[ { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 84.84, "hit_rate": 0.8232, "ssd_mb_per_tok": 117.29, "m5_decode_1th_tok_s": 23.77, "m5_decode_4th_tok_s": 34.19 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lfu", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 134.67, "hit_rate": 0.7194, "ssd_mb_per_tok": 186.17, "m5_decode_1th_tok_s": 16.74, "m5_decode_4th_tok_s": 25.39 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 96.54, "hit_rate": 0.7989, "ssd_mb_per_tok": 133.45, "m5_decode_1th_tok_s": 21.64, "m5_decode_4th_tok_s": 31.62 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 100.78, "hit_rate": 0.79, "ssd_mb_per_tok": 139.32, "m5_decode_1th_tok_s": 20.95, "m5_decode_4th_tok_s": 30.78 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 97.8, "hit_rate": 0.7963, "ssd_mb_per_tok": 135.2, "m5_decode_1th_tok_s": 21.43, "m5_decode_4th_tok_s": 31.37 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "belady_min", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 46.13, "hit_rate": 0.9039, "ssd_mb_per_tok": 63.77, "m5_decode_1th_tok_s": 35.27, "m5_decode_4th_tok_s": 46.78 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 76.09, "hit_rate": 0.8415, "ssd_mb_per_tok": 105.18, "m5_decode_1th_tok_s": 25.66, "m5_decode_4th_tok_s": 36.41 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 76.09, "hit_rate": 0.8415, "ssd_mb_per_tok": 105.18, "m5_decode_1th_tok_s": 25.66, "m5_decode_4th_tok_s": 36.41 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 70.07, "hit_rate": 0.854, "ssd_mb_per_tok": 96.86, "m5_decode_1th_tok_s": 27.15, "m5_decode_4th_tok_s": 38.1 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 70.07, "hit_rate": 0.854, "ssd_mb_per_tok": 96.86, "m5_decode_1th_tok_s": 27.15, "m5_decode_4th_tok_s": 38.1 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 66.05, "hit_rate": 0.8624, "ssd_mb_per_tok": 91.31, "m5_decode_1th_tok_s": 28.24, "m5_decode_4th_tok_s": 39.33 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 66.06, "hit_rate": 0.8624, "ssd_mb_per_tok": 91.32, "m5_decode_1th_tok_s": 28.24, "m5_decode_4th_tok_s": 39.33 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 90.97, "hit_rate": 0.8105, "ssd_mb_per_tok": 125.76, "m5_decode_1th_tok_s": 22.6, "m5_decode_4th_tok_s": 32.79 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 90.97, "hit_rate": 0.8105, "ssd_mb_per_tok": 125.76, "m5_decode_1th_tok_s": 22.6, "m5_decode_4th_tok_s": 32.79 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 84.03, "hit_rate": 0.8249, "ssd_mb_per_tok": 116.16, "m5_decode_1th_tok_s": 23.93, "m5_decode_4th_tok_s": 34.38 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 84.03, "hit_rate": 0.8249, "ssd_mb_per_tok": 116.16, "m5_decode_1th_tok_s": 23.93, "m5_decode_4th_tok_s": 34.38 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 79.32, "hit_rate": 0.8347, "ssd_mb_per_tok": 109.66, "m5_decode_1th_tok_s": 24.93, "m5_decode_4th_tok_s": 35.55 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 79.33, "hit_rate": 0.8347, "ssd_mb_per_tok": 109.66, "m5_decode_1th_tok_s": 24.93, "m5_decode_4th_tok_s": 35.55 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 88.21, "hit_rate": 0.8162, "ssd_mb_per_tok": 121.95, "m5_decode_1th_tok_s": 23.11, "m5_decode_4th_tok_s": 33.41 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 88.21, "hit_rate": 0.8162, "ssd_mb_per_tok": 121.95, "m5_decode_1th_tok_s": 23.11, "m5_decode_4th_tok_s": 33.41 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 81.49, "hit_rate": 0.8302, "ssd_mb_per_tok": 112.66, "m5_decode_1th_tok_s": 