{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2021:L7OYSFSVBQGOGCT5BUTOA7RHKQ","short_pith_number":"pith:L7OYSFSV","schema_version":"1.0","canonical_sha256":"5fdd8916550c0ce30a7d0d26e07e27543421862204fbafbedae2fc0771ebfc72","source":{"kind":"arxiv","id":"2107.02027","version":2},"attestation_state":"computed","paper":{"title":"Efficient Sequence Packing without Cross-contamination: Accelerating Large Language Models without Impacting Performance","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.CC","cs.IT","cs.LG","math.IT"],"primary_cat":"cs.CL","authors_text":"Andrew Fitzgibbon, Mario Michael Krell, Matej Kosec, Sergio P. Perez","submitted_at":"2021-06-29T04:37:23Z","abstract_excerpt":"Effective training of today's large language models (LLMs) depends on large batches and long sequences for throughput and accuracy. To handle variable-length sequences on hardware accelerators, it is common practice to introduce padding tokens, so that all sequences in a batch have the same length. We show in this paper that the variation in sequence lengths in common NLP datasets is such that up to 50% of all tokens can be padding. In less common, but not extreme, cases (e.g. GLUE-cola with sequence length 128), the ratio is up to 89%. Existing methods to address the resulting inefficiency ar"},"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":"2107.02027","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2021-06-29T04:37:23Z","cross_cats_sorted":["cs.CC","cs.IT","cs.LG","math.IT"],"title_canon_sha256":"a35ac4ef54910d0d4db133514dbe866a811a9c7012d8c391786118426128934b","abstract_canon_sha256":"d78706899687440daa186035fd2e634b3d5d6e2e1b75f396c7f0c8126dc10670"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T05:03:49.410615Z","signature_b64":"m0e8BfHOcJ6YwJns0HAWaEAmK6fbkvQ/Q15764PiNxuwXPCUYornM3fWLPo9X59cnbdAOaXUDmyfR3r3yooOAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"5fdd8916550c0ce30a7d0d26e07e27543421862204fbafbedae2fc0771ebfc72","last_reissued_at":"2026-07-05T05:03:49.410131Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T05:03:49.410131Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Efficient Sequence Packing without Cross-contamination: Accelerating Large Language Models without Impacting Performance","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"","cross_cats":["cs.CC","cs.IT","cs.LG","math.IT"],"primary_cat":"cs.CL","authors_text":"Andrew Fitzgibbon, Mario Michael Krell, Matej Kosec, Sergio P. Perez","submitted_at":"2021-06-29T04:37:23Z","abstract_excerpt":"Effective training of today's large language models (LLMs) depends on large batches and long sequences for throughput and accuracy. To handle variable-length sequences on hardware accelerators, it is common practice to introduce padding tokens, so that all sequences in a batch have the same length. We show in this paper that the variation in sequence lengths in common NLP datasets is such that up to 50% of all tokens can be padding. In less common, but not extreme, cases (e.g. GLUE-cola with sequence length 128), the ratio is up to 89%. Existing methods to address the resulting inefficiency ar"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2107.02027","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/2107.02027/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":"2107.02027","created_at":"2026-07-05T05:03:49.410187+00:00"},{"alias_kind":"arxiv_version","alias_value":"2107.02027v2","created_at":"2026-07-05T05:03:49.410187+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2107.02027","created_at":"2026-07-05T05:03:49.410187+00:00"},{"alias_kind":"pith_short_12","alias_value":"L7OYSFSVBQGO","created_at":"2026-07-05T05:03:49.410187+00:00"},{"alias_kind":"pith_short_16","alias_value":"L7OYSFSVBQGOGCT5","created_at":"2026-07-05T05:03:49.410187+00:00"},{"alias_kind":"pith_short_8","alias_value":"L7OYSFSV","created_at":"2026-07-05T05:03:49.410187+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":21,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2607.07675","citing_title":"Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence","ref_index":50,"is_internal_anchor":true},{"citing_arxiv_id":"2607.05196","citing_title":"Unified Audio Intelligence Without Regressing on Text Intelligence","ref_index":146,"is_internal_anchor":true},{"citing_arxiv_id":"2606.19989","citing_title":"Online Dynamic Batching with