{ "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": 29979236, "effective_router_decisions": 239833888, "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": 64617208, "total_decisions": 239833888, "sparsity": 0.2694248445824303, "full_attention_usage": 0.7305751554175697, "per_layer": { "0": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "1": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "2": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "3": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "4": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "5": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "6": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "7": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "8": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "9": { "sparse_decisions": 155096, "total_decisions": 8565496, "sparsity": 0.018107065837168097 }, "10": { "sparse_decisions": 0, "total_decisions": 8565496, "sparsity": 0.0 }, "11": { "sparse_decisions": 33416, "total_decisions": 8565496, "sparsity": 0.0039012335070847037 }, "12": { "sparse_decisions": 14248, "total_decisions": 8565496, "sparsity": 0.0016634179736935257 }, "13": { "sparse_decisions": 723056, "total_decisions": 8565496, "sparsity": 0.08441495974080193 }, "14": { "sparse_decisions": 623240, "total_decisions": 8565496, "sparsity": 0.07276169412722859 }, "15": { "sparse_decisions": 4935736, "total_decisions": 8565496, "sparsity": 0.5762346979089127 }, "16": { "sparse_decisions": 652168, "total_decisions": 8565496, "sparsity": 0.07613896498229641 }, "17": { "sparse_decisions": 5110480, "total_decisions": 8565496, "sparsity": 0.5966356180657839 }, "18": { "sparse_decisions": 3184560, "total_decisions": 8565496, "sparsity": 0.3717893277867388 }, "19": { "sparse_decisions": 3518472, "total_decisions": 8565496, "sparsity": 0.410772709484658 }, "20": { "sparse_decisions": 5684152, "total_decisions": 8565496, "sparsity": 0.6636103735265302 }, "21": { "sparse_decisions": 4218912, "total_decisions": 8565496, "sparsity": 0.49254730841039446 }, "22": { "sparse_decisions": 6686616, "total_decisions": 8565496, "sparsity": 0.7806455107795275 }, "23": { "sparse_decisions": 7276744, "total_decisions": 8565496, "sparsity": 0.8495414626310024 }, "24": { "sparse_decisions": 6210728, "total_decisions": 8565496, "sparsity": 0.7250867900703006 }, "25": { "sparse_decisions": 5428184, "total_decisions": 8565496, "sparsity": 0.6337267567459024 }, "26": { "sparse_decisions": 1597024, "total_decisions": 8565496, "sparsity": 0.186448513898086 }, "27": { "sparse_decisions": 8564376, "total_decisions": 8565496, "sparsity": 0.9998692428319388 } }, "by_phase": { "decode": { "sparse_decisions": 1078112, "total_decisions": 2764160, "sparsity": 0.390032414910859, "full_attention_usage": 0.6099675850891411 }, "prefill": { "sparse_decisions": 63539096, "total_decisions": 237069728, "sparsity": 0.26801859746513057, "full_attention_usage": 0.7319814025348694 } } }