{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:N3HS76Y2IWUSH3K6Y47ZQQ7ICN","short_pith_number":"pith:N3HS76Y2","schema_version":"1.0","canonical_sha256":"6ecf2ffb1a45a923ed5ec73f9843e8135dadb1946080a84463ed33a67b1a7d52","source":{"kind":"arxiv","id":"2502.07864","version":5},"attestation_state":"computed","paper":{"title":"TransMLA: Multi-Head Latent Attention Is All You Need","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Fanxu Meng, Muhan Zhang, Pingzhi Tang, Xiaojuan Tang, Xing Sun, Zengwei Yao","submitted_at":"2025-02-11T18:20:18Z","abstract_excerpt":"In this paper, we present TransMLA, a framework that seamlessly converts any GQA-based pre-trained model into an MLA-based model. Our approach enables direct compatibility with DeepSeek's codebase, allowing these models to fully leverage DeepSeek-specific optimizations such as vLLM and SGlang. By compressing 93% of the KV cache in LLaMA-2-7B, TransMLA achieves a 10.6x inference speedup at an 8K context length while preserving meaningful output quality. Additionally, the model requires only 6 billion tokens for fine-tuning to regain performance on par with the original across multiple benchmark"},"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":"2502.07864","kind":"arxiv","version":5},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2025-02-11T18:20:18Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"78b9c6bf6c293e52e8097feffe0782d14a535ec9040f99c17d61f710dea7351e","abstract_canon_sha256":"05481715f94cc16c04b85adb37f3435e67980c1708aea6839fc6cf8ed6aad9e9"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:20:09.074157Z","signature_b64":"Q8dANKOTqJkfoF5zCfiDLoj7SPvVATSWvTr0ngHfondaYqB8e1PImKuuZ147WCiH7P4htl083BKKQth2C5sSCA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"6ecf2ffb1a45a923ed5ec73f9843e8135dadb1946080a84463ed33a67b1a7d52","last_reissued_at":"2026-07-05T11:20:09.073499Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:20:09.073499Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"TransMLA: Multi-Head Latent Attention Is All You Need","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Fanxu Meng, Muhan Zhang, Pingzhi Tang, Xiaojuan Tang, Xing Sun, Zengwei Yao","submitted_at":"2025-02-11T18:20:18Z","abstract_excerpt":"In this paper, we present TransMLA, a framework that seamlessly converts any GQA-based pre-trained model into an MLA-based model. Our approach enables direct compatibility with DeepSeek's codebase, allowing these models to fully leverage DeepSeek-specific optimizations such as vLLM and SGlang. By compressing 93% of the KV cache in LLaMA-2-7B, TransMLA achieves a 10.6x inference speedup at an 8K context length while preserving meaningful output quality. Additionally, the model requires only 6 billion tokens for fine-tuning to regain performance on par with the original across multiple benchmark"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2502.07864","kind":"arxiv","version":5},"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/2502.07864/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":"2502.07864","created_at":"2026-07-05T11:20:09.073590+00:00"},{"alias_kind":"arxiv_version","alias_value":"2502.07864v5","created_at":"2026-07-05T11:20:09.073590+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2502.07864","created_at":"2026-07-05T11:20:09.073590+00:00"},{"alias_kind":"pith_short_12","alias_value":"N3HS76Y2IWUS","created_at":"2026-07-05T11:20:09.073590+00:00"},{"alias_kind":"pith_short_16","alias_value":"N3HS76Y2IWUSH3K6","created_at":"2026-07-05T11:20:09.073590+00:00"},{"alias_kind":"pith_short_8","alias_value":"N3HS76Y2","created_at":"2026-07-05T11:20:09.073590+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":6,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.08032","citing_title":"What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents","ref_index":79,"is_internal_anchor":true},{"citing_arxiv_id":"2606.15079","citing_title":"Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale","ref_index":231,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05868","citing_title":"YouZhi: Towards High-Concurrency Financial LLMs via Adaptive GQA-to-MLA Transition","ref_index":33,"is_internal_anchor":false},{"citing_arxiv_id":"2605.30351","citing_title":"VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion","ref_index":18,"is_internal_anchor":false},{"citing_arxiv_id":"2509.09505","citing_title":"Combating the Memory Walls: Optimization Pathways for Long-Context Agentic LLM Inference","ref_index":45,"is_internal_anchor":false},{"citing_arxiv_id":"2605.11999","citing_title":"The Illusion of Power Capping in LLM Decode: A Phase-Aware Energy Characterisation Across Attention Architectures","ref_index":12,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN","json":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN.json","graph_json":"https://pith.science/api/pith-number/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/graph.json","events_json":"https://pith.science/api/pith-number/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/events.json","paper":"https://pith.science/paper/N3HS76Y2"},"agent_actions":{"view_html":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN","download_json":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN.json","view_paper":"https://pith.science/paper/N3HS76Y2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2502.07864&json=true","fetch_graph":"https://pith.science/api/pith-number/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/graph.json","fetch_events":"https://pith.science/api/pith-number/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/action/timestamp_anchor","attest_storage":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/action/storage_attestation","attest_author":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/action/author_attestation","sign_citation":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/action/citation_signature","submit_replication":"https://pith.science/pith/N3HS76Y2IWUSH3K6Y47ZQQ7ICN/action/replication_record"}},"created_at":"2026-07-05T11:20:09.073590+00:00","updated_at":"2026-07-05T11:20:09.073590+00:00"}