Something went wrong. Try again.
strata but optimized for m5 (and maybe more later)
Something went wrong. Try again.
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442[ { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 84.61, "hit_rate": 0.8237, "ssd_mb_per_tok": 116.96, "m5_decode_1th_tok_s": 23.81, "m5_decode_4th_tok_s": 34.25 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lfu", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 155.43, "hit_rate": 0.6762, "ssd_mb_per_tok": 214.87, "m5_decode_1th_tok_s": 14.9, "m5_decode_4th_tok_s": 22.94 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 87.91, "hit_rate": 0.8168, "ssd_mb_per_tok": 121.53, "m5_decode_1th_tok_s": 23.17, "m5_decode_4th_tok_s": 33.48 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 91.25, "hit_rate": 0.8099, "ssd_mb_per_tok": 126.14, "m5_decode_1th_tok_s": 22.55, "m5_decode_4th_tok_s": 32.73 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 88.87, "hit_rate": 0.8148, "ssd_mb_per_tok": 122.86, "m5_decode_1th_tok_s": 22.99, "m5_decode_4th_tok_s": 33.26 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "belady_min", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 42.81, "hit_rate": 0.9108, "ssd_mb_per_tok": 59.18, "m5_decode_1th_tok_s": 36.8, "m5_decode_4th_tok_s": 48.3 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 74.61, "hit_rate": 0.8446, "ssd_mb_per_tok": 103.14, "m5_decode_1th_tok_s": 26.01, "m5_decode_4th_tok_s": 36.81 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 74.61, "hit_rate": 0.8446, "ssd_mb_per_tok": 103.14, "m5_decode_1th_tok_s": 26.01, "m5_decode_4th_tok_s": 36.81 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 67.74, "hit_rate": 0.8589, "ssd_mb_per_tok": 93.64, "m5_decode_1th_tok_s": 27.77, "m5_decode_4th_tok_s": 38.8 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 67.74, "hit_rate": 0.8589, "ssd_mb_per_tok": 93.64, "m5_decode_1th_tok_s": 27.77, "m5_decode_4th_tok_s": 38.8 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 63.2, "hit_rate": 0.8683, "ssd_mb_per_tok": 87.36, "m5_decode_1th_tok_s": 29.07, "m5_decode_4th_tok_s": 40.25 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 63.21, "hit_rate": 0.8683, "ssd_mb_per_tok": 87.37, "m5_decode_1th_tok_s": 29.07, "m5_decode_4th_tok_s": 40.24 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 81.47, "hit_rate": 0.8303, "ssd_mb_per_tok": 112.63, "m5_decode_1th_tok_s": 24.46, "m5_decode_4th_tok_s": 35.01 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 81.47, "hit_rate": 0.8303, "ssd_mb_per_tok": 112.63, "m5_decode_1th_tok_s": 24.46, "m5_decode_4th_tok_s": 35.01 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 74.59, "hit_rate": 0.8446, "ssd_mb_per_tok": 103.11, "m5_decode_1th_tok_s": 26.01, "m5_decode_4th_tok_s": 36.81 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 74.59, "hit_rate": 0.8446, "ssd_mb_per_tok": 103.12, "m5_decode_1th_tok_s": 26.01, "m5_decode_4th_tok_s": 36.81 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 69.99, "hit_rate": 0.8542, "ssd_mb_per_tok": 96.76, "m5_decode_1th_tok_s": 27.17, "m5_decode_4th_tok_s": 38.13 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 70.0, "hit_rate": 0.8542, "ssd_mb_per_tok": 96.77, "m5_decode_1th_tok_s": 27.16, "m5_decode_4th_tok_s": 38.12 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 79.22, "hit_rate": 0.835, "ssd_mb_per_tok": 109.51, "m5_decode_1th_tok_s": 24.95, "m5_decode_4th_tok_s": 35.58 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 79.22, "hit_rate": 0.835, "ssd_mb_per_tok": 109.51, "m5_decode_1th_tok_s": 24.95, "m5_decode_4th_tok_s": 35.58 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 72.48, "hit_rate": 