{ "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": 97002416, "total_decisions": 425294912, "sparsity": 0.22808270981619455, "full_attention_usage": 0.7719172901838054, "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": 7000, "total_decisions": 15189104, "sparsity": 0.0004608566772602255 }, "10": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "11": { "sparse_decisions": 10232, "total_decisions": 15189104, "sparsity": 0.0006736407888180896 }, "12": { "sparse_decisions": 31080, "total_decisions": 15189104, "sparsity": 0.0020462036470354013 }, "13": { "sparse_decisions": 234160, "total_decisions": 15189104, "sparsity": 0.015416314221036342 }, "14": { "sparse_decisions": 589968, "total_decisions": 15189104, "sparsity": 0.03884152745283725 }, "15": { "sparse_decisions": 6455432, "total_decisions": 15189104, "sparsity": 0.4250041345427617 }, "16": { "sparse_decisions": 364448, "total_decisions": 15189104, "sparsity": 0.02399404204487638 }, "17": { "sparse_decisions": 5393848, "total_decisions": 15189104, "sparsity": 0.35511298098953037 }, "18": { "sparse_decisions": 3350136, "total_decisions": 15189104, "sparsity": 0.2205617921899804 }, "19": { "sparse_decisions": 2640848, "total_decisions": 15189104, "sparsity": 0.17386463348990172 }, "20": { "sparse_decisions": 4845696, "total_decisions": 15189104, "sparsity": 0.3190244796533094 }, "21": { "sparse_decisions": 5868528, "total_decisions": 15189104, "sparsity": 0.3863643306412281 }, "22": { "sparse_decisions": 13004184, "total_decisions": 15189104, "sparsity": 0.856152146960084 }, "23": { "sparse_decisions": 13248752, "total_decisions": 15189104, "sparsity": 0.8722536892235382 }, "24": { "sparse_decisions": 11923056, "total_decisions": 15189104, "sparsity": 0.7849742815639421 }, "25": { "sparse_decisions": 10083496, "total_decisions": 15189104, "sparsity": 0.6638637802466821 }, "26": { "sparse_decisions": 3764448, "total_decisions": 15189104, "sparsity": 0.24783871385698591 }, "27": { "sparse_decisions": 15187104, "total_decisions": 15189104, "sparsity": 0.9998683266636399 } }, "by_phase": { "decode": { "sparse_decisions": 121680, "total_decisions": 840000, "sparsity": 0.14485714285714285, "full_attention_usage": 0.8551428571428572 }, "prefill": { "sparse_decisions": 96880736, "total_decisions": 424454912, "sparsity": 0.228247413944405, "full_attention_usage": 0.7717525860555949 } } }