The submission consists of an abstract describing REACH, a Transformer-based reinforcement learning scheduler for community GPU platforms, alongside a full text that is entirely a different paper titled 'E3RG: Building Explicit Emotion-driven Empathetic Response Generation System with Multimodal Large Language Model' by Lin et al. The abstract claims that REACH improves task completion rates by up to 17%, more than doubles high-priority task success, and reduces bandwidth penalties by over 80% in extensive simulations, with robustness to GPU churn and network congestion. The full text instead discusses multimodal empathetic response generation and an ACM MM 2025 competition entry, containing no mention of REACH, scheduling, community GPU platforms, simulations, baselines, or any related technical content.

Significance. If the REACH results were presented with adequate methodology, the claims of substantial gains in task completion, high-priority success, and bandwidth efficiency would be of practical interest to the community GPU scheduling literature. However, the manuscript as submitted provides no evidence for these claims: the abstract is not supported by any corresponding body text, methodology, experiments, or comparisons. The E3RG paper may itself be a valid contribution to multimodal empathetic response generation, but it is irrelevant to the REACH claims. The novelty and significance of REACH cannot be assessed from the submitted material.

major comments (3)
  1. [Abstract vs. full text] The full text of the submission is an unrelated paper titled 'E3RG: Building Explicit Emotion-driven Empathetic Response Generation System with Multimodal Large Language Model' (ACM MM 2025), which contains no reference to REACH, community GPU platforms, task scheduling, simulation environments, or any of the quantitative results in the abstract. The central claims of the abstract are therefore entirely unsupported by the manuscript as submitted.
  2. [Methodology] No description of the REACH framework, its Transformer-based sequence scoring formulation, state representation, reward design, training procedure, or hyperparameters appears anywhere in the submitted text. Similarly, the simulation environment, baseline schedulers, evaluation metrics, and stress-test protocols referenced in the abstract are all absent. Without these components, the reported improvements of 17% task completion, doubled high-priority success, and 80% bandwidth penalty reduction cannot be verified or reproduced.
  3. [Results and reproducibility] The abstract reports 'extensive simulation results' and 'stress tests' but the submitted document contains no figures, tables, error bars, statistical analyses, or code artifacts for these experiments. The complete detachment between the abstract and the body text means that the empirical claims are unverifiable, and no amount of revision to the current text can remedy this without essentially rewriting the paper.
minor comments (2)
  1. [Manuscript metadata] The arXiv identifier in the header footer (arXiv:2508.12854v1 [cs.AI]) differs from the abstract's stated identifier (arXiv:2508.12857), which is confusing and may indicate a compilation or submission error; this should be resolved if a corrected manuscript is submitted.
  2. [Title and scope] The title, keywords, and abstract of the manuscript describe REACH, while the body reports E3RG with different keywords and CCS concepts; this inconsistency should be corrected by submitting the intended REACH manuscript instead of the current mismatched content.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation identified: the manuscript body is an unrelated paper, so REACH's claims have no derivation chain to audit.

full rationale

The circularity analysis requires a load-bearing derivation chain whose outputs reduce to its own inputs by construction, by fitted parameter renaming, or by a self-citation chain. The supplied full text is 'E3RG: Building Explicit Emotion-driven Empathetic Response Generation System with Multimodal Large Language Model' by Lin et al., which contains no mention of REACH, community GPU platforms, task scheduling, simulation environments, baselines, or stress tests. The abstract's claim that 'Extensive simulation results show that REACH improves task completion rates by up to 17%' therefore has no supporting equations, model description, or experimental specification anywhere in the manuscript. Because no derivation chain is present, no specific reduction can be quoted, and no circularity step meets the required evidence standard. The abstract/body mismatch is a severe manuscript-integrity and completeness problem, and the claimed results are unverifiable from the submitted text, but it is not an instance of circularity as defined here. Accordingly, the honest non-finding is score 0.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

Only the abstract is available; the full text is a different paper. The listed domain assumption is implicit in the abstract's simulation-based claims.