0.849, "ssd_mb_per_tok": 100.2, "m5_decode_1th_tok_s": 26.53, "m5_decode_4th_tok_s": 37.4 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 72.48, "hit_rate": 0.849, "ssd_mb_per_tok": 100.2, "m5_decode_1th_tok_s": 26.53, "m5_decode_4th_tok_s": 37.4 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 67.97, "hit_rate": 0.8584, "ssd_mb_per_tok": 93.96, "m5_decode_1th_tok_s": 27.71, "m5_decode_4th_tok_s": 38.73 }, { "variant": "K512_unpruned", "n_experts": 512, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 67.97, "hit_rate": 0.8584, "ssd_mb_per_tok": 93.97, "m5_decode_1th_tok_s": 27.71, "m5_decode_4th_tok_s": 38.73 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 64.17, "hit_rate": 0.8663, "ssd_mb_per_tok": 88.71, "m5_decode_1th_tok_s": 28.78, "m5_decode_4th_tok_s": 39.93 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lfu", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 133.06, "hit_rate": 0.7228, "ssd_mb_per_tok": 183.94, "m5_decode_1th_tok_s": 16.9, "m5_decode_4th_tok_s": 25.61 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 68.47, "hit_rate": 0.8573, "ssd_mb_per_tok": 94.66, "m5_decode_1th_tok_s": 27.57, "m5_decode_4th_tok_s": 38.58 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 69.66, "hit_rate": 0.8549, "ssd_mb_per_tok": 96.3, "m5_decode_1th_tok_s": 27.25, "m5_decode_4th_tok_s": 38.22 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 67.9, "hit_rate": 0.8585, "ssd_mb_per_tok": 93.87, "m5_decode_1th_tok_s": 27.72, "m5_decode_4th_tok_s": 38.75 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "belady_min", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 30.41, "hit_rate": 0.9366, "ssd_mb_per_tok": 42.04, "m5_decode_1th_tok_s": 43.9, "m5_decode_4th_tok_s": 55.0 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 56.7, "hit_rate": 0.8819, "ssd_mb_per_tok": 78.38, "m5_decode_1th_tok_s": 31.16, "m5_decode_4th_tok_s": 42.51 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 56.7, "hit_rate": 0.8819, "ssd_mb_per_tok": 78.38, "m5_decode_1th_tok_s": 31.16, "m5_decode_4th_tok_s": 42.51 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 51.63, "hit_rate": 0.8924, "ssd_mb_per_tok": 71.37, "m5_decode_1th_tok_s": 33.0, "m5_decode_4th_tok_s": 44.45 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 51.63, "hit_rate": 0.8924, "ssd_mb_per_tok": 71.37, "m5_decode_1th_tok_s": 33.0, "m5_decode_4th_tok_s": 44.45 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 48.28, "hit_rate": 0.8994, "ssd_mb_per_tok": 66.74, "m5_decode_1th_tok_s": 34.35, "m5_decode_4th_tok_s": 45.84 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 48.28, "hit_rate": 0.8994, "ssd_mb_per_tok": 66.74, "m5_decode_1th_tok_s": 34.35, "m5_decode_4th_tok_s": 45.84 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 62.49, "hit_rate": 0.8698, "ssd_mb_per_tok": 86.39, "m5_decode_1th_tok_s": 29.28, "m5_decode_4th_tok_s": 40.48 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 62.49, "hit_rate": 0.8698, "ssd_mb_per_tok": 86.39, "m5_decode_1th_tok_s": 29.28, "m5_decode_4th_tok_s": 40.48 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 57.5, "hit_rate": 0.8802, "ssd_mb_per_tok": 79.49, "m5_decode_1th_tok_s": 30.88, "m5_decode_4th_tok_s": 42.21 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 57.51, "hit_rate": 0.8802, "ssd_mb_per_tok": 79.49, "m5_decode_1th_tok_s": 30.88, "m5_decode_4th_tok_s": 42.21 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 54.16, "hit_rate": 0.8872, "ssd_mb_per_tok": 74.87, "m5_decode_1th_tok_s": 32.05, "m5_decode_4th_tok_s": 43.46 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 54.16, "hit_rate": 0.8872, "ssd_mb_per_tok": 74.87, "m5_decode_1th_tok_s": 