Qwen3-30B parity on M3U96a (bit-identical, 121.7 vs 121.5 tok/s), make_tables.py, cold-NVMe sweep script, LOG

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Simon-Pierre Boucher committed 2 days ago (Nov 1, 2026) parent 258ba81 Browse files

23 changed files +2,727 −9

added benchmarks/processed/gateC/sim_olmoe_v2_ssd043_m8.json +2309 −0

++{
++ "cmd": "benchmarks/scripts/simulate_cache.py --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --capacities 16,24,32,48 --policies lru --m 8 --t-read-ms 0.43 --out benchmarks/processed/gateC/sim_olmoe_v2_ssd043_m8.json",
++ "ts": "2026-11-01T01:47:09",
++ "traces": "benchmarks/raw/gateC/traces_olmoe_v2_laptop",
++ "L": 16,
++ "E": 64,
++ "K": 8,
++ "t_read_ms": 0.43,
++ "t_layer_ms": 0.128,
++ "loaders": 1,
++ "results": [
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "none",
++   "m": 8,
++   "hit_rate": 0.5768595950704225,
++   "misses_per_token": 54.16197183098591,
++   "stall_ms_per_token": 23.289647887323987,
++   "prefetch_loads_per_token": 0.0,
++   "prefetch_precision": null,
++   "loads_per_token": 54.16197183098591,
++   "pred_recall_at_k": null,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.5229952830188679,
++     "stall_ms_per_token": 26.254339622641496
++    },
++    "code": {
++     "hit_rate": 0.6450788361508453,
++     "stall_ms_per_token": 19.534860858257467
++    },
++    "reasoning": {
++     "hit_rate": 0.534992465932914,
++     "stall_ms_per_token": 25.594014675052406
++    },
++    "multilingual": {
++     "hit_rate": 0.5699602711210398,
++     "stall_ms_per_token": 23.669386677497993
++    },
++    "longctx": {
++     "hit_rate": 0.5257540899795501,
++     "stall_ms_per_token": 26.10249488752555
++    },
++    "drift": {
++     "hit_rate": 0.5936745776347546,
++     "stall_ms_per_token": 22.364151246983084
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "none",
++   "m": 8,
++   "hit_rate": 0.5807302010923535,
++   "misses_per_token": 53.66653426017875,
++   "stall_ms_per_token": 23.076609731876843,
++   "prefetch_loads_per_token": 0.0,
++   "prefetch_precision": null,
++   "loads_per_token": 53.66653426017875,
++   "pred_recall_at_k": null,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.4989222212294496,
++     "stall_ms_per_token": 27.579320943531116
++    },
++    "code": {
++     "hit_rate": 0.6525027395115842,
++     "stall_ms_per_token": 19.126249217282393
++    },
++    "reasoning": {
++     "hit_rate": 0.5798861994219653,
++     "stall_ms_per_token": 23.123063583815036
++    },
++    "multilingual": {
++     "hit_rate": 0.5800225332068312,
++     "stall_ms_per_token": 23.115559772296024
++    },
++    "longctx": {
++     "hit_rate": 0.5904132443531828,
++     "stall_ms_per_token": 22.543655030800807
++    },
++    "drift": {
++     "hit_rate": 0.5726747776679841,
++     "stall_ms_per_token": 23.519980237154165
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "prev_rank",
++   "m": 8,
++   "hit_rate": 0.5450747016823161,
++   "misses_per_token": 58.230438184663534,
++   "stall_ms_per_token": 25.03908841940522,
++   "prefetch_loads_per_token": 101.44154929577465,
++   "prefetch_precision": 0.17654808654274107,
++   "loads_per_token": 159.6719874804382,
++   "pred_recall_at_k": null,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.4931080900676397,
++     "stall_ms_per_token": 27.899330722677103
++    },
++    "code": {
++     "hit_rate": 0.6040047951885565,
++     "stall_ms_per_token": 21.795576072821852
++    },
++    "reasoning": {
++     "hit_rate": 0.5079844732704403,
++     "stall_ms_per_token": 27.08053459119503
++    },
++    "multilingual": {
++     "hit_rate": 0.547024141957758,
++     "stall_ms_per_token": 24.931791226644997
++    },
++    "longctx": {
++     "hit_rate": 0.5017574130879345,
++     "stall_ms_per_token": 27.423271983640134
++    },
++    "drift": {
++     "hit_rate": 0.55595082461786,
++     "stall_ms_per_token": 24.440466613033017
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "prev_rank",
++   "m": 8,
++   "hit_rate": 0.5563927507447864,
++   "misses_per_token": 56.78172790466733,
++   "stall_ms_per_token": 24.41614299900682,
++   "prefetch_loads_per_token": 100.51956305858987,
++   "prefetch_precision": 0.17786831477368825,
++   "loads_per_token": 157.3012909632572,
++   "pred_recall_at_k": null,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.49301956754824877,
++     "stall_ms_per_token": 27.904203002144374
++    },
++    "code": {
++     "hit_rate": 0.6179897855353789,
++     "stall_ms_per_token": 21.025842204132797
++    },
++    "reasoning": {
++     "hit_rate": 0.5361121146435452,
++     "stall_ms_per_token": 25.53238921001927
++    },
++    "multilingual": {
++     "hit_rate": 0.5729512571157496,
++     "stall_ms_per_token": 23.504762808349145
++    },
++    "longctx": {
++     "hit_rate": 0.5559708675564682,
++     "stall_ms_per_token": 24.43936344969197
++    },
++    "drift": {
++     "hit_rate": 0.5411623023715415,
++     "stall_ms_per_token": 25.25442687747036
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "transition",
++   "m": 8,
++   "hit_rate": 0.6797400723787167,
++   "misses_per_token": 40.993270735524256,
++   "stall_ms_per_token": 17.323041314554185,
++   "prefetch_loads_per_token": 47.78051643192488,
++   "prefetch_precision": 0.5297419898957643,
++   "loads_per_token": 88.77378716744914,
++   "pred_recall_at_k": 0.5303931924882629,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.6455338866144535,
++     "stall_ms_per_token": 19.18719473122103
++    },
++    "code": {
++     "hit_rate": 0.7263146131339402,
++     "stall_ms_per_token": 14.791964109232758
++    },
++    "reasoning": {
++     "hit_rate": 0.6476267688679245,
++     "stall_ms_per_token": 19.093540880503163
++    },
++    "multilingual": {
++     "hit_rate": 0.6693015079203899,
++     "stall_ms_per_token": 17.88471242891963
++    },
++    "longctx": {
++     "hit_rate": 0.6531984918200409,
++     "stall_ms_per_token": 18.75508520790728
++    },
++    "drift": {
++     "hit_rate": 0.6896181114239742,
++     "stall_ms_per_token": 16.77829927594529
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "transition",
++   "m": 8,
++   "hit_rate": 0.6861811072492552,
++   "misses_per_token": 40.16881827209533,
++   "stall_ms_per_token": 17.002457199602762,
++   "prefetch_loads_per_token": 46.16464746772592,
++   "prefetch_precision": 0.5307672120427295,
++   "loads_per_token": 86.33346573982125,
++   "pred_recall_at_k": 0.5340830850711685,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.6342811829878484,
++     "stall_ms_per_token": 19.886887776983528
++    },
++    "code": {
++     "hit_rate": 0.7338026377582968,
++     "stall_ms_per_token": 14.39574326862868
++    },
++    "reasoning": {
++     "hit_rate": 0.6778961946050096,
++     "stall_ms_per_token": 17.441518304431565
++    },
++    "multilingual": {
++     "hit_rate": 0.6813626660341556,
++     "stall_ms_per_token": 17.254110056925995
++    },
++    "longctx": {
++     "hit_rate": 0.7016651694045175,
++     "stall_ms_per_token": 16.071305954825455
++    },
++    "drift": {
++     "hit_rate": 0.6779891304347826,
++     "stall_ms_per_token": 17.43838735177865
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "transition2",
++   "m": 8,
++   "hit_rate": 0.6632561864241001,
++   "misses_per_token": 43.10320813771518,
++   "stall_ms_per_token": 17.99225242566519,
++   "prefetch_loads_per_token": 43.67801251956182,
++   "prefetch_precision": 0.5143074676866026,
++   "loads_per_token": 86.78122065727699,
++   "pred_recall_at_k": 0.5199705175497429,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.6251084683161268,
++     "stall_ms_per_token": 20.05331363474548
++    },
++    "code": {
++     "hit_rate": 0.7141153283485046,
++     "stall_ms_per_token": 15.27096332899875
++    },
++    "reasoning": {
++     "hit_rate": 0.629373034591195,
++     "stall_ms_per_token": 19.87898951781966
++    },
++    "multilingual": {
++     "hit_rate": 0.652372309098294,
++     "stall_ms_per_token": 18.537619821283528
++    },
++    "longctx": {
++     "hit_rate": 0.6350066036128152,
++     "stall_ms_per_token": 19.499581458759362
++    },
++    "drift": {
++     "hit_rate": 0.6730440466613034,
++     "stall_ms_per_token": 17.445142397425595
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "transition2",
++   "m": 8,
++   "hit_rate": 0.6716360476663357,
++   "misses_per_token": 42.030585898709035,
++   "stall_ms_per_token": 17.603962661370414,
++   "prefetch_loads_per_token": 42.25759682224429,
++   "prefetch_precision": 0.5179233621755254,
++   "loads_per_token": 84.28818272095333,
++   "pred_recall_at_k": 0.5253883529578663,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.6152162258756254,
++     "stall_ms_per_token": 20.749392423159403
++    },
++    "code": {
++     "hit_rate": 0.7224385566687539,
++     "stall_ms_per_token": 14.84929993738261
++    },
++    "reasoning": {
++     "hit_rate": 0.6646194605009634,
++     "stall_ms_per_token": 17.95819653179191
++    },
++    "multilingual": {
++     "hit_rate": 0.66760555028463,
++     "stall_ms_per_token": 17.772789373814035
++    },
++    "longctx": {
++     "hit_rate": 0.6891362936344969,
++     "stall_ms_per_token": 16.517511293634488
++    },
++    "drift": {
++     "hit_rate": 0.6618391798418972,
++     "stall_ms_per_token": 18.107960474308307
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "transition3",
++   "m": 8,
++   "hit_rate": 0.6510508362676056,
++   "misses_per_token": 44.66549295774648,
++   "stall_ms_per_token": 18.48885117370898,
++   "prefetch_loads_per_token": 40.56651017214398,
++   "prefetch_precision": 0.5009181390324821,
++   "loads_per_token": 85.23200312989046,
++   "pred_recall_at_k": 0.5104414951245937,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.6096058428266287,
++     "stall_ms_per_token": 20.7317344250623
++    },
++    "code": {
++     "hit_rate": 0.7057847041612484,
++     "stall_ms_per_token": 15.57211911573469
++    },
++    "reasoning": {
++     "hit_rate": 0.6173021488469602,
++     "stall_ms_per_token": 20.341979035639437
++    },
++    "multilingual": {
++     "hit_rate": 0.6389812652315191,
++     "stall_ms_per_token": 19.107618196588138
++    },
++    "longctx": {
++     "hit_rate": 0.6204999573960464,
++     "stall_ms_per_token": 20.097956373551472
++    },
++    "drift": {
++     "hit_rate": 0.6612655872888173,
++     "stall_ms_per_token": 17.895713596138364
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "transition3",
++   "m": 8,
++   "hit_rate": 0.6597722194637538,
++   "misses_per_token": 43.54915590863952,
++   "stall_ms_per_token": 18.101746176762692,
++   "prefetch_loads_per_token": 39.104071499503476,
++   "prefetch_precision": 0.5073112261223329,
++   "loads_per_token": 82.653227408143,
++   "pred_recall_at_k": 0.5172217554044763,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.6017411990707648,
++     "stall_ms_per_token": 21.38684631879916
++    },
++    "code": {
++     "hit_rate": 0.7129040779586725,
++     "stall_ms_per_token": 15.208086412022535
++    },
++    "reasoning": {
++     "hit_rate": 0.6510266136801541,
++     "stall_ms_per_token": 18.514223506743733
++    },
++    "multilingual": {
++     "hit_rate": 0.6537891366223909,
++     "stall_ms_per_token": 18.379984819734347
++    },
++    "longctx": {
++     "hit_rate": 0.6780993326488707,
++     "stall_ms_per_token": 16.879236139630407
++    },
++    "drift": {
++     "hit_rate": 0.650089550395257,
++     "stall_ms_per_token": 18.61552964426879
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "probe",
++   "m": 8,
++   "hit_rate": 0.49547877543035995,
++   "misses_per_token": 64.57871674491393,
++   "stall_ms_per_token": 27.66435492957752,
++   "prefetch_loads_per_token": 74.49397496087637,
++   "prefetch_precision": 0.12610592279889457,
++   "loads_per_token": 139.0726917057903,
++   "pred_recall_at_k": 0.15737415232133542,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.45225447223211107,
++     "stall_ms_per_token": 30.049350658597398
++    },
++    "code": {
++     "hit_rate": 0.546966433680104,
++     "stall_ms_per_token": 24.83196827048118
++    },
++    "reasoning": {
++     "hit_rate": 0.4698637316561845,
++     "stall_ms_per_token": 29.060226415094373
++    },
++    "multilingual": {
++     "hit_rate": 0.4859552701056052,
++     "stall_ms_per_token": 28.199543460601106
++    },
++    "longctx": {
++     "hit_rate": 0.47049143660531695,
++     "stall_ms_per_token": 29.01728561690526
++    },
++    "drift": {
++     "hit_rate": 0.5019044147224457,
++     "stall_ms_per_token": 27.307982300885016
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "probe",
++   "m": 8,
++   "hit_rate": 0.5040823609731877,
++   "misses_per_token": 63.477457795431974,
++   "stall_ms_per_token": 27.196110029791424,
++   "prefetch_loads_per_token": 73.92115193644489,
++   "prefetch_precision": 0.12284218134140083,
++   "loads_per_token": 137.39860973187686,
++   "pred_recall_at_k": 0.1567345249917246,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.44355343102215866,
++     "stall_ms_per_token": 30.54273624017161
++    },
++    "code": {
++     "hit_rate": 0.5584151925485284,
++     "stall_ms_per_token": 24.21209392611144
++    },
++    "reasoning": {
++     "hit_rate": 0.49816353564547206,
++     "stall_ms_per_token": 27.495052023121385
++    },
++    "multilingual": {
++     "hit_rate": 0.5094728415559773,
++     "stall_ms_per_token": 26.895631878557854
++    },
++    "longctx": {
++     "hit_rate": 0.5100102669404517,
++     "stall_ms_per_token": 26.84471457905543
++    },
++    "drift": {
++     "hit_rate": 0.4947041749011858,
++     "stall_ms_per_token": 27.705743083003927
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "oracle",
++   "m": 8,
++   "hit_rate": 0.9378728971048513,
++   "misses_per_token": 7.952269170579029,
++   "stall_ms_per_token": 2.961490140845083,
++   "prefetch_loads_per_token": 53.00367762128325,
++   "prefetch_precision": 1.0,
++   "loads_per_token": 60.95594679186228,
++   "pred_recall_at_k": 1.0,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.9337981710573158,
++     "stall_ms_per_token": 3.1469683161267357
++    },
++    "code": {
++     "hit_rate": 0.9422586150845254,
++     "stall_ms_per_token": 2.7841321196358835
++    },
++    "reasoning": {
++     "hit_rate": 0.9387775157232704,
++     "stall_ms_per_token": 2.8992788259958058
++    },
++    "multilingual": {
++     "hit_rate": 0.938861317018684,
++     "stall_ms_per_token": 2.8865028432169013
++    },
++    "longctx": {
++     "hit_rate": 0.9329040985003408,
++     "stall_ms_per_token": 3.1930865712338177
++    },
++    "drift": {
++     "hit_rate": 0.9367269207562349,
++     "stall_ms_per_token": 3.013903459372493
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 16,
++   "policy": "lru",
++   "predictor": "oracle",
++   "m": 8,
++   "hit_rate": 0.9400229642502482,
++   "misses_per_token": 7.6770605759682224,
++   "stall_ms_per_token": 2.8805791459781513,
++   "prefetch_loads_per_token": 52.345382323733865,
++   "prefetch_precision": 1.0,
++   "loads_per_token": 60.022442899702085,
++   "pred_recall_at_k": 1.0,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.9348474356683345,
++     "stall_ms_per_token": 3.1205675482487574
++    },
++    "code": {
++     "hit_rate": 0.9450630087664371,
++     "stall_ms_per_token": 2.6702692548528533
++    },
++    "reasoning": {
++     "hit_rate": 0.9424373795761078,
++     "stall_ms_per_token": 2.736400770712911
++    },
++    "multilingual": {
++     "hit_rate": 0.943622509487666,
++     "stall_ms_per_token": 2.6937571157495235
++    },
++    "longctx": {
++     "hit_rate": 0.9356070328542094,
++     "stall_ms_per_token": 3.06346611909651
++    },
++    "drift": {
++     "hit_rate": 0.9364500988142292,
++     "stall_ms_per_token": 3.047256916996053
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "none",
++   "m": 8,
++   "hit_rate": 0.7056955447965572,
++   "misses_per_token": 37.67097026604069,
++   "stall_ms_per_token": 16.198517214397622,
++   "prefetch_loads_per_token": 0.0,
++   "prefetch_precision": null,
++   "loads_per_token": 37.67097026604069,
++   "pred_recall_at_k": null,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.6627970363118547,
++     "stall_ms_per_token": 18.559651121395515
++    },
++    "code": {
++     "hit_rate": 0.7612321196358908,
++     "stall_ms_per_token": 13.14178413524059
++    },
++    "reasoning": {
++     "hit_rate": 0.6531299135220126,
++     "stall_ms_per_token": 19.09172955974842
++    },
++    "multilingual": {
++     "hit_rate": 0.6999961921202275,
++     "stall_ms_per_token": 16.5122095857027
++    },
++    "longctx": {
++     "hit_rate": 0.6662619291070211,
++     "stall_ms_per_token": 18.368943421949588
++    },
++    "drift": {
++     "hit_rate": 0.7290200120675785,
++     "stall_ms_per_token": 14.914738535800488
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "none",
++   "m": 8,
++   "hit_rate": 0.7060638033763654,
++   "misses_per_token": 37.623833167825225,
++   "stall_ms_per_token": 16.17824826216484,
++   "prefetch_loads_per_token": 0.0,
++   "prefetch_precision": null,
++   "loads_per_token": 37.623833167825225,
++   "pred_recall_at_k": null,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.6400442280200143,
++     "stall_ms_per_token": 19.811965689778436
++    },
++    "code": {
++     "hit_rate": 0.7618434956167814,
++     "stall_ms_per_token": 13.10813400125233
++    },
++    "reasoning": {
++     "hit_rate": 0.6932803468208093,
++     "stall_ms_per_token": 16.881849710982664
++    },
++    "multilingual": {
++     "hit_rate": 0.7020279886148008,
++     "stall_ms_per_token": 16.40037950664138
++    },
++    "longctx": {
++     "hit_rate": 0.7278939938398358,
++     "stall_ms_per_token": 14.976714579055447
++    },
++    "drift": {
++     "hit_rate": 0.7088531373517787,
++     "stall_ms_per_token": 16.024723320158095
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "prev_rank",
++   "m": 8,
++   "hit_rate": 0.665095608372457,
++   "misses_per_token": 42.86776212832551,
++   "stall_ms_per_token": 18.43313771518011,
++   "prefetch_loads_per_token": 71.26275430359938,
++   "prefetch_precision": 0.2513774543282481,
++   "loads_per_token": 114.13051643192489,
++   "pred_recall_at_k": null,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.612879360982556,
++     "stall_ms_per_token": 21.307119971520095
++    },
++    "code": {
++     "hit_rate": 0.7267941319895969,
++     "stall_ms_per_token": 15.037250975292599
++    },
++    "reasoning": {
++     "hit_rate": 0.6330008516771488,
++     "stall_ms_per_token": 20.199633123689722
++    },
++    "multilingual": {
++     "hit_rate": 0.6600420136068237,
++     "stall_ms_per_token": 18.71128757108042
++    },
++    "longctx": {
++     "hit_rate": 0.6173472648261759,
++     "stall_ms_per_token": 21.06120654396727
++    },
++    "drift": {
++     "hit_rate": 0.6832386363636364,
++     "stall_ms_per_token": 17.43454545454546
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "prev_rank",
++   "m": 8,
++   "hit_rate": 0.6703093967229394,
++   "misses_per_token": 42.20039721946375,
++   "stall_ms_per_token": 18.146170804369323,
++   "prefetch_loads_per_token": 69.97715988083417,
++   "prefetch_precision": 0.2525295528403366,
++   "loads_per_token": 112.17755710029792,
++   "pred_recall_at_k": null,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.598206308077198,
++     "stall_ms_per_token": 22.11472480343104
++    },
++    "code": {
++     "hit_rate": 0.7332351675015655,
++     "stall_ms_per_token": 14.682736380713855
++    },
++    "reasoning": {
++     "hit_rate": 0.6695869460500964,
++     "stall_ms_per_token": 18.18593448940269
++    },
++    "multilingual": {
++     "hit_rate": 0.684638875711575,
++     "stall_ms_per_token": 17.357476280834916
++    },
++    "longctx": {
++     "hit_rate": 0.6707039271047228,
++     "stall_ms_per_token": 18.12445585215607
++    },
++    "drift": {
++     "hit_rate": 0.6564970355731226,
++     "stall_ms_per_token": 18.90640316205532
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "transition",
++   "m": 8,
++   "hit_rate": 0.7695171899452269,
++   "misses_per_token": 29.501799687010955,
++   "stall_ms_per_token": 12.539625508607376,
++   "prefetch_loads_per_token": 28.68740219092332,
++   "prefetch_precision": 0.5961022843504944,