assumptions (1)
  • domain assumption The simulation environment used in the evaluation faithfully represents real community GPU platform dynamics (hardware/software diversity, availability volatility, network conditions).
    The abstract's performance claims are based on simulation; their real-world validity depends on this representational assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REACH: Reinforcement Learning for Efficient Allocation in Community and Heterogeneous Networks." pith.science (2026). https://pith.science/paper/R6L6YJY4

@misc{pith2026250812857,
  author       = {Pith},
  title        = {Pith review of: REACH: Reinforcement Learning for Efficient Allocation in Community and Heterogeneous Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6L6YJY4}},
  note         = {Machine review of arXiv:2508.12857}
}
read the original abstract

Community GPU platforms are emerging as a cost-effective and democratized alternative to centralized GPU clusters for AI workloads, aggregating idle consumer GPUs from globally distributed and heterogeneous environments. However, their extreme hardware/software diversity, volatile availability, and variable network conditions render traditional schedulers ineffective, leading to suboptimal task completion. In this work, we present REACH (Reinforcement Learning for Efficient Allocation in Community and Heterogeneous Networks), a Transformer-based reinforcement learning framework that redefines task scheduling as a sequence scoring problem to balance performance, reliability, cost, and network efficiency. By modeling both global GPU states and task requirements, REACH learns to adaptively co-locate computation with data, prioritize critical jobs, and mitigate the impact of unreliable resources. Extensive simulation results show that REACH improves task completion rates by up to 17%, more than doubles the success rate for high-priority tasks, and reduces bandwidth penalties by over 80% compared to state-of-the-art baselines. Stress tests further demonstrate its robustness to GPU churn and network congestion, while scalability experiments confirm its effectiveness in large-scale, high-contention scenarios.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 46 canonical work pages

  1. [1]

    11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEcompsocconfonly \@IEEEauthordefaulttextstyle \@IEEEcompsocnotconfonly \@IEEEauthor...

    work page

  2. [2]

    Pricing for gpu instances

    Amazon Web Services . Pricing for gpu instances. https://aws.amazon.com/ec2/pricing/on-demand/, 2025. Accessed July 2025

    work page 2025

  3. [3]

    Volunteer computing

    David P Anderson. Volunteer computing . arXiv preprint arXiv:1903.01699 , 2020. A comprehensive overview of the VC paradigm, its potential, and challenges

    arXiv 1903

  4. [4]

    Anderson, Jeff Cobb, Eric Korpela, Matt Lebofsky, and Dan Werthimer

    David P. Anderson, Jeff Cobb, Eric Korpela, Matt Lebofsky, and Dan Werthimer. SETI@home : An experiment in public-resource computing. In Communications of the ACM , volume 45, pages 56--61, 2002

    work page 2002

  5. [5]

    R. et. Anil. Palm2 technical report. arXiv preprint arXiv:2305.10403 , 2023. Improved multilingual, reasoning and coding capabilities; more compute‑efficient than PaLM

    arXiv 2023

  6. [6]

    Distributed Deep Learning Using Volunteer Computing-Like Paradigm

    Medha Atre, Birendra Jha, and Ashwini Rao. Distributed Deep Learning Using Volunteer Computing-Like Paradigm . In 2021 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) , pages 826--835. IEEE, 2021

    work page 2021

  7. [7]

    How much VRAM do you need for Blender ?, October 2023

    Matt Bach. How much VRAM do you need for Blender ?, October 2023. Accessed on: 2024-05-16

    work page 2023

  8. [8]

    The untapped potential of idle gpus

    Kurt Bains. The untapped potential of idle gpus. https://semianalysis.com/semianalysis-data-products/#gpu-rental-pricing, 2025. Accessed July 2025

    work page 2025

Show all 61 references
  1. [9]

    slurm on kubernetes

    benchess. Comment on "slurm on kubernetes". Hacker News , Jan 2021. URL: https://news.ycombinator.com/item?id=25908166

    2021

  2. [10]

