{ "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": 5391064, "effective_router_decisions": 43128512, "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": 5716296, "total_decisions": 43128512, "sparsity": 0.13254099747285508, "full_attention_usage": 0.8674590025271449, "per_layer": { "0": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "1": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "2": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "3": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "4": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "5": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "6": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "7": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "8": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "9": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "10": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "11": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "12": { "sparse_decisions": 0, "total_decisions": 1540304, "sparsity": 0.0 }, "13": { "sparse_decisions": 32, "total_decisions": 1540304, "sparsity": 2.0775119716627365e-05 }, "14": { "sparse_decisions": 1640, "total_decisions": 1540304, "sparsity": 0.0010647248854771525 }, "15": { "sparse_decisions": 402208, "total_decisions": 1540304, "sparsity": 0.2611224797182894 }, "16": { "sparse_decisions": 4664, "total_decisions": 1540304, "sparsity": 0.0030279736986984387 }, "17": { "sparse_decisions": 287640, "total_decisions": 1540304, "sparsity": 0.18674235735283426 }, "18": { "sparse_decisions": 87720, "total_decisions": 1540304, "sparsity": 0.05694979692320477 }, "19": { "sparse_decisions": 321632, "total_decisions": 1540304, "sparsity": 0.20881072827182168 }, "20": { "sparse_decisions": 374168, "total_decisions": 1540304, "sparsity": 0.24291828106659463 }, "21": { "sparse_decisions": 333200, "total_decisions": 1540304, "sparsity": 0.21632093404938246 }, "22": { "sparse_decisions": 508832, "total_decisions": 1540304, "sparsity": 0.33034517861409174 }, "23": { "sparse_decisions": 857744, "total_decisions": 1540304, "sparsity": 0.5568666964443383 }, "24": { "sparse_decisions": 574504, "total_decisions": 1540304, "sparsity": 0.3729809180525403 }, "25": { "sparse_decisions": 375984, "total_decisions": 1540304, "sparsity": 0.24409726911051324 }, "26": { "sparse_decisions": 46264, "total_decisions": 1540304, "sparsity": 0.030035629330314017 }, "27": { "sparse_decisions": 1540064, "total_decisions": 1540304, "sparsity": 0.9998441866021253 } }, "by_phase": { "decode": { "sparse_decisions": 156624, "total_decisions": 2358720, "sparsity": 0.0664021164021164, "full_attention_usage": 0.9335978835978835 }, "prefill": { "sparse_decisions": 5559672, "total_decisions": 40769792, "sparsity": 0.1363674359682777, "full_attention_usage": 0.8636325640317223 } } }