++   "loads_per_token": 58.18920187793427,
++   "pred_recall_at_k": 0.5303931924882629,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.7396621351014596,
++     "stall_ms_per_token": 14.126401566393742
++    },
++    "code": {
++     "hit_rate": 0.8083042100130039,
++     "stall_ms_per_token": 10.47729882964887
++    },
++    "reasoning": {
++     "hit_rate": 0.740811713836478,
++     "stall_ms_per_token": 14.119475890985337
++    },
++    "multilingual": {
++     "hit_rate": 0.7617726949634444,
++     "stall_ms_per_token": 12.939423233143774
++    },
++    "longctx": {
++     "hit_rate": 0.7455691888207225,
++     "stall_ms_per_token": 13.80746557600546
++    },
++    "drift": {
++     "hit_rate": 0.7826390285599356,
++     "stall_ms_per_token": 11.832458567980705
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "transition",
++   "m": 8,
++   "hit_rate": 0.7710495283018868,
++   "misses_per_token": 29.30566037735849,
++   "stall_ms_per_token": 12.458053624627631,
++   "prefetch_loads_per_token": 28.439126117179743,
++   "prefetch_precision": 0.6003310263913235,
++   "loads_per_token": 57.74478649453823,
++   "pred_recall_at_k": 0.5340830850711685,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.7193028502501787,
++     "stall_ms_per_token": 15.229162258756267
++    },
++    "code": {
++     "hit_rate": 0.8130479023168441,
++     "stall_ms_per_token": 10.21995241077018
++    },
++    "reasoning": {
++     "hit_rate": 0.7692979287090559,
++     "stall_ms_per_token": 12.571815028901753
++    },
++    "multilingual": {
++     "hit_rate": 0.7678783206831119,
++     "stall_ms_per_token": 12.61445920303606
++    },
++    "longctx": {
++     "hit_rate": 0.7913244353182751,
++     "stall_ms_per_token": 11.335162217659146
++    },
++    "drift": {
++     "hit_rate": 0.7671535326086957,
++     "stall_ms_per_token": 12.661308300395252
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "transition2",
++   "m": 8,
++   "hit_rate": 0.7583577856025039,
++   "misses_per_token": 30.9302034428795,
++   "stall_ms_per_token": 13.046912050078358,
++   "prefetch_loads_per_token": 25.87175273865415,
++   "prefetch_precision": 0.5829525074022883,
++   "loads_per_token": 56.80195618153365,
++   "pred_recall_at_k": 0.5199705175497429,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.7251663180847276,
++     "stall_ms_per_token": 14.76521965112141
++    },
++    "code": {
++     "hit_rate": 0.8003169700910273,
++     "stall_ms_per_token": 10.867247854356316
++    },
++    "reasoning": {
++     "hit_rate": 0.72661982442348,
++     "stall_ms_per_token": 14.80265199161428
++    },
++    "multilingual": {
++     "hit_rate": 0.7508313870836718,
++     "stall_ms_per_token": 13.408433793663667
++    },
++    "longctx": {
++     "hit_rate": 0.7335655248807089,
++     "stall_ms_per_token": 14.32950238582145
++    },
++    "drift": {
++     "hit_rate": 0.7720987530168946,
++     "stall_ms_per_token": 12.328875301689468
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "transition2",
++   "m": 8,
++   "hit_rate": 0.7605294190665343,
++   "misses_per_token": 30.652234359483614,
++   "stall_ms_per_token": 12.922579543197587,
++   "prefetch_loads_per_token": 25.65183714001986,
++   "prefetch_precision": 0.5894376611410919,
++   "loads_per_token": 56.30407149950348,
++   "pred_recall_at_k": 0.5253883529578663,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.7039012240886348,
++     "stall_ms_per_token": 15.872744817726966
++    },
++    "code": {
++     "hit_rate": 0.805377269881027,
++     "stall_ms_per_token": 10.588763932373194
++    },
++    "reasoning": {
++     "hit_rate": 0.7580834537572254,
++     "stall_ms_per_token": 13.089668593448962
++    },
++    "multilingual": {
++     "hit_rate": 0.758657495256167,
++     "stall_ms_per_token": 13.022633776091077
++    },
++    "longctx": {
++     "hit_rate": 0.7839771560574949,
++     "stall_ms_per_token": 11.64388501026695
++    },
++    "drift": {
++     "hit_rate": 0.7574419466403162,
++     "stall_ms_per_token": 13.086806324110679
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "transition3",
++   "m": 8,
++   "hit_rate": 0.7494186473004695,
++   "misses_per_token": 32.07441314553991,
++   "stall_ms_per_token": 13.452467136150394,
++   "prefetch_loads_per_token": 23.852347417840374,
++   "prefetch_precision": 0.5681963566936651,
++   "loads_per_token": 55.926760563380284,
++   "pred_recall_at_k": 0.5104414951245937,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.7135657707369171,
++     "stall_ms_per_token": 15.286467782128895
++    },
++    "code": {
++     "hit_rate": 0.7938454973992197,
++     "stall_ms_per_token": 11.183056957087087
++    },
++    "reasoning": {
++     "hit_rate": 0.7177672955974843,
++     "stall_ms_per_token": 15.189937106918237
++    },
++    "multilingual": {
++     "hit_rate": 0.7413973649471974,
++     "stall_ms_per_token": 13.822506092607616
++    },
++    "longctx": {
++     "hit_rate": 0.7228186775732788,
++     "stall_ms_per_token": 14.804788002726658
++    },
++    "drift": {
++     "hit_rate": 0.7645879424778761,
++     "stall_ms_per_token": 12.665449718423186
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "transition3",
++   "m": 8,
++   "hit_rate": 0.7522964250248262,
++   "misses_per_token": 31.706057596822244,
++   "stall_ms_per_token": 13.296355908639498,
++   "prefetch_loads_per_token": 23.54438927507448,
++   "prefetch_precision": 0.5790241762691276,
++   "loads_per_token": 55.25044687189672,
++   "pred_recall_at_k": 0.5172217554044763,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.6938829074338814,
++     "stall_ms_per_token": 16.299172265904204
++    },
++    "code": {
++     "hit_rate": 0.799100853162179,
++     "stall_ms_per_token": 10.899272385723242
++    },
++    "reasoning": {
++     "hit_rate": 0.7492021917148363,
++     "stall_ms_per_token": 13.512396917148378
++    },
++    "multilingual": {
++     "hit_rate": 0.7502964895635673,
++     "stall_ms_per_token": 13.38154079696395
++    },
++    "longctx": {
++     "hit_rate": 0.7759721509240246,
++     "stall_ms_per_token": 11.994591375770028
++    },
++    "drift": {
++     "hit_rate": 0.7485486660079052,
++     "stall_ms_per_token": 13.502173913043492
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "probe",
++   "m": 8,
++   "hit_rate": 0.6263888888888889,
++   "misses_per_token": 47.82222222222222,
++   "stall_ms_per_token": 20.515099843505446,
++   "prefetch_loads_per_token": 54.35453834115806,
++   "prefetch_precision": 0.1747510620441056,
++   "loads_per_token": 102.17676056338028,
++   "pred_recall_at_k": 0.15737415232133542,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.5880206479174084,
++     "stall_ms_per_token": 22.625401210395133
++    },
++    "code": {
++     "hit_rate": 0.6773833712613784,
++     "stall_ms_per_token": 17.719434590377126
++    },
++    "reasoning": {
++     "hit_rate": 0.602307717505241,
++     "stall_ms_per_token": 21.826995807127904
++    },
++    "multilingual": {
++     "hit_rate": 0.6117644445572705,
++     "stall_ms_per_token": 21.314311129163283
++    },
++    "longctx": {
++     "hit_rate": 0.5963861196319018,
++     "stall_ms_per_token": 22.154528970688496
++    },
++    "drift": {
++     "hit_rate": 0.638211484312148,
++     "stall_ms_per_token": 19.86928077232501
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "probe",
++   "m": 8,
++   "hit_rate": 0.6299264523336644,
++   "misses_per_token": 47.369414101290964,
++   "stall_ms_per_token": 20.323469712015903,
++   "prefetch_loads_per_token": 53.83118172790467,
++   "prefetch_precision": 0.1742547225501771,
++   "loads_per_token": 101.20059582919563,
++   "pred_recall_at_k": 0.1567345249917246,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.5662973552537527,
++     "stall_ms_per_token": 23.825338098641907
++    },
++    "code": {
++     "hit_rate": 0.6815953741390106,
++     "stall_ms_per_token": 17.489403882279262
++    },
++    "reasoning": {
++     "hit_rate": 0.6366359585741811,
++     "stall_ms_per_token": 19.94579190751447
++    },
++    "multilingual": {
++     "hit_rate": 0.6333461812144212,
++     "stall_ms_per_token": 20.127434535104378
++    },
++    "longctx": {
++     "hit_rate": 0.6382186858316222,
++     "stall_ms_per_token": 19.85908829568788
++    },
++    "drift": {
++     "hit_rate": 0.6243515316205533,
++     "stall_ms_per_token": 20.62459288537549
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "oracle",
++   "m": 8,
++   "hit_rate": 0.9589391383020345,
++   "misses_per_token": 5.255790297339593,
++   "stall_ms_per_token": 2.014346009389659,
++   "prefetch_loads_per_token": 36.14765258215962,
++   "prefetch_precision": 1.0,
++   "loads_per_token": 41.40344287949922,
++   "pred_recall_at_k": 1.0,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.9532112184051263,
++     "stall_ms_per_token": 2.269038803844785
++    },
++    "code": {
++     "hit_rate": 0.9637678803641092,
++     "stall_ms_per_token": 1.8191771131339411
++    },
++    "reasoning": {
++     "hit_rate": 0.9594470649895178,
++     "stall_ms_per_token": 1.9843522012578618
++    },
++    "multilingual": {
++     "hit_rate": 0.9606487357839155,
++     "stall_ms_per_token": 1.9062542648253529
++    },
++    "longctx": {
++     "hit_rate": 0.9543498636673483,
++     "stall_ms_per_token": 2.221071574642139
++    },
++    "drift": {
++     "hit_rate": 0.9585868362831859,
++     "stall_ms_per_token": 2.035634754625913
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 24,
++   "policy": "lru",
++   "predictor": "oracle",
++   "m": 8,
++   "hit_rate": 0.9590988083416088,
++   "misses_per_token": 5.2353525322740815,
++   "stall_ms_per_token": 2.0118276067527248,
++   "prefetch_loads_per_token": 36.057199602780535,
++   "prefetch_precision": 1.0,
++   "loads_per_token": 41.29255213505462,
++   "pred_recall_at_k": 1.0,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.9526056558255898,
++     "stall_ms_per_token": 2.3025375268048696
++    },
++    "code": {
++     "hit_rate": 0.9640634783969944,
++     "stall_ms_per_token": 1.8043406386975698
++    },
++    "reasoning": {
++     "hit_rate": 0.961283718689788,
++     "stall_ms_per_token": 1.8951676300578073
++    },
++    "multilingual": {
++     "hit_rate": 0.9611153937381404,
++     "stall_ms_per_token": 1.889795066413664
++    },
++    "longctx": {
++     "hit_rate": 0.9581301334702259,
++     "stall_ms_per_token": 2.0661273100616055
++    },
++    "drift": {
++     "hit_rate": 0.9579730731225297,
++     "stall_ms_per_token": 2.0574150197628462
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "none",
++   "m": 8,
++   "hit_rate": 0.8038921899452269,
++   "misses_per_token": 25.101799687010956,
++   "stall_ms_per_token": 10.793773865414801,
++   "prefetch_loads_per_token": 0.0,
++   "prefetch_precision": null,
++   "loads_per_token": 25.101799687010956,
++   "pred_recall_at_k": null,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.7753843672125311,
++     "stall_ms_per_token": 12.362844428622287
++    },
++    "code": {
++     "hit_rate": 0.8468689044213265,
++     "stall_ms_per_token": 8.428335500650158
++    },
++    "reasoning": {
++     "hit_rate": 0.7583120414046122,
++     "stall_ms_per_token": 13.302505241090165
++    },
++    "multilingual": {
++     "hit_rate": 0.7954153127538587,
++     "stall_ms_per_token": 11.26034118602761
++    },
++    "longctx": {
++     "hit_rate": 0.7717439928425358,
++     "stall_ms_per_token": 12.563210633946827
++    },
++    "drift": {
++     "hit_rate": 0.8250892497988737,
++     "stall_ms_per_token": 9.627087691069988
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "none",
++   "m": 8,
++   "hit_rate": 0.8059241559086395,
++   "misses_per_token": 24.84170804369414,
++   "stall_ms_per_token": 10.68193445878844,
++   "prefetch_loads_per_token": 0.0,
++   "prefetch_precision": null,
++   "loads_per_token": 24.84170804369414,
++   "pred_recall_at_k": null,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.7592476769120801,
++     "stall_ms_per_token": 13.25100786275913
++    },
++    "code": {
++     "hit_rate": 0.8463525360050094,
++     "stall_ms_per_token": 8.456756418284268
++    },
++    "reasoning": {
++     "hit_rate": 0.79532454238921,
++     "stall_ms_per_token": 11.265337186897893
++    },
++    "multilingual": {
++     "hit_rate": 0.7923535341555977,
++     "stall_ms_per_token": 11.428861480075915
++    },
++    "longctx": {
++     "hit_rate": 0.8272747689938398,
++     "stall_ms_per_token": 9.506796714579071
++    },
++    "drift": {
++     "hit_rate": 0.8118360918972332,
++     "stall_ms_per_token": 10.35654150197629
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "prev_rank",
++   "m": 8,
++   "hit_rate": 0.7698002249608764,
++   "misses_per_token": 29.465571205007826,
++   "stall_ms_per_token": 12.670195618153517,
++   "prefetch_loads_per_token": 50.767605633802816,
++   "prefetch_precision": 0.32121884681185553,
++   "loads_per_token": 80.23317683881064,
++   "pred_recall_at_k": null,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.7298360181559274,
++     "stall_ms_per_token": 14.869825560697738
++    },
++    "code": {
++     "hit_rate": 0.8158058355006502,
++     "stall_ms_per_token": 10.138046814044172
++    },
++    "reasoning": {
++     "hit_rate": 0.7362257599580713,
++     "stall_ms_per_token": 14.518134171907757
++    },
++    "multilingual": {
++     "hit_rate": 0.7695534626320065,
++     "stall_ms_per_token": 12.68377741673434
++    },
++    "longctx": {
++     "hit_rate": 0.7352111025903204,
++     "stall_ms_per_token": 14.57398091342876
++    },
++    "drift": {
++     "hit_rate": 0.7848828439259855,
++     "stall_ms_per_token": 11.84004827031376
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "prev_rank",
++   "m": 8,
++   "hit_rate": 0.7714064051638531,
++   "misses_per_token": 29.259980139026812,
++   "stall_ms_per_token": 12.581791459781547,
++   "prefetch_loads_per_token": 50.12591857000993,
++   "prefetch_precision": 0.323047419804742,
++   "loads_per_token": 79.38589870903674,
++   "pred_recall_at_k": null,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.7131991601143675,
++     "stall_ms_per_token": 15.785518227305232
++    },
++    "code": {
++     "hit_rate": 0.819789057608015,
++     "stall_ms_per_token": 9.918810269254823
++    },
++    "reasoning": {
++     "hit_rate": 0.765655105973025,
++     "stall_ms_per_token": 12.898342967244727
++    },
++    "multilingual": {
++     "hit_rate": 0.7742231973434535,
++     "stall_ms_per_token": 12.426755218216348
++    },
++    "longctx": {
++     "hit_rate": 0.7876347535934292,
++     "stall_ms_per_token": 11.688583162217679
++    },
++    "drift": {
++     "hit_rate": 0.766983695652174,
++     "stall_ms_per_token": 12.825217391304362
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "transition",
++   "m": 8,
++   "hit_rate": 0.8389139524647887,
++   "misses_per_token": 20.619014084507043,
++   "stall_ms_per_token": 8.791339123630697,
++   "prefetch_loads_per_token": 16.70774647887324,
++   "prefetch_precision": 0.6440042149631191,
++   "loads_per_token": 37.32676056338028,
++   "pred_recall_at_k": 0.5303931924882629,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.8136375266998932,
++     "stall_ms_per_token": 10.128524741901023
++    },
++    "code": {
++     "hit_rate": 0.8715661573472041,
++     "stall_ms_per_token": 7.039603641092297
++    },
++    "reasoning": {
++     "hit_rate": 0.8099122117400419,
++     "stall_ms_per_token": 10.381123689727469
++    },
++    "multilingual": {
++     "hit_rate": 0.8317107534524777,
++     "stall_ms_per_token": 9.183303005686419
++    },
++    "longctx": {
++     "hit_rate": 0.8196734406952966,
++     "stall_ms_per_token": 9.83477982276755
++    },
++    "drift": {
++     "hit_rate": 0.854265134754626,
++     "stall_ms_per_token": 7.960181818181832
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "transition",
++   "m": 8,
++   "hit_rate": 0.8400276191658391,
++   "misses_per_token": 20.476464746772592,
++   "stall_ms_per_token": 8.728867924528226,
++   "prefetch_loads_per_token": 16.575372393247267,
++   "prefetch_precision": 0.64816612147573,
++   "loads_per_token": 37.05183714001986,
++   "pred_recall_at_k": 0.5340830850711685,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.7972714885632595,
++     "stall_ms_per_token": 11.02706647605432
++    },
++    "code": {
++     "hit_rate": 0.8732584533500313,
++     "stall_ms_per_token": 6.942993112085163
++    },
++    "reasoning": {
++     "hit_rate": 0.8345827312138728,
++     "stall_ms_per_token": 9.030084778420045
++    },
++    "multilingual": {
++     "hit_rate": 0.8367973197343453,
++     "stall_ms_per_token": 8.903495256166993
++    },
++    "longctx": {
++     "hit_rate": 0.8607225359342916,
++     "stall_ms_per_token": 7.609585215605754
++    },
++    "drift": {
++     "hit_rate": 0.842391304347826,
++     "stall_ms_per_token": 8.597628458498033
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "transition2",
++   "m": 8,
++   "hit_rate": 0.8318466353677622,
++   "misses_per_token": 21.523630672926448,
++   "stall_ms_per_token": 9.12297464788734,
++   "prefetch_loads_per_token": 14.8,
++   "prefetch_precision": 0.6313337985873197,
++   "loads_per_token": 36.32363067292645,
++   "pred_recall_at_k": 0.5199705175497429,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.8044677821288715,
++     "stall_ms_per_token": 10.531702385190426
++    },
++    "code": {
++     "hit_rate": 0.8670757477243173,
++     "stall_ms_per_token": 7.2714756827048035
++    },
++    "reasoning": {
++     "hit_rate": 0.8014200078616353,
++     "stall_ms_per_token": 10.799427672955982
++    },
++    "multilingual": {
++     "hit_rate": 0.8243774116571893,
++     "stall_ms_per_token": 9.527141348497146
++    },
++    "longctx": {
++     "hit_rate": 0.8110195126107703,
++     "stall_ms_per_token": 10.224362644853437
++    },
++    "drift": {
++     "hit_rate": 0.8474708366854384,
++     "stall_ms_per_token": 8.279665325824634
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "transition2",
++   "m": 8,
++   "hit_rate": 0.8331662735849057,
++   "misses_per_token": 21.354716981132075,
++   "stall_ms_per_token": 9.044283217477576,
++   "prefetch_loads_per_token": 14.600198609731876,
++   "prefetch_precision": 0.6373517248884536,
++   "loads_per_token": 35.954915590863955,
++   "pred_recall_at_k": 0.5253883529578663,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.7867003216583274,
++     "stall_ms_per_token": 11.485112223016426
++    },
++    "code": {
++     "hit_rate": 0.8686648794614903,
++     "stall_ms_per_token": 7.173221039448974
++    },
++    "reasoning": {
++     "hit_rate": 0.829329238921002,
++     "stall_ms_per_token": 9.266951830443181
++    },
++    "multilingual": {
++     "hit_rate": 0.8303041982922201,
++     "stall_ms_per_token": 9.208413662239096
++    },
++    "longctx": {
++     "hit_rate": 0.8547227926078029,
++     "stall_ms_per_token": 7.913527720739232
++    },
++    "drift": {
++     "hit_rate": 0.8357676630434783,
++     "stall_ms_per_token": 8.890098814229262
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "transition3",
++   "m": 8,
++   "hit_rate": 0.8264151750782472,
++   "misses_per_token": 22.21885758998435,
++   "stall_ms_per_token": 9.380527543035994,
++   "prefetch_loads_per_token": 13.466118935837246,
++   "prefetch_precision": 0.6154668588063709,
++   "loads_per_token": 35.6849765258216,
++   "pred_recall_at_k": 0.5104414951245937,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.7967609692061232,
++     "stall_ms_per_token": 10.88457173371306
++    },
++    "code": {
++     "hit_rate": 0.8634529421326398,
++     "stall_ms_per_token": 7.463318075422596
++    },
++    "reasoning": {
++     "hit_rate": 0.7947048611111112,
++     "stall_ms_per_token": 11.115494758909856
++    },
++    "multilingual": {
++     "hit_rate": 0.8191986951665313,
++     "stall_ms_per_token": 9.754780666125104
++    },
++    "longctx": {
++     "hit_rate": 0.8043306918882073,
++     "stall_ms_per_token": 10.542693933196997
++    },
++    "drift": {
++     "hit_rate": 0.8435551588897828,
++     "stall_ms_per_token": 8.467689460981514
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "transition3",
++   "m": 8,
++   "hit_rate": 0.8281234483614697,
++   "misses_per_token": 22.00019860973188,
++   "stall_ms_per_token": 9.279868520357407,
++   "prefetch_loads_per_token": 13.218669314796426,
