{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:ZBV2RIF3MBGQLBNMM5W7PPTZM4","short_pith_number":"pith:ZBV2RIF3","schema_version":"1.0","canonical_sha256":"c86ba8a0bb604d0585ac676df7be7967015793015f64bd2c92cc1648eb7e4d37","source":{"kind":"arxiv","id":"2406.07056","version":1},"attestation_state":"computed","paper":{"title":"Effectively Compress KV Heads for LLM","license":"http://creativecommons.org/publicdomain/zero/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Hao Yu, Jianxin Wu, Shen Li, Yong Li, Zelan Yang","submitted_at":"2024-06-11T08:37:33Z","abstract_excerpt":"The advent of pre-trained large language models (LLMs) has revolutionized various natural language processing tasks. These models predominantly employ an auto-regressive decoding mechanism that utilizes Key-Value (KV) caches to eliminate redundant calculations for previous tokens. Nevertheless, as context lengths and batch sizes increase, the linear expansion in memory footprint of KV caches becomes a key bottleneck of LLM deployment, which decreases generation speeds significantly. To mitigate this issue, previous techniques like multi-query attention (MQA) and grouped-query attention (GQA) h"},"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":"2406.07056","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/publicdomain/zero/1.0/","primary_cat":"cs.CL","submitted_at":"2024-06-11T08:37:33Z","cross_cats_sorted":[],"title_canon_sha256":"c27470cbadc8e7396557ae4724b3500bfc3f0ddf97ed9d5079d4d1658f2f18de","abstract_canon_sha256":"2296e336ae95d5b185f148ba6c75af6f9da25efe9757c398cdfb3d9f19312503"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:30:17.663804Z","signature_b64":"aDaQ8t57wL8ixxGK9iq7OGS/merz9kLhCAM24S7utQIRBEBT/iSs3vv2mCgZssfS1jCAzHD+m417zCFd/mNfAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"c86ba8a0bb604d0585ac676df7be7967015793015f64bd2c92cc1648eb7e4d37","last_reissued_at":"2026-07-05T08:30:17.663254Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:30:17.663254Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Effectively Compress KV Heads for LLM","license":"http://creativecommons.org/publicdomain/zero/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Hao Yu, Jianxin Wu, Shen Li, Yong Li, Zelan Yang","submitted_at":"2024-06-11T08:37:33Z","abstract_excerpt":"The advent of pre-trained large language models (LLMs) has revolutionized various natural language processing tasks. These models predominantly employ an auto-regressive decoding mechanism that utilizes Key-Value (KV) caches to eliminate redundant calculations for previous tokens. Nevertheless, as context lengths and batch sizes increase, the linear expansion in memory footprint of KV caches becomes a key bottleneck of LLM deployment, which decreases generation speeds significantly. To mitigate this issue, previous techniques like multi-query attention (MQA) and grouped-query attention (GQA) h"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2406.07056","kind":"arxiv","version":1},"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/2406.07056/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":"2406.07056","created_at":"2026-07-05T08:30:17.663352+00:00"},{"alias_kind":"arxiv_version","alias_value":"2406.07056v1","created_at":"2026-07-05T08:30:17.663352+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2406.07056","created_at":"2026-07-05T08:30:17.663352+00:00"},{"alias_kind":"pith_short_12","alias_value":"ZBV2RIF3MBGQ","created_at":"2026-07-05T08:30:17.663352+00:00"},{"alias_kind":"pith_short_16","alias_value":"ZBV2RIF3MBGQLBNM","created_at":"2026-07-05T08:30:17.663352+00:00"},{"alias_kind":"pith_short_8","alias_value":"ZBV2RIF3","created_at":"2026-07-05T08:30:17.663352+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.00573","citing_title":"LASER: Loss-Aware Singular-value Decomposition and Rank Allocation for Efficient Low-Precision Vision-Language Models","ref_index":55,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00535","citing_title":"DREAM-S: Speculative Decoding with Searchable Drafting and Target-Aware Refinement for Multimodal Generation","ref_index":58,"is_internal_anchor":false},{"citing_arxiv_id":"2410.13846","citing_title":"LightTransfer: Your Long-Context LLM is Secretly a Hybrid Model with Effortless Adaptation","ref_index":36,"is_internal_anchor":false},{"citing_arxiv_id":"2604.02570","citing_title":"WSVD: Weighted Low-Rank Approximation for Fast and Efficient Execution of Low-Precision Vision-Language Models","ref_index":17,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4","json":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4.json","graph_json":"https://pith.science/api/pith-number/ZBV2RIF3MBGQLBNMM5W7PPTZM4/graph.json","events_json":"https://pith.science/api/pith-number/ZBV2RIF3MBGQLBNMM5W7PPTZM4/events.json","paper":"https://pith.science/paper/ZBV2RIF3"},"agent_actions":{"view_html":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4","download_json":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4.json","view_paper":"https://pith.science/paper/ZBV2RIF3","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2406.07056&json=true","fetch_graph":"https://pith.science/api/pith-number/ZBV2RIF3MBGQLBNMM5W7PPTZM4/graph.json","fetch_events":"https://pith.science/api/pith-number/ZBV2RIF3MBGQLBNMM5W7PPTZM4/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4/action/storage_attestation","attest_author":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4/action/author_attestation","sign_citation":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4/action/citation_signature","submit_replication":"https://pith.science/pith/ZBV2RIF3MBGQLBNMM5W7PPTZM4/action/replication_record"}},"created_at":"2026-07-05T08:30:17.663352+00:00","updated_at":"2026-07-05T08:30:17.663352+00:00"}