    Brown, Benjamin Mann, and Nick Ryder etc

    Tom B. Brown, Benjamin Mann, and Nick Ryder etc. Language models are few-shot learners. CoRR , abs/2005.14165, 2020

    2005 arXiv

  3. [11]

    Gandiva fair : A fair GPU cluster scheduler for deep learning workloads

    Shubham Chaudhary, Ramachandran Ramjee, Muthian Sivathanu, Nipun Kwatra, and Srinidhi Viswanatha. Gandiva fair : A fair GPU cluster scheduler for deep learning workloads. In Proceedings of the Fifteenth European Conference on Computer Systems (EuroSys '20) , pages 1--16. ACM, 2020

    2020

  4. [12]

    Elastic deep learning in multi-tenant GPU clusters

    Jiarui Cheng, Shen Zhang, Huanle Xu, Bingsheng He, Wei Chen, and Wing Cheong Lau. Elastic deep learning in multi-tenant GPU clusters. In IEEE Transactions on Parallel and Distributed Systems , volume 32, pages 3010--3023. IEEE, 2021

    2021

  5. [13]

    Ng S. T. Chong. GPU-as-a-Service , harvesting idle capacity, and the rise of alternative processing units, February 2025. United Nations Univ. Blog, accessed 2025-07-13

    2025

  6. [14]

    Palm: Scaling language modeling with pathways

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 , 2022

    2022 arXiv

  7. [15]

    Coppock, Brian Zhang, Eliot H

    Patrick H. Coppock, Brian Zhang, Eliot H. Solomon, Vasilis Kypriotis, Leon Yang, Bikash Sharma, Dan Schatzberg, Todd C. Mowry, and Dimitrios Skarlatos. LithOS : An operating system for efficient machine learning on GPUs . arXiv preprint arXiv:2504.15465 , 2025

    2025 arXiv

  8. [16]

    The promise of analog deep learning: Recent advances, challenges and opportunities, 2024

    Aditya Datar and Pramit Saha. The promise of analog deep learning: Recent advances, challenges and opportunities, 2024

    2024

  9. [17]

    QLoRA : Efficient finetuning of quantized LLMs , 2023

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA : Efficient finetuning of quantized LLMs , 2023

    2023

  10. [18]

    BERT : Pre-training of deep bidirectional transformers for language understanding, 2018

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT : Pre-training of deep bidirectional transformers for language understanding, 2018

    2018

  11. [19]

    Cross-timeslot optimization for distributed gpu inference using reinforcement learning

    Chengze Du et al. Cross-timeslot optimization for distributed gpu inference using reinforcement learning. arXiv preprint arXiv:2507.10259 , 2025

    2025

  12. [20]

    Measuring GPU utilization one level deeper

    Paul Elvinger, Foteini Strati, Natalie Enright Jerger, and Ana Klimovic. Measuring GPU utilization one level deeper. arXiv preprint arXiv:2501.16909 , 2025

    2025

  13. [21]

    Anderson

    Trilce Estrada, Michela Taufer, Kevin Reed, and David P. Anderson. Scheduling strategies for BOINC projects: a case study with SimBA . In 2006 2nd IEEE International Conference on e-Science and Grid Computing (e-Science) , pages 109--109. IEEE, 2006

    2006

  14. [22]

    Garcia, Rafael Mayo, Enrique S

    Jose M. Garcia, Rafael Mayo, Enrique S. Quintana-Orti, and Federico Silla. Topology-aware GPU scheduling for learning workloads in cloud environments. In 2018 IEEE/ACM International Conference on Utility and Cloud Computing (UCC) , pages 189--198. IEEE, 2018

    2018

  15. [23]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, and …. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 , 2024. Multilingual, coding, reasoning, tool use; up to 405B params; compositional multimodal capability

    2024 arXiv

  16. [24]

    Deep Reinforcement Learning for Job Scheduling and Resource Management in Cloud Computing: An Algorithm-Level Review