++   "prefetch_precision": 0.6253831360057696,
++   "loads_per_token": 35.218867924528304,
++   "pred_recall_at_k": 0.5172217554044763,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.780317414224446,
++     "stall_ms_per_token": 11.764421729806983
++    },
++    "code": {
++     "hit_rate": 0.8646779117094552,
++     "stall_ms_per_token": 7.373107075767064
++    },
++    "reasoning": {
++     "hit_rate": 0.8230069845857418,
++     "stall_ms_per_token": 9.569302504816964
++    },
++    "multilingual": {
++     "hit_rate": 0.8256196631878557,
++     "stall_ms_per_token": 9.422288425047457
++    },
++    "longctx": {
++     "hit_rate": 0.8514822895277208,
++     "stall_ms_per_token": 8.053774127310074
++    },
++    "drift": {
++     "hit_rate": 0.8303020009881423,
++     "stall_ms_per_token": 9.163355731225302
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "probe",
++   "m": 8,
++   "hit_rate": 0.7370629156885758,
++   "misses_per_token": 33.655946791862284,
++   "stall_ms_per_token": 14.447689045383642,
++   "prefetch_loads_per_token": 39.34530516431925,
++   "prefetch_precision": 0.2312656488337082,
++   "loads_per_token": 73.00125195618153,
++   "pred_recall_at_k": 0.15737415232133542,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.7085511970452119,
++     "stall_ms_per_token": 16.01468494126022
++    },
++    "code": {
++     "hit_rate": 0.7734110858257477,
++     "stall_ms_per_token": 12.454475942782848
++    },
++    "reasoning": {
++     "hit_rate": 0.71187106918239,
++     "stall_ms_per_token": 15.826549266247383
++    },
++    "multilingual": {
++     "hit_rate": 0.7287615505686433,
++     "stall_ms_per_token": 14.90068155970755
++    },
++    "longctx": {
++     "hit_rate": 0.7159168370824812,
++     "stall_ms_per_token": 15.608365371506494
++    },
++    "drift": {
++     "hit_rate": 0.7497925884955752,
++     "stall_ms_per_token": 13.746804505229283
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "probe",
++   "m": 8,
++   "hit_rate": 0.7380291087388282,
++   "misses_per_token": 33.53227408142999,
++   "stall_ms_per_token": 14.393735451837136,
++   "prefetch_loads_per_token": 38.80754716981132,
++   "prefetch_precision": 0.2311050379741653,
++   "loads_per_token": 72.3398212512413,
++   "pred_recall_at_k": 0.1567345249917246,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.689448936740529,
++     "stall_ms_per_token": 17.063543959971398
++    },
++    "code": {
++     "hit_rate": 0.7764216108328115,
++     "stall_ms_per_token": 12.286919223544121
++    },
++    "reasoning": {
++     "hit_rate": 0.7364523121387283,
++     "stall_ms_per_token": 14.477795761079
++    },
++    "multilingual": {
++     "hit_rate": 0.7363911290322581,
++     "stall_ms_per_token": 14.48110056925996
++    },
++    "longctx": {
++     "hit_rate": 0.7545880390143738,
++     "stall_ms_per_token": 13.481716632443536
++    },
++    "drift": {
++     "hit_rate": 0.7385591650197628,
++     "stall_ms_per_token": 14.362130434782609
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "oracle",
++   "m": 8,
++   "hit_rate": 0.9727534477699531,
++   "misses_per_token": 3.487558685446009,
++   "stall_ms_per_token": 1.363878090766773,
++   "prefetch_loads_per_token": 23.683255086071988,
++   "prefetch_precision": 1.0,
++   "loads_per_token": 27.170813771517995,
++   "pred_recall_at_k": 1.0,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.9685191571733713,
++     "stall_ms_per_token": 1.556039159843367
++    },
++    "code": {
++     "hit_rate": 0.9760992360208063,
++     "stall_ms_per_token": 1.2257482444733523
++    },
++    "reasoning": {
++     "hit_rate": 0.9700029481132075,
++     "stall_ms_per_token": 1.4821153039832327
++    },
++    "multilingual": {
++     "hit_rate": 0.9746775995125914,
++     "stall_ms_per_token": 1.2524874086108941
++    },
++    "longctx": {
++     "hit_rate": 0.9693038513974097,
++     "stall_ms_per_token": 1.5213796864349076
++    },
++    "drift": {
++     "hit_rate": 0.9743438254223652,
++     "stall_ms_per_token": 1.3009010458567998
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 32,
++   "policy": "lru",
++   "predictor": "oracle",
++   "m": 8,
++   "hit_rate": 0.9722644612711023,
++   "misses_per_token": 3.550148957298908,
++   "stall_ms_per_token": 1.3869715988083386,
++   "prefetch_loads_per_token": 23.42542204568024,
++   "prefetch_precision": 1.0,
++   "loads_per_token": 26.975571002979144,
++   "pred_recall_at_k": 1.0,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.9654887419585418,
++     "stall_ms_per_token": 1.687325232308795
++    },
++    "code": {
++     "hit_rate": 0.9769538587977458,
++     "stall_ms_per_token": 1.1829392611145948
++    },
++    "reasoning": {
++     "hit_rate": 0.9724981936416185,
++     "stall_ms_per_token": 1.366709055876688
++    },
++    "multilingual": {
++     "hit_rate": 0.9747390891840607,
++     "stall_ms_per_token": 1.2623605313093011
++    },
++    "longctx": {
++     "hit_rate": 0.9736749229979466,
++     "stall_ms_per_token": 1.3395934291581142
++    },
++    "drift": {
++     "hit_rate": 0.9720232213438735,
++     "stall_ms_per_token": 1.396664031620556
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "none",
++   "m": 8,
++   "hit_rate": 0.9310189260563381,
++   "misses_per_token": 8.829577464788732,
++   "stall_ms_per_token": 3.796718309859156,
++   "prefetch_loads_per_token": 0.0,
++   "prefetch_precision": null,
++   "loads_per_token": 8.829577464788732,
++   "pred_recall_at_k": null,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.9204315370238519,
++     "stall_ms_per_token": 4.379448202207161
++    },
++    "code": {
++     "hit_rate": 0.9503332249674903,
++     "stall_ms_per_token": 2.7336592977892895
++    },
++    "reasoning": {
++     "hit_rate": 0.8983719863731656,
++     "stall_ms_per_token": 5.593605870020981
++    },
++    "multilingual": {
++     "hit_rate": 0.9266887946791227,
++     "stall_ms_per_token": 4.035048740861087
++    },
++    "longctx": {
++     "hit_rate": 0.9194146216768916,
++     "stall_ms_per_token": 4.435419222903903
++    },
++    "drift": {
++     "hit_rate": 0.9425281576830249,
++     "stall_ms_per_token": 3.1632502011263255
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "none",
++   "m": 8,
++   "hit_rate": 0.9327256082423039,
++   "misses_per_token": 8.611122144985105,
++   "stall_ms_per_token": 3.7027825223435427,
++   "prefetch_loads_per_token": 0.0,
++   "prefetch_precision": null,
++   "loads_per_token": 8.611122144985105,
++   "pred_recall_at_k": null,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.9169105164403145,
++     "stall_ms_per_token": 4.573245175125113
++    },
++    "code": {
++     "hit_rate": 0.9516427285535379,
++     "stall_ms_per_token": 2.661584220413288
++    },
++    "reasoning": {
++     "hit_rate": 0.9211524566473989,
++     "stall_ms_per_token": 4.33976878612718
++    },
++    "multilingual": {
++     "hit_rate": 0.9146110056925996,
++     "stall_ms_per_token": 4.699810246679331
++    },
++    "longctx": {
++     "hit_rate": 0.9499005390143738,
++     "stall_ms_per_token": 2.7574743326488793
++    },
++    "drift": {
++     "hit_rate": 0.9309535573122529,
++     "stall_ms_per_token": 3.8003162055336053
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "prev_rank",
++   "m": 8,
++   "hit_rate": 0.9188019610719875,
++   "misses_per_token": 10.393348982785602,
++   "stall_ms_per_token": 4.469140062597789,
++   "prefetch_loads_per_token": 20.301486697965572,
++   "prefetch_precision": 0.4753963145540811,
++   "loads_per_token": 30.694835680751172,
++   "pred_recall_at_k": null,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.9061359692061232,
++     "stall_ms_per_token": 5.166276254894943
++    },
++    "code": {
++     "hit_rate": 0.9401373537061118,
++     "stall_ms_per_token": 3.2948400520155627
++    },
++    "reasoning": {
++     "hit_rate": 0.8910835953878407,
++     "stall_ms_per_token": 5.9947589098532745
++    },
++    "multilingual": {
++     "hit_rate": 0.9138372004467912,
++     "stall_ms_per_token": 4.742400487408596
++    },
++    "longctx": {
++     "hit_rate": 0.9049505794137696,
++     "stall_ms_per_token": 5.2315201090661505
++    },
++    "drift": {
++     "hit_rate": 0.928882994770716,
++     "stall_ms_per_token": 3.914279967819814
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "prev_rank",
++   "m": 8,
++   "hit_rate": 0.9203062934458789,
++   "misses_per_token": 10.200794438927508,
++   "stall_ms_per_token": 4.386341608738748,
++   "prefetch_loads_per_token": 20.139622641509433,
++   "prefetch_precision": 0.47329960652051717,
++   "loads_per_token": 30.340417080436943,
++   "pred_recall_at_k": null,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.897812053252323,
++     "stall_ms_per_token": 5.624424588992169
++    },
++    "code": {
++     "hit_rate": 0.9410809329993738,
++     "stall_ms_per_token": 3.2429054477144827
++    },
++    "reasoning": {
++     "hit_rate": 0.9123916184971098,
++     "stall_ms_per_token": 4.821965317919084
++    },
++    "multilingual": {
++     "hit_rate": 0.9120908444022771,
++     "stall_ms_per_token": 4.838519924098684
++    },
++    "longctx": {
++     "hit_rate": 0.9331686344969199,
++     "stall_ms_per_token": 3.6783983572895407
++    },
++    "drift": {
++     "hit_rate": 0.9212265316205533,
++     "stall_ms_per_token": 4.335691699604754
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "transition",
++   "m": 8,
++   "hit_rate": 0.9404513888888889,
++   "misses_per_token": 7.622222222222222,
++   "stall_ms_per_token": 3.262892644757401,
++   "prefetch_loads_per_token": 4.385993740219092,
++   "prefetch_precision": 0.7281679838724064,
++   "loads_per_token": 12.008215962441314,
++   "pred_recall_at_k": 0.5303931924882629,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.929006096475614,
++     "stall_ms_per_token": 3.8840370238518847
++    },
++    "code": {
++     "hit_rate": 0.9579669213263979,
++     "stall_ms_per_token": 2.3106710013003733
++    },
++    "reasoning": {
++     "hit_rate": 0.9237175707547169,
++     "stall_ms_per_token": 4.18677777777779
++    },
++    "multilingual": {
++     "hit_rate": 0.934215703696182,
++     "stall_ms_per_token": 3.59804792851341
++    },
++    "longctx": {
++     "hit_rate": 0.9306194614860259,
++     "stall_ms_per_token": 3.7997709611452093
++    },
++    "drift": {
++     "hit_rate": 0.948932773531778,
++     "stall_ms_per_token": 2.7981769911504606
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "transition",
++   "m": 8,
++   "hit_rate": 0.942279046673287,
++   "misses_per_token": 7.388282025819265,
++   "stall_ms_per_token": 3.1628011916583154,
++   "prefetch_loads_per_token": 4.312611717974181,
++   "prefetch_precision": 0.7169107488256424,
++   "loads_per_token": 11.700893743793445,
++   "pred_recall_at_k": 0.5340830850711685,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.9223161633309507,
++     "stall_ms_per_token": 4.245799857040755
++    },
++    "code": {
++     "hit_rate": 0.9588828663118347,
++     "stall_ms_per_token": 2.2598810269255014
++    },
++    "reasoning": {
++     "hit_rate": 0.9390354046242775,
++     "stall_ms_per_token": 3.3480924855491416
++    },
++    "multilingual": {
++     "hit_rate": 0.9324893263757116,
++     "stall_ms_per_token": 3.6927134724857797
++    },
++    "longctx": {
++     "hit_rate": 0.95877181724846,
++     "stall_ms_per_token": 2.265782340862433
++    },
++    "drift": {
++     "hit_rate": 0.9427186264822134,
++     "stall_ms_per_token": 3.139612648221349
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "transition2",
++   "m": 8,
++   "hit_rate": 0.9383478824334899,
++   "misses_per_token": 7.891471048513302,
++   "stall_ms_per_token": 3.3680330203442703,
++   "prefetch_loads_per_token": 3.728325508607199,
++   "prefetch_precision": 0.7141747817327065,
++   "loads_per_token": 11.6197965571205,
++   "pred_recall_at_k": 0.5199705175497429,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.9262916073335706,
++     "stall_ms_per_token": 4.015260946956198
++    },
++    "code": {
++     "hit_rate": 0.9566076072821846,
++     "stall_ms_per_token": 2.383989596879038
++    },
++    "reasoning": {
++     "hit_rate": 0.9204910246331237,
++     "stall_ms_per_token": 4.3616834381551515
++    },
++    "multilingual": {
++     "hit_rate": 0.9317691409423233,
++     "stall_ms_per_token": 3.712066612510164
++    },
++    "longctx": {
++     "hit_rate": 0.9280472477845945,
++     "stall_ms_per_token": 3.928868438991166
++    },
++    "drift": {
++     "hit_rate": 0.9480025643604183,
++     "stall_ms_per_token": 2.843403057119886
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "transition2",
++   "m": 8,
++   "hit_rate": 0.9403488083416087,
++   "misses_per_token": 7.635352532274082,
++   "stall_ms_per_token": 3.26093187686189,
++   "prefetch_loads_per_token": 3.6307845084409136,
++   "prefetch_precision": 0.7060335867840928,
++   "loads_per_token": 11.266137040714995,
++   "pred_recall_at_k": 0.5253883529578663,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.9200545032165832,
++     "stall_ms_per_token": 4.351708363116534
++    },
++    "code": {
++     "hit_rate": 0.9575131105197244,
++     "stall_ms_per_token": 2.3338296806512324
++    },
++    "reasoning": {
++     "hit_rate": 0.9367473506743738,
++     "stall_ms_per_token": 3.4700809248554942
++    },
++    "multilingual": {
++     "hit_rate": 0.9301915322580645,
++     "stall_ms_per_token": 3.808740037950669
++    },
++    "longctx": {
++     "hit_rate": 0.9565419661190965,
++     "stall_ms_per_token": 2.3887761806981627
++    },
++    "drift": {
++     "hit_rate": 0.9409739377470355,
++     "stall_ms_per_token": 3.225521739130442
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "transition3",
++   "m": 8,
++   "hit_rate": 0.9371662265258216,
++   "misses_per_token": 8.042723004694835,
++   "stall_ms_per_token": 3.4264309859154753,
++   "prefetch_loads_per_token": 3.3140845070422533,
++   "prefetch_precision": 0.703782405439864,
++   "loads_per_token": 11.35680751173709,
++   "pred_recall_at_k": 0.5104414951245937,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.9251485181559274,
++     "stall_ms_per_token": 4.06746813812743
++    },
++    "code": {
++     "hit_rate": 0.9560021131339402,
++     "stall_ms_per_token": 2.4175490247073825
++    },
++    "reasoning": {
++     "hit_rate": 0.917886857966457,
++     "stall_ms_per_token": 4.495758909853259
++    },
++    "multilingual": {
++     "hit_rate": 0.9304522491876523,
++     "stall_ms_per_token": 3.7737863525588975
++    },
++    "longctx": {
++     "hit_rate": 0.926284509202454,
++     "stall_ms_per_token": 4.018560327198388
++    },
++    "drift": {
++     "hit_rate": 0.9469969328238134,
++     "stall_ms_per_token": 2.8910297666934945
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "transition3",
++   "m": 8,
++   "hit_rate": 0.9391090491559086,
++   "misses_per_token": 7.794041708043694,
++   "stall_ms_per_token": 3.3219992055610024,
++   "prefetch_loads_per_token": 3.208738828202582,
++   "prefetch_precision": 0.6996162416439713,
++   "loads_per_token": 11.002780536246275,
++   "pred_recall_at_k": 0.5172217554044763,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.9188427001429592,
++     "stall_ms_per_token": 4.406786275911391
++    },
++    "code": {
++     "hit_rate": 0.956974992172824,
++     "stall_ms_per_token": 2.362806512210409
++    },
++    "reasoning": {
++     "hit_rate": 0.934971098265896,
++     "stall_ms_per_token": 3.5673526011560726
++    },
++    "multilingual": {
++     "hit_rate": 0.9272118121442126,
++     "stall_ms_per_token": 3.9523415559772417
++    },
++    "longctx": {
++     "hit_rate": 0.9553708932238193,
++     "stall_ms_per_token": 2.447712525667362
++    },
++    "drift": {
++     "hit_rate": 0.939738759881423,
++     "stall_ms_per_token": 3.2833873517786603
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "probe",
++   "m": 8,
++   "hit_rate": 0.9027955056729264,
++   "misses_per_token": 12.442175273865415,
++   "stall_ms_per_token": 5.343785446009357,
++   "prefetch_loads_per_token": 17.129890453834115,
++   "prefetch_precision": 0.36841311894756074,
++   "loads_per_token": 29.57206572769953,
++   "pred_recall_at_k": 0.15737415232133542,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.8955116367034532,
++     "stall_ms_per_token": 5.745389106443556
++    },
++    "code": {
++     "hit_rate": 0.9188089239271782,
++     "stall_ms_per_token": 4.464062938881634
++    },
++    "reasoning": {
++     "hit_rate": 0.8817970387840671,
++     "stall_ms_per_token": 6.4974381551362885
++    },
++    "multilingual": {
++     "hit_rate": 0.894753376320065,
++     "stall_ms_per_token": 5.784663688058469
++    },
++    "longctx": {
++     "hit_rate": 0.893372955010225,
++     "stall_ms_per_token": 5.861074301295194
++    },
++    "drift": {
++     "hit_rate": 0.9128871681415929,
++     "stall_ms_per_token": 4.788305711987153
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "probe",
++   "m": 8,
++   "hit_rate": 0.9033841236345581,
++   "misses_per_token": 12.366832174776563,
++   "stall_ms_per_token": 5.311433167825154,
++   "prefetch_loads_per_token": 16.753525322740813,
++   "prefetch_precision": 0.35981696185124595,
++   "loads_per_token": 29.12035749751738,
++   "pred_recall_at_k": 0.1567345249917246,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.8891507326661902,
++     "stall_ms_per_token": 6.093732666190155
++    },
++    "code": {
++     "hit_rate": 0.9174086177207263,
++     "stall_ms_per_token": 4.542062617407669
++    },
++    "reasoning": {
++     "hit_rate": 0.8984826589595376,
++     "stall_ms_per_token": 5.579375722543363
++    },
++    "multilingual": {
++     "hit_rate": 0.8927300759013282,
++     "stall_ms_per_token": 5.895878557874778
++    },
++    "longctx": {
++     "hit_rate": 0.9118005646817249,
++     "stall_ms_per_token": 4.846349075975369
++    },
++    "drift": {
++     "hit_rate": 0.9064970355731226,
++     "stall_ms_per_token": 5.140837944664038
++    }
++   },
++   "split": "heldout"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "oracle",
++   "m": 8,
++   "hit_rate": 0.9908786922926448,
++   "misses_per_token": 1.167527386541471,
++   "stall_ms_per_token": 0.47129874804381033,
++   "prefetch_loads_per_token": 8.13262910798122,
++   "prefetch_precision": 1.0,
++   "loads_per_token": 9.300156494522692,
++   "pred_recall_at_k": 1.0,
++   "tokens": 12780,
++   "families": {
++    "general": {
++     "hit_rate": 0.9891003248487006,
++     "stall_ms_per_token": 0.5620968316126759
++    },
++    "code": {
++     "hit_rate": 0.9925674577373212,
++     "stall_ms_per_token": 0.3921425227568275
++    },
++    "reasoning": {
++     "hit_rate": 0.9900173611111112,
++     "stall_ms_per_token": 0.5243542976939221
++    },
++    "multilingual": {
++     "hit_rate": 0.9914576563769293,
++     "stall_ms_per_token": 0.4303460601137342
++    },
++    "longctx": {
++     "hit_rate": 0.9885395364689843,
++     "stall_ms_per_token": 0.5844526244035498
++    },
++    "drift": {
++     "hit_rate": 0.9919486625100563,
++     "stall_ms_per_token": 0.4178133547868083
++    }
++   },
++   "split": "fit"
++  },
++  {
++   "capacity": 48,
++   "policy": "lru",
++   "predictor": "oracle",
++   "m": 8,
++   "hit_rate": 0.9908360228401192,
++   "misses_per_token": 1.1729890764647468,
++   "stall_ms_per_token": 0.47738709036743304,
++   "prefetch_loads_per_token": 7.858987090367428,
++   "prefetch_precision": 1.0,
++   "loads_per_token": 9.031976166832175,
++   "pred_recall_at_k": 1.0,
++   "tokens": 5035,
++   "families": {
++    "general": {
++     "hit_rate": 0.9886693620443173,
++     "stall_ms_per_token": 0.5807276626161585
++    },
++    "code": {
++     "hit_rate": 0.9924516671884784,
++     "stall_ms_per_token": 0.3983882279273665
++    },
++    "reasoning": {
++     "hit_rate": 0.9911638969171483,
++     "stall_ms_per_token": 0.468088631984588
++    },
++    "multilingual": {
++     "hit_rate": 0.9900675996204934,
++     "stall_ms_per_token": 0.5177760910815962
++    },
++    "longctx": {
++     "hit_rate": 0.9914976899383984,
++     "stall_ms_per_token": 0.4448377823408645
++    },
++    "drift": {
++     "hit_rate": 0.9915544713438735,
++     "stall_ms_per_token": 0.43979841897233335
++    }
++   },
++   "split": "heldout"
++  }
++ ]
++}
+ \ No newline at end of file