32.05, "m5_decode_4th_tok_s": 43.46 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 60.59, "hit_rate": 0.8738, "ssd_mb_per_tok": 83.76, "m5_decode_1th_tok_s": 29.87, "m5_decode_4th_tok_s": 41.12 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 60.59, "hit_rate": 0.8738, "ssd_mb_per_tok": 83.76, "m5_decode_1th_tok_s": 29.87, "m5_decode_4th_tok_s": 41.12 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 55.59, "hit_rate": 0.8842, "ssd_mb_per_tok": 76.85, "m5_decode_1th_tok_s": 31.54, "m5_decode_4th_tok_s": 42.92 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 55.59, "hit_rate": 0.8842, "ssd_mb_per_tok": 76.85, "m5_decode_1th_tok_s": 31.54, "m5_decode_4th_tok_s": 42.92 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 52.23, "hit_rate": 0.8912, "ssd_mb_per_tok": 72.2, "m5_decode_1th_tok_s": 32.77, "m5_decode_4th_tok_s": 44.21 }, { "variant": "K320_reap_l2_sum", "n_experts": 320, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 52.23, "hit_rate": 0.8912, "ssd_mb_per_tok": 72.21, "m5_decode_1th_tok_s": 32.77, "m5_decode_4th_tok_s": 44.21 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 50.06, "hit_rate": 0.8957, "ssd_mb_per_tok": 69.2, "m5_decode_1th_tok_s": 33.62, "m5_decode_4th_tok_s": 45.09 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lfu", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 110.52, "hit_rate": 0.7697, "ssd_mb_per_tok": 152.79, "m5_decode_1th_tok_s": 19.54, "m5_decode_4th_tok_s": 29.01 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 54.11, "hit_rate": 0.8873, "ssd_mb_per_tok": 74.8, "m5_decode_1th_tok_s": 32.07, "m5_decode_4th_tok_s": 43.48 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 54.42, "hit_rate": 0.8866, "ssd_mb_per_tok": 75.23, "m5_decode_1th_tok_s": 31.96, "m5_decode_4th_tok_s": 43.36 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 53.42, "hit_rate": 0.8887, "ssd_mb_per_tok": 73.85, "m5_decode_1th_tok_s": 32.33, "m5_decode_4th_tok_s": 43.74 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "belady_min", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 22.95, "hit_rate": 0.9522, "ssd_mb_per_tok": 31.72, "m5_decode_1th_tok_s": 49.67, "m5_decode_4th_tok_s": 60.01 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 44.04, "hit_rate": 0.9082, "ssd_mb_per_tok": 60.88, "m5_decode_1th_tok_s": 36.22, "m5_decode_4th_tok_s": 47.73 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 44.04, "hit_rate": 0.9082, "ssd_mb_per_tok": 60.88, "m5_decode_1th_tok_s": 36.22, "m5_decode_4th_tok_s": 47.72 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 40.03, "hit_rate": 0.9166, "ssd_mb_per_tok": 55.34, "m5_decode_1th_tok_s": 38.18, "m5_decode_4th_tok_s": 49.66 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 40.03, "hit_rate": 0.9166, "ssd_mb_per_tok": 55.34, "m5_decode_1th_tok_s": 38.18, "m5_decode_4th_tok_s": 49.66 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 37.41, "hit_rate": 0.9221, "ssd_mb_per_tok": 51.72, "m5_decode_1th_tok_s": 39.59, "m5_decode_4th_tok_s": 51.01 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 37.41, "hit_rate": 0.9221, "ssd_mb_per_tok": 51.72, "m5_decode_1th_tok_s": 39.59, "m5_decode_4th_tok_s": 51.01 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 48.71, "hit_rate": 0.8985, "ssd_mb_per_tok": 67.34, "m5_decode_1th_tok_s": 34.17, "m5_decode_4th_tok_s": 45.66 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 48.71, "hit_rate": 0.8985, "ssd_mb_per_tok": 67.34, "m5_decode_1th_tok_s": 34.17, "m5_decode_4th_tok_s": 45.66 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 44.79, "hit_rate": 0.9067, "ssd_mb_per_tok": 61.92, "m5_decode_1th_tok_s": 35.87, "m5_decode_4th_tok_s": 47.38 