{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2023:CFNEGUD6Y5RYZH4FUUJMH2WXZO","short_pith_number":"pith:CFNEGUD6","schema_version":"1.0","canonical_sha256":"115a43507ec7638c9f85a512c3ead7cbbd315793c85725c5af4306562852ccdf","source":{"kind":"arxiv","id":"2311.00257","version":2},"attestation_state":"computed","paper":{"title":"AMSP: Reducing Communication Overhead of ZeRO for Efficient LLM Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.DC","authors_text":"Guoteng Wang, Hang Yan, Peng Sun, Qiaoling Chen, Qinghao Hu, Tianwei Zhang, Ting Huang, Xun Chen, Yang Gao, Yingtong Xiong, Yonggang Wen","submitted_at":"2023-11-01T03:14:48Z","abstract_excerpt":"Training large language models (LLMs) encounters challenges in GPU memory consumption due to the high memory requirements of model states. The widely used Zero Redundancy Optimizer (ZeRO) addresses this issue through strategic sharding but introduces communication challenges at scale. To tackle this problem, we propose AMSP, a system designed to optimize ZeRO for scalable LLM training. AMSP incorporates three flexible sharding strategies: Full-Replica, Full-Sharding, and Partial-Sharding, and allows each component within the model states (Parameters, Gradients, Optimizer States) to independent"},"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":"2311.00257","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.DC","submitted_at":"2023-11-01T03:14:48Z","cross_cats_sorted":[],"title_canon_sha256":"9cd6a21c9547d7b64b65881df87d8c27668abbdda11d4b3a08f9000c0f684c8b","abstract_canon_sha256":"ead24029ce5aefd09bf159695aca34482b3bf26d2484ec4dbb695ead859d1838"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T07:55:25.841918Z","signature_b64":"g7xpN1slwLgG9XqjLxiYmokAot6Eyn9uySrtvm41YsSxvBpOgUGBlmPoHULQvEC3/BprSoMpcZf1fCZ0DqyDBQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"115a43507ec7638c9f85a512c3ead7cbbd315793c85725c5af4306562852ccdf","last_reissued_at":"2026-07-05T07:55:25.841468Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T07:55:25.841468Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"AMSP: Reducing Communication Overhead of ZeRO for Efficient LLM Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.DC","authors_text":"Guoteng Wang, Hang Yan, Peng Sun, Qiaoling Chen, Qinghao Hu, Tianwei Zhang, Ting Huang, Xun Chen, Yang Gao, Yingtong Xiong, Yonggang Wen","submitted_at":"2023-11-01T03:14:48Z","abstract_excerpt":"Training large language models (LLMs) encounters challenges in GPU memory consumption due to the high memory requirements of model states. The widely used Zero Redundancy Optimizer (ZeRO) addresses this issue through strategic sharding but introduces communication challenges at scale. To tackle this problem, we propose AMSP, a system designed to optimize ZeRO for scalable LLM training. AMSP incorporates three flexible sharding strategies: Full-Replica, Full-Sharding, and Partial-Sharding, and allows each component within the model states (Parameters, Gradients, Optimizer States) to independent"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2311.00257","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/2311.00257/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":"2311.00257","created_at":"2026-07-05T07:55:25.841523+00:00"},{"alias_kind":"arxiv_version","alias_value":"2311.00257v2","created_at":"2026-07-05T07:55:25.841523+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2311.00257","created_at":"2026-07-05T07:55:25.841523+00:00"},{"alias_kind":"pith_short_12","alias_value":"CFNEGUD6Y5RY","created_at":"2026-07-05T07:55:25.841523+00:00"},{"alias_kind":"pith_short_16","alias_value":"CFNEGUD6Y5RYZH4F","created_at":"2026-07-05T07:55:25.841523+00:00"},{"alias_kind":"pith_short_8","alias_value":"CFNEGUD6","created_at":"2026-07-05T07:55:25.841523+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2501.04266","citing_title":"Scaling Large Language Model Training on Frontier with Low-Bandwidth Partitioning","ref_index":11,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO","json":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO.json","graph_json":"https://pith.science/api/pith-number/CFNEGUD6Y5RYZH4FUUJMH2WXZO/graph.json","events_json":"https://pith.science/api/pith-number/CFNEGUD6Y5RYZH4FUUJMH2WXZO/events.json","paper":"https://pith.science/paper/CFNEGUD6"},"agent_actions":{"view_html":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO","download_json":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO.json","view_paper":"https://pith.science/paper/CFNEGUD6","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2311.00257&json=true","fetch_graph":"https://pith.science/api/pith-number/CFNEGUD6Y5RYZH4FUUJMH2WXZO/graph.json","fetch_events":"https://pith.science/api/pith-number/CFNEGUD6Y5RYZH4FUUJMH2WXZO/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO/action/timestamp_anchor","attest_storage":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO/action/storage_attestation","attest_author":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO/action/author_attestation","sign_citation":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO/action/citation_signature","submit_replication":"https://pith.science/pith/CFNEGUD6Y5RYZH4FUUJMH2WXZO/action/replication_record"}},"created_at":"2026-07-05T07:55:25.841523+00:00","updated_at":"2026-07-05T07:55:25.841523+00:00"}