gahow/aituner
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Evaluate vLLM 0.20 profiles against Frontier

Browse Source
This commit is contained in:
Gahow Wang
2026-07-16 23:59:10 +08:00
parent 008324e70c
commit 76107d3e87
33 changed files with 13973 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
runs/frontier-qwen30-vllm020-profile-v1/frozen/profile-v2/attention_true_mixed_fused.csv Normal file
View File

@@ -0,0 +1,31 @@
num_tensor_parallel_workers,batch_spec,num_prefill_seqs,num_decode_seqs,total_prefill_tokens,total_decode_tokens,decode_avg_kv_cache_size,attention_core_mean_ms,attention_core_mean_as_median_ms,kv_cache_update_median_ms,pure_prefill_reference_mean_ms,pure_decode_reference_mean_ms,projected_prefill_mean_ms,projected_decode_mean_ms,projection_sum_error_ms,representation
1,q64_8q1s512,1,8,64,8,512.0,0.06604800000786781,0.06604800000786781,0.021359999664127827,0.05810240097343922,0.0600041144660541,0.032492258074592426,0.03355574193327539,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q128_8q1s1k,1,8,128,8,1024.0,0.07483199909329416,0.07483199909329416,0.024255999363958836,0.05587520003318787,0.06090559959411622,0.03580428402241854,0.03902771507087562,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q128_16q1s1k,1,16,128,16,1024.0,0.10597119852900506,0.10597119852900506,0.018112000077962875,0.05587520003318787,0.06891520097851753,0.04744885718421094,0.05852234134479412,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q256_16q1s2k,1,16,256,16,2048.0,0.13697599917650224,0.13697599917650224,0.01673599984496832,0.056396800279617305,0.11392960175871848,0.04535414343408448,0.09162185574241775,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q256_32q1s2k,1,32,256,32,2048.0,0.2163648009300232,0.2163648009300232,0.01515199989080429,0.056396800279617305,0.13807999789714814,0.06274415549817247,0.15362064543185072,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q512_32q1s4k,1,32,512,32,4096.0,0.3746495962142944,0.3746495962142944,0.016191999427974224,0.07432000041007995,0.2231455981731415,0.09360396051475776,0.2810456356995366,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q512_64q1s4k,1,64,512,64,4096.0,0.6718560099601746,0.6718560099601746,0.024848000146448612,0.07432000041007995,0.4058080047369003,0.10399797220840973,0.5678580377517648,-1.1102230246251565e-16,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q1k_64q1s8k,1,64,1024,64,8192.0,1.3472000002861022,1.3472000002861022,0.019600000232458115,0.1307712011039257,0.7464000046253203,0.2008444429250931,1.1463555573610091,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q2k_32q1s4k,1,32,2048,32,4096.0,0.6351647913455962,0.6351647913455962,0.028256000019609928,0.33396480381488797,0.2231455981731415,0.38075520430942184,0.25440958703617433,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1,q4k_16q1s4k,1,16,4096,16,4096.0,1.2515872120857237,1.2515872120857237,0.04395199939608574,1.0999775886535645,0.1389280006289482,1.1112371236754128,0.14035008841031085,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q64_8q1s512,1,8,64,8,512.0,0.05652800053358078,0.05652800053358078,0.019600000232458115,0.05961279980838299,0.055968457407185014,0.029155180178492036,0.027372820355088746,6.938893903907228e-18,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q128_8q1s1k,1,8,128,8,1024.0,0.07362559959292413,0.07362559959292413,0.014431999996304512,0.05882879942655563,0.05778240002691746,0.03714313592014963,0.036482463672774496,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q128_16q1s1k,1,16,128,16,1024.0,0.0736224003136158,0.0736224003136158,0.014479999896138906,0.05882879942655563,0.06270079985260964,0.03563837490653034,0.037984025407085474,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q256_16q1s2k,1,16,256,16,2048.0,0.10047360062599181,0.10047360062599181,0.01425600005313754,0.05459520071744919,0.06628479920327664,0.045378692890289625,0.05509490773570219,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q256_32q1s2k,1,32,256,32,2048.0,0.14032640159130094,0.14032640159130094,0.01566399959847331,0.05459520071744919,0.09493440166115762,0.05123499252968345,0.0890914090616175,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q512_32q1s4k,1,32,512,32,4096.0,0.22105600088834762,0.22105600088834762,0.016032000072300434,0.06650560013949873,0.13783999979496003,0.07194410845270183,0.14911189243564577,-2.7755575615628914e-17,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q512_64q1s4k,1,64,512,64,4096.0,0.36867519915103913,0.36867519915103913,0.015263999812304974,0.06650560013949873,0.22715839892625808,0.0834932625520734,0.2851819365989658,5.551115123125783e-17,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q1k_64q1s8k,1,64,1024,64,8192.0,0.6853824079036711,0.6853824079036711,0.015856000129133463,0.08607039973139763,0.40332479774951924,0.12053885718696096,0.5648435507167102,1.1102230246251565e-16,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q2k_32q1s4k,1,32,2048,32,4096.0,0.3503839999437332,0.3503839999437332,0.018432000651955605,0.19546559900045393,0.13783999979496003,0.2054811520619412,0.14490284788179197,-5.551115123125783e-17,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
