SymTensor: Symbolic and Adaptive Tensor Partitioning by Unified Parallelism for Deep Learning

SymTensor: Symbolic and Adaptive Tensor Partitioning by Unified Parallelism for Deep Learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article SymTensor: Symbolic and Adaptive Tensor Partitioning by Unified Parallelism for Deep Learning Hongxing Wang, Zhengdao Yu, Chong Li, Serge Petiton This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7744659/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract The rapid expansion of deep learning models in scale and structural diversity has made distributed training essential. Designing efficient parallelization strategies requires balancing computation, communication, and memory. However, existing methods struggle to coordinate multiple parallelization strategies across different model components and to adapt to changing models. This paper proposes SymTensor, a strategy generation method based on a principled tensor-level cost model without relying on predefined rules.SymTensor unifies different forms of parallelism into a single system and formulates a symbolic model to jointly analyze computation, communication, and memory costs.It employs an adaptive tensor partitioning algorithm to minimize total cost.Our proposal adapts to changes such as model architectures, operator types, and input shapes. Our experiments on representative foundation models validated that SymTensor-generated strategies achieve up to more than 2x of the training performance compared to those generated by the state-of-the-art Megatron-LM.Our tensor-based symbolic-cost-driven solution provides strong efficiency, adaptability, and practicality over large-scale distributed training. Parallel Programming Symbolic Cost Model Tensor Partitioning Memory-Aware Parallelization Distributed Deep Learning Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 04 Feb, 2026 Reviewers agreed at journal 16 Jan, 2026 Reviews received at journal 09 Dec, 2025 Reviews received at journal 27 Nov, 2025 Reviewers agreed at journal 03 Nov, 2025 Reviewers agreed at journal 31 Oct, 2025 Reviewers invited by journal 29 Oct, 2025 Editor assigned by journal 28 Oct, 2025 Submission checks completed at journal 04 Oct, 2025 First submitted to journal 29 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7744659","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":541855709,"identity":"a4065e20-4e7b-4ada-bc92-71a6ab449bda","order_by":0,"name":"Hongxing Wang","email":"","orcid":"","institution":"Huawei Technologies (France)","correspondingAuthor":false,"prefix":"","firstName":"Hongxing","middleName":"","lastName":"Wang","suffix":""},{"id":541855710,"identity":"5e76e08a-a0fd-4855-b79f-652fc68d4a23","order_by":1,"name":"Zhengdao Yu","email":"","orcid":"","institution":"Huawei Technologies (France)","correspondingAuthor":false,"prefix":"","firstName":"Zhengdao","middleName":"","lastName":"Yu","suffix":""},{"id":541855711,"identity":"323a4f6e-08d4-4a68-a0f9-8f052ec96a57","order_by":2,"name":"Chong Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAp0lEQVRIiWNgGAWjYDACCQbGAx8bQCzGBqK1MBycSbKWw7zEKgYD/tkNDIdtdxyWNzje3MDwo2IbEZbcOcBwOPfMYcMNZw42MPacuU1Yi4FEAlBL22HGbTcSG5gZ24jVYtl22H7b/YekaGFsO5y47QYjkVokgO452NuWnrz/DJBBlF/4ZyQffPCzzdp2Zvvxhw9+VBChBSUGDxCjfhSMglEwCkYBEQAALd1CzAZbSW0AAAAASUVORK5CYII=","orcid":"","institution":"Huawei Technologies (France)","correspondingAuthor":true,"prefix":"","firstName":"Chong","middleName":"","lastName":"Li","suffix":""},{"id":541855712,"identity":"b65b4cc8-d910-4fc8-901a-00fe5d788408","order_by":3,"name":"Serge Petiton","email":"","orcid":"","institution":"University of Lille","correspondingAuthor":false,"prefix":"","firstName":"Serge","middleName":"","lastName":"Petiton","suffix":""}],"badges":[],"createdAt":"2025-09-29 18:23:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7744659/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7744659/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95542381,"identity":"2810e560-2df1-44db-905a-1535c5c4fc9c","added_by":"auto","created_at":"2025-11-10 11:51:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":501570,"visible":true,"origin":"","legend":"","description":"","filename":"IJPP2025paper1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7744659/v1/37db643fb0236818386dbbc5.pdf"},{"id":95542380,"identity":"8bfbfc4d-9a6d-4f3a-98b6-9727f346366a","added_by":"auto","created_at":"2025-11-10 