24.46, "m5_decode_4th_tok_s": 35.0 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 81.5, "hit_rate": 0.8302, "ssd_mb_per_tok": 112.66, "m5_decode_1th_tok_s": 24.46, "m5_decode_4th_tok_s": 35.0 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 76.92, "hit_rate": 0.8398, "ssd_mb_per_tok": 106.33, "m5_decode_1th_tok_s": 25.47, "m5_decode_4th_tok_s": 36.18 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 76.92, "hit_rate": 0.8398, "ssd_mb_per_tok": 106.33, "m5_decode_1th_tok_s": 25.47, "m5_decode_4th_tok_s": 36.18 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 68.12, "hit_rate": 0.8581, "ssd_mb_per_tok": 94.17, "m5_decode_1th_tok_s": 27.67, "m5_decode_4th_tok_s": 38.69 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lfu", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 110.0, "hit_rate": 0.7708, "ssd_mb_per_tok": 152.06, "m5_decode_1th_tok_s": 19.61, "m5_decode_4th_tok_s": 29.1 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 78.49, "hit_rate": 0.8365, "ssd_mb_per_tok": 108.51, "m5_decode_1th_tok_s": 25.11, "m5_decode_4th_tok_s": 35.77 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 80.32, "hit_rate": 0.8327, "ssd_mb_per_tok": 111.04, "m5_decode_1th_tok_s": 24.71, "m5_decode_4th_tok_s": 35.3 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 78.07, "hit_rate": 0.8374, "ssd_mb_per_tok": 107.93, "m5_decode_1th_tok_s": 25.21, "m5_decode_4th_tok_s": 35.88 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "belady_min", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 34.39, "hit_rate": 0.9284, "ssd_mb_per_tok": 47.54, "m5_decode_1th_tok_s": 41.34, "m5_decode_4th_tok_s": 52.65 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 61.08, "hit_rate": 0.8727, "ssd_mb_per_tok": 84.44, "m5_decode_1th_tok_s": 29.72, "m5_decode_4th_tok_s": 40.95 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 61.08, "hit_rate": 0.8727, "ssd_mb_per_tok": 84.44, "m5_decode_1th_tok_s": 29.72, "m5_decode_4th_tok_s": 40.95 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 56.29, "hit_rate": 0.8827, "ssd_mb_per_tok": 77.82, "m5_decode_1th_tok_s": 31.3, "m5_decode_4th_tok_s": 42.66 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 56.29, "hit_rate": 0.8827, "ssd_mb_per_tok": 77.82, "m5_decode_1th_tok_s": 31.3, "m5_decode_4th_tok_s": 42.66 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 53.11, "hit_rate": 0.8894, "ssd_mb_per_tok": 73.42, "m5_decode_1th_tok_s": 32.44, "m5_decode_4th_tok_s": 43.87 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 53.11, "hit_rate": 0.8894, "ssd_mb_per_tok": 73.42, "m5_decode_1th_tok_s": 32.44, "m5_decode_4th_tok_s": 43.87 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 72.64, "hit_rate": 0.8487, "ssd_mb_per_tok": 100.42, "m5_decode_1th_tok_s": 26.49, "m5_decode_4th_tok_s": 37.36 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 72.64, "hit_rate": 0.8487, "ssd_mb_per_tok": 100.42, "m5_decode_1th_tok_s": 26.49, "m5_decode_4th_tok_s": 37.36 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 67.2, "hit_rate": 0.86, "ssd_mb_per_tok": 92.9, "m5_decode_1th_tok_s": 27.92, "m5_decode_4th_tok_s": 38.97 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 67.2, "hit_rate": 0.86, "ssd_mb_per_tok": 92.9, "m5_decode_1th_tok_s": 27.92, "m5_decode_4th_tok_s": 38.97 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 63.52, "hit_rate": 0.8677, "ssd_mb_per_tok": 87.81, "m5_decode_1th_tok_s": 28.97, "m5_decode_4th_tok_s": 40.14 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 63.52, "hit_rate": 0.8677, "ssd_mb_per_tok": 