{ "definition": "hard AHA gate zeros / all token x KV-head x layer decisions, token-weighted over prefill and decode", "router_granularity": "token", "native_router_decisions": 53161864, "effective_router_decisions": 425294912, "native_router_note": "Native decisions count learned gate logits before broadcast; effective decisions always use token x KV-head x layer for matched sparsity.", "sparse_decisions": 54488120, "total_decisions": 425294912, "sparsity": 0.12811843843549203, "full_attention_usage": 0.8718815615645079, "per_layer": { "0": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "1": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "2": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "3": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "4": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "5": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "6": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "7": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "8": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "9": { "sparse_decisions": 368, "total_decisions": 15189104, "sparsity": 2.4227893890251854e-05 }, "10": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "11": { "sparse_decisions": 1088, "total_decisions": 15189104, "sparsity": 7.163029497987504e-05 }, "12": { "sparse_decisions": 2664, "total_decisions": 15189104, "sparsity": 0.00017538888403160582 }, "13": { "sparse_decisions": 40904, "total_decisions": 15189104, "sparsity": 0.002692983075236038 }, "14": { "sparse_decisions": 68952, "total_decisions": 15189104, "sparsity": 0.004539569944349581 }, "15": { "sparse_decisions": 2841368, "total_decisions": 15189104, "sparsity": 0.18706620219336176 }, "16": { "sparse_decisions": 28696, "total_decisions": 15189104, "sparsity": 0.0018892490300942044 }, "17": { "sparse_decisions": 1226768, "total_decisions": 15189104, "sparsity": 0.08076631774988176 }, "18": { "sparse_decisions": 403184, "total_decisions": 15189104, "sparsity": 0.026544291223498107 }, "19": { "sparse_decisions": 374392, "total_decisions": 15189104, "sparsity": 0.02464872187325862 }, "20": { "sparse_decisions": 1019824, "total_decisions": 15189104, "sparsity": 0.06714181429003317 }, "21": { "sparse_decisions": 1591304, "total_decisions": 15189104, "sparsity": 0.10476615342155798 }, "22": { "sparse_decisions": 8488256, "total_decisions": 15189104, "sparsity": 0.5588384936991675 }, "23": { "sparse_decisions": 10478168, "total_decisions": 15189104, "sparsity": 0.6898476697506317 }, "24": { "sparse_decisions": 6924024, "total_decisions": 15189104, "sparsity": 0.45585467055857937 }, "25": { "sparse_decisions": 4631320, "total_decisions": 15189104, "sparsity": 0.3049106780755468 }, "26": { "sparse_decisions": 1179736, "total_decisions": 15189104, "sparsity": 0.07766988757203848 }, "27": { "sparse_decisions": 15187104, "total_decisions": 15189104, "sparsity": 0.9998683266636399 } }, "by_phase": { "decode": { "sparse_decisions": 55088, "total_decisions": 840000, "sparsity": 0.06558095238095238, "full_attention_usage": 0.9344190476190476 }, "prefill": { "sparse_decisions": 54433032, "total_decisions": 424454912, "sparsity": 0.12824220066983227, "full_attention_usage": 0.8717577993301677 } } }