{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:MEAED32TIIUNQLVBK5JNU75GBL","short_pith_number":"pith:MEAED32T","schema_version":"1.0","canonical_sha256":"610041ef534228d82ea15752da7fa60adf46064a44dfd02b335d77462b61fde0","source":{"kind":"arxiv","id":"2404.01847","version":3},"attestation_state":"computed","paper":{"title":"Accelerating Transformer Pre-training with 2:4 Sparsity","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jianfei Chen, Jun Zhu, Kang Zhao, Weiyu Huang, Yuezhou Hu","submitted_at":"2024-04-02T11:12:42Z","abstract_excerpt":"Training large transformers is slow, but recent innovations on GPU architecture give us an advantage. NVIDIA Ampere GPUs can execute a fine-grained 2:4 sparse matrix multiplication twice as fast as its dense equivalent. In the light of this property, we comprehensively investigate the feasibility of accelerating feed-forward networks (FFNs) of transformers in pre-training. First, we define a ``flip rate'' to monitor the stability of a 2:4 training process. Utilizing this metric, we propose three techniques to preserve accuracy: to modify the sparse-refined straight-through estimator by applyin"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2404.01847","kind":"arxiv","version":3},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2024-04-02T11:12:42Z","cross_cats_sorted":[],"title_canon_sha256":"dfec37736c4f46436e8acf20250af28f45f32a19fa648919181fbd817502f5d1","abstract_canon_sha256":"4d634a0538956f31d72d56a0d1f47ce5fd0c90448b244bdb3f35ddfd56a19d7e"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:26:25.368414Z","signature_b64":"zb/pGRQOZ0ss2iEuPHBBPDne/O0LBWamzF/0L+ccIqbwq3XxQeRcXl/3QPm8qljN60FzvOfFrXLUdNuDzNw7CA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"610041ef534228d82ea15752da7fa60adf46064a44dfd02b335d77462b61fde0","last_reissued_at":"2026-07-05T09:26:25.367931Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:26:25.367931Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Accelerating Transformer Pre-training with 2:4 Sparsity","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Jianfei Chen, Jun Zhu, Kang Zhao, Weiyu Huang, Yuezhou Hu","submitted_at":"2024-04-02T11:12:42Z","abstract_excerpt":"Training large transformers is slow, but recent innovations on GPU architecture give us an advantage. NVIDIA Ampere GPUs can execute a fine-grained 2:4 sparse matrix multiplication twice as fast as its dense equivalent. In the light of this property, we comprehensively investigate the feasibility of accelerating feed-forward networks (FFNs) of transformers in pre-training. First, we define a ``flip rate'' to monitor the stability of a 2:4 training process. Utilizing this metric, we propose three techniques to preserve accuracy: to modify the sparse-refined straight-through estimator by applyin"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2404.01847","kind":"arxiv","version":3},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2404.01847/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2404.01847","created_at":"2026-07-05T09:26:25.367989+00:00"},{"alias_kind":"arxiv_version","alias_value":"2404.01847v3","created_at":"2026-07-05T09:26:25.367989+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2404.01847","created_at":"2026-07-05T09:26:25.367989+00:00"},{"alias_kind":"pith_short_12","alias_value":"MEAED32TIIUN","created_at":"2026-07-05T09:26:25.367989+00:00"},{"alias_kind":"pith_short_16","alias_value":"MEAED32TIIUNQLVB","created_at":"2026-07-05T09:26:25.367989+00:00"},{"alias_kind":"pith_short_8","alias_value":"MEAED32T","created_at":"2026-07-05T09:26:25.367989+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2509.09682","citing_title":"Faster and Memory-Efficient Training of Sequential Recommendation Models for Large Catalogs","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2507.03052","citing_title":"From 2:4 to 8:16 sparsity patterns in LLMs for Outliers and Weights with Variance Correction","ref_index":15,"is_internal_anchor":false},{"citing_arxiv_id":"2509.22166","citing_title":"Motivating Next-Gen Accelerators with Flexible (N:M) Activation Sparsity via Benchmarking Lightweight Post-Training Sparsification Approaches","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2605.03667","citing_title":"ELAS: Efficient Pre-Training of Low-Rank Large Language Models via 2:4 Activation Sparsity","ref_index":3,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL","json":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL.json","graph_json":"https://pith.science/api/pith-number/MEAED32TIIUNQLVBK5JNU75GBL/graph.json","events_json":"https://pith.science/api/pith-number/MEAED32TIIUNQLVBK5JNU75GBL/events.json","paper":"https://pith.science/paper/MEAED32T"},"agent_actions":{"view_html":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL","download_json":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL.json","view_paper":"https://pith.science/paper/MEAED32T","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2404.01847&json=true","fetch_graph":"https://pith.science/api/pith-number/MEAED32TIIUNQLVBK5JNU75GBL/graph.json","fetch_events":"https://pith.science/api/pith-number/MEAED32TIIUNQLVBK5JNU75GBL/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL/action/timestamp_anchor","attest_storage":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL/action/storage_attestation","attest_author":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL/action/author_attestation","sign_citation":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL/action/citation_signature","submit_replication":"https://pith.science/pith/MEAED32TIIUNQLVBK5JNU75GBL/action/replication_record"}},"created_at":"2026-07-05T09:26:25.367989+00:00","updated_at":"2026-07-05T09:26:25.367989+00:00"}