87.82, "m5_decode_1th_tok_s": 28.97, "m5_decode_4th_tok_s": 40.14 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 70.32, "hit_rate": 0.8535, "ssd_mb_per_tok": 97.2, "m5_decode_1th_tok_s": 27.08, "m5_decode_4th_tok_s": 38.03 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 70.32, "hit_rate": 0.8535, "ssd_mb_per_tok": 97.2, "m5_decode_1th_tok_s": 27.08, "m5_decode_4th_tok_s": 38.03 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 64.87, "hit_rate": 0.8649, "ssd_mb_per_tok": 89.68, "m5_decode_1th_tok_s": 28.58, "m5_decode_4th_tok_s": 39.7 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 64.87, "hit_rate": 0.8648, "ssd_mb_per_tok": 89.68, "m5_decode_1th_tok_s": 28.58, "m5_decode_4th_tok_s": 39.7 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 61.19, "hit_rate": 0.8725, "ssd_mb_per_tok": 84.58, "m5_decode_1th_tok_s": 29.68, "m5_decode_4th_tok_s": 40.92 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 61.19, "hit_rate": 0.8725, "ssd_mb_per_tok": 84.59, "m5_decode_1th_tok_s": 29.68, "m5_decode_4th_tok_s": 40.92 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 54.29, "hit_rate": 0.8869, "ssd_mb_per_tok": 75.05, "m5_decode_1th_tok_s": 32.01, "m5_decode_4th_tok_s": 43.41 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lfu", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 88.0, "hit_rate": 0.8167, "ssd_mb_per_tok": 121.65, "m5_decode_1th_tok_s": 23.15, "m5_decode_4th_tok_s": 33.46 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 62.69, "hit_rate": 0.8694, "ssd_mb_per_tok": 86.66, "m5_decode_1th_tok_s": 29.22, "m5_decode_4th_tok_s": 40.41 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 63.43, "hit_rate": 0.8679, "ssd_mb_per_tok": 87.68, "m5_decode_1th_tok_s": 29.0, "m5_decode_4th_tok_s": 40.17 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 62.08, "hit_rate": 0.8707, "ssd_mb_per_tok": 85.81, "m5_decode_1th_tok_s": 29.41, "m5_decode_4th_tok_s": 40.62 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "belady_min", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 26.11, "hit_rate": 0.9456, "ssd_mb_per_tok": 36.09, "m5_decode_1th_tok_s": 47.05, "m5_decode_4th_tok_s": 57.78 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 48.52, "hit_rate": 0.8989, "ssd_mb_per_tok": 67.07, "m5_decode_1th_tok_s": 34.25, "m5_decode_4th_tok_s": 45.74 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 48.52, "hit_rate": 0.8989, "ssd_mb_per_tok": 67.07, "m5_decode_1th_tok_s": 34.25, "m5_decode_4th_tok_s": 45.74 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 44.65, "hit_rate": 0.907, "ssd_mb_per_tok": 61.73, "m5_decode_1th_tok_s": 35.94, "m5_decode_4th_tok_s": 47.44 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 44.65, "hit_rate": 0.907, "ssd_mb_per_tok": 61.73, "m5_decode_1th_tok_s": 35.94, "m5_decode_4th_tok_s": 47.44 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 42.11, "hit_rate": 0.9123, "ssd_mb_per_tok": 58.21, "m5_decode_1th_tok_s": 37.14, "m5_decode_4th_tok_s": 48.64 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 42.11, "hit_rate": 0.9123, "ssd_mb_per_tok": 58.21, "m5_decode_1th_tok_s": 37.14, "m5_decode_4th_tok_s": 48.64 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 57.02, "hit_rate": 0.8812, "ssd_mb_per_tok": 78.82, "m5_decode_1th_tok_s": 31.05, "m5_decode_4th_tok_s": 42.39 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 57.02, "hit_rate": 0.8812, "ssd_mb_per_tok": 78.82, "m5_decode_1th_tok_s": 31.05, "m5_decode_4th_tok_s": 42.39 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 