    Yan Gu, Zhaoze Liu, Shuhong Dai, Cong Liu, Ying Wang, Shen Wang, Georgios Theodoropoulos, and Long Cheng. Deep Reinforcement Learning for Job Scheduling and Resource Management in Cloud Computing: An Algorithm-Level Review . arXiv preprint arXiv:2501.01007 , 2025

    2025 arXiv

  17. [25]

    Henderson, Mathieu Lacage, George F

    Thomas R. Henderson, Mathieu Lacage, George F. Riley, Christoph Dowell, and Sally Floyd. Network simulations with the ns‑3 simulator. In Proceedings of the 2008 ACM CoNEXT Conference (also SIGCOMM Posters/Demos) , pages 1--9, 2008. Early demonstration of ns‑3, a discrete-event...

    2008

  18. [26]

    Ark: Gpu-driven execution for distributed deep learning

    In-Chul Hwang, Gwang-Tae Lee, Jaeyong Lee, and Jae-Wook Lee. Ark: Gpu-driven execution for distributed deep learning. 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23) , pages 85--104, 2023

    2023

  19. [27]

    A survey on resource scheduling approaches in multi-access edge computing environment: a deep reinforcement learning study

    Ahmed A Ismail, Nour Eldeen Khalifa, and Reda A El-Khoribi. A survey on resource scheduling approaches in multi-access edge computing environment: a deep reinforcement learning study . Cluster Computing , 28(3), 2025

    2025

  20. [28]

    Analysis of large-scale multi-tenant gpu clusters for dnn training workloads

    Myeongjae Jeon, Shivaram Venkataraman, Amar Phanishayee, Junjie Qian, Wencong Xiao, and Fan Yang. Analysis of large-scale multi-tenant gpu clusters for dnn training workloads. In USENIX Annual Technical Conference (ATC) , pages 947--960, 2019

    2019

  21. [29]

    A genetic algorithm-based scheduling method for optimizing GPU utilization in multi-tenant cloud environments

    Jihye Kim, Sangmin Lee, and Chanik Park. A genetic algorithm-based scheduling method for optimizing GPU utilization in multi-tenant cloud environments. Applied Sciences , 14(11):4697, 2024

    2024

  22. [30]

    Eric J. Korpela. SETI@home , BOINC , and volunteer distributed computing. Annual Review of Earth and Planetary Sciences , 40:41--56, 2012

    2012

  23. [31]

    Mininet: An instant virtual network on your laptop (or other pc)

    Bob Lantz, Brandon Heller, and Nick McKeown. Mininet: An instant virtual network on your laptop (or other pc). Technical report / white paper, 2010. Introduces Mininet network emulator for SDN prototyping

    2010

  24. [32]

    Volunteer computing on mobile devices: State of the art and future research directions

    Anas L'Hanafi, Anas El-Hanafi, Abdellah Hair, Hicham Guermah, and Abderrahim El-Maouhab. Volunteer computing on mobile devices: State of the art and future research directions . International Journal of Advanced Computer Science and Applications , 7(1), 2016

    2016

  25. [33]

    Optimizing mixture-of-experts inference time combining model deployment and communication scheduling

    Jialong Li, Shreyansh Tripathi, Lakshay Rastogi, Yiming Lei, Rui Pan, and Yiting Xia. Optimizing mixture-of-experts inference time combining model deployment and communication scheduling. arXiv preprint arXiv:2410.17043 , 2024

    arXiv

  26. [34]

    Astraea: A Fair Deep Learning Scheduler for Multi-Tenant GPU Clusters

    Xiao Li, Zhi Zhou, Chen Xu, Jian He, and Keqin Li. Astraea: A Fair Deep Learning Scheduler for Multi-Tenant GPU Clusters . IEEE Transactions on Parallel and Distributed Systems , 33(11):3194--3207, 2022

    2022

  27. [35]

    Feng Liang, Zhen Zhang, Haifeng Lu, Chengming Li, Victor C. M. Leung, Yanyi Guo, and Xiping Hu. Resource allocation and workload scheduling for large-scale distributed deep learning: A survey. arXiv preprint arXiv:2406.08115 , 2024

