{ "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": 24280016, "total_decisions": 425294912, "sparsity": 0.057089834171352605, "full_attention_usage": 0.9429101658286474, "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": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "10": { "sparse_decisions": 0, "total_decisions": 15189104, "sparsity": 0.0 }, "11": { "sparse_decisions": 32, "total_decisions": 15189104, "sparsity": 2.106773381761031e-06 }, "12": { "sparse_decisions": 32, "total_decisions": 15189104, "sparsity": 2.106773381761031e-06 }, "13": { "sparse_decisions": 2704, "total_decisions": 15189104, "sparsity": 0.0001780223507588071 }, "14": { "sparse_decisions": 1744, "total_decisions": 15189104, "sparsity": 0.00011481914930597618 }, "15": { "sparse_decisions": 515936, "total_decisions": 15189104, "sparsity": 0.0339675072341331 }, "16": { "sparse_decisions": 504, "total_decisions": 15189104, "sparsity": 3.3181680762736235e-05 }, "17": { "sparse_decisions": 41672, "total_decisions": 15189104, "sparsity": 0.0027435456363983026 }, "18": { "sparse_decisions": 3176, "total_decisions": 15189104, "sparsity": 0.0002090972581397823 }, "19": { "sparse_decisions": 5448, "total_decisions": 15189104, "sparsity": 0.0003586781682448155 }, "20": { "sparse_decisions": 22504, "total_decisions": 15189104, "sparsity": 0.001481588380723445 }, "21": { "sparse_decisions": 66592, "total_decisions": 15189104, "sparsity": 0.004384195407444705 }, "22": { "sparse_decisions": 1687440, "total_decisions": 15189104, "sparsity": 0.11109542735371356 }, "23": { "sparse_decisions": 4707840, "total_decisions": 15189104, "sparsity": 0.3099484999246829 }, "24": { "sparse_decisions": 1144760, "total_decisions": 15189104, "sparsity": 0.07536718426577367 }, "25": { "sparse_decisions": 708656, "total_decisions": 15189104, "sparsity": 0.04665554992578891 }, "26": { "sparse_decisions": 183888, "total_decisions": 15189104, "sparsity": 0.012106573238289764 }, "27": { "sparse_decisions": 15187088, "total_decisions": 15189104, "sparsity": 0.999867273276949 } }, "by_phase": { "decode": { "sparse_decisions": 35696, "total_decisions": 840000, "sparsity": 0.04249523809523809, "full_attention_usage": 0.9575047619047619 }, "prefill": { "sparse_decisions": 24244320, "total_decisions": 424454912, "sparsity": 0.057118717005211615, "full_attention_usage": 0.9428812829947884 } } }