{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:BFJLOAP54USTGBFE2FWPBJSMLD","short_pith_number":"pith:BFJLOAP5","schema_version":"1.0","canonical_sha256":"0952b701fde5253304a4d16cf0a64c58c63f4a61608d5a1fb81b670ec53b0d2e","source":{"kind":"arxiv","id":"2407.21118","version":2},"attestation_state":"computed","paper":{"title":"Palu: Compressing KV-Cache with Low-Rank Projection","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Chi-Chih Chang, Chien-Yu Lin, Chong-Yan Chen, Kai-Chiang Wu, Luis Ceze, Mohamed S. Abdelfattah, Ning-Chi Huang, Pei-Shuo Wang, Wei-Cheng Lin, Yu-Fang Hu","submitted_at":"2024-07-30T18:19:38Z","abstract_excerpt":"Post-training KV-Cache compression methods typically either sample a subset of effectual tokens or quantize the data into lower numerical bit width. However, these methods cannot exploit redundancy in the hidden dimension of the KV tensors. This paper presents a hidden dimension compression approach called Palu, a KV-Cache compression framework that utilizes low-rank projection to reduce inference-time LLM memory usage. Palu decomposes the linear layers into low-rank matrices, caches compressed intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, comp"},"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":"2407.21118","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.AI","submitted_at":"2024-07-30T18:19:38Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"76a1374c7de699cc9c1887c9c00ee66aaa96826201b59d4fba9be4e0c3fb6bc3","abstract_canon_sha256":"70b07c65f43d88a06f5dc469a1fe41ee99dd354264195866245f6f987557d31b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T09:30:19.636354Z","signature_b64":"QfP+ewPDpkO3IMgcvtEE8SwIkJ1wO+ZJLWeYEnZjEKsxe+vF2qtxSlooX3hQ3mMMMH/QTs8jKKLP/djFazxBDQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0952b701fde5253304a4d16cf0a64c58c63f4a61608d5a1fb81b670ec53b0d2e","last_reissued_at":"2026-07-05T09:30:19.635801Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T09:30:19.635801Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Palu: Compressing KV-Cache with Low-Rank Projection","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.LG"],"primary_cat":"cs.AI","authors_text":"Chi-Chih Chang, Chien-Yu Lin, Chong-Yan Chen, Kai-Chiang Wu, Luis Ceze, Mohamed S. Abdelfattah, Ning-Chi Huang, Pei-Shuo Wang, Wei-Cheng Lin, Yu-Fang Hu","submitted_at":"2024-07-30T18:19:38Z","abstract_excerpt":"Post-training KV-Cache compression methods typically either sample a subset of effectual tokens or quantize the data into lower numerical bit width. However, these methods cannot exploit redundancy in the hidden dimension of the KV tensors. This paper presents a hidden dimension compression approach called Palu, a KV-Cache compression framework that utilizes low-rank projection to reduce inference-time LLM memory usage. Palu decomposes the linear layers into low-rank matrices, caches compressed intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, comp"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2407.21118","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/2407.21118/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":"2407.21118","created_at":"2026-07-05T09:30:19.635871+00:00"},{"alias_kind":"arxiv_version","alias_value":"2407.21118v2","created_at":"2026-07-05T09:30:19.635871+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2407.21118","created_at":"2026-07-05T09:30:19.635871+00:00"},{"alias_kind":"pith_short_12","alias_value":"BFJLOAP54UST","created_at":"2026-07-05T09:30:19.635871+00:00"},{"alias_kind":"pith_short_16","alias_value":"BFJLOAP54USTGBFE","created_at":"2026-07-05T09:30:19.635871+00:00"},{"alias_kind":"pith_short_8","alias_value":"BFJLOAP5","created_at":"2026-07-05T09:30:19.635871+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":13,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.11164","citing_title":"ReasonAlloc: Hierarchical Decoding-Time KV Cache Budget Allocation for Reasoning Models","ref_index":11,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08565","citing_title":"EinSort: Sorting is All We Need for Tensorizing LLM","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.31105","citing_title":"GRKV: Global Regression for Training-Free KV Cache Compression in Long-Context LLMs","ref_index":5,"is_internal_anchor":false},{"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":6,"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":56,"is_internal_anchor":false},{"citing_arxiv_id":"2505.12942","citing_title":"A3 : an Analytical Low-Rank Approximation Framework for Attention","ref_index":4,"is_internal_anchor":false},{"citing_arxiv_id":"2312.05821","citing_title":"ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2605.17757","citing_title":"OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2605.19218","citing_title":"Rotation-Aligned Key Channel Pruning for Efficient Vision-Language Model Inference","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2509.21623","citing_title":"OjaKV: Context-Aware Online Low-Rank KV Cache Compression","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2603.22910","citing_title":"EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction","ref_index":3,"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":3,"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":8,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD","json":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD.json","graph_json":"https://pith.science/api/pith-number/BFJLOAP54USTGBFE2FWPBJSMLD/graph.json","events_json":"https://pith.science/api/pith-number/BFJLOAP54USTGBFE2FWPBJSMLD/events.json","paper":"https://pith.science/paper/BFJLOAP5"},"agent_actions":{"view_html":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD","download_json":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD.json","view_paper":"https://pith.science/paper/BFJLOAP5","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2407.21118&json=true","fetch_graph":"https://pith.science/api/pith-number/BFJLOAP54USTGBFE2FWPBJSMLD/graph.json","fetch_events":"https://pith.science/api/pith-number/BFJLOAP54USTGBFE2FWPBJSMLD/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD/action/storage_attestation","attest_author":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD/action/author_attestation","sign_citation":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD/action/citation_signature","submit_replication":"https://pith.science/pith/BFJLOAP54USTGBFE2FWPBJSMLD/action/replication_record"}},"created_at":"2026-07-05T09:30:19.635871+00:00","updated_at":"2026-07-05T09:30:19.635871+00:00"}