{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2020:UYPVJX2HLVCUA5M3Y3ZYVP4L46","short_pith_number":"pith:UYPVJX2H","schema_version":"1.0","canonical_sha256":"a61f54df475d4540759bc6f38abf8be78052dfdbc50273eb2b58994fe58ec0ff","source":{"kind":"arxiv","id":"2006.09503","version":3},"attestation_state":"computed","paper":{"title":"Memory-Efficient Pipeline-Parallel DNN Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.DC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Amar Phanishayee, Deepak Narayanan, Kaiyu Shi, Matei Zaharia, Xie Chen","submitted_at":"2020-06-16T20:33:54Z","abstract_excerpt":"Many state-of-the-art ML results have been obtained by scaling up the number of parameters in existing models. However, parameters and activations for such large models often do not fit in the memory of a single accelerator device; this means that it is necessary to distribute training of large models over multiple accelerators. In this work, we propose PipeDream-2BW, a system that supports memory-efficient pipeline parallelism. PipeDream-2BW uses a novel pipelining and weight gradient coalescing strategy, combined with the double buffering of weights, to ensure high throughput, low memory foo"},"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":"2006.09503","kind":"arxiv","version":3},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2020-06-16T20:33:54Z","cross_cats_sorted":["cs.DC","stat.ML"],"title_canon_sha256":"d6e59f45561a0fe8af913ac8c27c7034ce4c0dc781f9a4d5949ebf589d6db0c0","abstract_canon_sha256":"e386ffdb742082225684495496b00499ae6248ef94b4782fc2f2a07a8d0a0c8a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T02:59:52.898689Z","signature_b64":"JmzXbGzpcomvkaaLqj6fCUaHoPmqxPCFzSs7hexhGTVc4LhFPNxBkMlfLhgRd/BYGv4ASYs/1GHg6ssX/3M5CQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"a61f54df475d4540759bc6f38abf8be78052dfdbc50273eb2b58994fe58ec0ff","last_reissued_at":"2026-07-05T02:59:52.898337Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T02:59:52.898337Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Memory-Efficient Pipeline-Parallel DNN Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.DC","stat.ML"],"primary_cat":"cs.LG","authors_text":"Amar Phanishayee, Deepak Narayanan, Kaiyu Shi, Matei Zaharia, Xie Chen","submitted_at":"2020-06-16T20:33:54Z","abstract_excerpt":"Many state-of-the-art ML results have been obtained by scaling up the number of parameters in existing models. However, parameters and activations for such large models often do not fit in the memory of a single accelerator device; this means that it is necessary to distribute training of large models over multiple accelerators. In this work, we propose PipeDream-2BW, a system that supports memory-efficient pipeline parallelism. PipeDream-2BW uses a novel pipelining and weight gradient coalescing strategy, combined with the double buffering of weights, to ensure high throughput, low memory foo"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2006.09503","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/2006.09503/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":"2006.09503","created_at":"2026-07-05T02:59:52.898395+00:00"},{"alias_kind":"arxiv_version","alias_value":"2006.09503v3","created_at":"2026-07-05T02:59:52.898395+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2006.09503","created_at":"2026-07-05T02:59:52.898395+00:00"},{"alias_kind":"pith_short_12","alias_value":"UYPVJX2HLVCU","created_at":"2026-07-05T02:59:52.898395+00:00"},{"alias_kind":"pith_short_16","alias_value":"UYPVJX2HLVCUA5M3","created_at":"2026-07-05T02:59:52.898395+00:00"},{"alias_kind":"pith_short_8","alias_value":"UYPVJX2H","created_at":"2026-07-05T02:59:52.898395+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07494","citing_title":"GIFT: Geometry-Informed Low-precision Gradient Communication for LLM Pretraining","ref_index":19,"is_internal_anchor":true},{"citing_arxiv_id":"2605.18750","citing_title":"A Readiness-Driven Runtime for Pipeline-Parallel Training under Runtime Variability","ref_index":32,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27085","citing_title":"Efficient Training on Multiple Consumer GPUs with RoundPipe","ref_index":34,"is_internal_anchor":false},{"citing_arxiv_id":"2605.05049","citing_title":"Piper: Efficient Large-Scale MoE Training via Resource Modeling and Pipelined Hybrid Parallelism","ref_index":27,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46","json":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46.json","graph_json":"https://pith.science/api/pith-number/UYPVJX2HLVCUA5M3Y3ZYVP4L46/graph.json","events_json":"https://pith.science/api/pith-number/UYPVJX2HLVCUA5M3Y3ZYVP4L46/events.json","paper":"https://pith.science/paper/UYPVJX2H"},"agent_actions":{"view_html":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46","download_json":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46.json","view_paper":"https://pith.science/paper/UYPVJX2H","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2006.09503&json=true","fetch_graph":"https://pith.science/api/pith-number/UYPVJX2HLVCUA5M3Y3ZYVP4L46/graph.json","fetch_events":"https://pith.science/api/pith-number/UYPVJX2HLVCUA5M3Y3ZYVP4L46/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46/action/timestamp_anchor","attest_storage":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46/action/storage_attestation","attest_author":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46/action/author_attestation","sign_citation":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46/action/citation_signature","submit_replication":"https://pith.science/pith/UYPVJX2HLVCUA5M3Y3ZYVP4L46/action/replication_record"}},"created_at":"2026-07-05T02:59:52.898395+00:00","updated_at":"2026-07-05T02:59:52.898395+00:00"}