added benchmarks/raw/gateC/cold_nvme_laptop/sweep.jsonl +3 −0

++{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "258ba81"}, "seed": 0, "warmup": true, "repeat": -1, "family": "reasoning", "prompt": "Q: A tank fills in 6 hours with pipe A and 9 hours with pipe", "ts": "2026-11-01T01:47:04", "prompt_tokens": 32, "gen_tokens": 64, "ttft_ms": 3041.2979170214385, "prompt_tps": 10.538982502473754, "gen_tps": 9.615565496623349, "decode_tps_wall": 9.465435344104892, "itl_ms_p50": 102.56666608620435, "itl_ms_p95": 147.70795905496925, "itl_ms_p99": 167.03716607298702, "itl_ms_mean": 105.64754431742051, "total_s": 9.69709320901893, "peak_mem_gb": 2.090173904, "text": "If pipe A fills the tank in 6 hours, then the rate of pipe A is 1 / 6 of the tank per hour. If pipe B fills the tank in 9 hours, then the rate of pipe B is 1 / 9 of the tank per hour. The combined rat", "cache_hit_rate": 0.4844649021864212, "cache_misses": 5824, "cache_hits": 5473, "stall_s": 8.191949911764823, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 20610809856, "ssd_read_s": 7.725407627178356}
++{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "258ba81"}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a C function that computes the dot product of two floa", "ts": "2026-11-01T01:47:11", "prompt_tokens": 24, "gen_tokens": 64, "ttft_ms": 1772.822791012004, "prompt_tps": 13.540661227523396, "gen_tps": 14.592826924160933, "decode_tps_wall": 14.364850714221733, "itl_ms_p50": 64.1560829244554, "itl_ms_p95": 111.17766599636525, "itl_ms_p99": 147.35658396966755, "itl_ms_mean": 69.61436772955552, "total_s": 6.158527957974002, "peak_mem_gb": 1.930492816, "text": "#include <arm_neon.h>\n\nfloat dot_product_neon(const float *a, const float *b, int len) {\n    // TODO: Implement the dot product computation using NEON intrinsics\n    // Start by loading the arrays int", "cache_hit_rate": 0.6523887973640856, "cache_misses": 3587, "cache_hits": 6732, "stall_s": 4.940353738376871, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 12694192128, "ssd_read_s": 4.6701676497468725}
++{"backend": "titan", "capacity": 16, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": true, "max_tokens": 64, "split": "heldout", "tag": "cold_nvme", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/e23844197887b031e7ddddbb0b8959c5a6853a7b --backend titan --capacity 16 --predictor none --m 8 --nocache --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 --tag cold_nvme --out benchmarks/raw/gateC/cold_nvme_laptop/sweep.jsonl", "env": {"host": "MacBookPro-de-simon-pierre-2.local", "chip": "Apple M5 Max", "ram_gb": 48, "macos": "27.0", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": "258ba81"}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Summarize the history of the Rideau Canal in one paragraph.", "ts": "2026-11-01T01:47:19", "prompt_tokens": 14, "gen_tokens": 64, "ttft_ms": 1448.2218329794705, "prompt_tps": 9.668972925355584, "gen_tps": 9.486046026204507, "decode_tps_wall": 9.337841435614223, "itl_ms_p50": 106.86812503263354, "itl_ms_p95": 156.7806249950081, "itl_ms_p99": 170.41387502104044, "itl_ms_mean": 107.09113095303081, "total_s": 8.194963083020411, "peak_mem_gb": 1.742661784, "text": "\n\nThe Rideau Canal, officially opened in 1832, was constructed by the British to connect Ottawa to Montreal via the Ottawa River. It was designed by Colonel John By and took approximately three years ", "cache_hit_rate": 0.48355739400206826, "cache_misses": 4994, "cache_hits": 4676, "stall_s": 6.844891174114309, "hits_device": 0, "layers_complete": 0, "prefetch": null, "ssd_bytes": 17673486336, "ssd_read_s": 6.502651879563928}

added benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl +48 −0

++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a Python function that parses a CSV file of timestamps", "ts": "2026-11-01T01:44:36", "prompt_tokens": 25, "gen_tokens": 64, "ttft_ms": 190.09033299516886, "prompt_tps": 248.33388555616366, "gen_tps": 121.36794961117941, "decode_tps_wall": 119.49292325368565, "itl_ms_p50": 8.360582985915244, "itl_ms_p95": 8.444542065262794, "itl_ms_p99": 8.543666917830706, "itl_ms_mean": 8.3686964279632, "total_s": 0.7173182079568505, "peak_mem_gb": 17.262479236, "text": "import csv\nfrom datetime import datetime\nimport statistics\n\ndef median_inter_arrival_time(csv_file_path):\n    \"\"\"\n    Parses a CSV file containing timestamps and returns the median inter-arrival time "}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Give five practical tips for a first-year university student", "ts": "2026-11-01T01:44:36", "prompt_tokens": 16, "gen_tokens": 64, "ttft_ms": 47.851667040959, "prompt_tps": 338.8367836252268, "gen_tps": 121.09034285180219, "decode_tps_wall": 119.20153883118209, "itl_ms_p50": 8.380957995541394, "itl_ms_p95": 8.534958004020154, "itl_ms_p99": 8.731834008358419, "itl_ms_mean": 8.389153443868198, "total_s": 0.5763683340046555, "peak_mem_gb": 17.226143856, "text": "1. **Create a Weekly Schedule:** Use a digital calendar or planner to block out time for classes, study sessions, assignments, and personal activities. Having a visual representation of your week help"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Write a short paragraph describing the climate of Gatineau, ", "ts": "2026-11-01T01:44:37", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 51.94775003474206, "prompt_tps": 370.8173313765438, "gen_tps": 120.87331925022238, "decode_tps_wall": 118.99180893614097, "itl_ms_p50": 8.380333078093827, "itl_ms_p95": 8.60504200682044, "itl_ms_p99": 8.897583931684494, "itl_ms_mean": 8.40393980846755, "total_s": 0.5813959579681978, "peak_mem_gb": 17.249301776, "text": "Include details about temperature, precipitation, and weather patterns.\n\nGatineau, Qu\u00e9bec, experiences a humid continental climate with distinct seasonal variations. In winter (December\u2013February), tem"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Explain in three sentences why mixture-of-experts models are", "ts": "2026-11-01T01:44:37", "prompt_tokens": 17, "gen_tokens": 64, "ttft_ms": 49.24083407968283, "prompt_tps": 349.63177556877565, "gen_tps": 121.55569446717989, "decode_tps_wall": 119.65929399141673, "itl_ms_p50": 8.350541931577027, "itl_ms_p95": 8.482167031615973, "itl_ms_p99": 8.48712504375726, "itl_ms_mean": 8.35706084035337, "total_s": 0.5757356670219451, "peak_mem_gb": 17.226964956, "text": "Mixture-of-Experts (MoE) models are memory-bound during decoding because they rely on storing and accessing a large number of expert parameters, which significantly increases memory requirements. Duri"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Implement an LRU cache class in Python with get and put meth", "ts": "2026-11-01T01:44:38", "prompt_tokens": 21, "gen_tokens": 64, "ttft_ms": 53.90341696329415, "prompt_tps": 394.2674782039338, "gen_tps": 121.6063822332558, "decode_tps_wall": 119.70997873886853, "itl_ms_p50": 8.343624998815358, "itl_ms_p95": 8.49016709253192, "itl_ms_p99": 8.694582968018949, "itl_ms_mean": 8.353522492735276, "total_s": 0.5801753340056166, "peak_mem_gb": 17.25364684, "text": "class LRUCache:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.cache = {}\n        self.order = []\n\n    def get(self, key):\n        if key in self.cache:\n            # "}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th Fibonacci number.\"\"", "ts": "2026-11-01T01:44:38", "prompt_tokens": 13, "gen_tokens": 64, "ttft_ms": 43.613624991849065, "prompt_tps": 302.6094225075362, "gen_tps": 121.57031801287954, "decode_tps_wall": 119.67484525660637, "itl_ms_p50": 8.331499993801117, "itl_ms_p95": 8.527584024704993, "itl_ms_p99": 8.608166011981666, "itl_ms_mean": 8.355974873882674, "total_s": 0.5700400420464575, "peak_mem_gb": 17.221640784, "text": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n == 1:\n        return 1\n    return fibonacci(n - 1) + fibonacci(n - 2)\n\n# Example usage:\nprint(f"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a Python function that parses a CSV file of timestamps", "ts": "2026-11-01T01:44:39", "prompt_tokens": 25, "gen_tokens": 64, "ttft_ms": 58.112749946303666, "prompt_tps": 434.565325733222, "gen_tps": 121.39687992141418, "decode_tps_wall": 119.50260375382749, "itl_ms_p50": 8.366625057533383, "itl_ms_p95": 8.409874979406595, "itl_ms_p99": 8.545209071598947, "itl_ms_mean": 8.368018508282683, "total_s": 0.5852979159681126, "peak_mem_gb": 17.26254994, "text": "import csv\nfrom datetime import datetime\nimport statistics\n\ndef median_inter_arrival_time(csv_file_path):\n    \"\"\"\n    Parses a CSV file containing timestamps and returns the median inter-arrival time "}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Explain in three sentences why mixture-of-experts models are", "ts": "2026-11-01T01:44:40", "prompt_tokens": 17, "gen_tokens": 64, "ttft_ms": 49.202000023797154, "prompt_tps": 349.57276418716367, "gen_tps": 121.89223393922533, "decode_tps_wall": 119.98965804328934, "itl_ms_p50": 8.334666024893522, "itl_ms_p95": 8.4102499531582, "itl_ms_p99": 8.483083103783429, "itl_ms_mean": 8.334051586672782, "total_s": 0.5742472499841824, "peak_mem_gb": 17.226956788, "text": "Mixture-of-Experts (MoE) models are memory-bound during decoding because they rely on storing and accessing a large number of expert parameters, which significantly increases memory requirements. Duri"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Write a short paragraph describing the climate of Gatineau, ", "ts": "2026-11-01T01:44:40", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 51.78712494671345, "prompt_tps": 370.81039172716237, "gen_tps": 121.22391645529108, "decode_tps_wall": 119.33175172953106, "itl_ms_p50": 8.359083090908825, "itl_ms_p95": 8.605082985013723, "itl_ms_p99": 8.688666974194348, "itl_ms_mean": 8.379999333844772, "total_s": 0.579727082978934, "peak_mem_gb": 17.249301772, "text": "Include details about temperature, precipitation, and weather patterns.\n\nGatineau, Qu\u00e9bec, experiences a humid continental climate with distinct seasonal variations. In winter (December\u2013February), tem"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Implement an LRU cache class in Python with get and put meth", "ts": "2026-11-01T01:44:41", "prompt_tokens": 21, "gen_tokens": 64, "ttft_ms": 53.91687504015863, "prompt_tps": 393.6823358259812, "gen_tps": 121.8112669918118, "decode_tps_wall": 119.91010556436554, "itl_ms_p50": 8.336875005625188, "itl_ms_p95": 8.415000047534704, "itl_ms_p99": 8.565249969251454, "itl_ms_mean": 8.339580682490672, "total_s": 0.5793104580370709, "peak_mem_gb": 17.25364682, "text": "class LRUCache:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.cache = {}\n        self.order = []\n\n    def get(self, key):\n        if key in self.cache:\n            # "}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Give five practical tips for a first-year university student", "ts": "2026-11-01T01:44:41", "prompt_tokens": 16, "gen_tokens": 64, "ttft_ms": 47.72816691547632, "prompt_tps": 339.2244104269285, "gen_tps": 121.67545188769039, "decode_tps_wall": 119.7779166732931, "itl_ms_p50": 8.335041929967701, "itl_ms_p95": 8.467959007248282, "itl_ms_p99": 8.659999934025109, "itl_ms_mean": 8.348784381745471, "total_s": 0.573701582965441, "peak_mem_gb": 17.22614392, "text": "1. **Create a Weekly Schedule:** Use a digital calendar or planner to block out time for classes, study sessions, assignments, and personal activities. Having a visual representation of your week help"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th Fibonacci number.\"\"", "ts": "2026-11-01T01:44:42", "prompt_tokens": 13, "gen_tokens": 64, "ttft_ms": 43.5285420389846, "prompt_tps": 302.6079571144651, "gen_tps": 121.40749208551493, "decode_tps_wall": 119.5127483416421, "itl_ms_p50": 8.359082974493504, "itl_ms_p95": 8.43141705263406, "itl_ms_p99": 8.60049994662404, "itl_ms_mean": 8.367308206664074, "total_s": 0.5706689590588212, "peak_mem_gb": 17.221640784, "text": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n == 1:\n        return 1\n    return fibonacci(n - 1) + fibonacci(n - 2)\n\n# Example usage:\nprint(f"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Implement an LRU cache class in Python with get and put meth", "ts": "2026-11-01T01:44:42", "prompt_tokens": 21, "gen_tokens": 64, "ttft_ms": 53.48683404736221, "prompt_tps": 396.8369599269693, "gen_tps": 121.25802358542144, "decode_tps_wall": 119.36565680352243, "itl_ms_p50": 8.361416985280812, "itl_ms_p95": 8.532540989108384, "itl_ms_p99": 8.71666707098484, "itl_ms_mean": 8.377619047043106, "total_s": 0.5812768340110779, "peak_mem_gb": 17.253646828, "text": "class LRUCache:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.cache = {}\n        self.order = []\n\n    def get(self, key):\n        if key in self.cache:\n            # "}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th Fibonacci number.\"\"", "ts": "2026-11-01T01:44:43", "prompt_tokens": 13, "gen_tokens": 64, "ttft_ms": 43.41387492604554, "prompt_tps": 303.4981079859437, "gen_tps": 121.88605345378718, "decode_tps_wall": 119.9840021389397, "itl_ms_p50": 8.336500031873584, "itl_ms_p95": 8.409750065766275, "itl_ms_p99": 8.488874998874962, "itl_ms_mean": 8.334444444035254, "total_s": 0.5684838749002665, "peak_mem_gb": 17.221640784, "text": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n == 1:\n        return 1\n    return fibonacci(n - 1) + fibonacci(n - 2)\n\n# Example usage:\nprint(f"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Explain in three sentences why mixture-of-experts models are", "ts": "2026-11-01T01:44:44", "prompt_tokens": 17, "gen_tokens": 64, "ttft_ms": 49.159541027620435, "prompt_tps": 350.5781123763205, "gen_tps": 121.71373900696982, "decode_tps_wall": 119.81625321442641, "itl_ms_p50": 8.33791692275554, "itl_ms_p95": 8.414624957367778, "itl_ms_p99": 8.695708005689085, "itl_ms_mean": 8.346113095444347, "total_s": 0.5749646660406142, "peak_mem_gb": 17.226960852, "text": "Mixture-of-Experts (MoE) models are memory-bound during decoding because they rely on storing and accessing a large number of expert parameters, which significantly increases memory requirements. Duri"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Give five practical tips for a first-year university student", "ts": "2026-11-01T01:44:44", "prompt_tokens": 16, "gen_tokens": 64, "ttft_ms": 47.806749935261905, "prompt_tps": 338.7148422536113, "gen_tps": 121.40541932367935, "decode_tps_wall": 119.51012231876976, "itl_ms_p50": 8.363333996385336, "itl_ms_p95": 8.515625027939677, "itl_ms_p99": 8.595749968662858, "itl_ms_mean": 8.367492063414483, "total_s": 0.5749587499303743, "peak_mem_gb": 17.226143856, "text": "1. **Create a Weekly Schedule:** Use a digital calendar or planner to block out time for classes, study sessions, assignments, and personal activities. Having a visual representation of your week help"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Write a short paragraph describing the climate of Gatineau, ", "ts": "2026-11-01T01:44:45", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 51.81904206983745, "prompt_tps": 370.7121119944423, "gen_tps": 121.67745683114906, "decode_tps_wall": 119.77846688130617, "itl_ms_p50": 8.337999926880002, "itl_ms_p95": 8.509374922141433, "itl_ms_p99": 8.653583005070686, "itl_ms_mean": 8.34874603121231, "total_s": 0.577790042036213, "peak_mem_gb": 17.249301776, "text": "Include details about temperature, precipitation, and weather patterns.\n\nGatineau, Qu\u00e9bec, experiences a humid continental climate with distinct seasonal variations. In winter (December\u2013February), tem"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a Python function that parses a CSV file of timestamps", "ts": "2026-11-01T01:44:45", "prompt_tokens": 25, "gen_tokens": 64, "ttft_ms": 58.237458928488195, "prompt_tps": 433.6403767509204, "gen_tps": 121.9397082106601, "decode_tps_wall": 120.03663974946353, "itl_ms_p50": 8.328375057317317, "itl_ms_p95": 8.410041918978095, "itl_ms_p99": 8.536917041055858, "itl_ms_mean": 8.330789682943196, "total_s": 0.5830772089539096, "peak_mem_gb": 17.262549952, "text": "import csv\nfrom datetime import datetime\nimport statistics\n\ndef median_inter_arrival_time(csv_file_path):\n    \"\"\"\n    Parses a CSV file containing timestamps and returns the median inter-arrival time "}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "code", "prompt": "Implement an LRU cache class in Python with get and put meth", "ts": "2026-11-01T01:44:46", "prompt_tokens": 21, "gen_tokens": 64, "ttft_ms": 53.03687497507781, "prompt_tps": 400.2061405267773, "gen_tps": 122.84042573142314, "decode_tps_wall": 120.92373271713936, "itl_ms_p50": 8.24875000398606, "itl_ms_p95": 8.425749954767525, "itl_ms_p99": 8.515499997884035, "itl_ms_mean": 8.26967525340262, "total_s": 0.5740264159394428, "peak_mem_gb": 17.253640676, "text": "class LRUCache:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.cache = {}\n        self.order = []\n\n    def get(self, key):\n        if key in self.cache:\n            # "}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "general", "prompt": "Explain in three sentences why mixture-of-experts models are", "ts": "2026-11-01T01:44:47", "prompt_tokens": 17, "gen_tokens": 64, "ttft_ms": 48.9684579661116, "prompt_tps": 351.7829711443551, "gen_tps": 121.93942754370109, "decode_tps_wall": 120.03831691113533, "itl_ms_p50": 8.33029206842184, "itl_ms_p95": 8.438290911726654, "itl_ms_p99": 8.636665996164083, "itl_ms_mean": 8.330673286099993, "total_s": 0.5738008749904111, "peak_mem_gb": 17.226956788, "text": "Mixture-of-Experts (MoE) models are memory-bound during decoding because they rely on storing and accessing a large number of expert parameters, which significantly increases memory requirements. Duri"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "code", "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th Fibonacci number.\"\"", "ts": "2026-11-01T01:44:47", "prompt_tokens": 13, "gen_tokens": 64, "ttft_ms": 43.77604101318866, "prompt_tps": 300.7414438468318, "gen_tps": 121.05787559700309, "decode_tps_wall": 119.16836550102333, "itl_ms_p50": 8.376457961276174, "itl_ms_p95": 8.59345798380673, "itl_ms_p99": 8.693208917975426, "itl_ms_mean": 8.39148876294198, "total_s": 0.5724398330785334, "peak_mem_gb": 17.221638736, "text": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n == 1:\n        return 1\n    return fibonacci(n - 1) + fibonacci(n - 2)\n\n# Example usage:\nprint(f"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "general", "prompt": "Write a short paragraph describing the climate of Gatineau, ", "ts": "2026-11-01T01:44:48", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 51.64116702508181, "prompt_tps": 372.6351792877575, "gen_tps": 121.64076273747433, "decode_tps_wall": 119.74421314955649, "itl_ms_p50": 8.346999995410442, "itl_ms_p95": 8.42333398759365, "itl_ms_p99": 8.572999970056117, "itl_ms_mean": 8.351134252734482, "total_s": 0.5777626249473542, "peak_mem_gb": 17.249301776, "text": "Include details about temperature, precipitation, and weather patterns.\n\nGatineau, Qu\u00e9bec, experiences a humid continental climate with distinct seasonal variations. In winter (December\u2013February), tem"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "general", "prompt": "Give five practical tips for a first-year university student", "ts": "2026-11-01T01:44:48", "prompt_tokens": 16, "gen_tokens": 64, "ttft_ms": 47.79970902018249, "prompt_tps": 338.6786920728613, "gen_tps": 121.35481265622165, "decode_tps_wall": 119.46146588396937, "itl_ms_p50": 8.368958020582795, "itl_ms_p95": 8.484416990540922, "itl_ms_p99": 8.614916005171835, "itl_ms_mean": 8.370900127504553, "total_s": 0.5751664170529693, "peak_mem_gb": 17.226143856, "text": "1. **Create a Weekly Schedule:** Use a digital calendar or planner to block out time for classes, study sessions, assignments, and personal activities. Having a visual representation of your week help"}
