{"paper":{"title":"TurboGR: An Accelerated Training System for Large-Scale Generative Recommendation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"TurboGR enables training of up to 0.2 billion parameter generative recommendation models on Ascend NPUs at 54.71 percent MFU with 0.97 scalability.","cross_cats":["cs.LG"],"primary_cat":"cs.DC","authors_text":"Hengfeng Wang, Huichao Chai, Lu Xu, Maojun Peng, Shiqing Fan, Wei Guo, Xuemiao Li, Yaoyuan Wang, Yibo Jin, Yongxiang Feng, Zhixin Wu","submitted_at":"2026-05-13T12:26:29Z","abstract_excerpt":"Generative recommendation (GR) has emerged as a promising paradigm that replaces fragmented, scenario-specific architectures with unified Transformer-based models, exhibiting scaling-law behavior where recommendation quality improves systematically with increased model capacity and training data. However, deploying GR at scale on Ascend NPUs faces fundamental system-level challenges. These challenges are further exacerbated on Ascend NPUs due to the absence of high-performance implementations for jagged operators and the architectural mismatch between irregular sparse primitives and NPU's dens"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Evaluated on the KuaiRand-27K dataset, TurboGR supports training at up to 0.2B parameters and achieves 54.71% MFU with near-linear scalability (0.97).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The semi-asynchronous training and jagged optimizations preserve model quality and convergence while the reported MFU and scalability numbers generalize beyond the specific KuaiRand-27K setup and Ascend hardware configuration.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TurboGR trains up to 0.2B-parameter generative recommendation models on Ascend NPUs at 54.71% MFU with 0.97 near-linear scalability via jagged acceleration, hierarchical parallelism, and negative sampling optimizations.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"TurboGR enables training of up to 0.2 billion parameter generative recommendation models on Ascend NPUs at 54.71 percent MFU with 0.97 scalability.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"77d23e31e58850d03cbc97b7b3494c92fc08020f60f0aa48f6ef20e52de40dfc"},"source":{"id":"2605.13433","kind":"arxiv","version":1},"verdict":{"id":"cca99fea-3770-46ce-8585-3dc1c982a71b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T18:13:34.300120Z","strongest_claim":"Evaluated on the KuaiRand-27K dataset, TurboGR supports training at up to 0.2B parameters and achieves 54.71% MFU with near-linear scalability (0.97).","one_line_summary":"TurboGR trains up to 0.2B-parameter generative recommendation models on Ascend NPUs at 54.71% MFU with 0.97 near-linear scalability via jagged acceleration, hierarchical parallelism, and negative sampling optimizations.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The semi-asynchronous training and jagged optimizations preserve model quality and convergence while the reported MFU and scalability numbers generalize beyond the specific KuaiRand-27K setup and Ascend hardware configuration.","pith_extraction_headline":"TurboGR enables training of up to 0.2 billion parameter generative recommendation models on Ascend NPUs at 54.71 percent MFU with 0.97 scalability."},"references":{"count":21,"sample":[{"doi":"","year":2023,"title":"Personalized news recommendation: Methods and challenges.ACM Transactions on Information Systems, 41(1):1–50, 2023","work_id":"b52b3f9e-d7ba-4223-b120-632261c0b3a8","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2000,"title":"Analysis of recommendation algorithms for e-commerce","work_id":"841957dc-6d8d-4f90-aa17-da83415dfb9e","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Analyzing user engagement with tiktok’s short format video recommendations using data donations","work_id":"d1ad05e1-6b87-4031-bf5e-83cf6e2284e7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2001,"title":"Item-based collaborative filtering recommen- dation algorithms","work_id":"d013d10a-a65c-488d-8754-5652e1073e36","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Self-attentive sequential recommendation","work_id":"5c939252-a610-423e-8bb5-92e0ac0a59f6","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":21,"snapshot_sha256":"a27f0f8b341d20a24d8812d4174557be1c51c8a2f742cc18bd88f42b8313925b","internal_anchors":1},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}