52.55, "hit_rate": 0.8905, "ssd_mb_per_tok": 72.65, "m5_decode_1th_tok_s": 32.65, "m5_decode_4th_tok_s": 44.09 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 52.55, "hit_rate": 0.8905, "ssd_mb_per_tok": 72.65, "m5_decode_1th_tok_s": 32.65, "m5_decode_4th_tok_s": 44.08 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 49.57, "hit_rate": 0.8967, "ssd_mb_per_tok": 68.52, "m5_decode_1th_tok_s": 33.82, "m5_decode_4th_tok_s": 45.3 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 49.57, "hit_rate": 0.8967, "ssd_mb_per_tok": 68.52, "m5_decode_1th_tok_s": 33.82, "m5_decode_4th_tok_s": 45.3 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 55.48, "hit_rate": 0.8844, "ssd_mb_per_tok": 76.69, "m5_decode_1th_tok_s": 31.58, "m5_decode_4th_tok_s": 42.96 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 55.48, "hit_rate": 0.8844, "ssd_mb_per_tok": 76.69, "m5_decode_1th_tok_s": 31.58, "m5_decode_4th_tok_s": 42.96 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 50.94, "hit_rate": 0.8939, "ssd_mb_per_tok": 70.42, "m5_decode_1th_tok_s": 33.27, "m5_decode_4th_tok_s": 44.73 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 50.94, "hit_rate": 0.8939, "ssd_mb_per_tok": 70.42, "m5_decode_1th_tok_s": 33.27, "m5_decode_4th_tok_s": 44.73 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 47.93, "hit_rate": 0.9001, "ssd_mb_per_tok": 66.26, "m5_decode_1th_tok_s": 34.5, "m5_decode_4th_tok_s": 45.99 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 47.93, "hit_rate": 0.9001, "ssd_mb_per_tok": 66.26, "m5_decode_1th_tok_s": 34.5, "m5_decode_4th_tok_s": 45.99 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 45.35, "hit_rate": 0.9055, "ssd_mb_per_tok": 62.7, "m5_decode_1th_tok_s": 35.62, "m5_decode_4th_tok_s": 47.12 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lfu", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 72.96, "hit_rate": 0.848, "ssd_mb_per_tok": 100.86, "m5_decode_1th_tok_s": 26.41, "m5_decode_4th_tok_s": 37.27 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 51.8, "hit_rate": 0.8921, "ssd_mb_per_tok": 71.61, "m5_decode_1th_tok_s": 32.94, "m5_decode_4th_tok_s": 44.38 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 52.16, "hit_rate": 0.8913, "ssd_mb_per_tok": 72.1, "m5_decode_1th_tok_s": 32.8, "m5_decode_4th_tok_s": 44.24 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 51.24, "hit_rate": 0.8933, "ssd_mb_per_tok": 70.83, "m5_decode_1th_tok_s": 33.15, "m5_decode_4th_tok_s": 44.61 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "belady_min", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 21.31, "hit_rate": 0.9556, "ssd_mb_per_tok": 29.46, "m5_decode_1th_tok_s": 51.14, "m5_decode_4th_tok_s": 61.23 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 40.58, "hit_rate": 0.9154, "ssd_mb_per_tok": 56.1, "m5_decode_1th_tok_s": 37.9, "m5_decode_4th_tok_s": 49.38 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 40.58, "hit_rate": 0.9154, "ssd_mb_per_tok": 56.1, "m5_decode_1th_tok_s": 37.9, "m5_decode_4th_tok_s": 49.38 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 37.33, "hit_rate": 0.9222, "ssd_mb_per_tok": 51.61, "m5_decode_1th_tok_s": 39.63, "m5_decode_4th_tok_s": 51.05 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 37.34, "hit_rate": 0.9222, "ssd_mb_per_tok": 51.61, "m5_decode_1th_tok_s": 39.63, "m5_decode_4th_tok_s": 51.05 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 35.27, "hit_rate": 0.9265, "ssd_mb_per_tok": 48.76, "m5_decode_1th_tok_s": 40.81, "m5_decode_4th_tok_s": 52.16 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 