11:51:06","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5484,"visible":true,"origin":"","legend":"","description":"","filename":"b7aa70ae6dde40dd98cb75ae8ec47e81.json","url":"https://assets-eu.researchsquare.com/files/rs-7744659/v1/430a821cc84181cce8c90b27.json"},{"id":95654390,"identity":"2c5b6a31-3b80-447b-ace7-4696d4522672","added_by":"auto","created_at":"2025-11-11 16:11:31","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":447480,"visible":true,"origin":"","legend":"","description":"","filename":"IJPP2025paper1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7744659/v1_covered_841168b2-f000-4531-977c-9a091b92e38a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"SymTensor: Symbolic and Adaptive Tensor Partitioning by Unified Parallelism for Deep Learning","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"international-journal-of-parallel-programming","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ijpp","sideBox":"Learn more about [International Journal of Parallel Programming](http://link.springer.com/journal/10766)","snPcode":"10766","submissionUrl":"https://submission.nature.com/new-submission/10766/3","title":"International Journal of Parallel Programming","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Parallel Programming, Symbolic Cost Model, Tensor Partitioning, Memory-Aware Parallelization, Distributed Deep Learning","lastPublishedDoi":"10.21203/rs.3.rs-7744659/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7744659/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"The rapid expansion of deep learning models in scale and structural diversity has made distributed training essential. Designing efficient parallelization strategies requires balancing computation, communication, and memory. However, existing methods struggle to coordinate multiple parallelization strategies across different model components and to adapt to changing models. This paper proposes SymTensor, a strategy generation method based on a principled tensor-level cost model without relying on predefined rules.SymTensor unifies different forms of parallelism into a single system and formulates a symbolic model to jointly analyze computation, communication, and memory costs.It employs an adaptive tensor partitioning algorithm to minimize total cost.Our proposal adapts to changes such as model architectures, operator types, and input shapes. Our experiments on representative foundation models validated that SymTensor-generated strategies achieve up to more than 2x of the training performance compared to those generated by the state-of-the-art Megatron-LM.Our tensor-based symbolic-cost-driven solution provides strong efficiency, adaptability, and practicality over large-scale distributed training.","manuscriptTitle":"SymTensor: Symbolic and Adaptive Tensor Partitioning by Unified Parallelism for Deep Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-10 11:51:02","doi":"10.21203/rs.3.rs-7744659/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-02-04T18:46:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"146678469633285254820533571710069210738","date":"2026-01-16T16:46:26+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-09T18:10:30+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-27T21:29:22+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"69684093294771361275451749107701218517","date":"2025-11-03T05:44:44+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"173633217443944681233876713290069190538","date":"2025-10-31T07:37:42+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-10-29T07:23:03+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-28T09:42:06+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-04T08:25:21+00:00","index":"","fulltext":""},{"type":"submitted","content":"International Journal of Parallel Programming","date":"2025-09-29T18:18:50+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"international-journal-of-parallel-programming","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ijpp","sideBox":"Learn more about [International Journal of Parallel Programming](http://link.springer.com/journal/10766)","snPcode":"10766","submissionUrl":"https://submission.nature.com/new-submission/10766/3","title":"International Journal of Parallel Programming","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"eea2984c-aaeb-4f04-a884-1e4541bbb1e9","owner":[],"postedDate":"November 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-02-24T05:24:35+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-10 11:51:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7744659","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7744659","identity":"rs-7744659","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}