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 44.79, "hit_rate": 0.9067, "ssd_mb_per_tok": 61.92, "m5_decode_1th_tok_s": 35.87, "m5_decode_4th_tok_s": 47.38 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 42.18, "hit_rate": 0.9121, "ssd_mb_per_tok": 58.31, "m5_decode_1th_tok_s": 37.11, "m5_decode_4th_tok_s": 48.6 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 42.18, "hit_rate": 0.9121, "ssd_mb_per_tok": 58.31, "m5_decode_1th_tok_s": 37.1, "m5_decode_4th_tok_s": 48.6 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 47.5, "hit_rate": 0.901, "ssd_mb_per_tok": 65.66, "m5_decode_1th_tok_s": 34.68, "m5_decode_4th_tok_s": 46.18 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 47.5, "hit_rate": 0.901, "ssd_mb_per_tok": 65.66, "m5_decode_1th_tok_s": 34.68, "m5_decode_4th_tok_s": 46.18 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 43.5, "hit_rate": 0.9094, "ssd_mb_per_tok": 60.13, "m5_decode_1th_tok_s": 36.47, "m5_decode_4th_tok_s": 47.98 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 43.5, "hit_rate": 0.9094, "ssd_mb_per_tok": 60.13, "m5_decode_1th_tok_s": 36.47, "m5_decode_4th_tok_s": 47.98 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 40.84, "hit_rate": 0.9149, "ssd_mb_per_tok": 56.46, "m5_decode_1th_tok_s": 37.77, "m5_decode_4th_tok_s": 49.25 }, { "variant": "K256_reap_l2_sum", "n_experts": 256, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 40.84, "hit_rate": 0.9149, "ssd_mb_per_tok": 56.46, "m5_decode_1th_tok_s": 37.77, "m5_decode_4th_tok_s": 49.25 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 41.35, "hit_rate": 0.9139, "ssd_mb_per_tok": 57.16, "m5_decode_1th_tok_s": 37.52, "m5_decode_4th_tok_s": 49.01 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lfu", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 94.37, "hit_rate": 0.8034, "ssd_mb_per_tok": 130.46, "m5_decode_1th_tok_s": 22.0, "m5_decode_4th_tok_s": 32.07 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 44.89, "hit_rate": 0.9065, "ssd_mb_per_tok": 62.05, "m5_decode_1th_tok_s": 35.83, "m5_decode_4th_tok_s": 47.34 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 44.94, "hit_rate": 0.9064, "ssd_mb_per_tok": 62.12, "m5_decode_1th_tok_s": 35.81, "m5_decode_4th_tok_s": 47.31 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 44.23, "hit_rate": 0.9079, "ssd_mb_per_tok": 61.14, "m5_decode_1th_tok_s": 36.13, "m5_decode_4th_tok_s": 47.64 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "belady_min", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 18.93, "hit_rate": 0.9606, "ssd_mb_per_tok": 26.17, "m5_decode_1th_tok_s": 53.44, "m5_decode_4th_tok_s": 63.1 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 36.36, "hit_rate": 0.9243, "ssd_mb_per_tok": 50.26, "m5_decode_1th_tok_s": 40.18, "m5_decode_4th_tok_s": 51.57 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 36.36, "hit_rate": 0.9243, "ssd_mb_per_tok": 50.26, "m5_decode_1th_tok_s": 40.18, "m5_decode_4th_tok_s": 51.57 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 33.11, "hit_rate": 0.931, "ssd_mb_per_tok": 45.77, "m5_decode_1th_tok_s": 42.13, "m5_decode_4th_tok_s": 53.39 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 33.11, "hit_rate": 0.931, "ssd_mb_per_tok": 45.77, "m5_decode_1th_tok_s": 42.13, "m5_decode_4th_tok_s": 53.39 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 31.02, "hit_rate": 0.9354, "ssd_mb_per_tok": 42.88, "m5_decode_1th_tok_s": 43.49, "m5_decode_4th_tok_s": 54.63 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "lru", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 31.02, "hit_rate": 0.9354, "ssd_mb_per_tok": 42.88, "m5_decode_1th_tok_s": 43.49, "m5_decode_4th_tok_s": 54.63 