{ "fixtureArrays": { "onnx_backend_attention_4d_input_q": [0.54881352186203, 0.7151893377304077, 0.6027633547782898, 0.5448831915855408, 0.42365479469299316, 0.6458941102027893, 0.4375872015953064, 0.891772985458374, 0.9636627435684204, 0.3834415078163147, 0.7917250394821167, 0.5288949012756348, 0.5680445432662964, 0.9255966544151306, 0.07103605568408966, 0.08712930232286453, 0.020218396559357643, 0.832619845867157, 0.7781567573547363, 0.8700121641159058, 0.978618323802948, 0.7991585731506348, 0.4614793658256531, 0.7805292010307312, 0.11827442795038223, 0.6399210095405579, 0.14335328340530396, 0.9446688890457153, 0.5218483209609985, 0.4146619439125061, 0.26455560326576233, 0.7742336988449097, 0.4561503231525421, 0.568433940410614, 0.018789799883961678, 0.6176354885101318, 0.6120957136154175, 0.6169340014457703, 0.9437480568885803, 0.681820273399353, 0.35950788855552673, 0.43703195452690125, 0.6976311802864075, 0.0602254718542099, 0.6667667031288147, 0.670637845993042, 0.21038256585597992, 0.12892629206180573, 0.31542834639549255, 0.36371076107025146, 0.5701967477798462, 0.4386015236377716, 0.9883738160133362, 0.10204481333494186, 0.20887675881385803, 0.16130951046943665, 0.6531082987785339, 0.25329160690307617, 0.4663107693195343, 0.24442559480667114, 0.15896958112716675, 0.11037514358758926, 0.6563295722007751, 0.13818295300006866, 0.1965823620557785, 0.3687251806259155, 0.8209932446479797, 0.09710127860307693, 0.8379449248313904, 0.0960984081029892, 0.9764594435691833, 0.4686512053012848, 0.9767611026763916, 0.6048455238342285, 0.7392635941505432, 0.03918779268860817, 0.28280696272850037, 0.12019655853509903, 0.296140193939209, 0.11872772127389908, 0.3179831802845001, 0.414262980222702, 0.06414749473333359, 0.6924721002578735, 0.5666014552116394, 0.26538950204849243, 0.5232480764389038, 0.09394051134586334, 0.5759465098381042, 0.9292961955070496, 0.3185689449310303, 0.6674103736877441, 0.13179786503314972, 0.7163271903991699, 0.28940609097480774, 0.18319135904312134, 0.5865129232406616, 0.02010754682123661, 0.8289400339126587, 0.004695476032793522, 0.6778165102005005, 0.2700079679489136, 0.7351940274238586, 0.9621885418891907, 0.2487531453371048, 0.5761573314666748, 0.5920419096946716, 0.5722519159317017, 0.22308163344860077, 0.9527490139007568, 0.4471253752708435, 0.8464086651802063, 0.6994792819023132, 0.2974369525909424, 0.8137978315353394, 0.396505743265152, 0.8811032176017761, 0.5812729001045227, 0.8817353844642639, 0.6925315856933594, 0.7252542972564697, 0.5013243556022644, 0.9560836553573608, 0.6439902186393738, 0.4238550364971161, 0.6063932180404663, 0.019193198531866074, 0.30157482624053955, 0.6601735353469849, 0.2900775969028473, 0.6180154085159302, 0.42876869440078735, 0.1354740709066391, 0.29828232526779175, 0.5699648857116699, 0.5908727645874023, 0.5743252635002136, 0.6532008051872253, 0.6521032452583313, 0.43141844868659973, 0.8965466022491455, 0.36756187677383423, 0.4358649253845215, 0.8919233679771423, 0.806194007396698, 0.7038885951042175, 0.10022688657045364, 0.9194825887680054, 0.7142413258552551, 0.9988470077514648, 0.14944830536842346, 0.8681260347366333, 0.16249293088912964, 0.6155595779418945, 0.1238199844956398, 0.8480082154273987, 0.8073189854621887, 0.5691007375717163, 0.40718328952789307, 0.06916699558496475, 0.6974287629127502, 0.45354267954826355, 0.7220556139945984, 0.8663823008537292, 0.9755215048789978, 0.855803370475769, 0.011714084073901176, 0.359978049993515, 0.729990541934967, 0.17162968218326569, 0.5210366249084473, 0.054337989538908005, 0.19999653100967407, 0.01852179504930973, 0.793697714805603, 0.2239246815443039, 0.3453516662120819, 0.9280812740325928, 0.704414427280426, 0.031838931143283844, 0.1646941602230072, 0.6214783787727356, 0.5772286057472229, 0.23789282143115997, 0.9342139959335327, 0.6139659285545349, 0.5356327891349792, 0.5899099707603455, 0.7301220297813416, 0.31194499135017395, 0.39822107553482056, 0.20984375476837158], "onnx_backend_attention_4d_input_k": [0.18619300425052643, 0.9443724155426025, 0.739550769329071, 0.49045881628990173, 0.22741462290287018, 0.2543564736843109, 0.058029159903526306, 0.43441662192344666, 0.3117958903312683, 0.6963434815406799, 0.37775182723999023, 0.1796036809682846, 0.024678727611899376, 0.06724963337182999, 0.6793927550315857, 0.4536968469619751, 0.5365791916847229, 0.8966712951660156, 0.990338921546936, 0.21689698100090027, 0.6630781888961792, 0.2633223831653595, 0.02065099962055683, 0.7583786249160767, 0.32001715898513794, 0.38346388936042786, 0.5883170962333679, 0.8310484290122986, 0.6289818286895752, 0.872650682926178, 0.27354204654693604, 0.7980468273162842, 0.18563593924045563, 0.9527916312217712, 0.6874882578849792, 0.21550767123699188, 0.9473705887794495, 0.7308558225631714, 0.2539416551589966, 0.21331197023391724, 0.518200695514679, 0.02566271834075451, 0.20747007429599762, 0.4246854782104492, 0.3741699755191803, 0.46357542276382446, 0.27762871980667114, 0.5867843627929688, 0.8638556003570557, 0.11753185838460922, 0.517379105091095, 0.13206811249256134, 0.7168596982955933, 0.39605969190597534, 0.5654212832450867, 0.1832798421382904, 0.14484776556491852, 0.4880562722682953, 0.35561272501945496, 0.9404319524765015, 0.7653252482414246, 0.748663604259491, 0.9037197232246399, 0.08342243731021881, 0.5521924495697021, 0.5844760537147522, 0.961936354637146, 0.29214751720428467, 0.24082878232002258, 0.10029394179582596, 0.016429629176855087, 0.9295293092727661, 0.669916570186615, 0.7851529121398926, 0.28173011541366577, 0.5864101648330688, 0.06395526975393295, 0.48562759160995483, 0.9774951338768005, 0.8765052556991577, 0.3381589651107788, 0.961570143699646, 0.23170162737369537, 0.9493188261985779, 0.9413776993751526, 0.799202561378479, 0.6304479241371155, 0.8742879629135132, 0.2930202782154083, 0.8489435315132141, 0.6178767085075378, 0.013236857950687408, 0.34723350405693054, 0.14814086258411407, 0.9818294048309326, 0.4783703088760376, 0.49739137291908264, 0.6394725441932678, 0.36858460307121277, 0.13690027594566345, 0.8221177458763123, 0.1898479163646698, 0.5113189816474915, 0.2243170291185379, 0.09784448146820068, 0.8621914982795715, 0.9729194641113281, 0.9608346819877625, 0.9065554738044739, 0.774047315120697, 0.3331451416015625, 0.08110138773918152, 0.40724116563796997, 0.2322341352701187, 0.13248763978481293, 0.053427182137966156, 0.7255943417549133, 0.011427458375692368, 0.7705807685852051, 0.14694663882255554, 0.07952208071947098, 0.08960303664207458, 0.6720477938652039, 0.24536721408367157, 0.4205394685268402, 0.557368814945221, 0.8605511784553528, 0.7270442843437195, 0.2703278958797455, 0.131482794880867, 0.05537432059645653, 0.3015986382961273, 0.2621181607246399, 0.45614057779312134, 0.6832813620567322, 0.6956254243850708, 0.28351885080337524, 0.3799269497394562, 0.18115095794200897, 0.7885454893112183, 0.05684807524085045, 0.6969972252845764, 0.7786954045295715, 0.7774075865745544, 0.25942257046699524, 0.3738131523132324, 0.5875996351242065, 0.27282190322875977, 0.3708527982234955, 0.19705428183078766, 0.4598558843135834, 0.044612299650907516, 0.7997958660125732, 0.07695644348859787, 0.5188351273536682, 0.3068101108074188, 0.5775429606437683, 0.9594333171844482, 0.6455702185630798, 0.03536243736743927, 0.4304024279117584, 0.5100168585777283, 0.5361775159835815, 0.6813924908638, 0.2775960862636566, 0.12886056303977966, 0.3926756680011749, 0.9564056992530823, 0.1871308982372284, 0.9039839506149292, 0.5438059568405151, 0.4569114148616791, 0.8820413947105408, 0.45860394835472107, 0.7241676449775696, 0.3990253210067749, 0.9040443897247314, 0.6900250315666199, 0.6996220350265503, 0.32772040367126465, 0.7567786574363708, 0.6360610723495483, 0.2400202751159668, 0.16053882241249084, 0.796391487121582, 0.9591665863990784, 0.4581388235092163, 0.5909841656684875, 0.8577226400375366, 0.45722344517707825, 0.9518744945526123, 0.5757511854171753, 0.8207671046257019, 0.9088436961174011, 0.8155238032341003, 0.15941447019577026, 0.6288984417915344, 0.39843425154685974, 0.06271295249462128, 0.4240322411060333, 0.25868406891822815, 0.849038302898407, 0.03330462798476219, 0.9589827060699463, 0.35536885261535645, 0.3567068874835968, 0.01632850244641304, 0.18523232638835907, 0.40125951170921326, 0.9292914271354675, 0.0996149331331253, 0.9453015327453613, 0.869488537311554, 0.4541623890399933, 0.326700896024704, 0.23274412751197815, 0.6144647002220154, 0.03307459130883217, 0.015606064349412918, 0.428795725107193, 0.06807407736778259, 0.2519409954547882, 0.2211609184741974, 0.253191202878952, 0.13105523586273193, 0.012036222964525223, 0.11548429727554321, 0.6184802651405334, 0.9742562174797058, 0.9903450012207031, 0.40905410051345825, 0.1629544198513031, 0.6387617588043213, 0.4903053343296051, 0.9894098043441772, 0.06530420482158661, 0.7832344174385071, 0.28839850425720215, 0.24141861498355865, 0.6625045537948608, 0.24606318771839142, 0.6658591032028198, 0.5173085331916809, 0.4240889847278595, 0.5546877980232239, 0.2870515286922455, 0.7065746784210205, 0.414856880903244, 0.3605455458164215, 0.8286569118499756, 0.9249669313430786, 0.04600730910897255, 0.2326269894838333, 0.34851935505867004, 0.8149664998054504, 0.9854914546012878, 0.9689717292785645, 0.904948353767395, 0.2965562641620636, 0.9920112490653992, 0.24942004680633545, 0.10590615123510361, 0.9509525895118713, 0.2334202527999878, 0.6897682547569275, 0.05835635960102081, 0.7307090759277344, 0.8817201852798462, 0.27243688702583313, 0.3790569007396698, 0.3742961883544922, 0.7487882375717163, 0.2378072440624237, 0.17185309529304504, 0.4492916464805603, 0.30446839332580566, 0.8391891121864319, 0.23774182796478271, 0.5023894309997559, 0.9425836205482483, 0.6339976787567139, 0.8672894239425659, 0.940209686756134, 0.7507648468017578, 0.6995750665664673, 0.9679655432701111, 0.9944007992744446, 0.4518216848373413], "onnx_backend_attention_4d_input_v": [0.07086978107690811, 0.29279401898384094, 0.15235470235347748, 0.41748636960983276, 0.13128933310508728, 0.6041178107261658, 0.38280805945396423, 0.8953858613967896, 0.96779465675354, 0.5468848943710327, 0.2748235762119293, 0.5922304391860962, 0.8967611789703369, 0.40673333406448364, 0.5520782470703125, 0.2716527581214905, 0.4554441571235657, 0.4017135500907898, 0.24841345846652985, 0.5058664083480835, 0.31038081645965576, 0.37303486466407776, 0.5249704718589783, 0.7505950331687927, 0.3335074782371521, 0.9241587519645691, 0.8623185753822327, 0.048690296709537506, 0.2536425292491913, 0.4461355209350586, 0.10462789237499237, 0.34847599267959595, 0.7400975227355957, 0.6805144548416138, 0.6223844289779663, 0.7105283737182617, 0.20492368936538696, 0.3416981101036072, 0.676242470741272, 0.879234790802002, 0.5436780452728271, 0.2826996445655823, 0.030235258862376213, 0.7103368043899536, 0.007884103804826736, 0.37267908453941345, 0.5305371880531311, 0.922111451625824, 0.08949454873800278, 0.40594232082366943, 0.024313200265169144, 0.3426109850406647, 0.6222310662269592, 0.2790679335594177, 0.2097499519586563, 0.11570323258638382, 0.5771402716636658, 0.6952700018882751, 0.6719571352005005, 0.9488610029220581, 0.002703213831409812, 0.6471966505050659, 0.60039222240448, 0.5887396335601807, 0.9627703428268433, 0.016871673986315727, 0.6964824199676514, 0.8136786222457886, 0.5098071694374084, 0.33396488428115845, 0.7908401489257812, 0.09724292904138565, 0.44203564524650574, 0.5199523568153381, 0.6939564347267151, 0.09088572859764099, 0.2277594953775406, 0.4103015661239624, 0.6232946515083313, 0.8869608044624329, 0.618826150894165, 0.13346147537231445, 0.9805801510810852, 0.8717857599258423, 0.5027207732200623, 0.9223479628562927, 0.5413808226585388, 0.9233060479164124, 0.8298973441123962, 0.968286395072937, 0.919782817363739, 0.03603381663560867, 0.1747720092535019, 0.3891346752643585, 0.9521427154541016, 0.300028920173645, 0.16046763956546783, 0.8863046765327454, 0.4463944137096405, 0.9078755974769592, 0.16023047268390656, 0.6611174941062927, 0.4402637481689453, 0.07648676633834839, 0.6964631676673889, 0.2473987489938736, 0.03961552307009697, 0.05994429811835289, 0.06107853725552559, 0.9077329635620117, 0.7398838996887207, 0.8980623483657837, 0.6725823283195496, 0.5289399027824402, 0.30444636940956116, 0.997962236404419, 0.36218905448913574, 0.47064894437789917, 0.37824517488479614, 0.979526937007904, 0.1746583878993988, 0.32798799872398376, 0.6803486943244934, 0.06320761889219284, 0.60724937915802, 0.47764649987220764, 0.2839999794960022, 0.2384132742881775, 0.5145127177238464, 0.36792758107185364, 0.4565199017524719, 0.3374773859977722, 0.9704936742782593, 0.13343943655490875, 0.09680395573377609, 0.3433917164802551, 0.5910269021987915, 0.6591764688491821, 0.3972567617893219, 0.9992780089378357, 0.35189300775527954, 0.7214066386222839, 0.6375827193260193, 0.8130538463592529, 0.9762256741523743, 0.8897936344146729, 0.7645619511604309, 0.6982485055923462, 0.335498183965683, 0.14768557250499725, 0.06263600289821625, 0.2419017106294632, 0.432281494140625, 0.521996259689331, 0.7730835676193237, 0.9587409496307373, 0.1173204779624939, 0.10700414329767227, 0.5896947383880615, 0.7453980445861816, 0.848150372505188, 0.9358320832252502, 0.9834262132644653, 0.39980170130729675, 0.3803351819515228, 0.14780867099761963, 0.6849344372749329, 0.6567619442939758, 0.8620625734329224, 0.09725799411535263, 0.49777689576148987, 0.5810819268226624, 0.2415570467710495, 0.16902540624141693, 0.8595808148384094, 0.05853492394089699, 0.47062090039253235, 0.11583399772644043, 0.45705875754356384, 0.9799623489379883, 0.4237063527107239, 0.857124924659729, 0.11731556057929993, 0.2712520658969879, 0.40379273891448975, 0.39981213212013245, 0.6713835000991821, 0.3447181284427643, 0.713766872882843, 0.6391869187355042, 0.399161159992218, 0.43176013231277466, 0.614527702331543, 0.0700421929359436, 0.8224067091941833, 0.65342116355896, 0.7263424396514893, 0.5369229912757874, 0.11047711223363876, 0.4050356149673462, 0.40537357330322266, 0.3210429847240448, 0.029950324445962906, 0.73725426197052, 0.10978446155786514, 0.6063081622123718, 0.7032175064086914, 0.6347863078117371, 0.95914226770401, 0.10329815745353699, 0.8671671748161316, 0.02919023483991623, 0.534916877746582, 0.4042436182498932, 0.5241838693618774, 0.36509987711906433, 0.19056691229343414, 0.01912289671599865, 0.5181497931480408, 0.8427768349647522, 0.3732159435749054, 0.2228638231754303, 0.080532006919384, 0.0853109210729599, 0.22139644622802734, 0.10001406073570251, 0.26503971219062805, 0.06614946573972702, 0.06560486555099487, 0.8562761545181274, 0.1621202677488327, 0.5596824288368225, 0.7734555602073669, 0.4564095735549927, 0.15336887538433075, 0.19959613680839539, 0.43298420310020447, 0.52823406457901, 0.3494403064250946, 0.7814795970916748, 0.7510216236114502, 0.9272118210792542, 0.028952548280358315, 0.8956912755966187, 0.39256879687309265, 0.8783724904060364, 0.690784752368927, 0.987348735332489, 0.7592824697494507, 0.3645446300506592, 0.5010631680488586, 0.37638914585113525, 0.364911824464798, 0.2609044909477234, 0.49597030878067017, 0.6817399263381958, 0.27734026312828064, 0.5243797898292542, 0.117380291223526, 0.1598452925682068, 0.04680635407567024, 0.9707314372062683, 0.0038603513967245817, 0.17857997119426727, 0.6128667593002319, 0.08136960119009018, 0.8818964958190918, 0.7196201682090759, 0.9663899540901184, 0.5076355338096619, 0.3004036843776703, 0.549500584602356, 0.9308187365531921, 0.5207614302635193, 0.2672070264816284, 0.8773987889289856, 0.3719187378883362, 0.0013833499979227781, 0.2476850152015686, 0.31823351979255676, 0.8587774634361267, 0.4585031569004059, 0.4445872902870178, 0.33610227704048157, 0.880678117275238, 0.9450267553329468, 0.9918903112411499, 0.3767412602901459], "onnx_backend_attention_4d_diff_heads_sizes_input_v": [0.07086978107690811, 0.29279401898384094, 0.15235470235347748, 0.41748636960983276, 0.13128933310508728, 0.6041178107261658, 0.38280805945396423, 0.8953858613967896, 0.96779465675354, 0.5468848943710327, 0.2748235762119293, 0.5922304391860962, 0.8967611789703369, 0.40673333406448364, 0.5520782470703125, 0.2716527581214905, 0.4554441571235657, 0.4017135500907898, 0.24841345846652985, 0.5058664083480835, 0.31038081645965576, 0.37303486466407776, 0.5249704718589783, 0.7505950331687927, 0.3335074782371521, 0.9241587519645691, 0.8623185753822327, 0.048690296709537506, 0.2536425292491913, 0.4461355209350586, 0.10462789237499237, 0.34847599267959595, 0.7400975227355957, 0.6805144548416138, 0.6223844289779663, 0.7105283737182617, 0.20492368936538696, 0.3416981101036072, 0.676242470741272, 0.879234790802002, 0.5436780452728271, 0.2826996445655823, 0.030235258862376213, 0.7103368043899536, 0.007884103804826736, 0.37267908453941345, 0.5305371880531311, 0.922111451625824, 0.08949454873800278, 0.40594232082366943, 0.024313200265169144, 0.3426109850406647, 0.6222310662269592, 0.2790679335594177, 0.2097499519586563, 0.11570323258638382, 0.5771402716636658, 0.6952700018882751, 0.6719571352005005, 0.9488610029220581, 0.002703213831409812, 0.6471966505050659, 0.60039222240448, 0.5887396335601807, 0.9627703428268433, 0.016871673986315727, 0.6964824199676514, 0.8136786222457886, 0.5098071694374084, 0.33396488428115845, 0.7908401489257812, 0.09724292904138565, 0.44203564524650574, 0.5199523568153381, 0.6939564347267151, 0.09088572859764099, 0.2277594953775406, 0.4103015661239624, 0.6232946515083313, 0.8869608044624329, 0.618826150894165, 0.13346147537231445, 0.9805801510810852, 0.8717857599258423, 0.5027207732200623, 0.9223479628562927, 0.5413808226585388, 0.9233060479164124, 0.8298973441123962, 0.968286395072937, 0.919782817363739, 0.03603381663560867, 0.1747720092535019, 0.3891346752643585, 0.9521427154541016, 0.300028920173645, 0.16046763956546783, 0.8863046765327454, 0.4463944137096405, 0.9078755974769592, 0.16023047268390656, 0.6611174941062927, 0.4402637481689453, 0.07648676633834839, 0.6964631676673889, 0.2473987489938736, 0.03961552307009697, 0.05994429811835289, 0.06107853725552559, 0.9077329635620117, 0.7398838996887207, 0.8980623483657837, 0.6725823283195496, 0.5289399027824402, 0.30444636940956116, 0.997962236404419, 0.36218905448913574, 0.47064894437789917, 0.37824517488479614, 0.979526937007904, 0.1746583878993988, 0.32798799872398376, 0.6803486943244934, 0.06320761889219284, 0.60724937915802, 0.47764649987220764, 0.2839999794960022, 0.2384132742881775, 0.5145127177238464, 0.36792758107185364, 0.4565199017524719, 0.3374773859977722, 0.9704936742782593, 0.13343943655490875, 0.09680395573377609, 0.3433917164802551, 0.5910269021987915, 0.6591764688491821, 0.3972567617893219, 0.9992780089378357, 0.35189300775527954, 0.7214066386222839, 0.6375827193260193, 0.8130538463592529, 0.9762256741523743, 0.8897936344146729, 0.7645619511604309, 0.6982485055923462, 0.335498183965683, 0.14768557250499725, 0.06263600289821625, 0.2419017106294632, 0.432281494140625, 0.521996259689331, 0.7730835676193237, 0.9587409496307373, 0.1173204779624939, 0.10700414329767227, 0.5896947383880615, 0.7453980445861816, 0.848150372505188, 0.9358320832252502, 0.9834262132644653, 0.39980170130729675, 0.3803351819515228, 0.14780867099761963, 0.6849344372749329, 0.6567619442939758, 0.8620625734329224, 0.09725799411535263, 0.49777689576148987, 0.5810819268226624, 0.2415570467710495, 0.16902540624141693, 0.8595808148384094, 0.05853492394089699, 0.47062090039253235, 0.11583399772644043, 0.45705875754356384, 0.9799623489379883, 0.4237063527107239, 0.857124924659729, 0.11731556057929993, 0.2712520658969879, 0.40379273891448975, 0.39981213212013245, 0.6713835000991821, 0.3447181284427643, 0.713766872882843, 0.6391869187355042, 0.399161159992218, 0.43176013231277466, 0.614527702331543, 0.0700421929359436, 0.8224067091941833, 0.65342116355896, 0.7263424396514893, 0.5369229912757874, 0.11047711223363876, 0.4050356149673462, 0.40537357330322266, 0.3210429847240448, 0.029950324445962906, 0.73725426197052, 0.10978446155786514, 0.6063081622123718, 0.7032175064086914, 0.6347863078117371, 0.95914226770401, 0.10329815745353699, 0.8671671748161316, 0.02919023483991623, 0.534916877746582, 0.4042436182498932, 0.5241838693618774, 0.36509987711906433, 0.19056691229343414, 0.01912289671599865, 0.5181497931480408, 0.8427768349647522, 0.3732159435749054, 0.2228638231754303, 0.080532006919384, 0.0853109210729599, 0.22139644622802734, 0.10001406073570251, 0.26503971219062805, 0.06614946573972702, 0.06560486555099487, 0.8562761545181274, 0.1621202677488327, 0.5596824288368225, 0.7734555602073669, 0.4564095735549927, 0.15336887538433075, 0.19959613680839539, 0.43298420310020447, 0.52823406457901, 0.3494403064250946, 0.7814795970916748, 0.7510216236114502, 0.9272118210792542, 0.028952548280358315, 0.8956912755966187, 0.39256879687309265, 0.8783724904060364, 0.690784752368927, 0.987348735332489, 0.7592824697494507, 0.3645446300506592, 0.5010631680488586, 0.37638914585113525, 0.364911824464798, 0.2609044909477234, 0.49597030878067017, 0.6817399263381958, 0.27734026312828064, 0.5243797898292542, 0.117380291223526, 0.1598452925682068, 0.04680635407567024, 0.9707314372062683, 0.0038603513967245817, 0.17857997119426727, 0.6128667593002319, 0.08136960119009018, 0.8818964958190918, 0.7196201682090759, 0.9663899540901184, 0.5076355338096619, 0.3004036843776703, 0.549500584602356, 0.9308187365531921, 0.5207614302635193, 0.2672070264816284, 0.8773987889289856, 0.3719187378883362, 0.0013833499979227781, 0.2476850152015686, 0.31823351979255676, 0.8587774634361267, 0.4585031569004059, 0.4445872902870178, 0.33610227704048157, 0.880678117275238, 0.9450267553329468, 0.9918903112411499, 0.3767412602901459, 0.9661474227905273, 0.7918795943260193, 0.675689160823822, 0.24488948285579681, 0.21645726263523102, 0.1660478264093399, 0.9227566123008728, 0.2940766513347626, 0.4530942440032959, 0.49395784735679626, 0.7781715989112854, 0.8442349433898926, 0.1390727013349533, 0.4269043505191803, 0.842854917049408, 0.8180332779884338, 0.10241375863552094, 0.15638335049152374, 0.304198682308197, 0.07535906881093979, 0.4246630072593689, 0.10761770606040955, 0.5682175755500793, 0.24655693769454956, 0.5964330434799194, 0.11752564460039139, 0.9758838415145874, 0.9325612187385559, 0.39179694652557373, 0.24217858910560608, 0.2503982186317444, 0.4833935499191284, 0.0399928018450737, 0.6397051215171814, 0.408302903175354, 0.3774065673351288, 0.8093649744987488, 0.7090354561805725, 0.9543338418006897, 0.3519362509250641, 0.8975427746772766, 0.7699671983718872, 0.35742464661598206, 0.6216654181480408, 0.28856995701789856, 0.8743999004364014, 0.11242731660604477, 0.21243436634540558, 0.183033287525177, 0.4030260145664215, 0.7452329397201538, 0.5269074440002441, 0.48767632246017456, 0.0005459649255499244, 0.4254017174243927, 0.06355377286672592, 0.20825324952602386, 0.9323939681053162, 0.21539820730686188, 0.8583376407623291, 0.8028934001922607, 0.1591462343931198, 0.6057119369506836, 0.1156618744134903, 0.7278881669044495, 0.6374622583389282, 0.8119385838508606, 0.47938454151153564, 0.914863109588623, 0.049348946660757065, 0.2928885519504547, 0.715052604675293], "onnx_backend_attention_4d_gqa_input_q": [0.54881352186203, 0.7151893377304077, 0.6027633547782898, 0.5448831915855408, 0.42365479469299316, 0.6458941102027893, 0.4375872015953064, 0.891772985458374, 0.9636627435684204, 0.3834415078163147, 0.7917250394821167, 0.5288949012756348, 0.5680445432662964, 0.9255966544151306, 0.07103605568408966, 0.08712930232286453, 0.020218396559357643, 0.832619845867157, 0.7781567573547363, 0.8700121641159058, 0.978618323802948, 0.7991585731506348, 0.4614793658256531, 0.7805292010307312, 0.11827442795038223, 0.6399210095405579, 0.14335328340530396, 0.9446688890457153, 0.5218483209609985, 0.4146619439125061, 0.26455560326576233, 0.7742336988449097, 0.4561503231525421, 0.568433940410614, 0.018789799883961678, 0.6176354885101318, 0.6120957136154175, 0.6169340014457703, 0.9437480568885803, 0.681820273399353, 0.35950788855552673, 0.43703195452690125, 0.6976311802864075, 0.0602254718542099, 0.6667667031288147, 0.670637845993042, 0.21038256585597992, 0.12892629206180573, 0.31542834639549255, 0.36371076107025146, 0.5701967477798462, 0.4386015236377716, 0.9883738160133362, 0.10204481333494186, 0.20887675881385803, 0.16130951046943665, 0.6531082987785339, 0.25329160690307617, 0.4663107693195343, 0.24442559480667114, 0.15896958112716675, 0.11037514358758926, 0.6563295722007751, 0.13818295300006866, 0.1965823620557785, 0.3687251806259155, 0.8209932446479797, 0.09710127860307693, 0.8379449248313904, 0.0960984081029892, 0.9764594435691833, 0.4686512053012848, 0.9767611026763916, 0.6048455238342285, 0.7392635941505432, 0.03918779268860817, 0.28280696272850037, 0.12019655853509903, 0.296140193939209, 0.11872772127389908, 0.3179831802845001, 0.414262980222702, 0.06414749473333359, 0.6924721002578735, 0.5666014552116394, 0.26538950204849243, 0.5232480764389038, 0.09394051134586334, 0.5759465098381042, 0.9292961955070496, 0.3185689449310303, 0.6674103736877441, 0.13179786503314972, 0.7163271903991699, 0.28940609097480774, 0.18319135904312134, 0.5865129232406616, 0.02010754682123661, 0.8289400339126587, 0.004695476032793522, 0.6778165102005005, 0.2700079679489136, 0.7351940274238586, 0.9621885418891907, 0.2487531453371048, 0.5761573314666748, 0.5920419096946716, 0.5722519159317017, 0.22308163344860077, 0.9527490139007568, 0.4471253752708435, 0.8464086651802063, 0.6994792819023132, 0.2974369525909424, 0.8137978315353394, 0.396505743265152, 0.8811032176017761, 0.5812729001045227, 0.8817353844642639, 0.6925315856933594, 0.7252542972564697, 0.5013243556022644, 0.9560836553573608, 0.6439902186393738, 0.4238550364971161, 0.6063932180404663, 0.019193198531866074, 0.30157482624053955, 0.6601735353469849, 0.2900775969028473, 0.6180154085159302, 0.42876869440078735, 0.1354740709066391, 0.29828232526779175, 0.5699648857116699, 0.5908727645874023, 0.5743252635002136, 0.6532008051872253, 0.6521032452583313, 0.43141844868659973, 0.8965466022491455, 0.36756187677383423, 0.4358649253845215, 0.8919233679771423, 0.806194007396698, 0.7038885951042175, 0.10022688657045364, 0.9194825887680054, 0.7142413258552551, 0.9988470077514648, 0.14944830536842346, 0.8681260347366333, 0.16249293088912964, 0.6155595779418945, 0.1238199844956398, 0.8480082154273987, 0.8073189854621887, 0.5691007375717163, 0.40718328952789307, 0.06916699558496475, 0.6974287629127502, 0.45354267954826355, 0.7220556139945984, 0.8663823008537292, 0.9755215048789978, 0.855803370475769, 0.011714084073901176, 0.359978049993515, 0.729990541934967, 0.17162968218326569, 0.5210366249084473, 0.054337989538908005, 0.19999653100967407, 0.01852179504930973, 0.793697714805603, 0.2239246815443039, 0.3453516662120819, 0.9280812740325928, 0.704414427280426, 0.031838931143283844, 0.1646941602230072, 0.6214783787727356, 0.5772286057472229, 0.23789282143115997, 0.9342139959335327, 0.6139659285545349, 0.5356327891349792, 0.5899099707603455, 0.7301220297813416, 0.31194499135017395, 0.39822107553482056, 0.20984375476837158, 0.18619300425052643, 0.9443724155426025, 0.739550769329071, 0.49045881628990173, 0.22741462290287018, 0.2543564736843109, 0.058029159903526306, 0.43441662192344666, 0.3117958903312683, 0.6963434815406799, 0.37775182723999023, 0.1796036809682846, 0.024678727611899376, 0.06724963337182999, 0.6793927550315857, 0.4536968469619751, 0.5365791916847229, 0.8966712951660156, 0.990338921546936, 0.21689698100090027, 0.6630781888961792, 0.2633223831653595, 0.02065099962055683, 0.7583786249160767, 0.32001715898513794, 0.38346388936042786, 0.5883170962333679, 0.8310484290122986, 0.6289818286895752, 0.872650682926178, 0.27354204654693604, 0.7980468273162842, 0.18563593924045563, 0.9527916312217712, 0.6874882578849792, 0.21550767123699188, 0.9473705887794495, 0.7308558225631714, 0.2539416551589966, 0.21331197023391724, 0.518200695514679, 0.02566271834075451, 0.20747007429599762, 0.4246854782104492, 0.3741699755191803, 0.46357542276382446, 0.27762871980667114, 0.5867843627929688, 