++{"backend": "mlx_lm", "capacity": null, "policy": null, "predictor": null, "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend mlx_lm --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "code", "prompt": "Write a Python function that parses a CSV file of timestamps", "ts": "2026-11-01T01:44:49", "prompt_tokens": 25, "gen_tokens": 64, "ttft_ms": 58.15908405929804, "prompt_tps": 434.2721836914425, "gen_tps": 121.31440349292212, "decode_tps_wall": 119.42142230634539, "itl_ms_p50": 8.356499951332808, "itl_ms_p95": 8.525499957613647, "itl_ms_p99": 8.59662494622171, "itl_ms_mean": 8.373707000698362, "total_s": 0.5857026251032948, "peak_mem_gb": 17.26254994, "text": "import csv\nfrom datetime import datetime\nimport statistics\n\ndef median_inter_arrival_time(csv_file_path):\n    \"\"\"\n    Parses a CSV file containing timestamps and returns the median inter-arrival time "}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Write a Python function that parses a CSV file of timestamps", "ts": "2026-11-01T01:44:54", "prompt_tokens": 25, "gen_tokens": 64, "ttft_ms": 1220.5764999380335, "prompt_tps": 27.00275251582446, "gen_tps": 119.63154427799184, "decode_tps_wall": 117.78116172783089, "itl_ms_p50": 8.376332931220531, "itl_ms_p95": 8.598791086114943, "itl_ms_p99": 14.143792097456753, "itl_ms_mean": 8.490322096761139, "total_s": 1.7554667920339853, "peak_mem_gb": 33.489018536, "text": "import csv\nfrom datetime import datetime\nimport statistics\n\ndef median_inter_arrival_time(csv_file_path):\n    \"\"\"\n    Parses a CSV file containing timestamps and returns the median inter-arrival time ", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 34176, "stall_s": 0.0, "hits_device": 34176, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Give five practical tips for a first-year university student", "ts": "2026-11-01T01:44:55", "prompt_tokens": 16, "gen_tokens": 64, "ttft_ms": 48.77100000157952, "prompt_tps": 332.56485232655496, "gen_tps": 121.4452169164887, "decode_tps_wall": 119.55078327663087, "itl_ms_p50": 8.360707899555564, "itl_ms_p95": 8.474875008687377, "itl_ms_p99": 8.506624959409237, "itl_ms_mean": 8.364646157826341, "total_s": 0.575743707944639, "peak_mem_gb": 17.218282524, "text": "1. **Create a Weekly Schedule:** Use a digital calendar or planner to block out time for classes, study sessions, assignments, and personal activities. Having a visual representation of your week help", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 30720, "stall_s": 0.0, "hits_device": 30720, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Write a short paragraph describing the climate of Gatineau, ", "ts": "2026-11-01T01:44:55", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 53.07695805095136, "prompt_tps": 361.7012308418058, "gen_tps": 121.23609680837703, "decode_tps_wall": 119.34598402398768, "itl_ms_p50": 8.373667020350695, "itl_ms_p95": 8.46533291041851, "itl_ms_p99": 8.867958909831941, "itl_ms_mean": 8.378999998852137, "total_s": 0.580953957978636, "peak_mem_gb": 17.239856048, "text": "Include details about temperature, precipitation, and weather patterns.\n\nGatineau, Qu\u00e9bec, experiences a humid continental climate with distinct seasonal variations. In winter (December\u2013February), tem", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 31872, "stall_s": 0.0, "hits_device": 31872, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "general", "prompt": "Explain in three sentences why mixture-of-experts models are", "ts": "2026-11-01T01:44:56", "prompt_tokens": 17, "gen_tokens": 64, "ttft_ms": 50.045416108332574, "prompt_tps": 343.29477135842853, "gen_tps": 121.54620061293649, "decode_tps_wall": 119.64870752061906, "itl_ms_p50": 8.346708957105875, "itl_ms_p95": 8.540916023775935, "itl_ms_p99": 8.558625006116927, "itl_ms_mean": 8.357800269824645, "total_s": 0.5765868331072852, "peak_mem_gb": 17.21856966, "text": "Mixture-of-Experts (MoE) models are memory-bound during decoding because they rely on storing and accessing a large number of expert parameters, which significantly increases memory requirements. Duri", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 31104, "stall_s": 0.0, "hits_device": 31104, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "Implement an LRU cache class in Python with get and put meth", "ts": "2026-11-01T01:44:57", "prompt_tokens": 21, "gen_tokens": 64, "ttft_ms": 55.06129201967269, "prompt_tps": 385.90881583195323, "gen_tps": 117.97438897158018, "decode_tps_wall": 116.13470519492213, "itl_ms_p50": 8.571584010496736, "itl_ms_p95": 8.948875009082258, "itl_ms_p99": 9.198999963700771, "itl_ms_mean": 8.61069047638762, "total_s": 0.5975347920320928, "peak_mem_gb": 17.243156824, "text": "class LRUCache:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.cache = {}\n        self.order = []\n\n    def get(self, key):\n        if key in self.cache:\n            # ", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 32640, "stall_s": 0.0, "hits_device": 32640, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": true, "repeat": -1, "family": "code", "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th Fibonacci number.\"\"", "ts": "2026-11-01T01:44:57", "prompt_tokens": 13, "gen_tokens": 64, "ttft_ms": 44.345750007778406, "prompt_tps": 296.6154511429461, "gen_tps": 120.9117500982621, "decode_tps_wall": 119.02583968272118, "itl_ms_p50": 8.385625085793436, "itl_ms_p95": 8.570082951337099, "itl_ms_p99": 8.722792030312121, "itl_ms_mean": 8.401537033182288, "total_s": 0.5736425830982625, "peak_mem_gb": 17.215356644, "text": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n == 1:\n        return 1\n    return fibonacci(n - 1) + fibonacci(n - 2)\n\n# Example usage:\nprint(f", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 29568, "stall_s": 0.0, "hits_device": 29568, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Write a Python function that parses a CSV file of timestamps", "ts": "2026-11-01T01:44:58", "prompt_tokens": 25, "gen_tokens": 64, "ttft_ms": 60.059874900616705, "prompt_tps": 420.1707153604413, "gen_tps": 121.58243342111969, "decode_tps_wall": 119.68656369037812, "itl_ms_p50": 8.351624943315983, "itl_ms_p95": 8.393040974624455, "itl_ms_p99": 8.665792061947286, "itl_ms_mean": 8.355156745805981, "total_s": 0.5864347498863935, "peak_mem_gb": 17.250497668, "text": "import csv\nfrom datetime import datetime\nimport statistics\n\ndef median_inter_arrival_time(csv_file_path):\n    \"\"\"\n    Parses a CSV file containing timestamps and returns the median inter-arrival time ", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 34176, "stall_s": 0.0, "hits_device": 34176, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Explain in three sentences why mixture-of-experts models are", "ts": "2026-11-01T01:44:58", "prompt_tokens": 17, "gen_tokens": 64, "ttft_ms": 49.783666036091745, "prompt_tps": 344.82525692735203, "gen_tps": 122.00261589794333, "decode_tps_wall": 120.09875733097927, "itl_ms_p50": 8.330084034241736, "itl_ms_p95": 8.390332921408117, "itl_ms_p99": 8.473166031762958, "itl_ms_mean": 8.326480824810764, "total_s": 0.5743519579991698, "peak_mem_gb": 17.218569684, "text": "Mixture-of-Experts (MoE) models are memory-bound during decoding because they rely on storing and accessing a large number of expert parameters, which significantly increases memory requirements. Duri", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 31104, "stall_s": 0.0, "hits_device": 31104, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Write a short paragraph describing the climate of Gatineau, ", "ts": "2026-11-01T01:44:59", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 52.693875040858984, "prompt_tps": 363.86588147938994, "gen_tps": 121.39097933124957, "decode_tps_wall": 119.49697464663882, "itl_ms_p50": 8.35495803039521, "itl_ms_p95": 8.45249998383224, "itl_ms_p99": 8.801499963738024, "itl_ms_mean": 8.368412697953838, "total_s": 0.5799038750119507, "peak_mem_gb": 17.239863212, "text": "Include details about temperature, precipitation, and weather patterns.\n\nGatineau, Qu\u00e9bec, experiences a humid continental climate with distinct seasonal variations. In winter (December\u2013February), tem", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 31872, "stall_s": 0.0, "hits_device": 31872, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "Implement an LRU cache class in Python with get and put meth", "ts": "2026-11-01T01:44:59", "prompt_tokens": 21, "gen_tokens": 64, "ttft_ms": 55.027499911375344, "prompt_tps": 385.0664480922441, "gen_tps": 121.53900657171316, "decode_tps_wall": 119.6423643508119, "itl_ms_p50": 8.351791999302804, "itl_ms_p95": 8.486833074130118, "itl_ms_p99": 8.650125004351139, "itl_ms_mean": 8.3582433816489, "total_s": 0.5815968329552561, "peak_mem_gb": 17.243156804, "text": "class LRUCache:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.cache = {}\n        self.order = []\n\n    def get(self, key):\n        if key in self.cache:\n            # ", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 32640, "stall_s": 0.0, "hits_device": 32640, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "general", "prompt": "Give five practical tips for a first-year university student", "ts": "2026-11-01T01:45:00", "prompt_tokens": 16, "gen_tokens": 64, "ttft_ms": 48.99120796471834, "prompt_tps": 330.0295995197668, "gen_tps": 119.90046397486195, "decode_tps_wall": 118.03147842516533, "itl_ms_p50": 8.421833044849336, "itl_ms_p95": 8.914500009268522, "itl_ms_p99": 9.067249950021505, "itl_ms_mean": 8.472316142629891, "total_s": 0.5827471249504015, "peak_mem_gb": 17.218282524, "text": "1. **Create a Weekly Schedule:** Use a digital calendar or planner to block out time for classes, study sessions, assignments, and personal activities. Having a visual representation of your week help", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 30720, "stall_s": 0.0, "hits_device": 30720, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 0, "family": "code", "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th Fibonacci number.\"\"", "ts": "2026-11-01T01:45:01", "prompt_tokens": 13, "gen_tokens": 64, "ttft_ms": 44.39308401197195, "prompt_tps": 296.37510710304804, "gen_tps": 121.82079258300936, "decode_tps_wall": 119.92054796815864, "itl_ms_p50": 8.337458013556898, "itl_ms_p95": 8.392707910388708, "itl_ms_p99": 8.558874949812889, "itl_ms_mean": 8.338854491104565, "total_s": 0.5697409169515595, "peak_mem_gb": 17.215356624, "text": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n == 1:\n        return 1\n    return fibonacci(n - 1) + fibonacci(n - 2)\n\n# Example usage:\nprint(f", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 29568, "stall_s": 0.0, "hits_device": 29568, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Implement an LRU cache class in Python with get and put meth", "ts": "2026-11-01T01:45:01", "prompt_tokens": 21, "gen_tokens": 64, "ttft_ms": 54.63941604830325, "prompt_tps": 387.86446016558983, "gen_tps": 121.27424188755138, "decode_tps_wall": 119.3818012321434, "itl_ms_p50": 8.351875003427267, "itl_ms_p95": 8.4734580013901, "itl_ms_p99": 8.981000049971044, "itl_ms_mean": 8.376486111609708, "total_s": 0.5823580410797149, "peak_mem_gb": 17.243156812, "text": "class LRUCache:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.cache = {}\n        self.order = []\n\n    def get(self, key):\n        if key in self.cache:\n            # ", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 32640, "stall_s": 0.0, "hits_device": 32640, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th Fibonacci number.\"\"", "ts": "2026-11-01T01:45:02", "prompt_tokens": 13, "gen_tokens": 64, "ttft_ms": 44.173542060889304, "prompt_tps": 297.75367736190805, "gen_tps": 121.6031860206553, "decode_tps_wall": 119.70498422740722, "itl_ms_p50": 8.344833040609956, "itl_ms_p95": 8.480250020511448, "itl_ms_p99": 8.510250016115606, "itl_ms_mean": 8.353871030969517, "total_s": 0.5704674170119688, "peak_mem_gb": 17.215356644, "text": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n == 1:\n        return 1\n    return fibonacci(n - 1) + fibonacci(n - 2)\n\n# Example usage:\nprint(f", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 29568, "stall_s": 0.0, "hits_device": 29568, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Explain in three sentences why mixture-of-experts models are", "ts": "2026-11-01T01:45:02", "prompt_tokens": 17, "gen_tokens": 64, "ttft_ms": 50.10233307257295, "prompt_tps": 342.8398592472836, "gen_tps": 121.59208703388818, "decode_tps_wall": 119.69455089682388, "itl_ms_p50": 8.344458998180926, "itl_ms_p95": 8.48733400925994, "itl_ms_p99": 8.629083051346242, "itl_ms_mean": 8.354599206959682, "total_s": 0.5764420831110328, "peak_mem_gb": 17.218569652, "text": "Mixture-of-Experts (MoE) models are memory-bound during decoding because they rely on storing and accessing a large number of expert parameters, which significantly increases memory requirements. Duri", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 31104, "stall_s": 0.0, "hits_device": 31104, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Give five practical tips for a first-year university student", "ts": "2026-11-01T01:45:03", "prompt_tokens": 16, "gen_tokens": 64, "ttft_ms": 48.36112505290657, "prompt_tps": 334.16526579787626, "gen_tps": 121.67285920481129, "decode_tps_wall": 119.77398856188454, "itl_ms_p50": 8.340667001903057, "itl_ms_p95": 8.455541916191578, "itl_ms_p99": 8.47962498664856, "itl_ms_mean": 8.349058188734547, "total_s": 0.574351790943183, "peak_mem_gb": 17.218290684, "text": "1. **Create a Weekly Schedule:** Use a digital calendar or planner to block out time for classes, study sessions, assignments, and personal activities. Having a visual representation of your week help", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 30720, "stall_s": 0.0, "hits_device": 30720, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "general", "prompt": "Write a short paragraph describing the climate of Gatineau, ", "ts": "2026-11-01T01:45:04", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 52.854583016596735, "prompt_tps": 362.84210416888635, "gen_tps": 121.97403551088667, "decode_tps_wall": 120.07064158084249, "itl_ms_p50": 8.331332937814295, "itl_ms_p95": 8.383834036067128, "itl_ms_p99": 8.41845793183893, "itl_ms_mean": 8.328430554164308, "total_s": 0.5775457079289481, "peak_mem_gb": 17.239863216, "text": "Include details about temperature, precipitation, and weather patterns.\n\nGatineau, Qu\u00e9bec, experiences a humid continental climate with distinct seasonal variations. In winter (December\u2013February), tem", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 31872, "stall_s": 0.0, "hits_device": 31872, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 1, "family": "code", "prompt": "Write a Python function that parses a CSV file of timestamps", "ts": "2026-11-01T01:45:04", "prompt_tokens": 25, "gen_tokens": 64, "ttft_ms": 59.81308396439999, "prompt_tps": 421.45989484963627, "gen_tps": 121.37909403443791, "decode_tps_wall": 119.48395266974292, "itl_ms_p50": 8.344041998498142, "itl_ms_p95": 8.558999979868531, "itl_ms_p99": 8.64954199641943, "itl_ms_mean": 8.369324730694412, "total_s": 0.587080541998148, "peak_mem_gb": 17.25049768, "text": "import csv\nfrom datetime import datetime\nimport statistics\n\ndef median_inter_arrival_time(csv_file_path):\n    \"\"\"\n    Parses a CSV file containing timestamps and returns the median inter-arrival time ", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 34176, "stall_s": 0.0, "hits_device": 34176, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "code", "prompt": "Implement an LRU cache class in Python with get and put meth", "ts": "2026-11-01T01:45:05", "prompt_tokens": 21, "gen_tokens": 64, "ttft_ms": 55.06124999374151, "prompt_tps": 385.09145933521063, "gen_tps": 121.23431697523493, "decode_tps_wall": 119.34340303931936, "itl_ms_p50": 8.368332986719906, "itl_ms_p95": 8.470666012726724, "itl_ms_p99": 8.576125022955239, "itl_ms_mean": 8.379181207615941, "total_s": 0.5829496660735458, "peak_mem_gb": 17.243156804, "text": "class LRUCache:\n    def __init__(self, capacity):\n        self.capacity = capacity\n        self.cache = {}\n        self.order = []\n\n    def get(self, key):\n        if key in self.cache:\n            # ", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 32640, "stall_s": 0.0, "hits_device": 32640, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "general", "prompt": "Explain in three sentences why mixture-of-experts models are", "ts": "2026-11-01T01:45:05", "prompt_tokens": 17, "gen_tokens": 64, "ttft_ms": 49.646333092823625, "prompt_tps": 345.721111045929, "gen_tps": 122.00395295728767, "decode_tps_wall": 120.09996893673218, "itl_ms_p50": 8.340833010151982, "itl_ms_p95": 8.404958061873913, "itl_ms_p99": 8.497833972796798, "itl_ms_mean": 8.326396824688548, "total_s": 0.5742093330482021, "peak_mem_gb": 17.218569684, "text": "Mixture-of-Experts (MoE) models are memory-bound during decoding because they rely on storing and accessing a large number of expert parameters, which significantly increases memory requirements. Duri", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 31104, "stall_s": 0.0, "hits_device": 31104, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "code", "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th Fibonacci number.\"\"", "ts": "2026-11-01T01:45:06", "prompt_tokens": 13, "gen_tokens": 64, "ttft_ms": 43.86004095431417, "prompt_tps": 299.62950087647863, "gen_tps": 121.56949062605008, "decode_tps_wall": 119.67134054503336, "itl_ms_p50": 8.344084024429321, "itl_ms_p95": 8.46070796251297, "itl_ms_p99": 8.677207981236279, "itl_ms_mean": 8.356219588128466, "total_s": 0.5703018750064075, "peak_mem_gb": 17.215356608, "text": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n == 1:\n        return 1\n    return fibonacci(n - 1) + fibonacci(n - 2)\n\n# Example usage:\nprint(f", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 29568, "stall_s": 0.0, "hits_device": 29568, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "general", "prompt": "Write a short paragraph describing the climate of Gatineau, ", "ts": "2026-11-01T01:45:06", "prompt_tokens": 19, "gen_tokens": 64, "ttft_ms": 53.19029197562486, "prompt_tps": 360.89625262896436, "gen_tps": 121.45359072568283, "decode_tps_wall": 119.55926290013397, "itl_ms_p50": 8.34145792759955, "itl_ms_p95": 8.495041984133422, "itl_ms_p99": 8.691542083397508, "itl_ms_mean": 8.364052903498449, "total_s": 0.5801256248960271, "peak_mem_gb": 17.239863216, "text": "Include details about temperature, precipitation, and weather patterns.\n\nGatineau, Qu\u00e9bec, experiences a humid continental climate with distinct seasonal variations. In winter (December\u2013February), tem", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 31872, "stall_s": 0.0, "hits_device": 31872, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "general", "prompt": "Give five practical tips for a first-year university student", "ts": "2026-11-01T01:45:07", "prompt_tokens": 16, "gen_tokens": 64, "ttft_ms": 48.36929100565612, "prompt_tps": 334.04317433962683, "gen_tps": 121.9046360404261, "decode_tps_wall": 120.0016190010754, "itl_ms_p50": 8.329583099111915, "itl_ms_p95": 8.395583950914443, "itl_ms_p99": 8.49329202901572, "itl_ms_mean": 8.333220904219955, "total_s": 0.5733622079715133, "peak_mem_gb": 17.2182825, "text": "1. **Create a Weekly Schedule:** Use a digital calendar or planner to block out time for classes, study sessions, assignments, and personal activities. Having a visual representation of your week help", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 30720, "stall_s": 0.0, "hits_device": 30720, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}
++{"backend": "titan", "capacity": 128, "policy": "lru", "predictor": "none", "m": 8, "prefetch_workers": 2, "nocache": false, "max_tokens": 64, "split": "fit", "tag": "", "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa", "cmd": "benchmarks/scripts/bench_decode.py --model /Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa --backend titan --capacity 128 --families general,code --repeats 3 --out benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl", "env": {"host": "MacStudnpierre2", "chip": "Apple M3 Ultra", "ram_gb": 96, "macos": "26.6", "python": "3.12.13", "mlx": "0.32.4", "mlx_lm": "0.32.2", "numpy": "2.5.4", "git": ""}, "seed": 0, "warmup": false, "repeat": 2, "family": "code", "prompt": "Write a Python function that parses a CSV file of timestamps", "ts": "2026-11-01T01:45:08", "prompt_tokens": 25, "gen_tokens": 64, "ttft_ms": 59.67658292502165, "prompt_tps": 422.439421145433, "gen_tps": 121.7417052208992, "decode_tps_wall": 119.84254969519033, "itl_ms_p50": 8.343541994690895, "itl_ms_p95": 8.473082911223173, "itl_ms_p99": 8.57179204467684, "itl_ms_mean": 8.344281747538064, "total_s": 0.5853663330199197, "peak_mem_gb": 17.250497668, "text": "import csv\nfrom datetime import datetime\nimport statistics\n\ndef median_inter_arrival_time(csv_file_path):\n    \"\"\"\n    Parses a CSV file containing timestamps and returns the median inter-arrival time ", "cache_hit_rate": 1.0, "cache_misses": 0, "cache_hits": 34176, "stall_s": 0.0, "hits_device": 34176, "layers_complete": 48, "prefetch": null, "ssd_bytes": 0, "ssd_read_s": 0.0}