    2024 arXiv

  28. [36]

    Resource Scheduling in Edge Computing: A Survey

    Siqi Luo and Jiankun Hu. Resource Scheduling in Edge Computing: A Survey . IEEE Access , 2023

    2023

  29. [37]

    Scheduling deep learning jobs in multi-tenant gpu clusters via wise resource sharing

    Yizhou Luo et al. Scheduling deep learning jobs in multi-tenant gpu clusters via wise resource sharing. arXiv preprint arXiv:2407.13088 , 2024

    2024 arXiv

  30. [38]

    Themis: Fair and efficient GPU cluster scheduling for ML training

    Kanthi Mahajan, Anirudh Gupta, Amar Phanishayee, Ashish Shah, and Mosharaf Chowdhury. Themis: Fair and efficient GPU cluster scheduling for ML training. In 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI '20) , pages 63--80. USENIX Association, 2020

    2020

  31. [39]

    Fast and Fair Training for Deep Learning in Heterogeneous GPU Clusters

    Zizhao Mo, Huanle Xu, and Wing Cheong Lau. Fast and Fair Training for Deep Learning in Heterogeneous GPU Clusters . In 2025 International Conference on Supercomputing (ICS ’25) . ACM, 2025

    2025

  32. [40]

    Gavel: Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads

    Deepak Narayanan, Keshav Santhanam, Fiodar Kazhamiaka, Amar Phanishayee, and Matei Zaharia. Gavel: Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads . In 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI '20) , pages 481--498, 2020

    2020

  33. [41]

    Gpt‑4 technical report

    OpenAI , Josh Achiam, and Adler et. Gpt‑4 technical report. arXiv preprint arXiv:2303.08774 , 2023. Large‑scale multimodal model; developed by OpenAI; passes bar exam in top 10

    2023 arXiv

  34. [42]

    Efficient flow scheduling in distributed deep learning training with echelon formation

    Rui Pan, Yiming Lei, Jialong Li, Zhiqiang Xie, Binhang Yuan, and Yiting Xia. Efficient flow scheduling in distributed deep learning training with echelon formation. In Proceedings of the 21st ACM Workshop on Hot Topics in Networks , pages 93--100, 2022

    2022

  35. [43]

    Robust speech recognition via large-scale weak supervision, 2023

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision, 2023

    2023

  36. [44]

    CASSINI : Network-aware job scheduling for ML training

    Vignesh Rajasekaran, Muhammad Zaheer, Sachin Katti, and Amar Phanishayee. CASSINI : Network-aware job scheduling for ML training. In 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) , pages 1--17. USENIX Association, 2024

    2024

  37. [45]

    High-resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022

    2022

  38. [46]

    Towards topology aware pre-emptive job scheduling with deep reinforcement learning

    Bon Ryu, Aijun An, Zana Rashidi, Junfeng Liu, and Yonggang Hu. Towards topology aware pre-emptive job scheduling with deep reinforcement learning. In Proceedings of the 30th Annual International Conference on Computer Science and Software Engineering , pages 1--5, 2020

    2020

  39. [47]

    DistilBERT , a distilled version of BERT : smaller, faster, cheaper and lighter, 2019

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. DistilBERT , a distilled version of BERT : smaller, faster, cheaper and lighter, 2019

    2019

  40. [48]

    Will LLMs scaling hit the wall? breaking barriers via distributed resources on massive edge devices

    Tao Shen, Zhangyang Wang, et al. Will LLMs scaling hit the wall? breaking barriers via distributed resources on massive edge devices. arXiv preprint arXiv:2503.08223 , 2025

    2025 arXiv

  41. [49]

    SDXL 1.0: A new era for generative ai, 2023

    Stability AI . SDXL 1.0: A new era for generative ai, 2023. Accessed on: 2024-05-16

    2023

  42. [50]

