{"paper":{"title":"GQLA: Group-Query Latent Attention for Hardware-Adaptive Large Language Model Decoding","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Group-Query Latent Attention exposes two equivalent decoding paths from one set of weights for hardware-specific LLM inference.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Fanxu Meng","submitted_at":"2026-05-14T15:50:01Z","abstract_excerpt":"Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly. Its trained weights, however, expose only one decoding path - an absorbed MQA form - which ties efficient inference to H100-class compute-bandwidth ratios, forfeits tensor parallelism along the head axis, and yields no Multi-Token Prediction (MTP) gain on commodity inference GPUs such as the export-restricted H20. We propose Group-Query Latent Attention (GQLA), a minimal modification of MLA whose trained weights expos"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"A single set of GQLA weights pins the rooflines of both H100 (MQA-absorb, s_q=1) and H20 (GQA + MTP, s_q=2), while supporting up to 8-way zero-redundancy tensor parallelism on the GQA path, and compresses the per-token KV cache to 28.125% of the GQA baseline on the MQA-absorb path.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The two decoding paths remain algebraically equivalent after the TransGQLA conversion from a pretrained GQA checkpoint, so that accuracy and the claimed cache compression are preserved without any retraining or additional fine-tuning steps.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GQLA exposes two algebraically equivalent decoding paths over one set of weights so a single model can hit roofline on both high-end and commodity GPUs while cutting KV cache size to 28% on the absorbed path.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Group-Query Latent Attention exposes two equivalent decoding paths from one set of weights for hardware-specific LLM inference.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0933b17f132a73117b521f516ceb760935bfc2dcdf995054757fcc61bee9dfa6"},"source":{"id":"2605.15250","kind":"arxiv","version":1},"verdict":{"id":"a4e98122-ca52-4802-b3b6-245c97842b2f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:17:06.763524Z","strongest_claim":"A single set of GQLA weights pins the rooflines of both H100 (MQA-absorb, s_q=1) and H20 (GQA + MTP, s_q=2), while supporting up to 8-way zero-redundancy tensor parallelism on the GQA path, and compresses the per-token KV cache to 28.125% of the GQA baseline on the MQA-absorb path.","one_line_summary":"GQLA exposes two algebraically equivalent decoding paths over one set of weights so a single model can hit roofline on both high-end and commodity GPUs while cutting KV cache size to 28% on the absorbed path.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The two decoding paths remain algebraically equivalent after the TransGQLA conversion from a pretrained GQA checkpoint, so that accuracy and the claimed cache compression are preserved without any retraining or additional fine-tuning steps.","pith_extraction_headline":"Group-Query Latent Attention exposes two equivalent decoding paths from one set of weights for hardware-specific LLM inference."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.15250/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T17:31:18.462331Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T17:26:19.710935Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T15:01:54.172784Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:33:22.817469Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"0556c384898048622e706d9fddd3da1f2d300e4ea0b608f6c1b2ae9410ff5045"},"references":{"count":18,"sample":[{"doi":"","year":1911,"title":"Fast Transformer Decoding: One Write-Head is All You Need","work_id":"160ea164-b1d4-4adb-8ccb-a4655d8a0bb4","ref_index":1,"cited_arxiv_id":"1911.02150","is_internal_anchor":true},{"doi":"","year":null,"title":"Proceedings of EMNLP , year=","work_id":"51df08b2-6fee-42c0-8f5d-c0b3c9766f25","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model","work_id":"1e1df141-cac8-47fd-b068-c4c96e51e331","ref_index":3,"cited_arxiv_id":"2405.04434","is_internal_anchor":true},{"doi":"","year":null,"title":"Advances in Neural Information Processing Systems , volume=","work_id":"547d76ec-c69c-4bfd-b1b2-c92f6be4b490","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models","work_id":"07c85cc5-4086-4abc-823b-6d0f4ff784d0","ref_index":5,"cited_arxiv_id":"2512.02556","is_internal_anchor":true}],"resolved_work":18,"snapshot_sha256":"195fc96cd6f57a942634cae074d705a9664ba6b60756790531f7130e054e3dc8","internal_anchors":7},"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"}