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 40.18, "hit_rate": 0.9163, "ssd_mb_per_tok": 55.55, "m5_decode_1th_tok_s": 38.11, "m5_decode_4th_tok_s": 49.58 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 40.18, "hit_rate": 0.9163, "ssd_mb_per_tok": 55.55, "m5_decode_1th_tok_s": 38.11, "m5_decode_4th_tok_s": 49.58 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 37.0, "hit_rate": 0.9229, "ssd_mb_per_tok": 51.15, "m5_decode_1th_tok_s": 39.82, "m5_decode_4th_tok_s": 51.22 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 37.0, "hit_rate": 0.9229, "ssd_mb_per_tok": 51.15, "m5_decode_1th_tok_s": 39.82, "m5_decode_4th_tok_s": 51.22 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 34.85, "hit_rate": 0.9274, "ssd_mb_per_tok": 48.17, "m5_decode_1th_tok_s": 41.06, "m5_decode_4th_tok_s": 52.4 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 34.85, "hit_rate": 0.9274, "ssd_mb_per_tok": 48.17, "m5_decode_1th_tok_s": 41.06, "m5_decode_4th_tok_s": 52.4 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 2, "misses_per_tok_48L": 39.28, "hit_rate": 0.9182, "ssd_mb_per_tok": 54.3, "m5_decode_1th_tok_s": 38.58, "m5_decode_4th_tok_s": 50.04 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.05, "protect_top_j": 3, "misses_per_tok_48L": 39.28, "hit_rate": 0.9182, "ssd_mb_per_tok": 54.3, "m5_decode_1th_tok_s": 38.58, "m5_decode_4th_tok_s": 50.04 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 2, "misses_per_tok_48L": 36.01, "hit_rate": 0.925, "ssd_mb_per_tok": 49.77, "m5_decode_1th_tok_s": 40.38, "m5_decode_4th_tok_s": 51.76 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 36.01, "hit_rate": 0.925, "ssd_mb_per_tok": 49.77, "m5_decode_1th_tok_s": 40.38, "m5_decode_4th_tok_s": 51.76 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 2, "misses_per_tok_48L": 33.83, "hit_rate": 0.9295, "ssd_mb_per_tok": 46.77, "m5_decode_1th_tok_s": 41.68, "m5_decode_4th_tok_s": 52.97 }, { "variant": "K224_reap_l2_sum", "n_experts": 224, "policy": "inverse_router_ema", "lambda_moce": 0.12, "protect_top_j": 3, "misses_per_tok_48L": 33.83, "hit_rate": 0.9295, "ssd_mb_per_tok": 46.77, "m5_decode_1th_tok_s": 41.68, "m5_decode_4th_tok_s": 52.97 }, { "variant": "K320_reap_hybrid", "n_experts": 320, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 61.7, "hit_rate": 0.8715, "ssd_mb_per_tok": 85.29, "m5_decode_1th_tok_s": 29.52, "m5_decode_4th_tok_s": 40.74 }, { "variant": "K320_reap_hybrid", "n_experts": 320, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 49.54, "hit_rate": 0.8968, "ssd_mb_per_tok": 68.48, "m5_decode_1th_tok_s": 33.83, "m5_decode_4th_tok_s": 45.31 }, { "variant": "K320_reap_hybrid", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 66.48, "hit_rate": 0.8615, "ssd_mb_per_tok": 91.9, "m5_decode_1th_tok_s": 28.12, "m5_decode_4th_tok_s": 39.19 }, { "variant": "K320_reap_hybrid", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 54.73, "hit_rate": 0.886, "ssd_mb_per_tok": 75.66, "m5_decode_1th_tok_s": 31.85, "m5_decode_4th_tok_s": 43.24 }, { "variant": "K256_reap_hybrid", "n_experts": 256, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 48.28, "hit_rate": 0.8994, "ssd_mb_per_tok": 66.75, "m5_decode_1th_tok_s": 34.35, "m5_decode_4th_tok_s": 45.84 }, { "variant": "K256_reap_hybrid", "n_experts": 256, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 38.45, "hit_rate": 0.9199, "ssd_mb_per_tok": 53.16, "m5_decode_1th_tok_s": 39.02, "m5_decode_4th_tok_s": 50.46 }, { "variant": "K256_reap_hybrid", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 51.77, "hit_rate": 0.8922, "ssd_mb_per_tok": 71.56, "m5_decode_1th_tok_s": 32.95, "m5_decode_4th_tok_s": 44.4 }, { "variant": "K256_reap_hybrid", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 42.34, "hit_rate": 0.9118, "ssd_mb_per_tok": 58.54, "m5_decode_1th_tok_s": 37.02, "m5_decode_4th_tok_s": 48.52 }, { "variant": "K224_reap_hybrid", "n_experts": 224, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 40.53, "hit_rate": 0.9156, "ssd_mb_per_tok": 56.03, "m5_decode_1th_tok_s": 37.93, "m5_decode_4th_tok_s": 49.41 }, { "variant": "K224_reap_hybrid", "n_experts": 224, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 32.44, "hit_rate": 0.9324, "ssd_mb_per_tok": 44.84, "m5_decode_1th_tok_s": 42.56, "m5_decode_4th_tok_s": 53.78 }, { "variant": "K224_reap_hybrid", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 43.28, "hit_rate": 0.9098, "ssd_mb_per_tok": 59.83, "m5_decode_1th_tok_s": 36.58, "m5_decode_4th_tok_s": 48.08 }, { "variant": "K224_reap_hybrid", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 35.53, "hit_rate": 0.926, "ssd_mb_per_tok": 49.11, "m5_decode_1th_tok_s": 40.66, "m5_decode_4th_tok_s": 52.02 }, { "variant": "K320_reap_shrink", "n_experts": 320, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 61.47, "hit_rate": 0.8719, "ssd_mb_per_tok": 84.97, "m5_decode_1th_tok_s": 29.6, "m5_decode_4th_tok_s": 40.82 }, { "variant": "K320_reap_shrink", "n_experts": 320, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 48.84, "hit_rate": 0.8983, "ssd_mb_per_tok": 67.51, "m5_decode_1th_tok_s": 34.12, "m5_decode_4th_tok_s": 45.6 }, { "variant": "K320_reap_shrink", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 65.44, "hit_rate": 0.8637, "ssd_mb_per_tok": 90.47, "m5_decode_1th_tok_s": 28.41, "m5_decode_4th_tok_s": 39.52 }, { "variant": "K320_reap_shrink", "n_experts": 320, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 53.38, "hit_rate": 0.8888, "ssd_mb_per_tok": 73.8, "m5_decode_1th_tok_s": 32.34, "m5_decode_4th_tok_s": 43.76 }, { "variant": "K256_reap_shrink", "n_experts": 256, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 48.22, "hit_rate": 0.8995, "ssd_mb_per_tok": 66.66, "m5_decode_1th_tok_s": 34.37, "m5_decode_4th_tok_s": 45.87 }, { "variant": "K256_reap_shrink", "n_experts": 256, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 37.87, "hit_rate": 0.9211, "ssd_mb_per_tok": 52.35, "m5_decode_1th_tok_s": 39.33, "m5_decode_4th_tok_s": 50.76 }, { "variant": "K256_reap_shrink", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 50.91, "hit_rate": 0.8939, "ssd_mb_per_tok": 70.38, "m5_decode_1th_tok_s": 33.28, "m5_decode_4th_tok_s": 44.74 }, { "variant": "K256_reap_shrink", "n_experts": 256, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 41.08, "hit_rate": 0.9144, "ssd_mb_per_tok": 56.79, "m5_decode_1th_tok_s": 37.65, "m5_decode_4th_tok_s": 49.14 }, { "variant": "K224_reap_shrink", "n_experts": 224, "policy": "lru", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 40.65, "hit_rate": 0.9153, "ssd_mb_per_tok": 56.19, "m5_decode_1th_tok_s": 37.87, "m5_decode_4th_tok_s": 49.35 }, { "variant": "K224_reap_shrink", "n_experts": 224, "policy": "lru", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 32.07, "hit_rate": 0.9332, "ssd_mb_per_tok": 44.33, "m5_decode_1th_tok_s": 42.8, "m5_decode_4th_tok_s": 54.0 }, { "variant": "K224_reap_shrink", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.0, "protect_top_j": 3, "misses_per_tok_48L": 42.64, "hit_rate": 0.9112, "ssd_mb_per_tok": 58.94, "m5_decode_1th_tok_s": 36.88, "m5_decode_4th_tok_s": 48.38 }, { "variant": "K224_reap_shrink", "n_experts": 224, "policy": "near_pick_ema", "lambda_moce": 0.09, "protect_top_j": 3, "misses_per_tok_48L": 34.47, "hit_rate": 0.9282, "ssd_mb_per_tok": 47.65, "m5_decode_1th_tok_s": 41.29, "m5_decode_4th_tok_s": 52.61 }]