{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:WQ4Z65EXM6SQBBFGIZOZQOLS5K","short_pith_number":"pith:WQ4Z65EX","schema_version":"1.0","canonical_sha256":"b4399f749767a50084a6465d983972ea802df571e997b915928eac274b8223ef","source":{"kind":"arxiv","id":"2412.20677","version":2},"attestation_state":"computed","paper":{"title":"Align Attention Heads Before Merging Them: An Effective Way for Converting MHA to GQA","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Feng Zhou, Qingyun Jin, Xiaohui Song, Zengchang Qin","submitted_at":"2024-12-30T03:05:45Z","abstract_excerpt":"Large language models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks. However, as the model size and the input sequence's length increase, the linearly increasing key-value (KV) cache significantly degrades inference throughput. Therefore, grouped-query attention (GQA), as an alternative to multi-head attention (MHA), has been widely introduced into LLMs. In this work, we propose a cost-effective method for converting MHA into GQA with any compression ratio of KV heads. The key point of our method lies in the application of Procrustes analysis"},"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":"2412.20677","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2024-12-30T03:05:45Z","cross_cats_sorted":[],"title_canon_sha256":"b87a688ebcc995dbb4622e66760063c0fd7329470e0a45b870702e001d8f35a4","abstract_canon_sha256":"dbe04b4c24319f5f9b77988f7c5916326ad433291fe7754e264d7dbf3819474d"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:43:30.364664Z","signature_b64":"CVxFFzS74WvQtC9N1cOR4aUwYOS51b3waHpNQfm1DHKtPgnaQPmGQE9LOAqJsG6NjVbsft9PM6jYEVJRqqzQAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"b4399f749767a50084a6465d983972ea802df571e997b915928eac274b8223ef","last_reissued_at":"2026-07-05T11:43:30.364013Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:43:30.364013Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Align Attention Heads Before Merging Them: An Effective Way for Converting MHA to GQA","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Feng Zhou, Qingyun Jin, Xiaohui Song, Zengchang Qin","submitted_at":"2024-12-30T03:05:45Z","abstract_excerpt":"Large language models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks. However, as the model size and the input sequence's length increase, the linearly increasing key-value (KV) cache significantly degrades inference throughput. Therefore, grouped-query attention (GQA), as an alternative to multi-head attention (MHA), has been widely introduced into LLMs. In this work, we propose a cost-effective method for converting MHA into GQA with any compression ratio of KV heads. The key point of our method lies in the application of Procrustes analysis"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2412.20677","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/2412.20677/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":"2412.20677","created_at":"2026-07-05T11:43:30.364080+00:00"},{"alias_kind":"arxiv_version","alias_value":"2412.20677v2","created_at":"2026-07-05T11:43:30.364080+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2412.20677","created_at":"2026-07-05T11:43:30.364080+00:00"},{"alias_kind":"pith_short_12","alias_value":"WQ4Z65EXM6SQ","created_at":"2026-07-05T11:43:30.364080+00:00"},{"alias_kind":"pith_short_16","alias_value":"WQ4Z65EXM6SQBBFG","created_at":"2026-07-05T11:43:30.364080+00:00"},{"alias_kind":"pith_short_8","alias_value":"WQ4Z65EX","created_at":"2026-07-05T11:43:30.364080+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2505.02351","citing_title":"Opt-GPTQ: An Optimized GPTQ Combining Sparse Attention and Quantization Techniques","ref_index":13,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K","json":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K.json","graph_json":"https://pith.science/api/pith-number/WQ4Z65EXM6SQBBFGIZOZQOLS5K/graph.json","events_json":"https://pith.science/api/pith-number/WQ4Z65EXM6SQBBFGIZOZQOLS5K/events.json","paper":"https://pith.science/paper/WQ4Z65EX"},"agent_actions":{"view_html":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K","download_json":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K.json","view_paper":"https://pith.science/paper/WQ4Z65EX","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2412.20677&json=true","fetch_graph":"https://pith.science/api/pith-number/WQ4Z65EXM6SQBBFGIZOZQOLS5K/graph.json","fetch_events":"https://pith.science/api/pith-number/WQ4Z65EXM6SQBBFGIZOZQOLS5K/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K/action/timestamp_anchor","attest_storage":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K/action/storage_attestation","attest_author":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K/action/author_attestation","sign_citation":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K/action/citation_signature","submit_replication":"https://pith.science/pith/WQ4Z65EXM6SQBBFGIZOZQOLS5K/action/replication_record"}},"created_at":"2026-07-05T11:43:30.364080+00:00","updated_at":"2026-07-05T11:43:30.364080+00:00"}