{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:3HBLRQPHZ5PT7QT7J6MNCD2PBT","short_pith_number":"pith:3HBLRQPH","schema_version":"1.0","canonical_sha256":"d9c2b8c1e7cf5f3fc27f4f98d10f4f0cef963974c13153f0f3313dbe7442f984","source":{"kind":"arxiv","id":"2310.03294","version":2},"attestation_state":"computed","paper":{"title":"DISTFLASHATTN: Distributed Memory-efficient Attention for Long-context LLMs Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.DC"],"primary_cat":"cs.LG","authors_text":"Anze Xie, Dacheng Li, Eric P. Xing, Hao Zhang, Ion Stoica, Joseph E. Gonzalez, Rulin Shao, Xuezhe Ma","submitted_at":"2023-10-05T03:47:57Z","abstract_excerpt":"FlashAttention (Dao, 2023) effectively reduces the quadratic peak memory usage to linear in training transformer-based large language models (LLMs) on a single GPU. In this paper, we introduce DISTFLASHATTN, a distributed memory-efficient attention mechanism optimized for long-context LLMs training. We propose three key techniques: token-level workload balancing, overlapping key-value communication, and a rematerialization-aware gradient checkpointing algorithm. We evaluate DISTFLASHATTN on Llama-7B and variants with sequence lengths from 32K to 512K. DISTFLASHATTN achieves 8x longer sequences"},"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":"2310.03294","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2023-10-05T03:47:57Z","cross_cats_sorted":["cs.AI","cs.DC"],"title_canon_sha256":"1e8a35ce44d5773751f8f000baf3afedff78c46e9b6019a392766d6e77929f0b","abstract_canon_sha256":"184a8c5bf20a96648214c8b1ecf0e441da3b0bac29299f046e9f659c5b182393"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:02:46.702707Z","signature_b64":"ezkHmf28zDH3ydHvzQHv3B7m5js9wFCYhg53RCTcrDRtZglzuHJb6k3nf1+wrQ5BsIxJCZakCIcI+X2CfSLlBA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d9c2b8c1e7cf5f3fc27f4f98d10f4f0cef963974c13153f0f3313dbe7442f984","last_reissued_at":"2026-07-05T08:02:46.702204Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:02:46.702204Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"DISTFLASHATTN: Distributed Memory-efficient Attention for Long-context LLMs Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI","cs.DC"],"primary_cat":"cs.LG","authors_text":"Anze Xie, Dacheng Li, Eric P. Xing, Hao Zhang, Ion Stoica, Joseph E. Gonzalez, Rulin Shao, Xuezhe Ma","submitted_at":"2023-10-05T03:47:57Z","abstract_excerpt":"FlashAttention (Dao, 2023) effectively reduces the quadratic peak memory usage to linear in training transformer-based large language models (LLMs) on a single GPU. In this paper, we introduce DISTFLASHATTN, a distributed memory-efficient attention mechanism optimized for long-context LLMs training. We propose three key techniques: token-level workload balancing, overlapping key-value communication, and a rematerialization-aware gradient checkpointing algorithm. We evaluate DISTFLASHATTN on Llama-7B and variants with sequence lengths from 32K to 512K. DISTFLASHATTN achieves 8x longer sequences"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2310.03294","kind":"arxiv","version":2},"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/2310.03294/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":"2310.03294","created_at":"2026-07-05T08:02:46.702256+00:00"},{"alias_kind":"arxiv_version","alias_value":"2310.03294v2","created_at":"2026-07-05T08:02:46.702256+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2310.03294","created_at":"2026-07-05T08:02:46.702256+00:00"},{"alias_kind":"pith_short_12","alias_value":"3HBLRQPHZ5PT","created_at":"2026-07-05T08:02:46.702256+00:00"},{"alias_kind":"pith_short_16","alias_value":"3HBLRQPHZ5PT7QT7","created_at":"2026-07-05T08:02:46.702256+00:00"},{"alias_kind":"pith_short_8","alias_value":"3HBLRQPH","created_at":"2026-07-05T08:02:46.702256+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":5,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2509.21275","citing_title":"InfiniPipe: Elastic Pipeline Parallelism for Efficient Variable-Length Long-Context LLM Training","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2402.08268","citing_title":"World Model on Million-Length Video And Language With Blockwise RingAttention","ref_index":17,"is_internal_anchor":false},{"citing_arxiv_id":"2312.06635","citing_title":"Gated Linear Attention Transformers with Hardware-Efficient Training","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2604.23798","citing_title":"ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14561","citing_title":"CoCoDiff: Optimizing Collective Communications for Distributed Diffusion Transformer Inference Under Ulysses Sequence Parallelism","ref_index":56,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT","json":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT.json","graph_json":"https://pith.science/api/pith-number/3HBLRQPHZ5PT7QT7J6MNCD2PBT/graph.json","events_json":"https://pith.science/api/pith-number/3HBLRQPHZ5PT7QT7J6MNCD2PBT/events.json","paper":"https://pith.science/paper/3HBLRQPH"},"agent_actions":{"view_html":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT","download_json":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT.json","view_paper":"https://pith.science/paper/3HBLRQPH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2310.03294&json=true","fetch_graph":"https://pith.science/api/pith-number/3HBLRQPHZ5PT7QT7J6MNCD2PBT/graph.json","fetch_events":"https://pith.science/api/pith-number/3HBLRQPHZ5PT7QT7J6MNCD2PBT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT/action/storage_attestation","attest_author":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT/action/author_attestation","sign_citation":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT/action/citation_signature","submit_replication":"https://pith.science/pith/3HBLRQPHZ5PT7QT7J6MNCD2PBT/action/replication_record"}},"created_at":"2026-07-05T08:02:46.702256+00:00","updated_at":"2026-07-05T08:02:46.702256+00:00"}