0.8638556003570557, 0.11753185838460922, 0.517379105091095, 0.13206811249256134, 0.7168596982955933, 0.39605969190597534, 0.5654212832450867, 0.1832798421382904, 0.14484776556491852, 0.4880562722682953, 0.35561272501945496, 0.9404319524765015, 0.7653252482414246, 0.748663604259491, 0.9037197232246399, 0.08342243731021881, 0.5521924495697021, 0.5844760537147522, 0.961936354637146, 0.29214751720428467, 0.24082878232002258, 0.10029394179582596, 0.016429629176855087, 0.9295293092727661, 0.669916570186615, 0.7851529121398926, 0.28173011541366577, 0.5864101648330688, 0.06395526975393295, 0.48562759160995483, 0.9774951338768005, 0.8765052556991577, 0.3381589651107788, 0.961570143699646, 0.23170162737369537, 0.9493188261985779, 0.9413776993751526, 0.799202561378479, 0.6304479241371155, 0.8742879629135132, 0.2930202782154083, 0.8489435315132141, 0.6178767085075378, 0.013236857950687408, 0.34723350405693054, 0.14814086258411407, 0.9818294048309326, 0.4783703088760376, 0.49739137291908264, 0.6394725441932678, 0.36858460307121277, 0.13690027594566345, 0.8221177458763123, 0.1898479163646698, 0.5113189816474915, 0.2243170291185379, 0.09784448146820068, 0.8621914982795715, 0.9729194641113281, 0.9608346819877625, 0.9065554738044739, 0.774047315120697, 0.3331451416015625, 0.08110138773918152, 0.40724116563796997, 0.2322341352701187, 0.13248763978481293, 0.053427182137966156, 0.7255943417549133, 0.011427458375692368, 0.7705807685852051, 0.14694663882255554, 0.07952208071947098, 0.08960303664207458, 0.6720477938652039, 0.24536721408367157, 0.4205394685268402, 0.557368814945221, 0.8605511784553528, 0.7270442843437195, 0.2703278958797455, 0.131482794880867, 0.05537432059645653, 0.3015986382961273, 0.2621181607246399, 0.45614057779312134, 0.6832813620567322, 0.6956254243850708, 0.28351885080337524, 0.3799269497394562, 0.18115095794200897, 0.7885454893112183, 0.05684807524085045, 0.6969972252845764, 0.7786954045295715, 0.7774075865745544, 0.25942257046699524, 0.3738131523132324, 0.5875996351242065, 0.27282190322875977, 0.3708527982234955, 0.19705428183078766, 0.4598558843135834, 0.044612299650907516, 0.7997958660125732, 0.07695644348859787, 0.5188351273536682, 0.3068101108074188, 0.5775429606437683, 0.9594333171844482, 0.6455702185630798, 0.03536243736743927, 0.4304024279117584, 0.5100168585777283, 0.5361775159835815, 0.6813924908638, 0.2775960862636566, 0.12886056303977966, 0.3926756680011749, 0.9564056992530823, 0.1871308982372284, 0.9039839506149292, 0.5438059568405151, 0.4569114148616791, 0.8820413947105408, 0.45860394835472107, 0.7241676449775696, 0.3990253210067749, 0.9040443897247314, 0.6900250315666199, 0.6996220350265503, 0.32772040367126465, 0.7567786574363708, 0.6360610723495483, 0.2400202751159668, 0.16053882241249084, 0.796391487121582, 0.9591665863990784, 0.4581388235092163, 0.5909841656684875, 0.8577226400375366, 0.45722344517707825, 0.9518744945526123, 0.5757511854171753, 0.8207671046257019, 0.9088436961174011, 0.8155238032341003, 0.15941447019577026, 0.6288984417915344, 0.39843425154685974, 0.06271295249462128, 0.4240322411060333, 0.25868406891822815, 0.849038302898407, 0.03330462798476219, 0.9589827060699463, 0.35536885261535645, 0.3567068874835968, 0.01632850244641304, 0.18523232638835907, 0.40125951170921326, 0.9292914271354675, 0.0996149331331253, 0.9453015327453613, 0.869488537311554, 0.4541623890399933, 0.326700896024704, 0.23274412751197815, 0.6144647002220154, 0.03307459130883217, 0.015606064349412918, 0.428795725107193, 0.06807407736778259, 0.2519409954547882, 0.2211609184741974, 0.253191202878952, 0.13105523586273193, 0.012036222964525223, 0.11548429727554321, 0.6184802651405334, 0.9742562174797058, 0.9903450012207031, 0.40905410051345825, 0.1629544198513031, 0.6387617588043213, 0.4903053343296051, 0.9894098043441772, 0.06530420482158661, 0.7832344174385071, 0.28839850425720215, 0.24141861498355865, 0.6625045537948608, 0.24606318771839142, 0.6658591032028198, 0.5173085331916809, 0.4240889847278595, 0.5546877980232239, 0.2870515286922455, 0.7065746784210205, 0.414856880903244, 0.3605455458164215, 0.8286569118499756, 0.9249669313430786, 0.04600730910897255, 0.2326269894838333, 0.34851935505867004, 0.8149664998054504, 0.9854914546012878, 0.9689717292785645, 0.904948353767395, 0.2965562641620636, 0.9920112490653992, 0.24942004680633545, 0.10590615123510361, 0.9509525895118713, 0.2334202527999878, 0.6897682547569275, 0.05835635960102081, 0.7307090759277344, 0.8817201852798462, 0.27243688702583313, 0.3790569007396698, 0.3742961883544922, 0.7487882375717163, 0.2378072440624237, 0.17185309529304504, 0.4492916464805603, 0.30446839332580566, 0.8391891121864319, 0.23774182796478271, 0.5023894309997559, 0.9425836205482483, 0.6339976787567139, 0.8672894239425659, 0.940209686756134, 0.7507648468017578, 0.6995750665664673, 0.9679655432701111, 0.9944007992744446, 0.4518216848373413, 0.07086978107690811, 0.29279401898384094, 0.15235470235347748, 0.41748636960983276, 0.13128933310508728, 0.6041178107261658, 0.38280805945396423, 0.8953858613967896, 0.96779465675354, 0.5468848943710327, 0.2748235762119293, 0.5922304391860962, 0.8967611789703369, 0.40673333406448364, 0.5520782470703125, 0.2716527581214905, 0.4554441571235657, 0.4017135500907898, 0.24841345846652985, 0.5058664083480835, 0.31038081645965576, 0.37303486466407776, 0.5249704718589783, 0.7505950331687927, 0.3335074782371521, 0.9241587519645691, 0.8623185753822327, 0.048690296709537506, 0.2536425292491913, 0.4461355209350586, 0.10462789237499237, 0.34847599267959595, 0.7400975227355957, 0.6805144548416138, 0.6223844289779663, 0.7105283737182617, 0.20492368936538696, 0.3416981101036072, 0.676242470741272, 0.879234790802002, 0.5436780452728271, 0.2826996445655823, 0.030235258862376213, 0.7103368043899536, 0.007884103804826736, 0.37267908453941345, 0.5305371880531311, 0.922111451625824, 0.08949454873800278, 0.40594232082366943, 0.024313200265169144, 0.3426109850406647, 0.6222310662269592, 0.2790679335594177, 0.2097499519586563, 0.11570323258638382, 0.5771402716636658, 0.6952700018882751, 0.6719571352005005, 0.9488610029220581, 0.002703213831409812, 0.6471966505050659, 0.60039222240448, 0.5887396335601807, 0.9627703428268433, 0.016871673986315727, 0.6964824199676514, 0.8136786222457886, 0.5098071694374084, 0.33396488428115845, 0.7908401489257812, 0.09724292904138565, 0.44203564524650574, 0.5199523568153381, 0.6939564347267151, 0.09088572859764099, 0.2277594953775406, 0.4103015661239624, 0.6232946515083313, 0.8869608044624329, 0.618826150894165, 0.13346147537231445, 0.9805801510810852, 0.8717857599258423, 0.5027207732200623, 0.9223479628562927, 0.5413808226585388, 0.9233060479164124, 0.8298973441123962, 0.968286395072937, 0.919782817363739, 0.03603381663560867, 0.1747720092535019, 0.3891346752643585, 0.9521427154541016, 0.300028920173645], "onnx_backend_attention_4d_gqa_input_k": [0.16046763956546783, 0.8863046765327454, 0.4463944137096405, 0.9078755974769592, 0.16023047268390656, 0.6611174941062927, 0.4402637481689453, 0.07648676633834839, 0.6964631676673889, 0.2473987489938736, 0.03961552307009697, 0.05994429811835289, 0.06107853725552559, 0.9077329635620117, 0.7398838996887207, 0.8980623483657837, 0.6725823283195496, 0.5289399027824402, 0.30444636940956116, 0.997962236404419, 0.36218905448913574, 0.47064894437789917, 0.37824517488479614, 0.979526937007904, 0.1746583878993988, 0.32798799872398376, 0.6803486943244934, 0.06320761889219284, 0.60724937915802, 0.47764649987220764, 0.2839999794960022, 0.2384132742881775, 0.5145127177238464, 0.36792758107185364, 0.4565199017524719, 0.3374773859977722, 0.9704936742782593, 0.13343943655490875, 0.09680395573377609, 0.3433917164802551, 0.5910269021987915, 0.6591764688491821, 0.3972567617893219, 0.9992780089378357, 0.35189300775527954, 0.7214066386222839, 0.6375827193260193, 0.8130538463592529, 0.9762256741523743, 0.8897936344146729, 0.7645619511604309, 0.6982485055923462, 0.335498183965683, 0.14768557250499725, 0.06263600289821625, 0.2419017106294632, 0.432281494140625, 0.521996259689331, 0.7730835676193237, 0.9587409496307373, 0.1173204779624939, 0.10700414329767227, 0.5896947383880615, 0.7453980445861816, 0.848150372505188, 0.9358320832252502, 0.9834262132644653, 0.39980170130729675, 0.3803351819515228, 0.14780867099761963, 0.6849344372749329, 0.6567619442939758, 0.8620625734329224, 0.09725799411535263, 0.49777689576148987, 0.5810819268226624, 0.2415570467710495, 0.16902540624141693, 0.8595808148384094, 0.05853492394089699, 0.47062090039253235, 0.11583399772644043, 0.45705875754356384, 0.9799623489379883, 0.4237063527107239, 0.857124924659729, 0.11731556057929993, 0.2712520658969879, 0.40379273891448975, 0.39981213212013245, 0.6713835000991821, 0.3447181284427643, 0.713766872882843, 0.6391869187355042, 0.399161159992218, 0.43176013231277466, 0.614527702331543, 0.0700421929359436, 0.8224067091941833, 0.65342116355896, 0.7263424396514893, 0.5369229912757874, 0.11047711223363876, 0.4050356149673462, 0.40537357330322266, 0.3210429847240448, 0.029950324445962906, 0.73725426197052, 0.10978446155786514, 0.6063081622123718, 0.7032175064086914, 0.6347863078117371, 0.95914226770401, 0.10329815745353699, 0.8671671748161316, 0.02919023483991623, 0.534916877746582, 0.4042436182498932, 0.5241838693618774, 0.36509987711906433, 0.19056691229343414, 0.01912289671599865, 0.5181497931480408, 0.8427768349647522, 0.3732159435749054, 0.2228638231754303, 0.080532006919384, 0.0853109210729599, 0.22139644622802734, 0.10001406073570251, 0.26503971219062805, 0.06614946573972702, 0.06560486555099487, 0.8562761545181274, 0.1621202677488327, 0.5596824288368225, 0.7734555602073669, 0.4564095735549927, 0.15336887538433075, 0.19959613680839539, 0.43298420310020447, 0.52823406457901, 0.3494403064250946, 0.7814795970916748, 0.7510216236114502, 0.9272118210792542, 0.028952548280358315, 0.8956912755966187, 0.39256879687309265, 0.8783724904060364, 0.690784752368927, 0.987348735332489, 0.7592824697494507, 0.3645446300506592, 0.5010631680488586, 0.37638914585113525, 0.364911824464798, 0.2609044909477234, 0.49597030878067017, 0.6817399263381958, 0.27734026312828064, 0.5243797898292542, 0.117380291223526, 0.1598452925682068, 0.04680635407567024, 0.9707314372062683, 0.0038603513967245817, 0.17857997119426727, 0.6128667593002319, 0.08136960119009018, 0.8818964958190918, 0.7196201682090759, 0.9663899540901184, 0.5076355338096619, 0.3004036843776703, 0.549500584602356, 0.9308187365531921, 0.5207614302635193, 0.2672070264816284, 0.8773987889289856, 0.3719187378883362, 0.0013833499979227781, 0.2476850152015686, 0.31823351979255676, 0.8587774634361267, 0.4585031569004059, 0.4445872902870178, 0.33610227704048157, 0.880678117275238, 0.9450267553329468, 0.9918903112411499, 0.3767412602901459, 0.9661474227905273, 0.7918795943260193, 0.675689160823822, 0.24488948285579681, 0.21645726263523102, 0.1660478264093399, 0.9227566123008728, 0.2940766513347626, 0.4530942440032959, 0.49395784735679626, 0.7781715989112854, 0.8442349433898926, 0.1390727013349533, 0.4269043505191803, 0.842854917049408, 0.8180332779884338, 0.10241375863552094, 0.15638335049152374, 0.304198682308197, 0.07535906881093979, 0.4246630072593689, 0.10761770606040955, 0.5682175755500793, 0.24655693769454956, 0.5964330434799194, 0.11752564460039139, 0.9758838415145874, 0.9325612187385559, 0.39179694652557373, 0.24217858910560608, 0.2503982186317444, 0.4833935499191284, 0.0399928018450737, 0.6397051215171814, 0.408302903175354, 0.3774065673351288, 0.8093649744987488, 0.7090354561805725, 0.9543338418006897, 0.3519362509250641, 0.8975427746772766, 0.7699671983718872, 0.35742464661598206, 0.6216654181480408, 0.28856995701789856, 0.8743999004364014, 0.11242731660604477, 0.21243436634540558, 0.183033287525177, 0.4030260145664215, 0.7452329397201538, 0.5269074440002441, 0.48767632246017456, 0.0005459649255499244, 0.4254017174243927, 0.06355377286672592, 0.20825324952602386, 0.9323939681053162, 0.21539820730686188, 0.8583376407623291, 0.8028934001922607, 0.1591462343931198, 0.6057119369506836, 0.1156618744134903, 0.7278881669044495, 0.6374622583389282, 0.8119385838508606, 0.47938454151153564, 0.914863109588623, 0.049348946660757065, 0.2928885519504547, 0.715052604675293, 0.4181092083454132, 0.1729513555765152, 0.10721074789762497, 0.8173391222953796, 0.47314298152923584, 0.8822836875915527, 0.7332891225814819, 0.40972620248794556, 0.37351101636886597, 0.5156383514404297, 0.8890599608421326, 0.7372785806655884, 0.005152964498847723, 0.6941578388214111, 0.9195073843002319, 0.7104557752609253, 0.17700578272342682, 0.483518123626709, 0.14031602442264557, 0.3589952886104584, 0.9371170401573181, 0.923305332660675, 0.2828368544578552, 0.339631050825119], "onnx_backend_attention_4d_gqa_input_v": [0.6002128720283508, 0.9631972908973694, 0.14780133962631226, 0.2569166421890259, 0.8735568523406982, 0.4918922185897827, 0.898961067199707, 0.18551789224147797, 0.5326685905456543, 0.32626962661743164, 0.3165425658226013, 0.44687697291374207, 0.43307745456695557, 0.35734689235687256, 0.9149707555770874, 0.7317441701889038, 0.7275469899177551, 0.28991344571113586, 0.577709436416626, 0.7791794538497925, 0.795590341091156, 0.34453046321868896, 0.7708727717399597, 0.7358939051628113, 0.14150647819042206, 0.8659454584121704, 0.441321462392807, 0.4864104390144348, 0.44836917519569397, 0.5678460001945496, 0.6211692690849304, 0.49817955493927, 0.8667885661125183, 0.6277347803115845, 0.4014279544353485, 0.41669175028800964, 0.8108386397361755, 0.3481919467449188, 0.21145479381084442, 0.05938318744301796, 0.8760268688201904, 0.9185464382171631, 0.12012018263339996, 0.33447372913360596, 0.17537206411361694, 0.11589846760034561, 0.8998667597770691, 0.0568772591650486, 0.9804856777191162, 0.09645085781812668, 0.8634706735610962, 0.5665060877799988, 0.3679174780845642, 0.3423423767089844, 0.7573641538619995, 0.3145732879638672, 0.6573188900947571, 0.5173260569572449, 0.48496565222740173, 0.9011621475219727, 0.5546450614929199, 0.8268616199493408, 0.7255735397338867, 0.03855724632740021, 0.773110032081604, 0.2168702483177185, 0.9031496644020081, 0.0429241918027401, 0.3330720365047455, 0.09973295032978058, 0.4755891263484955, 0.820022463798523, 0.29818734526634216, 0.15093488991260529, 0.3302670419216156, 0.8138801455497742, 0.14038395881652832, 0.22736245393753052, 0.06885196268558502, 0.7057100534439087, 0.39523324370384216, 0.310839980840683, 0.7186263799667358, 0.33597755432128906, 0.7277712821960449, 0.8151993751525879, 0.21766284108161926, 0.9738187193870544, 0.16235794126987457, 0.29084089398384094, 0.1797952950000763, 0.34550565481185913, 0.480060875415802, 0.5221758484840393, 0.8536060452461243, 0.8894479274749756, 0.22010385990142822, 0.6228940486907959, 0.11149605363607407, 0.4589698612689972, 0.32233354449272156, 0.3165007531642914, 0.48258423805236816, 0.7298276424407959, 0.06918265670537949, 0.8791733384132385, 0.7348137497901917, 0.1764993816614151, 0.9391608834266663, 0.5063122510910034, 0.9998085498809814, 0.19725947082042694, 0.5349081754684448, 0.2902480363845825, 0.3041735589504242, 0.5910654067993164, 0.9217190742492676, 0.805263876914978, 0.7239413857460022, 0.5591737627983093, 0.9222984910011292, 0.4923613965511322, 0.8738321661949158, 0.8339816331863403, 0.21383534371852875, 0.7712254524230957, 0.012171156704425812, 0.32282954454421997, 0.22956743836402893, 0.5068629384040833, 0.7368531823158264, 0.0976763665676117, 0.5149222016334534, 0.9384120106697083, 0.22864654660224915, 0.6771411299705505, 0.5928802490234375, 0.010063695721328259, 0.4758262038230896, 0.7087703943252563, 0.043975431472063065, 0.8795214891433716, 0.5200814008712769, 0.030661048367619514, 0.22441361844539642, 0.9536756873130798, 0.5823197364807129, 0.10747256875038147, 0.28754448890686035, 0.45670363306999207, 0.020950069651007652, 0.4116155207157135, 0.4894586205482483, 0.24367786943912506, 0.5886390209197998, 0.7532401084899902, 0.23583422601222992, 0.6204999089241028, 0.6396222710609436, 0.9485403299331665, 0.7782761454582214, 0.8483452796936035, 0.4904198944568634, 0.18534858524799347, 0.9958152770996094, 0.1293557584285736, 0.4714573323726654, 0.06809309870004654, 0.9438508749008179, 0.9649249315261841, 0.7193890810012817, 0.3499928414821625, 0.2543824017047882, 0.2653033137321472, 0.12729401886463165, 0.5258089303970337, 0.14181727170944214, 0.31673067808151245, 0.6267064809799194, 0.7275435924530029, 0.024272704496979713, 0.4301159977912903, 0.6521245837211609, 0.8532459735870361, 0.4753247797489166, 0.9692058563232422, 0.2656325399875641, 0.013508706353604794, 0.48375287652015686, 0.25611379742622375, 0.8237176537513733, 0.23277267813682556, 0.3106292188167572, 0.7912274599075317, 0.7151432633399963, 0.5580512285232544, 0.7049480676651001, 0.41863685846328735, 0.005310047417879105, 0.011355128139257431, 0.5112217664718628, 0.08329097926616669, 0.05107548087835312, 0.9655166268348694, 0.8590026497840881, 0.15202723443508148, 0.0006642185617238283, 0.9416677951812744, 0.27832528948783875, 0.1858976036310196, 0.6915081143379211, 0.10890373587608337, 0.2646495997905731, 0.9750946760177612, 0.6394627690315247, 0.5206778049468994, 0.39791861176490784, 0.7745009660720825, 0.14095747470855713, 0.9673377871513367, 0.8611230254173279, 0.6176570057868958, 0.04290619120001793, 0.7008556723594666, 0.9132843613624573, 0.5245770812034607, 0.35422483086586, 0.12027734518051147, 0.754901111125946, 0.8850218653678894, 0.10025174170732498, 0.7589845657348633, 0.01706048659980297, 0.9670549035072327, 0.6150580048561096, 0.5524390339851379, 0.2959498465061188, 0.9292916655540466, 0.26590561866760254, 0.8281466364860535, 0.9851086735725403, 0.7833966612815857, 0.5189899206161499, 0.06607426702976227, 0.4724137783050537, 0.4382559359073639, 0.20279604196548462, 0.4235876500606537, 0.35775789618492126, 0.16368426382541656, 0.44137415289878845, 0.2627999484539032, 0.5220624208450317, 0.03516006097197533, 0.9062314033508301, 0.8163642883300781, 0.5525813102722168, 0.8518086075782776, 0.9623950719833374, 0.11052229255437851, 0.6308318376541138, 0.9979940056800842, 0.9878891706466675, 0.6033229827880859, 0.12802086770534515, 0.5831928253173828, 0.0020646355114877224, 0.19891133904457092, 0.9561231732368469, 0.33044058084487915, 0.6383901238441467, 0.28085950016975403, 0.9478219151496887, 0.7285587191581726, 0.3296511471271515, 0.7917613983154297, 0.10816552489995956, 0.3923189342021942, 0.22121812403202057, 0.68372642993927, 0.10244628041982651, 0.39702582359313965, 0.27664974331855774, 0.5063429474830627, 0.34989768266677856, 0.7064105868339539, 0.024577023461461067, 0.6339869499206543], "onnx_backend_attention_4d_attn_mask_4d_input_attn_mask": [0.9661474227905273, 0.7918795943260193, 0.675689160823822, 0.24488948285579681, 0.21645726263523102, 0.1660478264093399, 0.9227566123008728, 0.2940766513347626, 0.4530942440032959, 0.49395784735679626, 0.7781715989112854, 0.8442349433898926, 0.1390727013349533, 0.4269043505191803, 0.842854917049408, 0.8180332779884338, 0.10241375863552094, 0.15638335049152374, 0.304198682308197, 0.07535906881093979, 0.4246630072593689, 0.10761770606040955, 0.5682175755500793, 0.24655693769454956, 0.5964330434799194, 0.11752564460039139, 0.9758838415145874, 0.9325612187385559, 0.39179694652557373, 0.24217858910560608, 0.2503982186317444, 0.4833935499191284, 0.0399928018450737, 0.6397051215171814, 0.408302903175354, 0.3774065673351288, 0.8093649744987488, 0.7090354561805725, 0.9543338418006897, 0.3519362509250641, 0.8975427746772766, 0.7699671983718872, 0.35742464661598206, 0.6216654181480408, 0.28856995701789856, 0.8743999004364014, 0.11242731660604477, 0.21243436634540558, 0.183033287525177, 0.4030260145664215, 0.7452329397201538, 0.5269074440002441, 0.48767632246017456, 0.0005459649255499244, 0.4254017174243927, 0.06355377286672592, 0.20825324952602386, 0.9323939681053162, 0.21539820730686188, 0.8583376407623291, 0.8028934001922607, 0.1591462343931198, 0.6057119369506836, 0.1156618744134903, 0.7278881669044495, 0.6374622583389282, 0.8119385838508606, 0.47938454151153564, 0.914863109588623, 0.049348946660757065, 0.2928885519504547, 0.715052604675293, 0.4181092083454132, 0.1729513555765152, 0.10721074789762497, 0.8173391222953796, 0.47314298152923584, 0.8822836875915527, 0.7332891225814819, 0.40972620248794556, 0.37351101636886597, 0.5156383514404297, 0.8890599608421326, 0.7372785806655884, 0.005152964498847723, 0.6941578388214111, 0.9195073843002319, 0.7104557752609253, 0.17700578272342682, 0.483518123626709, 0.14031602442264557, 0.3589952886104584, 0.9371170401573181, 0.923305332660675, 0.2828368544578552, 0.339631050825119, 0.6002128720283508, 0.9631972908973694, 0.14780133962631226, 0.2569166421890259, 0.8735568523406982, 0.4918922185897827, 0.898961067199707, 0.18551789224147797, 0.5326685905456543, 0.32626962661743164, 0.3165425658226013, 0.44687697291374207, 0.43307745456695557, 0.35734689235687256, 0.9149707555770874, 0.7317441701889038, 0.7275469899177551, 0.28991344571113586, 0.577709436416626, 0.7791794538497925, 0.795590341091156, 0.34453046321868896, 0.7708727717399597, 0.7358939051628113, 0.14150647819042206, 0.8659454584121704, 0.441321462392807, 0.4864104390144348, 0.44836917519569397, 0.5678460001945496, 0.6211692690849304, 0.49817955493927, 0.8667885661125183, 0.6277347803115845, 0.4014279544353485, 0.41669175028800964, 0.8108386397361755, 0.3481919467449188, 0.21145479381084442, 0.05938318744301796, 0.8760268688201904, 0.9185464382171631, 0.12012018263339996, 0.33447372913360596, 0.17537206411361694, 0.11589846760034561, 0.8998667597770691, 0.0568772591650486], "scalar_mask_all_neg_inf_row_zero_f32_input_attn_mask": [0, -0.25, -0.5, -0.75, "-Infinity", "-Infinity", "-Infinity", "-Infinity", -0.1, 0, -0.2, -0.3, -0.4, -0.05, 0, -0.15] }, "cases": [ { "name": "masked_decode_splitk_boundary_gqa_f32", "provenance": { "notes": "Masked decode with a long key/value context; a rank-2 all-zero additive mask broadcasts across query heads." }, "attrs": { "is_causal": 0, "scale": 0.1767766952966369 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 1, 32], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.2 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1024], "data": { "kind": "constant", "value": 0.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 1, 32], "tolerance": 0.00003, "relTolerance": 0.00002 } } }, { "name": "masked_decode_splitk_boundary_gqa_f32_offset_value_scale", "provenance": { "notes": "Offsetting V around 1.0 keeps the 1,024-key average away from zero, making scale errors in the per-split denominators and cross-split merge observable on masked decode split-K." }, "attrs": { "is_causal": 0, "scale": 0.1767766952966369 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 1, 32], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.2, "offset": 1.0 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1024], "data": { "kind": "constant", "value": 0.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 1, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "masked_decode_splitk_bool_rank4_gqa_f32", "provenance": { "notes": "Boolean rank-4 broadcast-mask parity for masked split-K decode; includes both attended and masked keys." }, "attrs": { "is_causal": 0, "scale": 0.1767766952966369 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 1, 32], "data": { "kind": "fillFloat32", "sinStep": 0.012, "cosStep": 0.018, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.024, "cosStep": 0.032, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.038, "cosStep": 0.044, "scale": 0.2 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 1024], "data": { "kind": "cycle", "values": [1, 1, 1, 0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 1, 32], "tolerance": 0.00003, "relTolerance": 0.00002 } } }, { "name": "splitk_bool_mask_masked_head_decode_zero_f32", "attrs": { "is_causal": 0, "scale": 0.1767766952966369 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 1, 32], "data": { "kind": "fillFloat32", "sinStep": 0.012, "cosStep": 0.018, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.024, "cosStep": 0.032, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.038, "cosStep": 0.044, "scale": 0.2 } }, "attn_mask": { "dtype": "bool", "shape": [1, 4, 1, 1024], "data": { "kind": "linspace", "start": 0, "end": 2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 1, 32], "tolerance": 0.00003, "relTolerance": 0.00002 } }, "provenance": { "notes": "A boolean mask over the key sequence is false for the first two heads and true for the other two. Fully masked heads must produce zero rows while the open heads retain their attention results." } }, { "name": "splitk_bool_mask_masked_head_decode_zero_f32_offset_value", "provenance": { "notes": "Offsetting V to oscillate about 1.0 makes an accidental mean-of-V fallback an O(1) error while the required zero rows of the two fully masked heads stay exactly zero. This prevents the split-K zero-denominator guard from passing only because a zero-mean V fixture nearly cancels." }, "attrs": { "is_causal": 0, "scale": 0.1767766952966369 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 1, 32], "data": { "kind": "fillFloat32", "sinStep": 0.012, "cosStep": 0.018, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.024, "cosStep": 0.032, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.038, "cosStep": 0.044, "scale": 0.2, "offset": 1.0 } }, "attn_mask": { "dtype": "bool", "shape": [1, 4, 1, 1024], "data": { "kind": "linspace", "start": 0, "end": 2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 1, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "splitk_bool_mask_masked_rows_short_query_zero_f32", "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 4, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.6 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 64], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.033, "scale": 0.6 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.045, "scale": 0.15 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 4, 1024], "data": { "kind": "linspace", "start": 0, "end": 2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 4, 64], "tolerance": 0.00003, "relTolerance": 0.00002 } }, "provenance": { "notes": "Logical-bool mask on the short-query split-K arm, false for the first half of the query rows and true for the rest. Its per-query mask rows use a different index path from decode, so a masked row must still carry zero probability mass through every split while its neighbours in the same tile keep their answers." } }, { "name": "cluster_splitkv_bool_mask_masked_rows_zero_f32", "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 16, 128], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.6 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.033, "scale": 0.6 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.045, "scale": 0.15 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 16, 1024], "data": { "kind": "linspace", "start": 0, "end": 2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 16, 128], "tolerance": 0.00003, "relTolerance": 0.00002 } }, "provenance": { "notes": "Logical-bool mask on the masked clustered split-key/value tile (sixteen queries over 1024 keys at head dim 128), false for the first eight query rows. Every split of a masked row carries zero mass and the merge emits a zero row, beside open rows of the same tile that must not be zeroed with it." } }, { "name": "masked_short_query_splitk_rank3_gqa_f16", "provenance": { "notes": "F16 short-query and rank-3 broadcast-mask parity for the masked split-K family." }, "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 4, 4, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.6 } }, "k": { "dtype": "float16", "shape": [1, 1, 1024, 64], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.033, "scale": 0.6 } }, "v": { "dtype": "float16", "shape": [1, 1, 1024, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.045, "scale": 0.15 } }, "attn_mask": { "dtype": "float16", "shape": [1, 4, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.009, "scale": 0.8 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 4, 4, 64], "tolerance": 0.00006, "relTolerance": 0.002 } } }, { "name": "masked_cluster_splitkv_rank2_additive_f32", "provenance": { "notes": "Sixteen masked queries over 1024 keys at head dim 128 take the masked clustered split-key/value tile (the per-query split-K route keeps rank-3 masks and up to sixteen queries elsewhere); rank-2 additive mask." }, "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 16, 128], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.15 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.033, "scale": 0.15 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.045, "scale": 0.15 } }, "attn_mask": { "dtype": "float32", "shape": [16, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.009, "scale": 0.05 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 16, 128], "tolerance": 0.000004, "relTolerance": 0.00002 } } }, { "name": "masked_cluster_splitkv_bool_rank4_f32", "provenance": { "notes": "Sixteen queries with a rank-4 logical-bool mask on the masked clustered split-key/value tile: rejected keys carry no mass across the split merge." }, "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 16, 128], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.15 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.033, "scale": 0.15 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.045, "scale": 0.15 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 16, 1024], "data": { "kind": "cycle", "values": [1, 1, 1, 0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 16, 128], "tolerance": 0.000004, "relTolerance": 0.00002 } } }, { "name": "masked_cluster_splitkv_rank2_additive_f16", "provenance": { "notes": "Sixteen f16 queries with a rank-2 additive mask on the masked clustered split-key/value tile." }, "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 4, 16, 128], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.15 } }, "k": { "dtype": "float16", "shape": [1, 1, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.033, "scale": 0.15 } }, "v": { "dtype": "float16", "shape": [1, 1, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.045, "scale": 0.15 } }, "attn_mask": { "dtype": "float16", "shape": [16, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.009, "scale": 0.05 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 4, 16, 128], "tolerance": 0.00002, "relTolerance": 0.006 } } }, { "name": "masked_short_query_splitk_rank3_gqa_f16_offset_value_scale", "provenance": { "notes": "Offsetting V around 1.0 makes normalization errors observable in masked short-query split-K with a rank-3 broadcast mask, including the per-split maximum and denominator that incorporate the additive mask." }, "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 4, 4, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.6 } }, "k": { "dtype": "float16", "shape": [1, 1, 1024, 64], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.033, "scale": 0.6 } }, "v": { "dtype": "float16", "shape": [1, 1, 1024, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.045, "scale": 0.15, "offset": 1.0 } }, "attn_mask": { "dtype": "float16", "shape": [1, 4, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.009, "scale": 0.8 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 4, 4, 64], "tolerance": 0.01, "relTolerance": 0.01 } } }, { "name": "masked_short_query_splitk_q16_upper_boundary_f32", "provenance": { "notes": "Masked short-query attention with a long key/value context and an additive mask." }, "attrs": { "is_causal": 0, "scale": 0.1767766952966369 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 16, 32], "data": { "kind": "fillFloat32", "sinStep": 0.014, "cosStep": 0.016, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.026, "cosStep": 0.034, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 1024, 32], "data": { "kind": "fillFloat32", "sinStep": 0.04, "cosStep": 0.046, "scale": 0.2 } }, "attn_mask": { "dtype": "float32", "shape": [16, 1024], "data": { "kind": "constant", "value": 0.