35.27, "hit_rate": 0.9265, "ssd_mb_per_tok": 48.76, "m5_decode_1th_tok_s": 40.81, "m5_decode_4th_tok_s": 52.16 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 46.83, "hit_rate": 0.9024, "ssd_mb_per_tok": 64.74, "m5_decode_1th_tok_s": 34.96, "m5_decode_4th_tok_s": 46.47 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 46.83, "hit_rate": 0.9024, "ssd_mb_per_tok": 64.74, "m5_decode_1th_tok_s": 34.96, "m5_decode_4th_tok_s": 46.47 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 43.16, "hit_rate": 0.9101, "ssd_mb_per_tok": 59.66, "m5_decode_1th_tok_s": 36.63, "m5_decode_4th_tok_s": 48.14 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 43.16, "hit_rate": 0.9101, "ssd_mb_per_tok": 59.66, "m5_decode_1th_tok_s": 36.63, "m5_decode_4th_tok_s": 48.14 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 40.69, "hit_rate": 0.9152, "ssd_mb_per_tok": 56.25, "m5_decode_1th_tok_s": 37.84, "m5_decode_4th_tok_s": 49.33 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 40.69, "hit_rate": 0.9152, "ssd_mb_per_tok": 56.26, "m5_decode_1th_tok_s": 37.84, "m5_decode_4th_tok_s": 49.33 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 45.67, "hit_rate": 0.9049, "ssd_mb_per_tok": 63.13, "m5_decode_1th_tok_s": 35.48, "m5_decode_4th_tok_s": 46.98 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 45.67, "hit_rate": 0.9049, "ssd_mb_per_tok": 63.13, "m5_decode_1th_tok_s": 35.48, "m5_decode_4th_tok_s": 46.98 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 41.88, "hit_rate": 0.9127, "ssd_mb_per_tok": 57.89, "m5_decode_1th_tok_s": 37.25, "m5_decode_4th_tok_s": 48.75 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 41.88, "hit_rate": 0.9127, "ssd_mb_per_tok": 57.9, "m5_decode_1th_tok_s": 37.25, "m5_decode_4th_tok_s": 48.75 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 39.4, "hit_rate": 0.9179, "ssd_mb_per_tok": 54.46, "m5_decode_1th_tok_s": 38.51, "m5_decode_4th_tok_s": 49.98 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 39.4, "hit_rate": 0.9179, "ssd_mb_per_tok": 54.46, "m5_decode_1th_tok_s": 38.51, "m5_decode_4th_tok_s": 49.98 }, { "variant": "K320_reap_hybrid", "n_experts": 320, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 65.32, "hit_rate": 0.8639, "ssd_mb_per_tok": 90.3, "m5_decode_1th_tok_s": 28.45, "m5_decode_4th_tok_s": 39.56 }, { "variant": "K320_reap_hybrid", "n_experts": 320, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 53.95, "hit_rate": 0.8876, "ssd_mb_per_tok": 74.59, "m5_decode_1th_tok_s": 32.13, "m5_decode_4th_tok_s": 43.54 }, { "variant": "K320_reap_hybrid", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 76.27, "hit_rate": 0.8411, "ssd_mb_per_tok": 105.44, "m5_decode_1th_tok_s": 25.62, "m5_decode_4th_tok_s": 36.36 }, { "variant": "K320_reap_hybrid", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 63.74, "hit_rate": 0.8672, "ssd_mb_per_tok": 88.12, "m5_decode_1th_tok_s": 28.91, "m5_decode_4th_tok_s": 40.07 }, { "variant": "K256_reap_hybrid", "n_experts": 256, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 51.73, "hit_rate": 0.8922, "ssd_mb_per_tok": 71.52, "m5_decode_1th_tok_s": 32.96, "m5_decode_4th_tok_s": 44.41 }, { "variant": "K256_reap_hybrid", "n_experts": 256, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 42.34, "hit_rate": 0.9118, "ssd_mb_per_tok": 58.54, "m5_decode_1th_tok_s": 37.02, "m5_decode_4th_tok_s": 48.52 }, { "variant": "K256_reap_hybrid", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 59.02, "hit_rate": 0.877, "ssd_mb_per_tok": 81.59, "m5_decode_1th_tok_s": 30.38, "m5_decode_4th_tok_s": 