{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:Z7GIEZMBGLOK56ZYI65E6URSX4","short_pith_number":"pith:Z7GIEZMB","schema_version":"1.0","canonical_sha256":"cfcc82658132dcaefb3847ba4f5232bf141a8e0830a0cad28d0dc76f6230512f","source":{"kind":"arxiv","id":"2502.15304","version":1},"attestation_state":"computed","paper":{"title":"SVDq: 1.25-bit and 410x Key Cache Compression for LLM Attention","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Hong Yankun, Liu Wulong, Li Xing, Yuan Mingxuan, Yu Xianzhi, Zhen Hui-Ling","submitted_at":"2025-02-21T08:55:21Z","abstract_excerpt":"For the efficient inference of Large Language Models (LLMs), the effective compression of key-value (KV) cache is essential. Three main types of KV cache compression techniques, namely sparsity, channel compression, and quantization, have been identified. This study presents SVDq, a Singular Value Decomposition (SVD) - based mixed precision quantization method for K cache. Initially, K cache is transformed into latent channels using SVD basis representations. Since the values in latent channels decay rapidly and become negligible after only a few latent channels, our method then incorporates i"},"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":"2502.15304","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2025-02-21T08:55:21Z","cross_cats_sorted":["cs.AI","cs.CL"],"title_canon_sha256":"635caee327d9fe3b5c45c272f1eaa325a7c88f07f86adca7427107fa69c4929e","abstract_canon_sha256":"6a4767a3fc0b358d2185dc3b62dde41aa3d6185831c48155e04978b91bfbc03a"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T10:17:59.471852Z","signature_b64":"uzMfc0kbtF1jxvLw7jlhNyWGoKTO2tOsXxt8o9DUWbxX2ROTH8jn0a0J5ZFU6xruCYovk/ti9PfmuZ4641LDBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"cfcc82658132dcaefb3847ba4f5232bf141a8e0830a0cad28d0dc76f6230512f","last_reissued_at":"2026-07-05T10:17:59.471358Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T10:17:59.471358Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"SVDq: 1.25-bit and 410x Key Cache Compression for LLM Attention","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Hong Yankun, Liu Wulong, Li Xing, Yuan Mingxuan, Yu Xianzhi, Zhen Hui-Ling","submitted_at":"2025-02-21T08:55:21Z","abstract_excerpt":"For the efficient inference of Large Language Models (LLMs), the effective compression of key-value (KV) cache is essential. Three main types of KV cache compression techniques, namely sparsity, channel compression, and quantization, have been identified. This study presents SVDq, a Singular Value Decomposition (SVD) - based mixed precision quantization method for K cache. Initially, K cache is transformed into latent channels using SVD basis representations. Since the values in latent channels decay rapidly and become negligible after only a few latent channels, our method then incorporates i"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.15304","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/2502.15304/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":"2502.15304","created_at":"2026-07-05T10:17:59.471441+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.15304v1","created_at":"2026-07-05T10:17:59.471441+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.15304","created_at":"2026-07-05T10:17:59.471441+00:00"},{"alias_kind":"pith_short_12","alias_value":"Z7GIEZMBGLOK","created_at":"2026-07-05T10:17:59.471441+00:00"},{"alias_kind":"pith_short_16","alias_value":"Z7GIEZMBGLOK56ZY","created_at":"2026-07-05T10:17:59.471441+00:00"},{"alias_kind":"pith_short_8","alias_value":"Z7GIEZMB","created_at":"2026-07-05T10:17:59.471441+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2607.00760","citing_title":"MosaicKV: Serving Long-Context LLM with Dynamic Two-D KV Cache Compression","ref_index":59,"is_internal_anchor":false},{"citing_arxiv_id":"2605.02905","citing_title":"eOptShrinkQ: Near-Lossless KV Cache Compression Through Optimal Spectral Denoising and Quantization","ref_index":39,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4","json":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4.json","graph_json":"https://pith.science/api/pith-number/Z7GIEZMBGLOK56ZYI65E6URSX4/graph.json","events_json":"https://pith.science/api/pith-number/Z7GIEZMBGLOK56ZYI65E6URSX4/events.json","paper":"https://pith.science/paper/Z7GIEZMB"},"agent_actions":{"view_html":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4","download_json":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4.json","view_paper":"https://pith.science/paper/Z7GIEZMB","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.15304&json=true","fetch_graph":"https://pith.science/api/pith-number/Z7GIEZMBGLOK56ZYI65E6URSX4/graph.json","fetch_events":"https://pith.science/api/pith-number/Z7GIEZMBGLOK56ZYI65E6URSX4/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4/action/timestamp_anchor","attest_storage":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4/action/storage_attestation","attest_author":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4/action/author_attestation","sign_citation":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4/action/citation_signature","submit_replication":"https://pith.science/pith/Z7GIEZMBGLOK56ZYI65E6URSX4/action/replication_record"}},"created_at":"2026-07-05T10:17:59.471441+00:00","updated_at":"2026-07-05T10:17:59.471441+00:00"}