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 16, 32], "tolerance": 0.00004, "relTolerance": 0.00002 } } }, { "name": "masked_short_query_splitk_q17_control_f32", "provenance": { "notes": "A query length of 17 lies outside the masked short-query split-K range and exercises the general masked attention path." }, "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 16, 17, 64], "data": { "kind": "fillFloat32", "sinStep": 0.015, "cosStep": 0.021, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 4, 1024, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.035, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 4, 1024, 64], "data": { "kind": "fillFloat32", "sinStep": 0.041, "cosStep": 0.047, "scale": 0.2 } }, "attn_mask": { "dtype": "float32", "shape": [17, 1024], "data": { "kind": "constant", "value": 0.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 16, 17, 64], "tolerance": 0.00005, "relTolerance": 0.00002 } } }, { "name": "rank4_vhead320_score_cache_f32", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 8, 16], "data": { "kind": "linspace", "start": -0.5, "end": 0.5 } }, "k": { "dtype": "float32", "shape": [1, 2, 8, 16], "data": { "kind": "linspace", "start": -0.25, "end": 0.75 } }, "v": { "dtype": "float32", "shape": [1, 2, 8, 320], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 8, 320], "tolerance": 0.00001 } } }, { "name": "rank4_vhead320_tiled_apply_f16", "provenance": { "notes": "Rank-four float16 attention with 1 query head and 1 key/value head, key/query head width 16 and value head width 320." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 1, 4, 16], "data": { "kind": "linspace", "start": -0.5, "end": 0.5 } }, "k": { "dtype": "float16", "shape": [1, 1, 4, 16], "data": { "kind": "linspace", "start": -0.25, "end": 0.75 } }, "v": { "dtype": "float16", "shape": [1, 1, 4, 320], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 4, 320], "tolerance": 0.006 } } }, { "name": "mha_rank4_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [10.0, 20.0, 30.0, 40.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 2, 2], "tolerance": 0.000002 } } }, { "name": "rank4_no_mask_small_tiled_4h_s32_d16", "provenance": { "notes": "Compact correctness case for the all-feature rank-4 tiled path." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 32, 16], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "k": { "dtype": "float32", "shape": [1, 4, 32, 16], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04 } }, "v": { "dtype": "float32", "shape": [1, 4, 32, 16], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 32, 16], "tolerance": 0.00001 } } }, { "name": "rank4_no_mask_tiled_4h_s32_d32", "provenance": { "notes": "Four heads with 32 queries and head dimension 32 check a short-query attention shape." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 32, 32], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 32, 32], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 32, 32], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 32, 32], "tolerance": 0.00002 } } }, { "name": "rank4_no_mask_flash_4h_s64_d60", "provenance": { "notes": "Head dimension 60 exercises the direct flash range for a realistic hidden size of 960." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 60], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 60], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 60], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 60], "tolerance": 0.00005, "relTolerance": 0.00002 } } }, { "name": "short_prompt_generic_flash_16h4kv_q16_kv64_d80", "provenance": { "notes": "A short-prompt GQA shape with 16 queries checks grouped heads and per-query output." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 16, 16, 80], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 80], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 80], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 16, 16, 80], "tolerance": 0.00003 } } }, { "name": "f32_large_logit_gap_tail_amplified_by_value_gpu_gap", "skipGpu": { "category": "permanent", "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: Metal flushes denormals to zero in the ALU; the softmax exp/normalize and value-weighting produce subnormal probabilities that flush on GPU. Permanent FTZ limitation." }, "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DDefault", "notes": "A large but finite score gap leaves a valid subnormal softmax tail; multiplying that tail by a large finite V turns it into an ordinary output contribution." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [0.0, -87.5] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [0.0, 1e+38] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0.0001 } } }, { "name": "ort_rank3_default_mha", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention3DDefault", "notes": "Small rank-3 packed-head MHA fixture using ORT's q_num_heads/kv_num_heads layout convention." }, "attrs": { "q_num_heads": 2, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 0.5, 0.5, 1.0, -1.0] } }, "k": { "dtype": "float32", "shape": [1, 3, 4], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 1.0, 0.0, 1.0, 1.0, -1.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 3, 4], "data": { "kind": "values", "values": [10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0, 80.0, 90.0, 100.0, 110.0, 120.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 4], "tolerance": 0.000003 } } }, { "name": "ort_rank3_gqa_packed_heads", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention3DGqaAttn", "notes": "Compact rank-3 GQA fixture: four packed query heads share two packed K/V heads." }, "attrs": { "q_num_heads": 4, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [2.0, 1.0, -1.0, 0.5, 0.0, -2.0, 1.0, 1.5] } }, "k": { "dtype": "float32", "shape": [1, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.5, 1.0, -1.0, 2.0] } }, "v": { "dtype": "float32", "shape": [1, 3, 2], "data": { "kind": "values", "values": [10.0, 100.0, 20.0, 200.0, 30.0, 300.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 4], "tolerance": 0.000003 } } }, { "name": "causal_gqa_f32", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 1.0, 1.0, -1.0, 1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [5.0, 9.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 2, 1], "tolerance": 0.000001 } } }, { "name": "bool_mask_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [5.0, 9.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [0, 1] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0.000001 } } }, { "name": "float_mask_all_large_negative_uniform_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, -1.0, 0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 3, 1], "data": { "kind": "values", "values": [10.0, 20.0, 40.0] } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 1, 3], "data": { "kind": "constant", "value": -3.402823e+38 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0.000001 } } }, { "name": "rank4_f16", "inputs": { "q": { "dtype": "float16", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "k": { "dtype": "float16", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "v": { "dtype": "float16", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [10.0, 20.0, 30.0, 40.0] } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 2, 2], "tolerance": 0.02 } } }, { "name": "float_mask_additive_bias_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [5.0, 50.0, 9.0, 90.0] } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [0.0, -10000.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 2], "tolerance": 0.000001 } } }, { "name": "gqa_four_query_heads_two_kv_heads", "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 1.0, 1.0, -1.0, 1.0, 0.5, 1.0, 1.0, -0.5, 2.0, 0.0, 0.0, 2.0] } }, "k": { "dtype": "float32", "shape": [1, 2, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 1.0, 1.0, -1.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 2, 2, 1], "data": { "kind": "values", "values": [5.0, 9.0, 20.0, 30.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 2, 1], "tolerance": 0.000003 } } }, { "name": "explicit_scale_causal_f32", "attrs": { "scale": 0.5, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 1.0, 1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 1.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 3, 1], "data": { "kind": "values", "values": [10.0, 20.0, 30.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 3, 1], "tolerance": 0.000001 } } }, { "name": "ort_softcap_changes_logits", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DSoftCap", "notes": "Raw logits [0.5, 2.0] are tanh-softcapped before softmax, so ignoring softcap changes the output." }, "attrs": { "softcap": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [0.5, 2.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [10.0, 20.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0.000001, "data": { "kind": "values", "values": [16.22908182510569] } } } }, { "name": "ort_softcap_neg_inf_mask_poison_v", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention_Unfused_Softcap_NegInfMask_PoisonV_CPU", "notes": "Compact ordering guard: two unmasked logits are softcapped before softmax, while a -inf-masked poison value must contribute zero." }, "attrs": { "softcap": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 3, 1], "data": { "kind": "values", "values": [0.5, 2.0, 3.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 3, 1], "data": { "kind": "values", "values": [10.0, 20.0, 1000.0] } }, "attn_mask": { "dtype": "float32", "shape": [1, 3], "data": { "kind": "values", "values": [0.0, 0.0, "-Infinity"] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0.000001, "data": { "kind": "values", "values": [16.22908182510569] } } } }, { "name": "ort_causal_cross_attention_upper_left", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DCausalCrossAttentionUpperLeft" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 3, 4], "data": { "kind": "values", "values": [1.0, 0.5, 0.3, 0.2, 0.4, 0.8, 0.1, 0.6, 0.7, 0.3, 0.9, 0.5] } }, "k": { "dtype": "float32", "shape": [1, 1, 5, 4], "data": { "kind": "values", "values": [0.2, 0.4, 0.6, 0.8, 0.1, 0.3, 0.5, 0.7, 0.9, 0.1, 0.2, 0.3, 0.5, 0.6, 0.7, 0.8, 0.3, 0.2, 0.1, 0.4] } }, "v": { "dtype": "float32", "shape": [1, 1, 5, 4], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.5, 0.5, 0.5, 0.5] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 3, 4], "tolerance": 0.000002 } } }, { "name": "ort_causal_square_no_past", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DCausalSquareNoPast" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 4, 4], "data": { "kind": "values", "values": [1.0, 0.5, 0.3, 0.2, 0.4, 0.8, 0.1, 0.6, 0.7, 0.3, 0.9, 0.5, 0.2, 0.6, 0.4, 0.8] } }, "k": { "dtype": "float32", "shape": [1, 1, 4, 4], "data": { "kind": "values", "values": [0.2, 0.4, 0.6, 0.8, 0.1, 0.3, 0.5, 0.7, 0.9, 0.1, 0.2, 0.3, 0.5, 0.6, 0.7, 0.8] } }, "v": { "dtype": "float32", "shape": [1, 1, 4, 4], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 4, 4], "tolerance": 0.000002 } } }, { "name": "bool_mask_masked_row_zero_f32", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DAttnMaskBoolAllFalse", "notes": "Q, K and V are onnxruntime's Attention4DAttnMaskBoolAllFalse tensors; the mask is false across the first query row and true across the second, so the fully masked row must emit zero while the open row keeps its softmax answer — a whole-output zero is then a visible failure rather than the expected value." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [2.0, -1.0, -1.0, 2.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 1.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 3, 2], "data": { "kind": "values", "values": [2.0, 10.0, 4.0, 20.0, 8.0, 40.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 2, 3], "data": { "kind": "values", "values": [0, 0, 0, 1, 1, 1] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 2, 2], "tolerance": 0.000001 } } }, { "name": "ort_causal_basic_two_batch_f32", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DAttnIsCausalBasic" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 1.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] } }, "k": { "dtype": "float32", "shape": [2, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] } }, "v": { "dtype": "float32", "shape": [2, 1, 3, 2], "data": { "kind": "values", "values": [0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 1, 3, 2], "tolerance": 0.000002 } } }, { "name": "ort_causal_basic_two_batch_f16", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DAttnIsCausalBasicFloat16" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 1.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] } }, "k": { "dtype": "float16", "shape": [2, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] } }, "v": { "dtype": "float16", "shape": [2, 1, 3, 2], "data": { "kind": "values", "values": [0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 1, 3, 2], "tolerance": 0.02 } } }, { "name": "ort_causal_basic_different_kv_length", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DAttnIsCausalBasicDifferentSequenceLength" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 1, 3, 2], "data": { "kind": "values", "values": [1.0, 1.0, 0.0, 1.0, 2.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] } }, "k": { "dtype": "float32", "shape": [2, 1, 4, 2], "data": { "kind": "values", "values": [1.0, 0.0, 1.0, 1.0, 1.0, 2.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 1.0, 2.0] } }, "v": { "dtype": "float32", "shape": [2, 1, 4, 2], "data": { "kind": "values", "values": [0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 2.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.0, 2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 1, 3, 2], "tolerance": 0.000002 } } }, { "name": "v_head_size_differs_from_qk", "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 3], "data": { "kind": "values", "values": [1.0, 2.0, 4.0, 8.0, 16.0, 32.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 2, 3], "tolerance": 0.000002 } } }, { "name": "onnx_backend_attention_4d", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d" }, "attrs": {}, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_v" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_causal", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_causal" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_v" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_diff_heads_sizes", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_diff_heads_sizes" }, "attrs": {}, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 10], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_diff_heads_sizes_input_v" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 10], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_diff_heads_sizes_causal", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_diff_heads_sizes_causal" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 10], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_diff_heads_sizes_input_v" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 10], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_diff_heads_sizes_scaled", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_diff_heads_sizes_scaled" }, "attrs": { "scale": 0.009999999776482582 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 10], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_diff_heads_sizes_input_v" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 10], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_gqa", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_gqa" }, "attrs": {}, "inputs": { "q": { "dtype": "float32", "shape": [2, 9, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_v" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 9, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_gqa_causal", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_gqa_causal" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 9, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_v" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 9, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_gqa_scaled", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_gqa_scaled" }, "attrs": { "scale": 0.009999999776482582 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 9, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_gqa_input_v" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 9, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_attn_mask_4d", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_attn_mask_4d" }, "attrs": {}, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_v" } } }, "attn_mask": { "dtype": "float32", "shape": [2, 3, 4, 6], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_attn_mask_4d_input_attn_mask" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_attn_mask_4d_causal", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_attn_mask_4d_causal" }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_q" } } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_k" } } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_input_v" } } }, "attn_mask": { "dtype": "float32", "shape": [2, 3, 4, 6], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_attention_4d_attn_mask_4d_input_attn_mask" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_attn_mask_3d", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_attn_mask_3d", "notes": "Uses the official rank-4 Q/K/V shape and the standard [B, 1, Q, K] additive mask, which broadcasts across query heads." }, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.5 } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.023, "scale": 0.5 } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.029, "scale": 0.5 } }, "attn_mask": { "dtype": "float32", "shape": [2, 1, 4, 6], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.031, "scale": 0.25 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_diff_heads_sizes_attn_mask_shape", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_diff_heads_sizes_attn_mask", "notes": "Uses the official rank-4 Q/K/V shape with V head size 10 and the standard [Q, K] additive mask broadcast across batch and heads." }, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.011, "scale": 0.5 } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.013, "scale": 0.5 } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 10], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.017, "scale": 0.5 } }, "attn_mask": { "dtype": "float32", "shape": [4, 6], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.037, "scale": 0.25 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 10], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_scaled_shape", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_scaled", "notes": "Uses the official rank-4 shape and scale attribute with deterministic generated data." }, "attrs": { "scale": 0.01 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 3, 4, 8], "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.031, "scale": 0.5 } }, "k": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.037, "scale": 0.5 } }, "v": { "dtype": "float32", "shape": [2, 3, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.041, "scale": 0.5 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 4, 8], "tolerance": 0.00005 } } }, { "name": "onnx_backend_attention_4d_fp16_shape", "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_attention_4d_fp16", "notes": "Uses the official rank-4 fp16 shape with deterministic generated data." }, "attrs": {}, "inputs": { "q": { "dtype": "float16", "shape": [2, 3, 4, 8], "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.031, "scale": 0.5 } }, "k": { "dtype": "float16", "shape": [2, 3, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.037, "scale": 0.5 } }, "v": { "dtype": "float16", "shape": [2, 3, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.041, "scale": 0.5 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 3, 4, 8], "tolerance": 0.005 } } }, { "name": "flash_noncausal_d64_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 128, 64], "tolerance": 0.00005 } } }, { "name": "flash_noncausal_gqa_d64_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 256, 64], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.029, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 2, 256, 64], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.019, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 2, 256, 64], "data": { "kind": "fillFloat32", "sinStep": 0.043, "cosStep": 0.021, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 256, 64], "tolerance": 0.00005 } } }, { "name": "flash_decode_gqa_d64_kv512_f32", "attrs": { "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 1, 64], "data": { "kind": "fillFloat32", "sinStep": 0.009, "cosStep": 0.021, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.035, "cosStep": 0.017, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.043, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 1, 64], "tolerance": 0.00005 } } }, { "name": "flash_noncausal_d128_multi_tile_f32", "provenance": { "notes": "The f32 Q32 broadcast specialization maps 32 vec4 head segments to a fixed 32-lane subgroup, exercising one segment per lane with a complete head." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.023, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.013, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.047, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 512, 128], "tolerance": 0.00005 } } }, { "name": "flash_noncausal_tail_tile_d96_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 300, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 300, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 96], "tolerance": 0.00005 } } }, { "name": "flash_noncausal_kv_shorter_than_tile_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 32, 64], "data": { "kind": "fillFloat32", "sinStep": 0.015, "cosStep": 0.025, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 8, 8, 64], "data": { "kind": "fillFloat32", "sinStep": 0.033, "cosStep": 0.045, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 8, 8, 64], "data": { "kind": "fillFloat32", "sinStep": 0.047, "cosStep": 0.035, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 32, 64], "tolerance": 0.00005 } } }, { "name": "flash_noncausal_explicit_scale_f32", "attrs": { "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.012, "cosStep": 0.026, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.034, "cosStep": 0.018, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.022, "cosStep": 0.042, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 128, 64], "tolerance": 0.00005 } } }, { "name": "flash_noncausal_batch2_gqa_d128_f32", "inputs": { "q": { "dtype": "float32", "shape": [2, 8, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.014, "cosStep": 0.028, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 2, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.036, "cosStep": 0.016, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 2, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.024, "cosStep": 0.044, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 8, 128, 128], "tolerance": 0.00005 } } }, { "name": "flash_noncausal_d64_f16", "inputs": { "q": { "dtype": "float16", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 4, 128, 64], "tolerance": 0.005 } } }, { "name": "flash_causal_d64_f32", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.016, "cosStep": 0.024, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.038, "cosStep": 0.014, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.026, "cosStep": 0.046, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 64], "tolerance": 0.00005 } } }, { "name": "ort_float_mask_qk_broadcast", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DGqaAttnMask", "notes": "Covers standard additive-mask broadcasting from [Q, K] to [B, QH, Q, K]." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 2, 4], "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.11, "scale": 0.5 } }, "k": { "dtype": "float32", "shape": [1, 1, 3, 4], "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.19, "scale": 0.5 } }, "v": { "dtype": "float32", "shape": [1, 1, 3, 4], "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.07, "scale": 0.5 } }, "attn_mask": { "dtype": "float32", "shape": [2, 3], "data": { "kind": "values", "values": [0.0, -1.0, 0.25, 0.5, 0.0, -2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 2, 4], "tolerance": 0.00005 } } }, { "name": "ort_float_mask_singleton_head_broadcast", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DGqaAttnMask", "notes": "Covers standard singleton-head additive-mask broadcasting from [B, 1, Q, K] to [B, QH, Q, K]." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 2, 4], "data": { "kind": "fillFloat32", "sinStep": 0.09, "cosStep": 0.21, "scale": 0.5 } }, "k": { "dtype": "float32", "shape": [1, 2, 3, 4], "data": { "kind": "fillFloat32", "sinStep": 0.15, "cosStep": 0.05, "scale": 0.5 } }, "v": { "dtype": "float32", "shape": [1, 2, 3, 4], "data": { "kind": "fillFloat32", "sinStep": 0.27, "cosStep": 0.13, "scale": 0.5 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 2, 3], "data": { "kind": "values", "values": [0.0, -2.0, 0.25, 0.5, 0.0, -1.5] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 2, 4], "tolerance": 0.00005 } } }, { "name": "fallback_float_mask_flash_shape_d64_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.018, "cosStep": 0.022, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.032, "cosStep": 0.012, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.028, "cosStep": 0.048, "scale": 0.3 } }, "attn_mask": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.041, "cosStep": 0.009, "scale": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 64], "tolerance": 0.00005 } } }, { "name": "flash_mask_rank4_broadcast_head_d64_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.3 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.011, "scale": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 128, 64], "tolerance": 0.00005 } } }, { "name": "flash_mask_rank2_qk_d64_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.014, "cosStep": 0.026, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.019, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.022, "cosStep": 0.043, "scale": 0.3 } }, "attn_mask": { "dtype": "float32", "shape": [128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.013, "scale": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 128, 64], "tolerance": 0.00005 } } }, { "name": "flash_mask_bool_rank4_d64_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.015, "cosStep": 0.025, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.018, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.024, "cosStep": 0.04, "scale": 0.3 } }, "attn_mask": { "dtype": "bool", "shape": [1, 8, 128, 128], "data": { "kind": "cycle", "values": [1, 1, 1, 0, 1] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 128, 64], "tolerance": 0.00005 } } }, { "name": "tiled_bool_mask_rank4_h4_s64_d32_f32", "provenance": { "notes": "Rank-four float32 attention with 4 heads, head width 32, a 64-token sequence, and a boolean attention mask that cycles true/true/false/true across keys." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 32], "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.4 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 32], "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.31, "scale": 0.4 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 32], "data": { "kind": "fillFloat32", "sinStep": 0.19, "cosStep": 0.13, "scale": 0.4 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 64, 64], "data": { "kind": "cycle", "values": [1, 1, 0, 1] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "tiled_bool_mask_rank4_masked_rows_h4_s64_d32_f32", "provenance": { "notes": "Rank-four float32 attention, 4 heads, head width 32, 64-token sequence. The first half of the query rows are masked out entirely and must produce zero rather than a divide-by-zero softmax denominator, while the second half keeps its normal result." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 32], "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.4 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 32], "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.31, "scale": 0.4 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 32], "data": { "kind": "fillFloat32", "sinStep": 0.19, "cosStep": 0.13, "scale": 0.4 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 64, 64], "data": { "kind": "linspace", "start": 0, "end": 2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "flash_mask_bool_rank4_d64_f16", "provenance": { "notes": "Rank-four float16 attention (8 heads, head width 64, 128-token sequence) with a boolean attention mask, which stays dtype-independent while a float mask would need to match Q/K/V's dtype. V is offset by 1.0 so the output stays near O(1) and float16 tolerance can catch a large normalization error." }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.015, "cosStep": 0.025, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.018, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.024, "cosStep": 0.04, "scale": 0.3, "offset": 1.0 } }, "attn_mask": { "dtype": "bool", "shape": [1, 8, 128, 128], "data": { "kind": "cycle", "values": [1, 1, 1, 0, 1] } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 128, 64], "tolerance": 0.01, "relTolerance": 0.01 } } }, { "name": "flash_mask_bool_rank2_d64_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.016, "cosStep": 0.024, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.028, "cosStep": 0.02, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.025, "cosStep": 0.039, "scale": 0.3 } }, "attn_mask": { "dtype": "bool", "shape": [128, 128], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 1, 1, 0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 128, 64], "tolerance": 0.00005 } } }, { "name": "flash_mask_gqa_rank4_d64_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.023, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 2, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.021, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 2, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.026, "cosStep": 0.038, "scale": 0.3 } }, "attn_mask": { "dtype": "float32", "shape": [1, 8, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.035, "cosStep": 0.015, "scale": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 128, 64], "tolerance": 0.00005 } } }, { "name": "flash_mask_rank4_broadcast_head_d64_f16", "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.3 } }, "attn_mask": { "dtype": "float16", "shape": [1, 1, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.011, "scale": 16.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 128, 64], "tolerance": 0.02 } } }, { "name": "flash_mask_rank4_broadcast_head_d64_f16_offset_value_scale", "provenance": { "notes": "Offsetting V around 1.0 makes normalization errors observable in masked float16 flash-cluster attention at headDim=64, including errors that preserve relative weights but change their sum." }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.3, "offset": 1.0 } }, "attn_mask": { "dtype": "float16", "shape": [1, 1, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.011, "scale": 16.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 128, 64], "tolerance": 0.01, "relTolerance": 0.01 } } }, { "name": "flash_decode_splitk_gqa_d64_kv2048_f32", "attrs": { "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 1, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.029, "scale": 1.2 } }, "k": { "dtype": "float32", "shape": [1, 2, 2048, 64], "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.007, "scale": 1.2 } }, "v": { "dtype": "float32", "shape": [1, 2, 2048, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.017, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 1, 64], "tolerance": 0.00008 } } }, { "name": "flash_causal_nonsquare_q128_kv512_d64", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.013, "cosStep": 0.027 } }, "k": { "dtype": "float32", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.011, "cosStep": 0.023 } }, "v": { "dtype": "float32", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.007, "cosStep": 0.031 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 128, 64], "tolerance": 0.0003 } } }, { "name": "flash_causal_nonsquare_q512_kv128_d64", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.013, "cosStep": 0.027 } }, "k": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.011, "cosStep": 0.023 } }, "v": { "dtype": "float32", "shape": [1, 4, 128, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.007, "cosStep": 0.031 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 512, 64], "tolerance": 0.0003 } } }, { "name": "flash_causal_nonsquare_batch2_q128_kv512", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 4, 128, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.013, "cosStep": 0.027 } }, "k": { "dtype": "float32", "shape": [2, 4, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.011, "cosStep": 0.023 } }, "v": { "dtype": "float32", "shape": [2, 4, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.007, "cosStep": 0.031 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 4, 128, 64], "tolerance": 0.0003 } } }, { "name": "causal_q1_kv4_upper_left_exact", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "notes": "Compact Attention-24 upper-left causal witness for qSeq=1 and kvSeq=4. The sole query may attend only the first key, so the output equals the first value rather than an average that includes future keys." }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 4, 1], "data": { "kind": "values", "values": [0.0, 0.0, 0.0, 0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 4, 1], "data": { "kind": "values", "values": [1.0, 2.0, 4.0, 8.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] }, "tolerance": 0, "relTolerance": 0 } } }, { "name": "flash_causal_q1_kv512_d64", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 64, 1, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.013, "cosStep": 0.027 } }, "k": { "dtype": "float32", "shape": [1, 64, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.011, "cosStep": 0.023 } }, "v": { "dtype": "float32", "shape": [1, 64, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.007, "cosStep": 0.031 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 1, 64], "tolerance": 0.0003 } } }, { "name": "flash_noncausal_d80_lpq4_f32", "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 128, 80], "data": { "kind": "fillFloat32", "sinStep": 0.012, "cosStep": 0.026, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 128, 80], "data": { "kind": "fillFloat32", "sinStep": 0.034, "cosStep": 0.018, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 128, 80], "data": { "kind": "fillFloat32", "sinStep": 0.022, "cosStep": 0.042, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 128, 80], "tolerance": 0.00005 } } }, { "name": "ort_rank4_float_mask_rank3_broadcast", "provenance": { "source": "onnxruntime/test/providers/cpu/llm/attention_op_test.cc", "test": "AttentionTest.Attention4DWithMask3DPastAndPresentQkMatmul", "notes": "A rank-3 [batch, query, key] additive mask broadcasts across heads without a cache input." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, 1.0, 1.0, -1.0, 1.0] } }, "k": { "dtype": "float32", "shape": [1, 2, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0, -1.0, 0.0, 0.5, 1.0, 1.0, -0.5, -1.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 2, 3, 1], "data": { "kind": "values", "values": [10.0, 20.0, 40.0, 100.0, 200.0, 400.0] } }, "attn_mask": { "dtype": "float32", "shape": [1, 2, 3], "data": { "kind": "values", "values": [0.0, 0.0, -10000.0, -10000.0, 0.0, 0.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 2, 1], "tolerance": 0.000003 } } }, { "name": "rank4_tiled_rank3_mask_q32_d16", "provenance": { "notes": "A rank-3 [1, query, key] additive mask over rank-4 attention inputs checks broadcasting across batch and heads." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 32, 16], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 32, 16], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 32, 16], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.3 } }, "attn_mask": { "dtype": "float32", "shape": [1, 32, 32], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.013, "scale": 0.5 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 32, 16], "tolerance": 0.00002 } } }, { "name": "empty_input_zero_dim", "inputs": { "q": { "dtype": "float32", "shape": [0, 1, 2, 2], "data": { "kind": "values", "values": [] } }, "k": { "dtype": "float32", "shape": [0, 1, 2, 2], "data": { "kind": "values", "values": [] } }, "v": { "dtype": "float32", "shape": [0, 1, 2, 2], "data": { "kind": "values", "values": [] } } }, "outputs": { "y": { "dtype": "float32", "shape": [0, 1, 2, 2], "tolerance": 0 } } }, { "name": "softcap_gemma2_prefill_causal_f32", "attrs": { "softcap": 50, "is_causal": 1, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 128], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.013 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 128], "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.019 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 128], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.029 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 128], "tolerance": 0.0002 } } }, { "name": "softcap_grok_decode_f16", "attrs": { "softcap": 30, "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 1, 128], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.007, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 1, 128], "tolerance": 0.003 } }, "provenance": { "notes": "Query and key amplitudes put the raw scores well past the softcap, so the tanh clamp is the thing under test rather than a near-linear pass-through. The tolerance bounds the f32 tanh rounding difference." } }, { "name": "softcap_grok_decode_f16_offset_value_scale", "provenance": { "notes": "Offsetting V around 1.0 keeps the 1,024-key average away from zero, exposing normalization errors when softcapped decode split-K merges partials with separate running maxima and denominators." }, "attrs": { "softcap": 30, "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 1, 128], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.007, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 1, 128], "tolerance": 0.01, "relTolerance": 0.01 } } }, { "name": "f16_decode_gqa_32h8kv_d128_kv2048", "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 1, 128], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 2048, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.007, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 2048, 128], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 1, 128], "tolerance": 0.0007 } } }, { "name": "f16_short_query_splitk_gqa_32h8kv_d128_q31_kv2048", "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 31, 128], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 2048, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.007, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 2048, 128], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 31, 128], "tolerance": 0.0007 } } }, { "name": "f16_short_query_splitk_gqa_32h8kv_d128_q2_kv2048", "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 2, 128], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 2048, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.007, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 2048, 128], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 2, 128], "tolerance": 0.0007 } } }, { "name": "f16_short_query_splitk_gqa_32h8kv_d128_q64_kv1024", "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 64, 128], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.007, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 64, 128], "tolerance": 0.004 } } }, { "name": "f16_short_query_splitk_gqa_32h8kv_d128_q64_kv1024_offset_value_scale", "provenance": { "notes": "Offsetting V around 1.0 makes normalization and query-by-split indexing errors observable in short-query split-K GQA at q64/kv1024." }, "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 64, 128], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.007, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 64, 128], "tolerance": 0.01, "relTolerance": 0.01 } } }, { "name": "f16_prefill_causal_gqa_8h2kv_d128_s128", "attrs": { "is_causal": 1, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "k": { "dtype": "float16", "shape": [1, 2, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.007 } }, "v": { "dtype": "float16", "shape": [1, 2, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 128, 128], "tolerance": 0.03 } } }, { "name": "f16_prefill_noncausal_gqa_8h2kv_d128_s128", "provenance": { "notes": "Amplitudes give the answer a magnitude of its own instead of a near-cancelling mean of V, which is what lets a tolerance this tight see a per-position error at f16." }, "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.021, "scale": 0.4 } }, "k": { "dtype": "float16", "shape": [1, 2, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.4 } }, "v": { "dtype": "float16", "shape": [1, 2, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.013, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 128, 128], "tolerance": 0.0002, "relTolerance": 0.002 } } }, { "name": "f16_prefill_noncausal_gqa_8h2kv_d128_s128_offset_value_scale", "provenance": { "notes": "Offsetting V around 1.0 keeps the 128-key average away from zero, exposing normalization errors in noncausal float16 GQA prefill at headDim=128 on both subgroup and reduced tiers." }, "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.021, "scale": 0.4 } }, "k": { "dtype": "float16", "shape": [1, 2, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.4 } }, "v": { "dtype": "float16", "shape": [1, 2, 128, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.013, "scale": 0.3, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 128, 128], "tolerance": 0.01, "relTolerance": 0.01 } } }, { "name": "f16_prefill_noncausal_8h_d96_s128", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 128, 96], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 128, 96], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 128, 96], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 128, 96], "tolerance": 0.008 } } }, { "name": "f16_prefill_noncausal_8h_d96_s128_offset_value_scale", "provenance": { "notes": "Head width 96 with eight heads and 128 tokens; offset V near 1.0 keeps the noncausal float16 output away from zero so normalization errors remain observable." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 128, 96], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 128, 96], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04, "scale": 1.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 128, 96], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 128, 96], "tolerance": 0.01, "relTolerance": 0.01 } } }, { "name": "scalar_gqa_softcap_vhead_gt_headsize_f32", "attrs": { "softcap": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 3, 4], "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.07 } }, "k": { "dtype": "float32", "shape": [1, 2, 5, 4], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.09 } }, "v": { "dtype": "float32", "shape": [1, 2, 5, 300], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 3, 300], "tolerance": 0.00003 } } }, { "name": "softcap_saturating_scores_rank4_f32", "provenance": { "notes": "Softcap-sensitive rank-4 case: max|score| is 3.6 times the cap, so cap*tanh(score/cap) materially changes the output. V is offset from zero to keep the result well-conditioned, while the softmax remains distributed rather than collapsing to one hot. This makes removal or misapplication of softcap observable." }, "attrs": { "softcap": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 4, 8], "data": { "kind": "fillFloat32", "sinStep": 0.37, "cosStep": 0.11, "scale": 2.0 } }, "k": { "dtype": "float32", "shape": [1, 2, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.19, "cosStep": 0.29, "scale": 2.0 } }, "v": { "dtype": "float32", "shape": [1, 2, 6, 8], "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.41, "scale": 0.5, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 4, 8], "tolerance": 0.00001, "relTolerance": 0.00001 } } }, { "name": "flash_float_mask_all_neg_inf_row_averages_v_f32", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "k": { "dtype": "float32", "shape": [1, 4, 80, 64], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04 } }, "v": { "dtype": "float32", "shape": [1, 4, 80, 64], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 64, 80], "data": { "kind": "constant", "value": -3.402823e+38 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 64], "tolerance": 0.00002 } } }, { "name": "flash_masked_row_trio_bool_all_false_row_zero_f32", "provenance": { "notes": "At heads=8, qSeq=32, kvSeq=16, and headDim=64, query row 5 is fully masked by a broadcast rank-4 boolean mask while every other row remains unmasked. Row 5 of y must be zero for every head because no key has probability mass." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 32, 64], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "k": { "dtype": "float32", "shape": [1, 8, 16, 64], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04 } }, "v": { "dtype": "float32", "shape": [1, 8, 16, 64], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 32, 16], "data": { "kind": "values", "values": [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 32, 64], "tolerance": 0.00002 } } }, { "name": "flash_masked_row_trio_float_neg_inf_row_zero_f32", "provenance": { "notes": "At heads=8, qSeq=32, kvSeq=16, and headDim=64, query row 5 receives -Infinity at every key from a broadcast rank-4 additive mask while every other row receives zero. Row 5 of y must be zero for every head, exercising the zero-denominator guard." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 32, 64], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "k": { "dtype": "float32", "shape": [1, 8, 16, 64], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04 } }, "v": { "dtype": "float32", "shape": [1, 8, 16, 64], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 32, 16], "data": { "kind": "values", "values": [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", "-Infinity", 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 32, 64], "tolerance": 0.00002 } } }, { "name": "flash_masked_row_trio_float_neg_flt_max_row_averages_v_f32", "provenance": { "notes": "At heads=8, qSeq=32, kvSeq=16, and headDim=64, query row 5 receives the finite value -3.402823e+38 at every key from a broadcast rank-4 additive mask. Every masked score rounds to the same finite value, so row 5 of y must equal the uniform average of its head's 16 value rows." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 8, 32, 64], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "k": { "dtype": "float32", "shape": [1, 8, 16, 64], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04 } }, "v": { "dtype": "float32", "shape": [1, 8, 16, 64], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 32, 16], "data": { "kind": "values", "values": [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, -3.402823e+38, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 8, 32, 64], "tolerance": 0.00002 } } }, { "name": "flash_bool_mask_masked_rows_zero_f32", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "k": { "dtype": "float32", "shape": [1, 4, 80, 64], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04 } }, "v": { "dtype": "float32", "shape": [1, 4, 80, 64], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 64, 80], "data": { "kind": "linspace", "start": 0, "end": 2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 64], "tolerance": 0.00002 } }, "provenance": { "notes": "Logical-bool mask at the flash cluster geometry (headDim 64, qSeq 64) with the first half of the query rows fully rejected. Unlike a finite additive float mask, a rejected bool position has zero probability mass, so those rows must emit zero rather than NaN or a mean of V while the open rows are unaffected." } }, { "name": "flash_bool_mask_masked_rows_causal_zero_f32", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.03, "cosStep": 0.04 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.05, "cosStep": 0.06 } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 64, 64], "data": { "kind": "linspace", "start": 0, "end": 2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 64], "tolerance": 0.00002 } }, "provenance": { "notes": "Composing the causal prefix with a logical-bool mask that rejects every key of the first half of the query rows leaves those flash-cluster rows no probability-bearing key at all; they must be zero, and the causal rows below them must still be right." } }, { "name": "rank3_mqa_causal_packed_kv1_f32", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 3, 8], "data": { "kind": "values", "values": [1.0, 0.0, -1.0, 0.5, 0.25, -0.5, 2.0, 1.0, 0.0, -2.0, 1.0, 1.5, -1.0, 0.5, 0.5, -0.25, 1.0, 1.0, -1.0, 0.0, 0.5, -0.5, 2.0, -2.0] } }, "k": { "dtype": "float32", "shape": [1, 3, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.5, 1.0, -1.0, 2.0] } }, "v": { "dtype": "float32", "shape": [1, 3, 2], "data": { "kind": "values", "values": [10.0, 100.0, 20.0, 200.0, 30.0, 300.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 3, 8], "tolerance": 0.000003 } } }, { "name": "scalar_2dfold_last_band_vhead_gt_256_f32", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 8200, 8], "data": { "kind": "fillFloat32", "sinStep": 0.001, "cosStep": 0.002 } }, "k": { "dtype": "float32", "shape": [1, 4, 8, 8], "data": { "kind": "fillFloat32", "sinStep": 0.003, "cosStep": 0.004 } }, "v": { "dtype": "float32", "shape": [1, 4, 8, 512], "data": { "kind": "fillFloat32", "sinStep": 0.0011, "cosStep": 0.0013 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 8200, 512], "tolerance": 0.00003 } } }, { "name": "materialized_f32_d72_specialized_tiles", "provenance": { "notes": "Locks the materialized f32 head-dimension tail at the minimum sequence boundary. On subgroup-matrix devices, D72 exercises guarded final K and output tiles; fallback tiers retain the scalar tiled implementation." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 512, 72], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 512, 72], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 512, 72], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.1 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 512, 72], "tolerance": 0.0001 } } }, { "name": "materialized_sgmat_gqa_d128_explicit_scale", "provenance": { "notes": "Rank-four float32 attention with 4 query heads grouped 4:1 over 1 key/value head, head width 128, a 512-token sequence, and an explicit scale of 0.125." }, "attrs": { "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 512, 128], "tolerance": 0.0001, "relTolerance": 0.00002 } } }, { "name": "materialized_gemm_d128_cross_attention_tails", "provenance": { "notes": "Cross-attention with query length 517 and key/value length 523 checks partial final query and key tiles and the resulting normalization." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 517, 128], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 523, 128], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 523, 128], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 517, 128], "tolerance": 0.0001, "relTolerance": 0.00002 } } }, { "name": "materialized_gemm_d128_cross_attention_tails_offset_value_scale", "provenance": { "notes": "Offsetting V around 1.0 keeps the 523-key average away from zero and exposes normalization errors at partial final query and key tiles." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 517, 128], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 523, 128], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 523, 128], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 517, 128], "tolerance": 0.0001, "relTolerance": 0.00002 } } }, { "name": "flash_noncausal_d72_lpq2_f32", "provenance": { "notes": "headDim=72 is divisible by 8 but not 16, exercising the LPQ2 clustered route with SLICE=9 and one subgroupShuffleXor reduction stage." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 72], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 300, 72], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 300, 72], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 72], "tolerance": 0.00005 } } }, { "name": "flash_causal_plus_mask_d64_cluster_f32", "provenance": { "notes": "With is_causal=1, an explicit attention mask, and headDim=64, the clustered route must restrict keys to maxKj=qIdx+1 and then apply the additive mask within that causal range." }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 2, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 2, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 64, 64], "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.023, "scale": 0.5 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 64, 64], "tolerance": 0.00005 } } }, { "name": "unmasked_2pass_vec4_qk_vhead300_f32", "provenance": { "notes": "Unmasked rank-four attention with four-aligned Q/K and a wider value head checks output normalization." }, "attrs": {}, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 4, 8], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "k": { "dtype": "float32", "shape": [1, 2, 6, 8], "data": { "kind": "linspace", "start": -0.75, "end": 1.25 } }, "v": { "dtype": "float32", "shape": [1, 2, 6, 300], "data": { "kind": "linspace", "start": -2.0, "end": 2.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 4, 300], "tolerance": 0.0001 } } }, { "name": "mask_2pass_additive_vhead300_f32", "provenance": { "notes": "A value-head width of 300 with a rank-4 additive float32 mask checks masked softmax and value application across a large value head." }, "attrs": {}, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 4, 8], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "k": { "dtype": "float32", "shape": [1, 2, 6, 8], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "v": { "dtype": "float32", "shape": [1, 2, 6, 300], "data": { "kind": "linspace", "start": -2.0, "end": 2.0 } }, "attn_mask": { "dtype": "float32", "shape": [1, 2, 4, 6], "data": { "kind": "linspace", "start": -1.5, "end": 0.5 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 4, 300], "tolerance": 0.0001 } } }, { "name": "mask_2pass_causal_bool_vhead512_f32", "provenance": { "notes": "A 512-wide value head with a boolean mask and causal attention. The all-true mask leaves causal key zero visible to every row, so no row is fully masked." }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 8, 8], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "k": { "dtype": "float32", "shape": [1, 2, 8, 8], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "v": { "dtype": "float32", "shape": [1, 2, 8, 512], "data": { "kind": "linspace", "start": -3.0, "end": 3.0 } }, "attn_mask": { "dtype": "bool", "shape": [1, 2, 8, 8], "data": { "kind": "constant", "value": 1 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 8, 512], "tolerance": 0.0001 } } }, { "name": "mask_2pass_bool_mask_masked_head_zero_f32", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 8, 8], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "k": { "dtype": "float32", "shape": [1, 2, 8, 8], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "v": { "dtype": "float32", "shape": [1, 2, 8, 512], "data": { "kind": "linspace", "start": -3.0, "end": 3.0 } }, "attn_mask": { "dtype": "bool", "shape": [1, 2, 8, 8], "data": { "kind": "linspace", "start": 0, "end": 2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 8, 512], "tolerance": 0.0001 } }, "provenance": { "notes": "Rank-four float32 attention with 2 heads, an 8-token sequence, and a large value head width of 512. The first head's boolean mask allows no key, so it must output zero rather than divide by zero or average V; the second head's rows carry the real attention result." } }, { "name": "rank3_flash_tiled_bsh_s32_h1_d64", "attrs": { "q_num_heads": 1, "kv_num_heads": 1, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 32, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 32, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 64], "tolerance": 0.0001 } } }, { "name": "f16_prefill_q31_boundary_gqa_8h2kv_d128_kv65", "provenance": { "notes": "The 31-query, 65-key boundary geometry is the point of the case; the amplitudes keep its answer above f16's own noise so the tolerance can stay a fraction of it." }, "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 31, 128], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.021, "scale": 0.4 } }, "k": { "dtype": "float16", "shape": [1, 2, 65, 128], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.4 } }, "v": { "dtype": "float16", "shape": [1, 2, 65, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.013, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 31, 128], "tolerance": 0.0005, "relTolerance": 0.002 } } }, { "name": "f16_prefill_q31_boundary_gqa_8h2kv_d128_kv65_offset_value_scale", "provenance": { "notes": "A partial 32-query tile and one-key KV tail exercise float16 GQA prefill boundaries. Offsetting V around 1.0 makes boundary-row normalization errors observable." }, "attrs": { "is_causal": 0, "scale": 0.08838834764831843 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 31, 128], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.021, "scale": 0.4 } }, "k": { "dtype": "float16", "shape": [1, 2, 65, 128], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.017, "scale": 0.4 } }, "v": { "dtype": "float16", "shape": [1, 2, 65, 128], "data": { "kind": "fillFloat32", "sinStep": 0.005, "cosStep": 0.013, "scale": 0.3, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 31, 128], "tolerance": 0.01, "relTolerance": 0.01 } } }, { "name": "rank4_past_key_value_cache", "provenance": { "notes": "One cached token and one new one, with q aligned to the cached key and orthogonal to the new one, so the two attention weights are unequal and y is a specific mixture of the two values rather than their mean. present_key/present_value are pinned to the join itself." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [0.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [10.0, 20.0] } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [2.0, 4.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 2], "tolerance": 0.000001 }, "present_key": { "dtype": "float32", "shape": [1, 1, 2, 2], "tolerance": 0, "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 2, 2], "tolerance": 0, "data": { "kind": "values", "values": [2.0, 4.0, 10.0, 20.0] } } } }, { "name": "rank4_causal_past_offset_exact", "provenance": { "source": "onnx/onnx docs/Operators.md Attention-24", "notes": "Two cached tokens precede a two-query block, so query i attends through key past_sequence_length+i. Zero logits make the expected rows exact means: [1,3,8] -> 4 and [1,3,8,12] -> 6." }, "attrs": { "is_causal": 1, "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "constant", "value": 0.0 } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [3.0, 4.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [8.0, 12.0] } }, "past_key": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [1.0, 2.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [1.0, 3.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [4.0, 6.0] }, "tolerance": 0, "relTolerance": 0 }, "present_key": { "dtype": "float32", "shape": [1, 1, 4, 1], "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0] }, "tolerance": 0, "relTolerance": 0 }, "present_value": { "dtype": "float32", "shape": [1, 1, 4, 1], "data": { "kind": "values", "values": [1.0, 3.0, 8.0, 12.0] }, "tolerance": 0, "relTolerance": 0 } } }, { "name": "rank4_causal_past_offset_bool_mask_y_only_f16_exact", "provenance": { "source": "onnx/onnx docs/Operators.md Attention-24", "notes": "Causal float16 attention over a 2-token cache-extended sequence with a boolean mask. Row 0 allows cached value 2 and current value 10 (mean 6); row 1 allows cached value 4 and both current values 10 and 16 (mean 10), matching the exact expected output." }, "attrs": { "is_causal": 1, "scale": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 1, 2, 1], "data": { "kind": "constant", "value": 0.0 } }, "k": { "dtype": "float16", "shape": [1, 1, 2, 1], "data": { "kind": "constant", "value": 0.0 } }, "v": { "dtype": "float16", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [10.0, 16.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 2, 4], "data": { "kind": "values", "values": [1, 0, 1, 1, 0, 1, 1, 1] } }, "past_key": { "dtype": "float16", "shape": [1, 1, 2, 1], "data": { "kind": "constant", "value": 0.0 } }, "past_value": { "dtype": "float16", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [2.0, 4.0] } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [6.0, 10.0] }, "tolerance": 0, "relTolerance": 0 } } }, { "name": "rank4_past_key_value_cache_f16", "provenance": { "notes": "Float16 counterpart of rank4_past_key_value_cache. It witnesses every optional cache port on the standard float16 T route and pins the key/value joins to directly representable values." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "k": { "dtype": "float16", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [0.0, 1.0] } }, "v": { "dtype": "float16", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [10.0, 20.0] } }, "past_key": { "dtype": "float16", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "past_value": { "dtype": "float16", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [2.0, 4.0] } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 1, 2], "tolerance": 0.02, "relTolerance": 0.005 }, "present_key": { "dtype": "float16", "shape": [1, 1, 2, 2], "tolerance": 0, "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "present_value": { "dtype": "float16", "shape": [1, 1, 2, 2], "tolerance": 0, "data": { "kind": "values", "values": [2.0, 4.0, 10.0, 20.0] } } } }, { "name": "rank4_past_key_value_cache_bool_mask", "provenance": { "notes": "A logical mask over the joined past_key/K cache rejects the cached token and allows only the new token, and the mask's key-sequence dimension covers past_key followed by K rather than K alone." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [0.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [10.0, 20.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [0, 1] } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1.0, 0.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [2.0, 4.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 2], "tolerance": 0.000001, "data": { "kind": "values", "values": [10.0, 20.0] } }, "present_key": { "dtype": "float32", "shape": [1, 1, 2, 2], "tolerance": 0, "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 2, 2], "tolerance": 0, "data": { "kind": "values", "values": [2.0, 4.0, 10.0, 20.0] } } } }, { "name": "rank4_cache_present_key_only", "provenance": { "notes": "With paired past inputs and only present_key requested, Y must still use past and current values, while present_value remains absent." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0.000001 }, "present_key": { "dtype": "float32", "shape": [1, 1, 2, 1], "tolerance": 0, "data": { "kind": "values", "values": [1.0, 0.0] } } } }, { "name": "rank4_cache_present_value_only_bool_mask", "provenance": { "notes": "Only present_value is requested from a paired past cache. The bool mask rejects the cached token, so Y pins that the hidden key join still feeds attention while present_value exposes the complete value join." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [0, 1] } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [10.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 2, 1], "tolerance": 0, "data": { "kind": "values", "values": [2.0, 10.0] } } } }, { "name": "rank4_cache_y_only_f16", "provenance": { "notes": "Paired float16 past inputs contribute to Y even when neither optional present output is requested; Y remains float16." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float16", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float16", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "past_key": { "dtype": "float16", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float16", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 1, 1], "tolerance": 0.02, "relTolerance": 0.005 } } }, { "name": "rank4_current_present_key_value", "provenance": { "notes": "Both optional present outputs are requested without past inputs. They must be exact copies of the current K and V tensors, and Y must still be computed from those direct inputs." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [3.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [7.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [7.0] } }, "present_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [3.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [7.0] } } } }, { "name": "rank4_current_present_key_only_bool_mask", "provenance": { "notes": "Only present_key is requested, with no past cache. An all-false boolean mask makes Y zero, while the requested present_key output still copies K exactly, separating cache output from the attention result." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [3.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [7.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 1], "data": { "kind": "constant", "value": 0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [0.0] } }, "present_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [3.0] } } } }, { "name": "rank4_current_present_value_only", "provenance": { "notes": "Only present_value is requested without a past cache. This exercises the positional output hole and requires an exact V copy while Y is computed normally." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [3.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [7.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [7.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [7.0] } } } }, { "name": "rank4_cache_present_value_only", "provenance": { "notes": "Only present_value is requested, while Y must still attend over both past and current keys." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0.000001 }, "present_value": { "dtype": "float32", "shape": [1, 1, 2, 1], "tolerance": 0, "data": { "kind": "values", "values": [2.0, 10.0] } } } }, { "name": "rank4_cache_present_key_only_bool_mask", "provenance": { "notes": "The boolean mask allows only the cached token, and present_value is not requested. Y depends on the hidden past_value join, while the requested present_key output independently verifies the past_key/K join." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [1, 0] } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [2.0] } }, "present_key": { "dtype": "float32", "shape": [1, 1, 2, 1], "tolerance": 0, "data": { "kind": "values", "values": [1.0, 0.0] } } } }, { "name": "rank4_cache_y_only_bool_mask", "provenance": { "notes": "Neither present_key nor present_value is requested. A boolean mask that allows only the new token verifies Y over the joined past and current sequence." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [0, 1] } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [10.0] } } } }, { "name": "rank4_past_key_value_cache_all_false_bool_mask", "provenance": { "notes": "Pins the zero-denominator rule for the masked rank-4 cache route while both joined caches remain public outputs." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "constant", "value": 0 } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [0.0] } }, "present_key": { "dtype": "float32", "shape": [1, 1, 2, 1], "tolerance": 0, "data": { "kind": "values", "values": [1.0, 0.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 2, 1], "tolerance": 0, "data": { "kind": "values", "values": [2.0, 10.0] } } } }, { "name": "rank4_cache_present_key_only_all_false_bool_mask", "provenance": { "notes": "A fully masked row must produce zero Y while the requested present_key remains correct." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "constant", "value": 0 } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [0.0] } }, "present_key": { "dtype": "float32", "shape": [1, 1, 2, 1], "tolerance": 0, "data": { "kind": "values", "values": [1.0, 0.0] } } } }, { "name": "rank4_cache_present_value_only_all_false_bool_mask", "provenance": { "notes": "A fully masked row must produce zero Y while the requested present_value remains correct." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "constant", "value": 0 } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [0.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 2, 1], "tolerance": 0, "data": { "kind": "values", "values": [2.0, 10.0] } } } }, { "name": "rank4_cache_y_only_all_false_bool_mask", "provenance": { "notes": "A fully masked row must produce zero Y even when neither optional present cache output is requested." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [0.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [10.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 2], "data": { "kind": "constant", "value": 0 } }, "past_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "past_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [2.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [0.0] } } } }, { "name": "rank4_current_present_key_value_bool_mask", "provenance": { "notes": "Both present outputs are requested with no past cache on the masked route. An all-false row makes Y zero while both exact current-cache copies remain observable." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [3.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [7.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 1], "data": { "kind": "constant", "value": 0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [0.0] } }, "present_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [3.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [7.0] } } } }, { "name": "rank4_current_present_key_only", "provenance": { "notes": "Unmasked no-past route with only present_key requested. The cache copy is exact and the direct K/V attention result remains independently checked." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [3.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [7.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [7.0] } }, "present_key": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [3.0] } } } }, { "name": "rank4_current_present_value_only_bool_mask", "provenance": { "notes": "Masked no-past route with only present_value requested. The all-false row must not suppress the exact V cache copy even though it makes Y zero." }, "attrs": { "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [1.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [3.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 1, 1], "data": { "kind": "values", "values": [7.0] } }, "attn_mask": { "dtype": "bool", "shape": [1, 1, 1, 1], "data": { "kind": "constant", "value": 0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [0.0] } }, "present_value": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [7.0] } } } }, { "name": "rank4_past_key_value_cache_grouped_multi_token", "provenance": { "notes": "With batch 2, four query heads over two KV heads, three cached tokens, and two new tokens, each query-head group must share the same independently joined five-token cache." }, "attrs": { "scale": 0.35 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 4, 3, 8], "data": { "kind": "fillFloat32", "sinStep": 0.19, "cosStep": 0.31, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 2, 2, 8], "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.17, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 2, 2, 8], "data": { "kind": "fillFloat32", "sinStep": 0.29, "cosStep": 0.41, "scale": 0.3 } }, "past_key": { "dtype": "float32", "shape": [2, 2, 3, 8], "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.37, "scale": 0.3 } }, "past_value": { "dtype": "float32", "shape": [2, 2, 3, 8], "data": { "kind": "fillFloat32", "sinStep": 0.43, "cosStep": 0.11, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 4, 3, 8], "tolerance": 0.00001, "relTolerance": 0.00001 }, "present_key": { "dtype": "float32", "shape": [2, 2, 5, 8], "tolerance": 0, "relTolerance": 0 }, "present_value": { "dtype": "float32", "shape": [2, 2, 5, 8], "tolerance": 0, "relTolerance": 0 } } }, { "name": "flash_noncausal_d256_register_boundary_f32", "provenance": { "notes": "A float32 noncausal attention head of width 256 checks a large head dimension over the same outputs on every feature tier." }, "attrs": { "is_causal": 0, "scale": 0.0625 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 64, 256], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.029, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 256], "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.023, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 256], "data": { "kind": "fillFloat32", "sinStep": 0.007, "cosStep": 0.041, "scale": 0.2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 64, 256], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "short_prompt_generic_flash_mask_16h4kv_q16_kv64_d80", "provenance": { "notes": "For a masked short prompt, head size 80 is a multiple of four but not 32, leaving the one-query-row-per-workgroup masked flash path on tiers without a fixed subgroup width." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 16, 16, 80], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 4, 64, 80], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 4, 64, 80], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.3 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 16, 64], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.013, "scale": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 16, 16, 80], "tolerance": 0.00003 } } }, { "name": "scalar_no_mask_vhead_over_online_cap_f32", "provenance": { "notes": "A 32-wide value head with a different QK head width checks unmasked attention across mismatched head dimensions." }, "tunables": { "ONLINE_VALUE_CAP": 16 }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 4, 8], "data": { "kind": "linspace", "start": -0.5, "end": 0.5 } }, "k": { "dtype": "float32", "shape": [1, 1, 4, 8], "data": { "kind": "linspace", "start": -0.25, "end": 0.75 } }, "v": { "dtype": "float32", "shape": [1, 1, 4, 32], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 4, 32], "tolerance": 0.00001 } } }, { "name": "scalar_mask_vhead_over_online_cap_f32", "provenance": { "notes": "A rank-4 additive mask with a value head wider than the query/key head checks masked attention over mismatched head widths." }, "tunables": { "ONLINE_VALUE_CAP": 16 }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 4, 8], "data": { "kind": "linspace", "start": -0.5, "end": 0.5 } }, "k": { "dtype": "float32", "shape": [1, 1, 4, 8], "data": { "kind": "linspace", "start": -0.25, "end": 0.75 } }, "v": { "dtype": "float32", "shape": [1, 1, 4, 32], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 4, 4], "data": { "kind": "linspace", "start": -1.0, "end": 0.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 4, 32], "tolerance": 0.00001 } } }, { "name": "scalar_mask_all_neg_inf_row_zero_f32", "provenance": { "notes": "Rank-four float32 attention, 1 head, 4-token sequence, value head width 32. An additive mask gives query row 1 negative infinity at every key while rows 0, 2 and 3 stay finite, so row 1 of Y must be zero and the remaining rows keep their finite result." }, "tunables": { "ONLINE_VALUE_CAP": 16 }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 4, 8], "data": { "kind": "linspace", "start": -0.5, "end": 0.5 } }, "k": { "dtype": "float32", "shape": [1, 1, 4, 8], "data": { "kind": "linspace", "start": -0.25, "end": 0.75 } }, "v": { "dtype": "float32", "shape": [1, 1, 4, 32], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "attn_mask": { "dtype": "float32", "shape": [1, 1, 4, 4], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/scalar_mask_all_neg_inf_row_zero_f32_input_attn_mask" } } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 4, 32], "tolerance": 0.00001 } } }, { "name": "cluster_splitkv_q100_kv4096_h2_d64_f32", "provenance": { "notes": "Rank-four float32 attention with 2 query heads over 1 key/value head, head width 64: a 100-token query against a 4096-token key/value cache, with query length not a multiple of any small tile." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 100, 64], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 4096, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 4096, 64], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 100, 64], "tolerance": 0.00003, "relTolerance": 0.00003 } } }, { "name": "cluster_splitkv_q100_kv4096_h2_d64_f32_offset_value_scale", "provenance": { "notes": "At query length 100, key/value length 4,096, and head width 64, offsetting V around 1.0 exposes errors in the maximum and denominator across long key ranges." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 100, 64], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 4096, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 4096, 64], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 100, 64], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "cluster_splitkv_q100_kv4096_h2_d128_f32", "provenance": { "notes": "Clustered split-KV prefill at headDim=128 uses the wide LPQ8 geometry without subgroups. Zero-mean V makes relative-weight and eight-term lane-fold errors observable, with tolerance scaled to the approximately 1e-4 output magnitude." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 100, 128], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 1, 4096, 128], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 1, 4096, 128], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 100, 128], "tolerance": 0.000001, "relTolerance": 0.00001 } } }, { "name": "cluster_splitkv_q100_kv4096_h2_d128_f32_offset_value_scale", "provenance": { "notes": "Covers the wide LPQ8 no-subgroup split-KV geometry at headDim 128, with GQA 2:1 heads and a query-tail tile. V oscillates about 1.0 so the 4096-key average remains O(1), making errors in the softmax denominator, cross-slice partial fold, or relative lane weights visible within the f32 tolerance." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 100, 128], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 1, 4096, 128], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 1, 4096, 128], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 100, 128], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "cluster_splitkv_q80_kv4096_h1_d80_f32", "provenance": { "notes": "HeadDim=80 is divisible by 16 but not 32, selecting the LPQ4 no-subgroup split-KV cluster geometry and exercising its reduction and padding behavior." }, "attrs": { "is_causal": 0, "scale": 0.11180339887498948 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 80, 80], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 1, 4096, 80], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.041, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 1, 4096, 80], "data": { "kind": "fillFloat32", "sinStep": 0.047, "cosStep": 0.053, "scale": 0.2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 80, 80], "tolerance": 0.00003, "relTolerance": 0.00003 } } }, { "name": "cluster_splitkv_96tiles_q241_kv4096_h6_d64_f32", "provenance": { "notes": "Rank-four float32 attention with 6 query heads grouped 2:1 over 3 key/value heads, head width 64: a 241-token query (not a multiple of 16) against a 4096-token key/value cache." }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 6, 241, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 3, 4096, 64], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.041, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 3, 4096, 64], "data": { "kind": "fillFloat32", "sinStep": 0.047, "cosStep": 0.053, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 6, 241, 64], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "batched_windows_b8_s128_f32", "provenance": { "notes": "Batch 8 x sequence 128 x 2 heads models many independent windowed attention blocks and exercises rank-4 indexing with substantial batch depth." }, "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [8, 2, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.019, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [8, 2, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.031, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [8, 2, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.043, "scale": 0.2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [8, 2, 128, 64], "tolerance": 0.00003 } } }, { "name": "batched_windows_b4_s128_f32", "provenance": { "notes": "Batch 4 x sequence 128 x 2 heads exercises windowed rank-4 attention and batch-dependent tile indexing." }, "attrs": { "is_causal": 0, "scale": 0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [4, 2, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.017, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [4, 2, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.029, "cosStep": 0.037, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [4, 2, 128, 64], "data": { "kind": "fillFloat32", "sinStep": 0.041, "cosStep": 0.047, "scale": 0.2 } } }, "outputs": { "y": { "dtype": "float32", "shape": [4, 2, 128, 64], "tolerance": 0.00003 } } }, { "name": "materialized_sgmat_unaligned_seq_runtime_direct_store", "provenance": { "notes": "A 522x522 score matrix is not divisible by the 64-wide subgroup-matrix tile on either axis. The 8x8 interior tiles publish through collectives, while the final row, column, and corner tiles use the guarded scratch epilogue." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 522, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 2, 522, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 2, 522, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 522, 64], "tolerance": 0.0001, "relTolerance": 0.00002 } } }, { "name": "materialized_sgmat_unaligned_seq_runtime_direct_store_offset_value_scale", "provenance": { "notes": "Rank-four float32 attention with 2 heads (no grouping), head width 64, and a 522-token sequence that is not a multiple of 64. V is offset near 1.0 so the 522-key average stays well-conditioned, exposing a missing or duplicated denominator." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 522, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 2, 522, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 2, 522, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 522, 64], "tolerance": 0.0001, "relTolerance": 0.00002 } } }, { "name": "materialized_sgmat_private_row_stats_d64_cross_attention_tails", "provenance": { "notes": "Query length 517 and key length 523 exercise partial final query and key tiles. Nonzero value offsets expose omitted tail weights or incorrect normalization." }, "tunables": { "MATERIALIZED_SGMAT_PRIVATE_ROW_STATS": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 517, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.8 } }, "k": { "dtype": "float32", "shape": [1, 1, 523, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.8 } }, "v": { "dtype": "float32", "shape": [1, 1, 523, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 517, 64], "tolerance": 0.0001, "relTolerance": 0.00002 } } }, { "name": "materialized_sgmat_score_cooperative_loader_gqa_d64_tails", "provenance": { "notes": "Rank-four attention with GQA 2:1, query length 513, key/value length 515, and head width 64 checks partial final query/key tiles. V is offset near 1.0 so normalization errors remain observable." }, "tunables": { "MATERIALIZED_SGMAT_SCORE_SUBGROUP_TILE_COLS": 64 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 513, 64], "data": { "kind": "fillFloat32", "sinStep": 0.019, "cosStep": 0.029, "scale": 0.1 } }, "k": { "dtype": "float32", "shape": [1, 1, 515, 64], "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.011, "scale": 0.1 } }, "v": { "dtype": "float32", "shape": [1, 1, 515, 64], "data": { "kind": "fillFloat32", "sinStep": 0.043, "cosStep": 0.023, "scale": 0.1, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 513, 64], "tolerance": 0.0001, "relTolerance": 0.00002 } } }, { "name": "explicit_zero_scale_is_not_omission", "provenance": { "notes": "ONNX scale has no schema default literal: omission means 1/sqrt(head_size), while an explicitly supplied zero is a valid multiplier. Zero logits make both keys equally likely, so Y must be the exact average of their values." }, "attrs": { "scale": 0, "softmax_precision": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 1, 1, 2], "data": { "kind": "values", "values": [100.0, -100.0] } }, "k": { "dtype": "float32", "shape": [1, 1, 2, 2], "data": { "kind": "values", "values": [1.0, 0.0, 0.0, 1.0] } }, "v": { "dtype": "float32", "shape": [1, 1, 2, 1], "data": { "kind": "values", "values": [2.0, 6.0] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1, 1], "tolerance": 0, "data": { "kind": "values", "values": [4.0] } } } }, { "name": "scalar_mask_vhead_over_online_cap_f16", "provenance": { "notes": "A float16 masked rank-4 case exercises the generic scalar attention path with half-precision inputs and output." }, "tunables": { "ONLINE_VALUE_CAP": 16 }, "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 1, 4, 8], "data": { "kind": "linspace", "start": -0.5, "end": 0.5 } }, "k": { "dtype": "float16", "shape": [1, 1, 4, 8], "data": { "kind": "linspace", "start": -0.25, "end": 0.75 } }, "v": { "dtype": "float16", "shape": [1, 1, 4, 32], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }, "attn_mask": { "dtype": "float16", "shape": [1, 1, 4, 4], "data": { "kind": "linspace", "start": -1.0, "end": 0.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 1, 4, 32], "tolerance": 0.005 } } }, { "name": "cluster_nosg_narrow_lanes_at_key_work_floor", "provenance": { "source": "hand-authored", "notes": "Rank-four float32 attention with 4 heads (no grouping), head width 64, and a 512-token sequence with an additive rank-2 mask shared across batch and heads." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.0173, "cosStep": 0.0291 } }, "k": { "dtype": "float32", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.0119, "cosStep": 0.0233 } }, "v": { "dtype": "float32", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.0071, "cosStep": 0.0317 } }, "attn_mask": { "dtype": "float32", "shape": [512, 512], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.0037, "cosStep": 0.0059 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 512, 64], "tolerance": 0.00001, "relTolerance": 0.00001 } } }, { "name": "cluster_nosg_wide_lanes_below_key_work_floor", "provenance": { "source": "hand-authored", "notes": "A 496-key context with four heads checks the adjacent long-context shape." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 496, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.0173, "cosStep": 0.0291 } }, "k": { "dtype": "float32", "shape": [1, 4, 496, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.0119, "cosStep": 0.0233 } }, "v": { "dtype": "float32", "shape": [1, 4, 496, 64], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.0071, "cosStep": 0.0317 } }, "attn_mask": { "dtype": "float32", "shape": [496, 496], "data": { "kind": "fillFloat32", "scale": 0.1, "sinStep": 0.0037, "cosStep": 0.0059 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 496, 64], "tolerance": 0.00001, "relTolerance": 0.00001 } } }, { "name": "materialized_sgmat_causal_square_f32", "provenance": { "notes": "Square causal attention without a past cache checks that future keys are masked and that each query normalizes over its visible prefix." }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.1 } }, "k": { "dtype": "float32", "shape": [1, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.1 } }, "v": { "dtype": "float32", "shape": [1, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 512, 64], "tolerance": 0.0001 } } }, { "name": "materialized_sgmat_causal_upper_left_wide_kv_f32", "provenance": { "notes": "kv longer than q under is_causal pins the UPPER-LEFT alignment: keys past the query length are never attended, which distinguishes this contract from the right-aligned cached routes." }, "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.1 } }, "k": { "dtype": "float32", "shape": [1, 2, 576, 64], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.1 } }, "v": { "dtype": "float32", "shape": [1, 2, 576, 64], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 512, 64], "tolerance": 0.0001 } } }, { "name": "materialized_sgmat_apply_wg256_gqa_d128_cross_attention_tails", "provenance": { "notes": "Head dimension 128 with two query heads per KV head, 517 queries and 523 keys exercises both sequence tails. Value offset 1.0 exposes lost normalization or tail weights at strict float32 tolerances." }, "inputs": { "q": { "dtype": "float32", "shape": [1, 2, 517, 128], "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.1 } }, "k": { "dtype": "float32", "shape": [1, 1, 523, 128], "data": { "kind": "fillFloat32", "sinStep": 0.031, "cosStep": 0.017, "scale": 0.1 } }, "v": { "dtype": "float32", "shape": [1, 1, 523, 128], "data": { "kind": "fillFloat32", "sinStep": 0.023, "cosStep": 0.041, "scale": 0.1, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 517, 128], "tolerance": 0.0001, "relTolerance": 0.00002 } } }, { "name": "rank3_query_tiles_d64_h63_c0", "attrs": { "q_num_heads": 63, "kv_num_heads": 63, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 4032], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 4032], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 4032], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 4032], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d64_h64_c0", "attrs": { "q_num_heads": 64, "kv_num_heads": 64, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 4096], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 4096], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 4096], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 4096], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d64_h123_c1", "attrs": { "q_num_heads": 123, "kv_num_heads": 123, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 7872], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 7872], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 7872], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 7872], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d64_h128_c1", "attrs": { "q_num_heads": 128, "kv_num_heads": 128, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 8192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 8192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 8192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 8192], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d96_h63_c0", "attrs": { "q_num_heads": 63, "kv_num_heads": 63, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 6048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 6048], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 6048], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 6048], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d96_h64_c0", "attrs": { "q_num_heads": 64, "kv_num_heads": 64, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 6144], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 6144], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 6144], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 6144], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d96_h123_c1", "attrs": { "q_num_heads": 123, "kv_num_heads": 123, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 11808], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 11808], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 11808], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 11808], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d96_h128_c1", "attrs": { "q_num_heads": 128, "kv_num_heads": 128, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 12288], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 12288], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 12288], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 12288], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d128_h63_c0", "attrs": { "q_num_heads": 63, "kv_num_heads": 63, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 8064], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 8064], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 8064], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 8064], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d128_h64_c0", "attrs": { "q_num_heads": 64, "kv_num_heads": 64, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 8192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 8192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 8192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 8192], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d128_h123_c1", "attrs": { "q_num_heads": 123, "kv_num_heads": 123, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 15744], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 15744], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 15744], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 15744], "tolerance": 0.0001 } } }, { "name": "rank3_query_tiles_d128_h128_c1", "attrs": { "q_num_heads": 128, "kv_num_heads": 128, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 32, 16384], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 16384], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 16384], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 16384], "tolerance": 0.0001 } } }, { "name": "rank3_memory_budget_float32_d64", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 31, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 31, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 128], "tolerance": 0.0001 } } }, { "name": "rank3_memory_budget_float32_d96", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 31, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 31, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 192], "tolerance": 0.0001 } } }, { "name": "rank3_memory_budget_float32_d128", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 31, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 31, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 256], "tolerance": 0.0001 } } }, { "name": "rank3_memory_budget_float32_d256", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 31, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 31, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 512], "tolerance": 0.0001 } } }, { "name": "rank3_empty_kv_float32", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 128], "data": { "kind": "constant", "value": 1.0 } }, "k": { "dtype": "float32", "shape": [1, 0, 128], "data": { "kind": "values", "values": [] } }, "v": { "dtype": "float32", "shape": [1, 0, 128], "data": { "kind": "values", "values": [] } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 128], "tolerance": 0 } } }, { "name": "rank3_memory_budget_float16_d64", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 31, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 31, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 128], "tolerance": 0.001 } } }, { "name": "rank3_memory_budget_float16_d96", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 31, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 31, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 192], "tolerance": 0.001 } } }, { "name": "rank3_memory_budget_float16_d128", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 31, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 31, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 256], "tolerance": 0.001 } } }, { "name": "rank3_memory_budget_float16_d256", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 31, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 31, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 512], "tolerance": 0.001 } } }, { "name": "rank3_empty_kv_float16", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 128], "data": { "kind": "constant", "value": 1.0 } }, "k": { "dtype": "float16", "shape": [1, 0, 128], "data": { "kind": "values", "values": [] } }, "v": { "dtype": "float16", "shape": [1, 0, 128], "data": { "kind": "values", "values": [] } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 128], "tolerance": 0 } } }, { "name": "rank3_query_tile_tunable_1", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 128], "tolerance": 0.0001 } }, "tunables": { "RANK3_QUERY_PACK_MIN_TILES": 1 } }, { "name": "rank3_query_tile_tunable_128", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 128], "tolerance": 0.0001 } }, "tunables": { "RANK3_QUERY_PACK_MIN_TILES": 128 } }, { "name": "rank3_query_tile_tunable_256", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 128], "tolerance": 0.0001 } }, "tunables": { "RANK3_QUERY_PACK_MIN_TILES": 256 } }, { "name": "rank3_portable_cluster_float32_d64", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 128], "tolerance": 0.0001 } } }, { "name": "rank3_portable_cluster_float32_d96", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 37, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 37, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 192], "tolerance": 0.0001 } } }, { "name": "rank3_portable_cluster_float32_d128", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 256], "tolerance": 0.0001 } } }, { "name": "rank3_portable_cluster_float32_d160", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 320], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 37, 320], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 37, 320], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 320], "tolerance": 0.0001 } } }, { "name": "rank3_portable_cluster_float32_d256", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 33, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [2, 37, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [2, 37, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 33, 512], "tolerance": 0.0001 } } }, { "name": "rank3_portable_short_keys_float32_d96", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [3, 65, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [3, 9, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [3, 9, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [3, 65, 192], "tolerance": 0.0001 } } }, { "name": "rank3_portable_short_keys_float32_d256", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [3, 65, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [3, 9, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [3, 9, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [3, 65, 512], "tolerance": 0.0001 } } }, { "name": "rank3_portable_cluster_float16_d64", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 37, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 128], "tolerance": 0.001 } } }, { "name": "rank3_portable_cluster_float16_d96", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 37, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 37, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 192], "tolerance": 0.001 } } }, { "name": "rank3_portable_cluster_float16_d128", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 256], "tolerance": 0.001 } } }, { "name": "rank3_portable_cluster_float16_d160", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 320], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 37, 320], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 37, 320], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 320], "tolerance": 0.001 } } }, { "name": "rank3_portable_cluster_float16_d256", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 33, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [2, 37, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [2, 37, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 33, 512], "tolerance": 0.001 } } }, { "name": "rank3_portable_short_keys_float16_d96", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [3, 65, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [3, 9, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [3, 9, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [3, 65, 192], "tolerance": 0.001 } } }, { "name": "rank3_portable_short_keys_float16_d256", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [3, 65, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float16", "shape": [3, 9, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float16", "shape": [3, 9, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float16", "shape": [3, 65, 512], "tolerance": 0.001 } } }, { "name": "rank3_portable_work_boundary_c0_4", "attrs": { "q_num_heads": 4, "kv_num_heads": 4, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 256], "tolerance": 0.0001 } }, "tunables": { "RANK3_PORTABLE_WIDE_LANES_MAX_TILES": 4 } }, { "name": "rank3_portable_work_boundary_c0_5", "attrs": { "q_num_heads": 4, "kv_num_heads": 4, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 256], "tolerance": 0.0001 } }, "tunables": { "RANK3_PORTABLE_WIDE_LANES_MAX_TILES": 5 } }, { "name": "rank3_portable_work_boundary_c1_2", "attrs": { "q_num_heads": 4, "kv_num_heads": 4, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 256], "tolerance": 0.0001 } }, "tunables": { "RANK3_PORTABLE_WIDE_LANES_MAX_TILES": 2 } }, { "name": "rank3_portable_work_boundary_c1_3", "attrs": { "q_num_heads": 4, "kv_num_heads": 4, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 256], "tolerance": 0.0001 } }, "tunables": { "RANK3_PORTABLE_WIDE_LANES_MAX_TILES": 3 } }, { "name": "rank3_portable_slice_budget_8", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 192], "tolerance": 0.0001 } }, "tunables": { "RANK3_PORTABLE_WIDE_LANES_MAX_TILES": 1, "RANK3_PORTABLE_MAX_VECS_PER_LANE": 8 } }, { "name": "rank3_portable_slice_budget_16", "attrs": { "q_num_heads": 2, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.3 } }, "k": { "dtype": "float32", "shape": [1, 37, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.3 } }, "v": { "dtype": "float32", "shape": [1, 37, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.3 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 192], "tolerance": 0.0001 } }, "tunables": { "RANK3_PORTABLE_WIDE_LANES_MAX_TILES": 1, "RANK3_PORTABLE_MAX_VECS_PER_LANE": 16 } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d31_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 124], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 124], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 124], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 124], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d32_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 128], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d33_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 132], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 132], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 132], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 132], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d33_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 132], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 132], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 132], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 132], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d63_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 252], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 252], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 252], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 252], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d65_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 260], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 260], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 260], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 260], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d96_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 384], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 384], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 384], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 384], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d127_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 508], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 508], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 508], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 508], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d128_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d129_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 516], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 516], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 516], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 516], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d129_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 516], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 516], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 516], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 516], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d255_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1020], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1020], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1020], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1020], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d255_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1020], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1020], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1020], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1020], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d256_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1024], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d256_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1024], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh4_d257_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1028], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1028], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1028], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1028], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q511_k512_h4_kvh4_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 511, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 511, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q512_k511_h4_kvh4_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 511, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 511, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q512_k512_h4_kvh4_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k513_h4_kvh4_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh1_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh1_d64_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h8_kvh2_d256_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 8, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 2048], "tolerance": 0.00001 } } }, { "name": "materialized_head_chunk_b1_q513_k527_h8_kvh2_d256_c0", "provenance": { "notes": "Two GQA groups with four heads each and query/key tails check independent attention output across head groups." }, "tunables": { "MATERIALIZED_SGMAT_HEAD_CHUNK": 4 }, "attrs": { "is_causal": 0, "q_num_heads": 8, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 2048], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h8_kvh2_d256_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 8, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 2048], "tolerance": 0.00001 } } }, { "name": "materialized_b2_q513_k527_h4_kvh2_d96_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 513, 384], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [2, 527, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [2, 527, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 513, 384], "tolerance": 0.00001 } } }, { "name": "materialized_b2_q513_k527_h4_kvh2_d96_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 513, 384], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [2, 527, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [2, 527, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 513, 384], "tolerance": 0.00001 } } }, { "name": "materialized_b3_q512_k512_h2_kvh1_d128_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [3, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [3, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [3, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [3, 512, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b3_q512_k512_h2_kvh1_d128_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [3, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [3, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [3, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [3, 512, 256], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c0_s0", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 2, "scale": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c1_s0", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 2, "scale": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c0_s-0.125", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 2, "scale": -0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c1_s-0.125", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 2, "scale": -0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c0_s0.0625", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 2, "scale": 0.0625 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c1_s0.0625", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 2, "scale": 0.0625 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c0_s1", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 2, "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "materialized_b1_q513_k527_h4_kvh2_d128_c1_s1", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 2, "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 512], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d257_c0", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 514], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 514], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d257_c1", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 514], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 514], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d288_c0", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 576], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 288], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 288], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 576], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d512_c0", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1024], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d513_c0", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1026], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1026], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d513_c1", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1026], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1026], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d1024_c0", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 2048], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d1025_c0", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 2050], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 2050], "tolerance": 0.00001 } } }, { "name": "wide_b1_q513_k527_h2_kvh1_d1025_c1", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 2050], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 2050], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d258_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 774], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 258], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 258], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 774], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d259_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 777], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 259], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 259], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 777], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d287_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 861], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 287], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 287], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 861], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d289_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 867], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 289], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 289], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 867], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d511_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 1533], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 511], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 511], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 1533], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d767_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 2301], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 767], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 767], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 2301], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d1023_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 3069], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 1023], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 1023], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 3069], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d1536_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 4608], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 1536], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 1536], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 4608], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d2048_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 6144], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 6144], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k513_h3_kvh1_d2048_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 6144], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 513, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 513, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 6144], "tolerance": 0.00001 } } }, { "name": "general_float32_b2_q512_k527_h3_kvh1_d257_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 512, 771], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [2, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [2, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 512, 771], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q513_k512_h1_kvh1_d257_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 1, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 257], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 512, 257], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 512, 257], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 257], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k512_h4_kvh4_d257_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 1028], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 512, 1028], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 512, 1028], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 1028], "tolerance": 0.00001 } } }, { "name": "general_float32_b2_q512_k527_h3_kvh1_d513_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 512, 1539], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [2, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [2, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 512, 1539], "tolerance": 0.00001 } } }, { "name": "general_float32_b2_q512_k527_h3_kvh1_d1025_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [2, 512, 3075], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [2, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [2, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [2, 512, 3075], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q513_k512_h1_kvh1_d1025_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 1, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 512, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 512, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 1025], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k512_h4_kvh4_d1025_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 4 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 4100], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 512, 4100], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 512, 4100], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 4100], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q511_k512_h2_kvh1_d513_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 511, 1026], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 512, 513], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 512, 513], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 511, 1026], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q512_k511_h2_kvh1_d513_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 512, 1026], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 511, 513], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 511, 513], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 1026], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q513_k527_h3_kvh1_d1025_c0_s0", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1, "scale": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 3075], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 3075], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q513_k527_h3_kvh1_d1025_c1_s0", "attrs": { "is_causal": 1, "q_num_heads": 3, "kv_num_heads": 1, "scale": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 3075], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 3075], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q513_k527_h3_kvh1_d1025_c0_s-0.125", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1, "scale": -0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 3075], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 3075], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q513_k527_h3_kvh1_d1025_c1_s-0.125", "attrs": { "is_causal": 1, "q_num_heads": 3, "kv_num_heads": 1, "scale": -0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 3075], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 3075], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q513_k527_h3_kvh1_d1025_c0_s1", "attrs": { "is_causal": 0, "q_num_heads": 3, "kv_num_heads": 1, "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 3075], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 3075], "tolerance": 0.00001 } } }, { "name": "general_float32_b1_q513_k527_h3_kvh1_d1025_c1_s1", "attrs": { "is_causal": 1, "q_num_heads": 3, "kv_num_heads": 1, "scale": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 3075], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 1025], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 3075], "tolerance": 0.00001 } } }, { "name": "materialized_narrow_r3_d1_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 4], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 2], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 2], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 4], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d1_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 4], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 2], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 2], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 4], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d7_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 28], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 14], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 14], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 28], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d7_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 28], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 14], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 14], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 28], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d8_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 32], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 16], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 16], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 32], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d8_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 32], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 16], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 16], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 32], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d15_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 60], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 30], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 30], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 60], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d15_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 60], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 30], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 30], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 60], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d16_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 32], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 32], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 64], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d16_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 32], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 32], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 64], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d17_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 68], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 34], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 34], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 68], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d17_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 68], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 34], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 34], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 68], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d31_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 124], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 62], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 62], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 124], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d31_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 124], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 62], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 62], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 124], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d32_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 128], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r3_d32_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 513, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 513, 128], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r4_d1_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 513, 1], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 2, 527, 1], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 2, 527, 1], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 513, 1], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r4_d1_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 513, 1], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 2, 527, 1], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 2, 527, 1], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 513, 1], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r4_d15_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 513, 15], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 2, 527, 15], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 2, 527, 15], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 513, 15], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r4_d15_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 513, 15], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 2, 527, 15], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 2, 527, 15], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 513, 15], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r4_d31_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 513, 31], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 2, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 2, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 513, 31], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_narrow_r4_d31_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 4, 513, 31], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 2, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 2, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 4, 513, 31], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d256_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 1024], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d256_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 1024], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d288_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 1152], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 576], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 576], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 1152], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d288_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 1152], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 576], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 576], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 1152], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d480_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 1920], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 960], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 960], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 1920], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d480_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 1920], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 960], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 960], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 1920], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d512_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 2048], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d512_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 2048], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d544_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 2176], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 1088], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 1088], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 2176], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_d544_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 2176], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 1088], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 1088], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 2176], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d256_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 1024], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d256_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 1024], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d288_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 1152], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 576], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 576], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 1152], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d288_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 1152], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 576], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 576], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 1152], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d480_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 1920], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 960], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 960], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 1920], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d480_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 1920], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 960], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 960], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 1920], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d512_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 2048], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d512_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 2048], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d544_c0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 2176], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 1088], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 1088], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 2176], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_d544_c1", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 2176], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 1088], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 1088], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 2176], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_full_float16_c0", "attrs": { "q_num_heads": 2, "kv_num_heads": 1, "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 512, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 1024], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_full_float16_c1", "attrs": { "q_num_heads": 2, "kv_num_heads": 1, "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 512, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 1024], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_explicit_scale_0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1, "scale": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 2048], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float16_explicit_scale_-0.125", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1, "scale": -0.125 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 33, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 33, 2048], "tolerance": 0.001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_explicit_scale_0", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1, "scale": 0 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 2048], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "rank3_cluster_float32_explicit_scale_-0.125", "attrs": { "q_num_heads": 4, "kv_num_heads": 2, "is_causal": 1, "scale": -0.125 }, "inputs": { "q": { "dtype": "float32", "shape": [1, 33, 2048], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float32", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float32", "shape": [1, 47, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float32", "shape": [1, 33, 2048], "tolerance": 0.00001, "relTolerance": 0 } } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d1_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 2], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 1], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 1], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 2], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 1; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d1_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 2], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 1], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 1], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 2], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 1; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d7_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 14], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 7], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 7], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 14], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 7; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d7_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 14], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 7], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 7], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 14], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 7; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d31_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 62], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 62], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 31; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d31_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 62], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 62], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 31; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d32_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 32], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 32], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 32; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d32_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 32], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 32], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 32; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 64; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d64_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 64; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d96_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 192], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 96; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d96_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 192], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 96; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d127_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 254], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 127], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 127], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 254], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 127; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d127_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 254], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 127], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 127], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 254], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 127; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d128_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 256], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 128; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d128_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 256], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 128; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d256_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 512], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 256; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d256_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 512], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 256; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d257_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 514], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 514], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 257; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d257_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 514], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 257], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 514], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 257; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d288_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 576], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 288], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 288], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 576], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 288; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d288_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 576], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 288], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 288], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 576], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 288; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d511_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 1022], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 511], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 511], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 1022], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 511; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d511_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 1022], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 511], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 511], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 1022], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 511; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d512_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 1024], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 512; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d512_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 1024], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 512; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d513_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 1026], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 1026], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 513; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d513_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 1026], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 513], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 1026], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 513; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r4_b1_q513_k527_h2_kvh1_d31_c0_sdefault", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 513, 31], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 513, 31], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 31; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r4_b1_q513_k527_h2_kvh1_d31_c1_sdefault", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 513, 31], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 527, 31], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 513, 31], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 31; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r4_b1_q513_k527_h2_kvh1_d64_c0_sdefault", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 513, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 513, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 64; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r4_b1_q513_k527_h2_kvh1_d64_c1_sdefault", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 513, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 513, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 64; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r4_b1_q513_k527_h2_kvh1_d256_c0_sdefault", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 513, 256], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 256; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r4_b1_q513_k527_h2_kvh1_d256_c1_sdefault", "attrs": { "is_causal": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 513, 256], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 256; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b2_q512_k512_h4_kvh2_d128_c0_s0", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 2, "scale": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 512, 512], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 4 query heads grouped 2:1 over 2 key/value heads, 2 batches, head width 128; scores and probabilities remain float32. An explicit scale of 0 overrides the default 1/sqrt(head_dim), zeroing the pre-softmax logits." } }, { "name": "materialized_half_r3_b2_q512_k512_h4_kvh2_d128_c0_s-0.125", "attrs": { "is_causal": 0, "q_num_heads": 4, "kv_num_heads": 2, "scale": -0.125 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 512, 512], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 4 query heads grouped 2:1 over 2 key/value heads, 2 batches, head width 128; scores and probabilities remain float32. An explicit scale of -0.125 overrides the default 1/sqrt(head_dim)." } }, { "name": "materialized_half_r3_b2_q512_k512_h4_kvh2_d128_c1_s0", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 2, "scale": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 512, 512], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 4 query heads grouped 2:1 over 2 key/value heads, 2 batches, head width 128; scores and probabilities remain float32. Causal masking restricts each query to keys at or before its own position; an explicit scale of 0 overrides the default 1/sqrt(head_dim), zeroing the pre-softmax logits." } }, { "name": "materialized_half_r3_b2_q512_k512_h4_kvh2_d128_c1_s-0.125", "attrs": { "is_causal": 1, "q_num_heads": 4, "kv_num_heads": 2, "scale": -0.125 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 512, 512], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 4 query heads grouped 2:1 over 2 key/value heads, 2 batches, head width 128; scores and probabilities remain float32. Causal masking restricts each query to keys at or before its own position; an explicit scale of -0.125 overrides the default 1/sqrt(head_dim)." } }, { "name": "materialized_half_r3_b1_q511_k512_h2_kvh1_d512_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 511, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 512, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 511, 1024], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 512; scores and probabilities remain float32. Query length 511 exercises a partial final query tile while key length 512 is tile-aligned." } }, { "name": "materialized_half_r3_b1_q512_k511_h2_kvh1_d512_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 512, 1024], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 511, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 511, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 1024], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 512; scores and probabilities remain float32. Key length 511 exercises a partial final key tile while query length 512 is tile-aligned." } }, { "name": "materialized_half_r3_b1_q2048_k2048_h2_kvh1_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2048, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2048, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2048, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2048, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 64; scores and probabilities remain float32." } }, { "name": "materialized_half_r3_b1_q2048_k2048_h2_kvh1_d64_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2048, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2048, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2048, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2048, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 64; scores and probabilities remain float32. Causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q2048_k2048_h2_kvh1_d96_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2048, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2048, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2048, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2048, 192], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 96; scores and probabilities remain float32." } }, { "name": "materialized_half_r3_b1_q2048_k2048_h2_kvh1_d96_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2048, 192], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2048, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2048, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2048, 192], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 96; scores and probabilities remain float32. Causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d160_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 320], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 160], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 160], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 320], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 160; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d160_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 320], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 160], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 160], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 320], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 160; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d192_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 384], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 384], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 192; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d192_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 384], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 192], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 192], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 384], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 192; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d224_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 448], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 224], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 224], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 448], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 224; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh1_d224_c1_sdefault", "attrs": { "is_causal": 1, "q_num_heads": 2, "kv_num_heads": 1 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 448], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 224], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 224], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 448], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 224; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles; causal masking restricts each query to keys at or before its own position." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh2_d64_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads and 2 key/value heads (no grouping), head width 64; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh2_d128_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 256], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads and 2 key/value heads (no grouping), head width 128; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_r3_b1_q513_k527_h2_kvh2_d256_c0_sdefault", "attrs": { "is_causal": 0, "q_num_heads": 2, "kv_num_heads": 2 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 513, 512], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 527, 512], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 513, 512], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-three float16 attention with 2 query heads and 2 key/value heads (no grouping), head width 256; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_rank4_grid_h4_kvh1_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 4, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 4 query heads grouped 4:1 over 1 key/value head, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_grid_h8_kvh2_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_grid_h16_kvh4_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 16, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 4, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 16, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 16 query heads grouped 4:1 over 4 key/value heads, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_grid_h8_kvh2_d64_q513_k527", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 513, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 527, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 513, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 64; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_rank4_grid_h4_kvh1_d96_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 4, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 4, 512, 96], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 4 query heads grouped 4:1 over 1 key/value head, head width 96, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_grid_h8_kvh2_d96_q513_k527", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 513, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 527, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 527, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 513, 96], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 96; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_rank4_grid_h8_kvh2_d128_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 128, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_grid_h16_kvh4_d128_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 16, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 4, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 4, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 16, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 16 query heads grouped 4:1 over 4 key/value heads, head width 128, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_grid_h32_kvh8_d128_q513_k527", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 513, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 513, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 32 query heads grouped 4:1 over 8 key/value heads, head width 128; scores and probabilities remain float32. Query length 513 and key length 527 exercise partial final query and key tiles." } }, { "name": "materialized_half_rank4_grid_h32_kvh8_d128_q512_k527", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 527, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 32 query heads grouped 4:1 over 8 key/value heads, head width 128; scores and probabilities remain float32. Key length 527 exercises a partial final key tile while query length 512 is tile-aligned." } }, { "name": "materialized_half_rank4_grid_h16_kvh4_d256_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 16, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 4, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 4, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 16, 512, 256], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 16 query heads grouped 4:1 over 4 key/value heads, head width 256, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_grid_h32_kvh8_d256_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 512, 256], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 32 query heads grouped 4:1 over 8 key/value heads, head width 256, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_boundary_h2_kvh1_d96_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 512, 96], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 96, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_boundary_h32_kvh8_d96_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 512, 96], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 32 query heads grouped 4:1 over 8 key/value heads, head width 96, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_boundary_h32_kvh8_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 32, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 8, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 8, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 32, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 32 query heads grouped 4:1 over 8 key/value heads, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_boundary_h24_kvh6_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 24, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 6, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 6, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 24 query heads grouped 4:1 over 6 key/value heads, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_boundary_h24_kvh6_d128_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 24, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 6, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 6, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 24 query heads grouped 4:1 over 6 key/value heads, head width 128, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_boundary_h24_kvh6_d256_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 24, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 6, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 6, 512, 256], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 24, 512, 256], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 24 query heads grouped 4:1 over 6 key/value heads, head width 256, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h12_kvh3_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 12, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 3, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 3, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 12, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 12 query heads grouped 4:1 over 3 key/value heads, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h12_kvh3_d128_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 12, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 3, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 3, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 12, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 12 query heads grouped 4:1 over 3 key/value heads, head width 128, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h20_kvh5_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 20, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 5, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 5, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 20, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 20 query heads grouped 4:1 over 5 key/value heads, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h20_kvh5_d128_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 20, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 5, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 5, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 20, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 20 query heads grouped 4:1 over 5 key/value heads, head width 128, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b2_h8_kvh2_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 8, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 2, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 8, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, 2 batches, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b2_h8_kvh2_d128_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 8, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 2, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 2, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 8, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, 2 batches, head width 128, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b2_h16_kvh4_d64_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 16, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 4, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 4, 512, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 16, 512, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 16 query heads grouped 4:1 over 4 key/value heads, 2 batches, head width 64, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b2_h16_kvh4_d96_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 16, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 4, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 4, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 16, 512, 96], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 16 query heads grouped 4:1 over 4 key/value heads, 2 batches, head width 96, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b2_h16_kvh4_d128_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [2, 16, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [2, 4, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [2, 4, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [2, 16, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 16 query heads grouped 4:1 over 4 key/value heads, 2 batches, head width 128, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d64_q768_k768", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 768, 64], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 768, 64], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 768, 64], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 768, 64], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 64, and 768 queries over 768 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d96_q768_k768", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 768, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 768, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 768, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 768, 96], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 96, and 768 queries over 768 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d128_q768_k768", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 768, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 768, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 768, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 768, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 128, and 768 queries over 768 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h16_kvh4_d128_q768_k768", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 16, 768, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 4, 768, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 4, 768, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 16, 768, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 16 query heads grouped 4:1 over 4 key/value heads, head width 128, and 768 queries over 768 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d96_q512_k1023", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 1023, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 1023, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 512, 96], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 96; scores and probabilities remain float32. Key length 1023 exercises a partial final key tile while query length 512 is tile-aligned." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d96_q512_k1024", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 512, 96], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 1024, 96], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 1024, 96], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 512, 96], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 96, and 512 queries over 1024 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d128_q512_k1023", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 1023, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 1023, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 128; scores and probabilities remain float32. Key length 1023 exercises a partial final key tile while query length 512 is tile-aligned." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d128_q512_k1024", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 512, 128], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 1024, 128], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 512, 128], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 128, and 512 queries over 1024 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h2_kvh1_d80_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 512, 80], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 512, 80], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 512, 80], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 512, 80], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 80, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d80_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 512, 80], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 512, 80], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 512, 80], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 512, 80], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 80, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h2_kvh1_d112_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 2, 512, 112], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 1, 512, 112], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 1, 512, 112], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 2, 512, 112], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 2 query heads grouped 2:1 over 1 key/value head, head width 112, and 512 queries over 512 keys; scores and probabilities remain float32." } }, { "name": "materialized_half_rank4_selection_b1_h8_kvh2_d112_q512_k512", "attrs": { "is_causal": 0 }, "inputs": { "q": { "dtype": "float16", "shape": [1, 8, 512, 112], "data": { "kind": "fillFloat32", "sinStep": 0.021, "cosStep": 0.033, "scale": 0.2 } }, "k": { "dtype": "float16", "shape": [1, 2, 512, 112], "data": { "kind": "fillFloat32", "sinStep": 0.039, "cosStep": 0.011, "scale": 0.2 } }, "v": { "dtype": "float16", "shape": [1, 2, 512, 112], "data": { "kind": "fillFloat32", "sinStep": 0.027, "cosStep": 0.049, "scale": 0.2, "offset": 1.0 } } }, "outputs": { "y": { "dtype": "float16", "shape": [1, 8, 512, 112], "tolerance": 0.001, "relTolerance": 0 } }, "provenance": { "source": "synthetic", "notes": "Rank-four float16 attention with 8 query heads grouped 4:1 over 2 key/value heads, head width 112, and 512 queries over 512 keys; scores and probabilities remain float32." } } ] }