added benchmarks/raw/tests/m3u96a/greedy_qwen3_moe_cap32_20261101T014433.json +14 −0

++[
++  {
++    "prompt": "The capital of Canada is",
++    "ref": "Ottawa, located in the province of Ontario. It is situated on the Ottawa River, which forms the border between Ontario and Quebec. Ottawa was chosen as the capital",
++    "titan": "Ottawa, located in the province of Ontario. It is situated on the Ottawa River, which forms the border between Ontario and Quebec. Ottawa was chosen as the capital",
++    "equal": true
++  },
++  {
++    "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th",
++    "ref": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n",
++    "titan": "   if n < 0:\n        raise ValueError(\"n must be non-negative\")\n    if n == 0:\n        return 0\n    if n",
++    "equal": true
++  }
++]
+ \ No newline at end of file

added benchmarks/raw/tests/m3u96a/numerical_qwen3_moe_cap128_2026-11-01T014417.json +50 −0

++{
++  "capacity": 128,
++  "prompts": [
++    {
++      "prompt": "The capital of Canada is",
++      "tokens": 5,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.0,
++      "misses": 1207,
++      "stall_s": 0.23856342188082635,
++      "layers_routed": 48
++    },
++    {
++      "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th",
++      "tokens": 13,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.3070747616658304,
++      "misses": 1381,
++      "stall_s": 0.27579883276484907,
++      "layers_routed": 48
++    },
++    {
++      "prompt": "Q: If a train travels 60 km in 45 minute",
++      "tokens": 27,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.6245239908606245,
++      "misses": 986,
++      "stall_s": 0.19227600295562297,
++      "layers_routed": 48
++    },
++    {
++      "prompt": "Les mod\u00e8les \u00e0 m\u00e9lange d'experts activent",
++      "tokens": 22,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.6931376780319378,
++      "misses": 711,
++      "stall_s": 0.14171795325819403,
++      "layers_routed": 48
++    }
++  ],
++  "tol": 0.001,
++  "ts": "2026-11-01T01:44:17",
++  "max_abs_dlogit_worst": 0.0,
++  "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa",
++  "model_type": "qwen3_moe"
++}
+ \ No newline at end of file

added benchmarks/raw/tests/m3u96a/numerical_qwen3_moe_cap16_2026-11-01T014425.json +50 −0

++{
++  "capacity": 16,
++  "prompts": [
++    {
++      "prompt": "The capital of Canada is",
++      "tokens": 5,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.20887728459530025,
++      "misses": 1212,
++      "stall_s": 0.2369246652815491,
++      "layers_routed": 111
++    },
++    {
++      "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th",
++      "tokens": 13,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.3497467642093416,
++      "misses": 2311,
++      "stall_s": 0.45020441873930395,
++      "layers_routed": 254
++    },
++    {
++      "prompt": "Q: If a train travels 60 km in 45 minute",
++      "tokens": 27,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.38492431263515603,
++      "misses": 3982,
++      "stall_s": 0.7945379961747676,
++      "layers_routed": 452
++    },
++    {
++      "prompt": "Les mod\u00e8les \u00e0 m\u00e9lange d'experts activent",
++      "tokens": 22,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.4454260997607215,
++      "misses": 3013,
++      "stall_s": 0.5720772070344537,
++      "layers_routed": 388
++    }
++  ],
++  "tol": 0.001,
++  "ts": "2026-11-01T01:44:25",
++  "max_abs_dlogit_worst": 0.0,
++  "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa",
++  "model_type": "qwen3_moe"
++}
+ \ No newline at end of file

added benchmarks/raw/tests/m3u96a/numerical_qwen3_moe_cap32_2026-11-01T014422.json +50 −0

++{
++  "capacity": 32,
++  "prompts": [
++    {
++      "prompt": "The capital of Canada is",
++      "tokens": 5,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.0008278145695364238,
++      "misses": 1207,
++      "stall_s": 0.236716914572753,
++      "layers_routed": 49
++    },
++    {
++      "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th",
++      "tokens": 13,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.4050682261208577,
++      "misses": 1526,
++      "stall_s": 0.2965041723800823,
++      "layers_routed": 96
++    },
++    {
++      "prompt": "Q: If a train travels 60 km in 45 minute",
++      "tokens": 27,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.4854474370112945,
++      "misses": 2369,
++      "stall_s": 0.4713788740336895,
++      "layers_routed": 164
++    },
++    {
++      "prompt": "Les mod\u00e8les \u00e0 m\u00e9lange d'experts activent",
++      "tokens": 22,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.452463768115942,
++      "misses": 1889,
++      "stall_s": 0.36406895343679935,
++      "layers_routed": 124
++    }
++  ],
++  "tol": 0.001,
++  "ts": "2026-11-01T01:44:22",
++  "max_abs_dlogit_worst": 0.0,
++  "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa",
++  "model_type": "qwen3_moe"
++}
+ \ No newline at end of file

added benchmarks/raw/tests/m3u96a/numerical_qwen3_moe_cap64_2026-11-01T014420.json +50 −0

++{
++  "capacity": 64,
++  "prompts": [
++    {
++      "prompt": "The capital of Canada is",
++      "tokens": 5,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.0,
++      "misses": 1207,
++      "stall_s": 0.2353273737244308,
++      "layers_routed": 48
++    },
++    {
++      "prompt": "def fibonacci(n):\n    \"\"\"Return the n-th",
++      "tokens": 13,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.3070747616658304,
++      "misses": 1381,
++      "stall_s": 0.2662418349646032,
++      "layers_routed": 48
++    },
++    {
++      "prompt": "Q: If a train travels 60 km in 45 minute",
++      "tokens": 27,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.6203536629375677,
++      "misses": 1052,
++      "stall_s": 0.20653866743668914,
++      "layers_routed": 54
++    },
++    {
++      "prompt": "Les mod\u00e8les \u00e0 m\u00e9lange d'experts activent",
++      "tokens": 22,
++      "max_abs_dlogit": 0.0,
++      "argmax_equal": true,
++      "hit_rate": 0.6107034958998705,
++      "misses": 902,
++      "stall_s": 0.1756609968142584,
++      "layers_routed": 48
++    }
++  ],
++  "tol": 0.001,
++  "ts": "2026-11-01T01:44:20",
++  "max_abs_dlogit_worst": 0.0,
++  "model": "/Users/simon-pierreboucher/.cache/huggingface/hub/models--mlx-community--Qwen3-30B-A3B-Instruct-2507-4bit/snapshots/e9675aa3ca5f900ccef55267914466d55ab325fa",
++  "model_type": "qwen3_moe"
++}
+ \ No newline at end of file

added benchmarks/raw/tests/m3u96a/pytest_qwen3_m3u96a_20261101T014415.txt +2 −0

		`@@ -0,0 +1,2 @@`
1	+......ss [100%]
2	+6 passed, 2 skipped in 18.17s

added benchmarks/scripts/run_cold_nvme_laptop.sh +17 −0

++#!/bin/bash
++# Cold-NVMe tier experiment on the laptop: purge the page cache before each configuration, open the checkpoint with F_NOCACHE,
++# compare demand loading with prefetch predictors at 16 and 32 slots on held-out prompts. Writes one JSONL (append-only).
++cd "$(dirname "$0")/../.."
++M=$(ls -d ~/.cache/huggingface/hub/models--mlx-community--OLMoE-1B-7B-0125-Instruct-4bit/snapshots/*)
++OUT=benchmarks/raw/gateC/cold_nvme_laptop/sweep.jsonl
++mkdir -p $(dirname $OUT)
++for cap in 16 32; do
++  for cfg in "none 8" "transition 8" "transition2 8" "transition3 8" "prev_rank 16" "transition2 16"; do
++    set -- $cfg; pr=$1; m=$2
++    echo "$SUDO_PW" | sudo -S purge 2>/dev/null; sleep 2
++    .venv/bin/python benchmarks/scripts/bench_decode.py --model $M --backend titan --capacity $cap --predictor $pr --m $m --nocache \
++      --traces benchmarks/raw/gateC/traces_olmoe_v2_laptop --families general,code,reasoning --split heldout --repeats 2 --max-tokens 64 \
++      --tag cold_nvme --out $OUT 2>&1 | grep -v Fetching | tail -2
++  done
++done
++echo DONE

added paper/main.pdf +0 −0

Binary file not shown.

modified paper/main.tex +1 −1

				`@@ -68,7 +68,7 @@`
68	\newcommand{\R}{\mathbb{R}}	68	\newcommand{\R}{\mathbb{R}}
69	\newcommand{\Ex}{\mathbb{E}}	69	\newcommand{\Ex}{\mathbb{E}}
70	\newcommand{\ind}{\mathbb{1}}	70	\newcommand{\ind}{\mathbb{1}}
71	−\newcommand{\toks}{\mathrm{tok/s}}	71	+\newcommand{\tps}{tok/s}
72	\newcommand{\code}[1]{\texttt{#1}}	72	\newcommand{\code}[1]{\texttt{#1}}
73	\newcommand{\titan}{TITAN-MoE}	73	\newcommand{\titan}{TITAN-MoE}
74	\newcommand{\gatherqmm}{\code{gather\_qmm}}	74	\newcommand{\gatherqmm}{\code{gather\_qmm}}

modified paper/sections/ablations.tex +1 −1

				`@@ -5,7 +5,7 @@ transport $\times$ kernels $\times$ batch) is executed gate by gate. In this ver`
5	\begin{itemize}	5	\begin{itemize}
6	\item \textbf{Install path} (per-miss synchronisation vs lazy batched installs): $+80$ to $+90\,\%$ decode throughput at 16 slots	6	\item \textbf{Install path} (per-miss synchronisation vs lazy batched installs): $+80$ to $+90\,\%$ decode throughput at 16 slots
7	(Section~\ref{sec:implementation}).	7	(Section~\ref{sec:implementation}).
8	− \item \textbf{Fast path} (device-side slot map when a layer is complete): 203 $\to$ 489~\toks{} at 64 slots.	8	+ \item \textbf{Fast path} (device-side slot map when a layer is complete): 203 $\to$ 489~\tps{} at 64 slots.
9	\item \textbf{Eviction policy} (LRU vs LFU): within 0.01 hit rate at every capacity in simulation.	9	\item \textbf{Eviction policy} (LRU vs LFU): within 0.01 hit rate at every capacity in simulation.
10	\item \textbf{Predictor} (none / prev-rank / transition$_d$ / probe / oracle) and \textbf{depth}: Table~\ref{tab:sim}, \textsc{pending} until the v2	10	\item \textbf{Predictor} (none / prev-rank / transition$_d$ / probe / oracle) and \textbf{depth}: Table~\ref{tab:sim}, \textsc{pending} until the v2
11	simulation and its online confirmation complete.	11	simulation and its online confirmation complete.

modified paper/sections/conclusion.tex +1 −1

				`@@ -1,7 +1,7 @@`
1	\section{Conclusion}	1	\section{Conclusion}
2	\label{sec:conclusion}	2	\label{sec:conclusion}
3	We have built an MLX-native MoE runtime whose bounded expert residency is semantically invisible (bit-identical logits to mlx-lm)	3	We have built an MLX-native MoE runtime whose bounded expert residency is semantically invisible (bit-identical logits to mlx-lm)
4	−and whose all-resident path costs nothing (489 vs 494~\toks{}), characterised the cluster it will run on as it actually is	4	+and whose all-resident path costs nothing (489 vs 494~\tps{}), characterised the cluster it will run on as it actually is
5	(gigabit Ethernet, no Thunderbolt), and obtained a first negative result that the formulation explains: one-layer-ahead	5	(gigabit Ethernet, no Thunderbolt), and obtained a first negative result that the formulation explains: one-layer-ahead
6	prefetching cannot pay when the fetch latency exceeds the per-layer compute time. The next steps are fixed by the protocol:	6	prefetching cannot pay when the fetch latency exceeds the per-layer compute time. The next steps are fixed by the protocol:
7	depth-$d$ and token-horizon prefetching against the cold-NVMe tier, the medium model on nodes that cannot hold it, the	7	depth-$d$ and token-horizon prefetching against the cold-NVMe tier, the medium model on nodes that cannot hold it, the

modified paper/sections/formulation.tex +1 −1

				`@@ -74,7 +74,7 @@ A prefetch can only reduce latency if`
74	r'\,\min(\lambda, d\,\tau_\ell) \;>\; m(1-q)\,\kappa\;/\;K(1-h),	74	r'\,\min(\lambda, d\,\tau_\ell) \;>\; m(1-q)\,\kappa\;/\;K(1-h),
75	\label{eq:condition}	75	\label{eq:condition}
76	\end{equation}	76	\end{equation}
77	−and it can \emph{remove} the stall only if $d\,\tau_\ell \ge \lambda$. On the laptop, $\tau_\ell = 0.128$~ms (489~\toks{} over	77	+and it can \emph{remove} the stall only if $d\,\tau_\ell \ge \lambda$. On the laptop, $\tau_\ell = 0.128$~ms (489~\tps{} over
78	16 layers), so a page-cache fetch ($\lambda = 0.21$~ms) needs $d \ge 2$ and a cold-NVMe fetch ($\lambda = 0.43$~ms) needs	78	16 layers), so a page-cache fetch ($\lambda = 0.21$~ms) needs $d \ge 2$ and a cold-NVMe fetch ($\lambda = 0.43$~ms) needs
79	$d \ge 4$; a LAN fetch ($\lambda \approx 28$~ms) can never be hidden within a token on this model, which is why weights are not	79	$d \ge 4$; a LAN fetch ($\lambda \approx 28$~ms) can never be hidden within a token on this model, which is why weights are not
80	moved over the LAN during decode in \titan{} and remote experts are instead executed where they reside. The one-layer online	80	moved over the LAN during decode in \titan{} and remote experts are instead executed where they reside. The one-layer online