41.67 }, { "variant": "K256_reap_hybrid", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 48.65, "hit_rate": 0.8986, "ssd_mb_per_tok": 67.26, "m5_decode_1th_tok_s": 34.19, "m5_decode_4th_tok_s": 45.68 }, { "variant": "K224_reap_hybrid", "n_experts": 224, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 43.66, "hit_rate": 0.909, "ssd_mb_per_tok": 60.36, "m5_decode_1th_tok_s": 36.4, "m5_decode_4th_tok_s": 47.9 }, { "variant": "K224_reap_hybrid", "n_experts": 224, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 35.72, "hit_rate": 0.9256, "ssd_mb_per_tok": 49.38, "m5_decode_1th_tok_s": 40.55, "m5_decode_4th_tok_s": 51.92 }, { "variant": "K224_reap_hybrid", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 48.3, "hit_rate": 0.8994, "ssd_mb_per_tok": 66.77, "m5_decode_1th_tok_s": 34.34, "m5_decode_4th_tok_s": 45.83 }, { "variant": "K224_reap_hybrid", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 39.58, "hit_rate": 0.9175, "ssd_mb_per_tok": 54.72, "m5_decode_1th_tok_s": 38.42, "m5_decode_4th_tok_s": 49.88 }, { "variant": "K320_reap_shrink", "n_experts": 320, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 63.8, "hit_rate": 0.8671, "ssd_mb_per_tok": 88.2, "m5_decode_1th_tok_s": 28.89, "m5_decode_4th_tok_s": 40.05 }, { "variant": "K320_reap_shrink", "n_experts": 320, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 52.26, "hit_rate": 0.8911, "ssd_mb_per_tok": 72.24, "m5_decode_1th_tok_s": 32.76, "m5_decode_4th_tok_s": 44.2 }, { "variant": "K320_reap_shrink", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 73.54, "hit_rate": 0.8468, "ssd_mb_per_tok": 101.66, "m5_decode_1th_tok_s": 26.27, "m5_decode_4th_tok_s": 37.11 }, { "variant": "K320_reap_shrink", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 60.88, "hit_rate": 0.8732, "ssd_mb_per_tok": 84.16, "m5_decode_1th_tok_s": 29.78, "m5_decode_4th_tok_s": 41.02 }, { "variant": "K256_reap_shrink", "n_experts": 256, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 50.51, "hit_rate": 0.8948, "ssd_mb_per_tok": 69.82, "m5_decode_1th_tok_s": 33.44, "m5_decode_4th_tok_s": 44.91 }, { "variant": "K256_reap_shrink", "n_experts": 256, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 40.76, "hit_rate": 0.9151, "ssd_mb_per_tok": 56.35, "m5_decode_1th_tok_s": 37.81, "m5_decode_4th_tok_s": 49.29 }, { "variant": "K256_reap_shrink", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 56.18, "hit_rate": 0.883, "ssd_mb_per_tok": 77.66, "m5_decode_1th_tok_s": 31.34, "m5_decode_4th_tok_s": 42.7 }, { "variant": "K256_reap_shrink", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 45.52, "hit_rate": 0.9052, "ssd_mb_per_tok": 62.93, "m5_decode_1th_tok_s": 35.54, "m5_decode_4th_tok_s": 47.05 }, { "variant": "K224_reap_shrink", "n_experts": 224, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 42.7, "hit_rate": 0.911, "ssd_mb_per_tok": 59.03, "m5_decode_1th_tok_s": 36.85, "m5_decode_4th_tok_s": 48.35 }, { "variant": "K224_reap_shrink", "n_experts": 224, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 34.22, "hit_rate": 0.9287, "ssd_mb_per_tok": 47.31, "m5_decode_1th_tok_s": 41.44, "m5_decode_4th_tok_s": 52.75 }, { "variant": "K224_reap_shrink", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 45.83, "hit_rate": 0.9045, "ssd_mb_per_tok": 63.36, "m5_decode_1th_tok_s": 35.4, "m5_decode_4th_tok_s": 46.91 }, { "variant": "K224_reap_shrink", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 36.8, "hit_rate": 0.9233, "ssd_mb_per_tok": 50.88, "m5_decode_1th_tok_s": 39.93, "m5_decode_4th_tok_s": 51.33 } ]