Formal Guarantees for LLM Training","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2607.01701","citing_title":"Arachne: Orchestrating Cascades for Efficient Text-to-Video Model Training","ref_index":24,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08476","citing_title":"FlashCP: Load-Balanced Communication-Efficient Context Parallelism for LLM Training","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30460","citing_title":"HSAP: A Hierarchical Sequence-aware Parallelism for Hybrid-Context Generative Models","ref_index":60,"is_internal_anchor":false},{"citing_arxiv_id":"2606.29814","citing_title":"Nemotron-Labs-Diffusion-Image: Advancing Masked Discrete Diffusion for High-Resolution Image Synthesis","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2606.30460","citing_title":"HSAP: A Hierarchical Sequence-aware Parallelism for Hybrid-Context Generative Models","ref_index":60,"is_internal_anchor":false},{"citing_arxiv_id":"2605.27918","citing_title":"Addressing Variable Heterogeneity in Distributed Multimodal Training with Entrain","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2605.31485","citing_title":"Graphical einops: bridging tensor networks and computation graphs","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2502.02189","citing_title":"deCIFer: Crystal Structure Prediction from Powder Diffraction Data using Autoregressive Language Models","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2504.09844","citing_title":"MegaScale-Data: Scaling Dataloader for Multisource Large Foundation Model Training","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2510.18830","citing_title":"MTraining: Distributed Dynamic Sparse Attention for Efficient Ultra-Long Context Training","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2412.13663","citing_title":"Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference","ref_index":158,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18710","citing_title":"Mosaic: Towards Efficient Training of Multimodal Models with Spatial Resource Multiplexing","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2509.21275","citing_title":"InfiniPipe: Elastic Pipeline Parallelism for Efficient Variable-Length Long-Context LLM Training","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2603.24562","citing_title":"Scaling Recurrence-aware Foundation Models for Clinical Records via Next-Visit Prediction","ref_index":28,"is_internal_anchor":false},{"citing_arxiv_id":"2402.17177","citing_title":"Sora: A Review on Background, Technology, Limitations, and Opportunities of Large Vision Models","ref_index":41,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11255","citing_title":"HEBATRON: A Hebrew-Specialized Open-Weight Mixture-of-Experts Language Model","ref_index":22,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06374","citing_title":"ResiHP: Taming LLM Training Failures with Dynamic Hybrid Parallelism","ref_index":27,"is_internal_anchor":false},{"citing_arxiv_id":"2605.06374","citing_title":"ResiHP: Taming LLM Training Failures with Dynamic Hybrid Parallelism","ref_index":27,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ","json":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ.json","graph_json":"https://pith.science/api/pith-number/L7OYSFSVBQGOGCT5BUTOA7RHKQ/graph.json","events_json":"https://pith.science/api/pith-number/L7OYSFSVBQGOGCT5BUTOA7RHKQ/events.json","paper":"https://pith.science/paper/L7OYSFSV"},"agent_actions":{"view_html":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ","download_json":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ.json","view_paper":"https://pith.science/paper/L7OYSFSV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2107.02027&json=true","fetch_graph":"https://pith.science/api/pith-number/L7OYSFSVBQGOGCT5BUTOA7RHKQ/graph.json","fetch_events":"https://pith.science/api/pith-number/L7OYSFSVBQGOGCT5BUTOA7RHKQ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ/action/storage_attestation","attest_author":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ/action/author_attestation","sign_citation":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ/action/citation_signature","submit_replication":"https://pith.science/pith/L7OYSFSVBQGOGCT5BUTOA7RHKQ/action/replication_record"}},"created_at":"2026-07-05T05:03:49.410187+00:00","updated_at":"2026-07-05T05:03:49.410187+00:00"}