modified paper/sections/implementation.tex +2 −2

				`@@ -12,8 +12,8 @@ experts per call 0.50~ms. The slot table with \gatherqmm{} is therefore both the`
12	(\code{benchmarks/raw/gateB/design/}).	12	(\code{benchmarks/raw/gateB/design/}).
13		13
14	\paragraph{No per-miss synchronisation.} The first version evaluated the table after every miss; removing that	14	\paragraph{No per-miss synchronisation.} The first version evaluated the table after every miss; removing that
15	−synchronisation and batching the reads of a layer's misses raised decode throughput at 16 slots from 28 to 51~\toks{}	15	+synchronisation and batching the reads of a layer's misses raised decode throughput at 16 slots from 28 to 51~\tps{}
16	−(general prompts) and from 41 to 73~\toks{} (code), with the stall halved	16	+(general prompts) and from 41 to 73~\tps{} (code), with the stall halved
17	(\code{benchmarks/raw/gateC/prefetch\_online\_laptop/}). The per-miss cost on the page-cache tier is now 0.21~ms: 0.13~ms	17	(\code{benchmarks/raw/gateC/prefetch\_online\_laptop/}). The per-miss cost on the page-cache tier is now 0.21~ms: 0.13~ms
18	of \code{pread}, 0.06~ms of array creation and the remainder in the lazy install.	18	of \code{pread}, 0.06~ms of array creation and the remainder in the lazy install.
19		19

modified paper/sections/results.tex +2 −2

				`@@ -25,7 +25,7 @@ page-cached checkpoint, \code{benchmarks/raw/gateB/capacity\_sweep\_laptop/sweep`
25	\label{tab:capacity}	25	\label{tab:capacity}
26	\begin{tabular}{llrrr}	26	\begin{tabular}{llrrr}
27	\toprule	27	\toprule
28	−Backend & Residency & decode \toks{} & hit rate & stall / 64 tokens \\	28	+Backend & Residency & decode \tps{} & hit rate & stall / 64 tokens \\
29	\midrule	29	\midrule
30	mlx-lm & all & 491--496 & -- & -- \\	30	mlx-lm & all & 491--496 & -- & -- \\
31	\titan{} & 64 slots, preloaded & 489 & 1.00 & 0 \\	31	\titan{} & 64 slots, preloaded & 489 & 1.00 & 0 \\
				`@@ -54,7 +54,7 @@ log-probabilities only 0.16 with these few traces. The v2 traces (65 prompts, 6`
54	\subsection{Prefetching (H1)}	54	\subsection{Prefetching (H1)}
55	\label{sec:results-prefetch}	55	\label{sec:results-prefetch}
56	With the one-layer horizon the online prefetcher raised the hit rate from 0.48 to 0.63 (general) and from 0.67 to 0.73	56	With the one-layer horizon the online prefetcher raised the hit rate from 0.48 to 0.63 (general) and from 0.67 to 0.73
57	−(code) but reduced throughput from 51 to 35 and from 73 to 56~\toks{} (Table~\ref{tab:capacity}): the lead time	57	+(code) but reduced throughput from 51 to 35 and from 73 to 56~\tps{} (Table~\ref{tab:capacity}): the lead time
58	$d\,\tau_\ell = 0.13$~ms is shorter than the 0.21~ms fetch, so predicted experts arrive late and are waited for, and the	58	$d\,\tau_\ell = 0.13$~ms is shorter than the 0.21~ms fetch, so predicted experts arrive late and are waited for, and the
59	50\,\% of wrong predictions add traffic. This is exactly the regime excluded by condition~\eqref{eq:condition}.	59	50\,\% of wrong predictions add traffic. This is exactly the regime excluded by condition~\eqref{eq:condition}.
60	Depth-$d$ and token-horizon prefetching, and the cold-NVMe tier: \textsc{pending} (next experiments of Gate C).	60	Depth-$d$ and token-horizon prefetching, and the cold-NVMe tier: \textsc{pending} (next experiments of Gate C).

modified paper/sections/setup.tex +1 −1

				`@@ -50,6 +50,6 @@ and $\ge 5$ in the final sweeps.`
50		50
51	\subsection{Tiers and constants used by the simulator}	51	\subsection{Tiers and constants used by the simulator}
52	Measured on the laptop: page-cache fetch of one OLMoE expert 0.21~ms (batched demand path), cold NVMe 0.43~ms (F\_NOCACHE after	52	Measured on the laptop: page-cache fetch of one OLMoE expert 0.21~ms (batched demand path), cold NVMe 0.43~ms (F\_NOCACHE after
53	−\code{purge}, single reader), 1~GbE transfer $\approx 28$~ms; per-layer decode compute $\tau_\ell = 0.128$~ms (489~\toks{}).	53	+\code{purge}, single reader), 1~GbE transfer $\approx 28$~ms; per-layer decode compute $\tau_\ell = 0.128$~ms (489~\tps{}).
54	The trace-driven simulator (\code{benchmarks/scripts/simulate\_cache.py}) is exact for hits and misses and uses these constants	54	The trace-driven simulator (\code{benchmarks/scripts/simulate\_cache.py}) is exact for hits and misses and uses these constants
55	for stall estimates.	55	for stall estimates.

added paper/tables/cold.tex +14 −0

++\begin{table}[t]
++\centering\small
++\caption{Cold-NVMe tier on the M5 Max laptop (purge + F\_NOCACHE), OLMoE 4-bit, held-out prompts, 64 tokens.}
++\label{tab:cold}
++\begin{tabular}{llrrrrrr}
++\toprule
++Backend & cfg & $n$ & decode tok/s (median) & p95 ITL ms & hit & stall s & TTFT ms \\
++\midrule
++\bottomrule
++\end{tabular}
++\begin{tablenotes}\footnotesize
++\item Sources: \code{benchmarks/raw/gateC/cold\_nvme\_laptop/sweep.jsonl}.
++\end{tablenotes}
++\end{table}

added paper/tables/qwen3.tex +16 −0

++\begin{table}[t]
++\centering\small
++\caption{Qwen3-30B-A3B-Instruct-2507 4-bit on M3U96a (M3 Ultra, 96 GB): mlx-lm vs TITAN all-resident (128 slots, preloaded), 64 tokens, 3 repeats.}
++\label{tab:qwen3}
++\begin{tabular}{llrrrrrr}
++\toprule
++Backend & cfg & $n$ & decode tok/s (median) & p95 ITL ms & hit & stall s & TTFT ms \\
++\midrule
++mlx\_lm & -- & 18 & 121.7 & 8.4 & -- & -- & 50 \\
++titan & C=128, none, m=8 & 18 & 121.6 & 8.5 & 1.000 & 0.00 & 51 \\
++\bottomrule
++\end{tabular}
++\begin{tablenotes}\footnotesize
++\item Sources: \code{benchmarks/raw/gateC/qwen3\_m3u96a/baseline.jsonl}.
++\end{tablenotes}
++\end{table}

added paper/tables/sim.tex +29 −0

++\begin{table}[t]
++\centering\small
++\caption{Trace-driven simulation (OLMoE, per-layer LRU tables, cold per prompt): hit rate and modelled stall per token by predictor, split and capacity $C$.}
++\label{tab:sim}
++\begin{tabular}{llrrrrrrrr}
++\toprule
++Predictor & split & \multicolumn{2}{c}{$C=16$} & \multicolumn{2}{c}{$C=24$} & \multicolumn{2}{c}{$C=32$} & \multicolumn{2}{c}{$C=48$} \\
++ & & hit & stall & hit & stall & hit & stall & hit & stall \\
++\midrule
++none & fit & 0.577 & 23.3 & 0.706 & 16.2 & 0.804 & 10.8 & 0.931 & 3.8 \\
++none & heldout & 0.581 & 23.1 & 0.706 & 16.2 & 0.806 & 10.7 & 0.933 & 3.7 \\
++prev\_rank & fit & 0.545 & 25.0 & 0.665 & 18.4 & 0.770 & 12.7 & 0.919 & 4.5 \\
++prev\_rank & heldout & 0.556 & 24.4 & 0.670 & 18.1 & 0.771 & 12.6 & 0.920 & 4.4 \\
++transition & fit & 0.680 & 17.3 & 0.770 & 12.5 & 0.839 & 8.8 & 0.940 & 3.3 \\
++transition & heldout & 0.686 & 17.0 & 0.771 & 12.5 & 0.840 & 8.7 & 0.942 & 3.2 \\
++transition2 & fit & 0.663 & 18.0 & 0.758 & 13.0 & 0.832 & 9.1 & 0.938 & 3.4 \\
++transition2 & heldout & 0.672 & 17.6 & 0.761 & 12.9 & 0.833 & 9.0 & 0.940 & 3.3 \\
++transition3 & fit & 0.651 & 18.5 & 0.749 & 13.5 & 0.826 & 9.4 & 0.937 & 3.4 \\
++transition3 & heldout & 0.660 & 18.1 & 0.752 & 13.3 & 0.828 & 9.3 & 0.939 & 3.3 \\
++probe & fit & 0.495 & 27.7 & 0.626 & 20.5 & 0.737 & 14.4 & 0.903 & 5.3 \\
++probe & heldout & 0.504 & 27.2 & 0.630 & 20.3 & 0.738 & 14.4 & 0.903 & 5.3 \\
++oracle & fit & 0.938 & 3.0 & 0.959 & 2.0 & 0.973 & 1.4 & 0.991 & 0.5 \\
++oracle & heldout & 0.940 & 2.9 & 0.959 & 2.0 & 0.972 & 1.4 & 0.991 & 0.5 \\
++\bottomrule
++\end{tabular}
++\begin{tablenotes}\footnotesize
++\item Source: \code{benchmarks/processed/gateC/sim\_olmoe\_v2\_ssd043\_m8.json}; t\_read 0.43 ms, t\_layer 0.128 ms, 1 loader(s); hit = per-layer hit rate, stall in ms per token (modelled).
++\end{tablenotes}
++\end{table}

modified research/LOG.md +9 −0

				`@@ -70,3 +70,12 @@ Chronological, append-only. Times are local (America/Toronto). Every entry names`
70	1-layer horizon*; deeper horizons (transition2/3/4) and the token horizon are being simulated (sim v2 on the 65-prompt traces).	70	1-layer horizon*; deeper horizons (transition2/3/4) and the token horizon are being simulated (sim v2 on the 65-prompt traces).
71	- Routing traces v2 collected (`benchmarks/raw/gateC/traces_olmoe_v2_laptop/`, 65 prompts, 6 families incl. long-context and drift,	71	- Routing traces v2 collected (`benchmarks/raw/gateC/traces_olmoe_v2_laptop/`, 65 prompts, 6 families incl. long-context and drift,
72	up to 384 generated tokens each).	72	up to 384 generated tokens each).
		73	+- Qwen3-30B-A3B-Instruct-2507-4bit on M3U96a (M3 Ultra 96 GB): checkpoint is in the stacked `switch_mlp` layout → registry
		74	+ extended (per-expert byte ranges = slices of the first axis; `tests/test_registry_stacked.py` byte-equal to `mx.load`). 2.53 MiB per
		75	+ expert, 16.3 GB of experts, 0.87 GB dense, 384 experts (0.97 GiB) per decode token; routers are 8-bit. Tests on M3U96a
		76	+ (`benchmarks/raw/tests/m3u96a/pytest_qwen3_m3u96a_20261101T014415.txt`, 6 passed): logits max \|Δ\| = 0.0 at 128/64/32/16 slots,
		77	+ greedy 32-token decode identical at 32 slots. Baselines (`benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl`, 64 tokens, 3 repeats):
		78	+ mlx-lm 121.3–121.6 tok/s decode, prefill 339–434 tok/s; TITAN all-resident (128 slots, preloaded) 121.5–121.9 tok/s — parity.
		79	+- Cluster venvs ready on M3U96a, M3U96b, M4M64a, M2U64, M4M64c (`scripts/cluster_setup.sh`, mlx 0.32.4 / mlx-lm 0.32.2 everywhere).
		80	+- Paper skeleton v0.1 compiles (24 pages; `tab:sim` pending the v2 simulation). Cold-NVMe prefetch sweep launched on the laptop
		81	+ (`benchmarks/scripts/run_cold_nvme_laptop.sh`: purge + F_NOCACHE, 16/32 slots, none/transition{1,2,3}/prev_rank, held-out).

added scripts/make_tables.py +57 −0

++#!/usr/bin/env python3
++"""Generate LaTeX table fragments under paper/tables/ from processed/raw logs. Every number in the paper's tables comes from here."""
++import json, os, sys, glob, statistics
++ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
++OUT = os.path.join(ROOT, "paper", "tables"); os.makedirs(OUT, exist_ok=True)
++
++def sim_table(path, label, caption):
++    d = json.load(open(path)); rows = d["results"]
++    caps = sorted({r["capacity"] for r in rows}); preds = []
++    for r in rows:
++        if r["predictor"] not in preds: preds.append(r["predictor"])
++    lines = ["\\begin{table}[t]", "\\centering\\small", f"\\caption{{{caption}}}", f"\\label{{{label}}}",
++             "\\begin{tabular}{ll" + "rr" * len(caps) + "}", "\\toprule",
++             "Predictor & split & " + " & ".join(f"\\multicolumn{{2}}{{c}}{{$C={c}$}}" for c in caps) + " \\\\",
++             " & & " + " & ".join("hit & stall" for _ in caps) + " \\\\", "\\midrule"]
++    for p in preds:
++        for split in ("fit", "heldout"):
++            cells = []
++            for c in caps:
++                m = [r for r in rows if r["predictor"] == p and r["split"] == split and r["capacity"] == c and r["policy"] == "lru"]
++                cells.append(f"{m[0]['hit_rate']:.3f} & {m[0]['stall_ms_per_token']:.1f}" if m else "-- & --")
++            lines.append(f"{p.replace('_', '\\_')} & {split} & " + " & ".join(cells) + " \\\\")
++    lines += ["\\bottomrule", "\\end{tabular}", "\\begin{tablenotes}\\footnotesize", f"\\item Source: \\code{{{os.path.relpath(path, ROOT).replace('_', '\\_')}}}; t\\_read {d['t_read_ms']} ms, t\\_layer {d['t_layer_ms']} ms, {d['loaders']} loader(s); hit = per-layer hit rate, stall in ms per token (modelled).", "\\end{tablenotes}", "\\end{table}"]
++    open(os.path.join(OUT, f"{label.split(':')[-1]}.tex"), "w").write("\n".join(lines) + "\n")
++    print("wrote", label)
++
++def bench_summary(paths, label, caption, keys=("backend", "capacity", "predictor", "m", "nocache", "split")):
++    """Median decode tok/s, hit rate and stall per configuration from bench_decode JSONL files (warm-up excluded)."""
++    recs = []
++    for p in paths:
++        recs += [json.loads(l) for l in open(p) if l.strip()]
++    recs = [r for r in recs if not r.get("warmup")]
++    groups = {}
++    for r in recs:
++        k = tuple(r.get(x) for x in keys); groups.setdefault(k, []).append(r)
++    lines = ["\\begin{table}[t]", "\\centering\\small", f"\\caption{{{caption}}}", f"\\label{{{label}}}",
++             "\\begin{tabular}{llrrrrrr}", "\\toprule", "Backend & cfg & $n$ & decode tok/s (median) & p95 ITL ms & hit & stall s & TTFT ms \\\\", "\\midrule"]
++    for k in sorted(groups, key=lambda x: tuple(str(v) for v in x)):
++        g = groups[k]
++        med = lambda f: statistics.median([r[f] for r in g if r.get(f) is not None]) if any(r.get(f) is not None for r in g) else None
++        cfg = f"C={k[1]}, {k[2]}, m={k[3]}" + (", cold" if k[4] else "") if k[0] == "titan" else "--"
++        f = lambda v, fmt: (fmt % v) if v is not None else "--"
++        lines.append(f"{k[0].replace('_', '\\_')} & {cfg.replace('_', '\\_')} & {len(g)} & {f(med('gen_tps'), '%.1f')} & {f(med('itl_ms_p95'), '%.1f')} & {f(med('cache_hit_rate'), '%.3f')} & {f(med('stall_s'), '%.2f')} & {f(med('ttft_ms'), '%.0f')} \\\\")
++    lines += ["\\bottomrule", "\\end{tabular}", "\\begin{tablenotes}\\footnotesize", "\\item Sources: " + ", ".join(f"\\code{{{os.path.relpath(p, ROOT).replace('_', '\\_')}}}" for p in paths) + ".", "\\end{tablenotes}", "\\end{table}"]
++    open(os.path.join(OUT, f"{label.split(':')[-1]}.tex"), "w").write("\n".join(lines) + "\n")
++    print("wrote", label)
++
++if __name__ == "__main__":
++    v2 = os.path.join(ROOT, "benchmarks/processed/gateC/sim_olmoe_v2_ssd043_m8.json")
++    v1 = os.path.join(ROOT, "benchmarks/processed/gateC/sim_olmoe_laptop_ssd_v1.json")
++    sim_table(v2 if os.path.exists(v2) else v1, "tab:sim", "Trace-driven simulation (OLMoE, per-layer LRU tables, cold per prompt): hit rate and modelled stall per token by predictor, split and capacity $C$.")
++    cold = os.path.join(ROOT, "benchmarks/raw/gateC/cold_nvme_laptop/sweep.jsonl")
++    if os.path.exists(cold):
++        bench_summary([cold], "tab:cold", "Cold-NVMe tier on the M5 Max laptop (purge + F\\_NOCACHE), OLMoE 4-bit, held-out prompts, 64 tokens.")
++    q = os.path.join(ROOT, "benchmarks/raw/gateC/qwen3_m3u96a/baseline.jsonl")
++    if os.path.exists(q):
++        bench_summary([q], "tab:qwen3", "Qwen3-30B-A3B-Instruct-2507 4-bit on M3U96a (M3 Ultra, 96 GB): mlx-lm vs TITAN all-resident (128 slots, preloaded), 64 tokens, 3 repeats.")

spb/wp33_uqo

Qwen3-30B parity on M3U96a (bit-identical, 121.7 vs 121.5 tok/s), make_tables.py, cold-NVMe sweep script, LOG

Theme