    Orion: Interference-aware, fine-grained GPU sharing for ML applications

    Foteini Strati, Xianzhe Ma, and Ana Klimovic. Orion: Interference-aware, fine-grained GPU sharing for ML applications. In Proceedings of the Nineteenth European Conference on Computer Systems (EuroSys '24) , pages 1--18. ACM, 2024

    2024

  43. [51]

    Hadar: Heterogeneity-Aware Optimization-Based misc Scheduling for Deep Learning Cluster

    Abeda Sultana, Fei Xu, Xu Yuan, Li Chen, and Nian-Feng Tzeng. Hadar: Heterogeneity-Aware Optimization-Based misc Scheduling for Deep Learning Cluster . In 2024 IEEE International Parallel and Distributed Processing Symposium (IPDPS) , pages 1--11. IEEE, 2024

    2024

  44. [52]

    Saladcloud: Rent and share gpus

    Salad Technologies. Saladcloud: Rent and share gpus. https://salad.com/compute, 2025. Accessed July 2025

    2025

  45. [53]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, and …. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 , 2023. Foundation models 7B–65B parameters, trained on public data, competitive with GPT‑3 / PaLM benchmarks

    2023 arXiv

  46. [54]

    GPU marketplace offerings data, 2025

    Vast.ai . GPU marketplace offerings data, 2025. Accessed: June 2025. URL: https://vast.ai/

    2025

  47. [55]

    Trustworthy distributed ai systems: Robustness, privacy, and governance

    Wenqi Wei and Ling Liu. Trustworthy distributed ai systems: Robustness, privacy, and governance. arXiv preprint arXiv:2402.01096 , 2024

    2024 arXiv

  48. [56]

    Taming GPU fragmentation in large-scale ML clusters with fragmentation gradient descent

    Qizhen Weng, Yizheng Zhu, Hong Wang, Zongyan Zhang, Wentao Li, and Jian Zhang. Taming GPU fragmentation in large-scale ML clusters with fragmentation gradient descent. In 2023 USENIX Annual Technical Conference (USENIX ATC 23) , pages 993--1007. USENIX Association, 2023

    2023

  49. [57]

    Yan and et al

    J. Yan and et al. Virtual‑time‑enabled mininet for scalable and accurate software‑defined network emulation. In Proceedings of SOSR , pages 1--6, 2015. Proposes VT‑Mininet to address performance fidelity issues in standard Mininet

    2015

  50. [58]

    GPU-Disaggregated Serving for Deep Learning Recommendation Models at Scale

    Lingyun Yang, Yongchen Wang, Yinghao Yu, Qizhen Weng, Jianbo Dong, Kan Liu, Chi Zhang, Yanyi Zi, Hao Li, Zechao Zhang, Nan Wang, Yu Dong, Menglei Zheng, Lanlan Xi, Xiaowei Lu, Liang Ye, Guodong Yang, Binzhang Fu, Tao Lan, Liping Zhang, Lin Qu, and Wei Wang. GPU-Disaggregated S...

    2025

  51. [59]

    Salus: Fine-grained GPU sharing primitives for deep learning applications

    Peifeng Yu and Mosharaf Chowdhury. Salus: Fine-grained GPU sharing primitives for deep learning applications. In Proceedings of the 3rd MLSys Conference , pages 383--396, 2020

    2020

  52. [60]

    TAG : An automatic framework for topology-aware and heterogeneity-aware distributed DNN training

    Si-Wei Zhang, Ziyue Li, Zirui Zhao, Yifei Wang, Ce Zhang, and Chuan Wu. TAG : An automatic framework for topology-aware and heterogeneity-aware distributed DNN training. IEEE Transactions on Parallel and Distributed Systems , 34(10):2846--2860, 2023

    2023

  53. [61]

    A survey on scheduling in deep learning systems

    Zibin Zheng, Jiacheng Chen, Zili Wang, Zhipeng Zhang, Shuying Chen, Ce Chen, and Wei Wang. A survey on scheduling in deep learning systems . arXiv preprint arXiv:2205.11913 , 2022

    2022 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.