2,q4k_16q1s4k,1,16,4096,16,4096.0,0.7185311973094941,0.7185311973094941,0.026575999334454536,0.5800191938877105,0.10921279862523078,0.6046757700946376,0.11385542721485654,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q64_8q1s512,1,8,64,8,512.0,0.0562208004295826,0.0562208004295826,0.015520000364631414,0.06054079942405224,0.0591108573866742,0.028446344104128624,0.027774456325453972,-6.938893903907228e-18,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q128_8q1s1k,1,8,128,8,1024.0,0.061199999228119854,0.061199999228119854,0.018240000121295452,0.05626560002565384,0.0588383998721838,0.029916029687899703,0.03128396954022015,-6.938893903907228e-18,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q128_16q1s1k,1,16,128,16,1024.0,0.06104319989681244,0.06104319989681244,0.016207999549806118,0.05626560002565384,0.05755840018391609,0.030174939058162802,0.03086826083864964,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q256_16q1s2k,1,16,256,16,2048.0,0.07356479987502099,0.07356479987502099,0.015584000386297703,0.0556256003677845,0.06238719932734965,0.0346749350032807,0.038889864871740294,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q256_32q1s2k,1,32,256,32,2048.0,0.09806400015950203,0.09806400015950203,0.015008000191301107,0.0556256003677845,0.06935679838061332,0.04364509673334097,0.05441890342616107,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q512_32q1s4k,1,32,512,32,4096.0,0.1509471982717514,0.1509471982717514,0.014864000026136637,0.0640383992344141,0.10621120035648347,0.05677791295527661,0.09416928531647478,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q512_64q1s4k,1,64,512,64,4096.0,0.21687040030956264,0.21687040030956264,0.015343999955803156,0.0640383992344141,0.14040640145540234,0.06793047918211377,0.1489399211274489,2.7755575615628914e-17,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q1k_64q1s8k,1,64,1024,64,8192.0,0.3848575979471207,0.3848575979471207,0.015647999942302704,0.08209280073642732,0.22809920012950893,0.10185316839884552,0.2830044295482752,0.0,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q2k_32q1s4k,1,32,2048,32,4096.0,0.2785151988267898,0.2785151988267898,0.01726400014013052,0.12431039959192276,0.10621120035648347,0.1501912864839173,0.1283239123428725,-5.551115123125783e-17,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4,q4k_16q1s4k,1,16,4096,16,4096.0,0.456512001156807,0.456512001156807,0.017680000513792038,0.3261695951223373,0.06970879957079888,0.3761264483787333,0.08038555277807365,-5.551115123125783e-17,one_fused_FA3_call_projected_for_Frontier_with_total_conservation
1 num_tensor_parallel_workers batch_spec num_prefill_seqs num_decode_seqs total_prefill_tokens total_decode_tokens decode_avg_kv_cache_size attention_core_mean_ms attention_core_mean_as_median_ms kv_cache_update_median_ms pure_prefill_reference_mean_ms pure_decode_reference_mean_ms projected_prefill_mean_ms projected_decode_mean_ms projection_sum_error_ms representation
2 1 q64_8q1s512 1 8 64 8 512.0 0.06604800000786781 0.06604800000786781 0.021359999664127827 0.05810240097343922 0.0600041144660541 0.032492258074592426 0.03355574193327539 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
3 1 q128_8q1s1k 1 8 128 8 1024.0 0.07483199909329416 0.07483199909329416 0.024255999363958836 0.05587520003318787 0.06090559959411622 0.03580428402241854 0.03902771507087562 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
4 1 q128_16q1s1k 1 16 128 16 1024.0 0.10597119852900506 0.10597119852900506 0.018112000077962875 0.05587520003318787 0.06891520097851753 0.04744885718421094 0.05852234134479412 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
5 1 q256_16q1s2k 1 16 256 16 2048.0 0.13697599917650224 0.13697599917650224 0.01673599984496832 0.056396800279617305 0.11392960175871848 0.04535414343408448 0.09162185574241775 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
6 1 q256_32q1s2k 1 32 256 32 2048.0 0.2163648009300232 0.2163648009300232 0.01515199989080429 0.056396800279617305 0.13807999789714814 0.06274415549817247 0.15362064543185072 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
7 1 q512_32q1s4k 1 32 512 32 4096.0 0.3746495962142944 0.3746495962142944 0.016191999427974224 0.07432000041007995 0.2231455981731415 0.09360396051475776 0.2810456356995366 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
8 1 q512_64q1s4k 1 64 512 64 4096.0 0.6718560099601746 0.6718560099601746 0.024848000146448612 0.07432000041007995 0.4058080047369003 0.10399797220840973 0.5678580377517648 -1.1102230246251565e-16 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
9 1 q1k_64q1s8k 1 64 1024 64 8192.0 1.3472000002861022 1.3472000002861022 0.019600000232458115 0.1307712011039257 0.7464000046253203 0.2008444429250931 1.1463555573610091 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
10 1 q2k_32q1s4k 1 32 2048 32 4096.0 0.6351647913455962 0.6351647913455962 0.028256000019609928 0.33396480381488797 0.2231455981731415 0.38075520430942184 0.25440958703617433 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
11 1 q4k_16q1s4k 1 16 4096 16 4096.0 1.2515872120857237 1.2515872120857237 0.04395199939608574 1.0999775886535645 0.1389280006289482 1.1112371236754128 0.14035008841031085 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
12 2 q64_8q1s512 1 8 64 8 512.0 0.05652800053358078 0.05652800053358078 0.019600000232458115 0.05961279980838299 0.055968457407185014 0.029155180178492036 0.027372820355088746 6.938893903907228e-18 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
13 2 q128_8q1s1k 1 8 128 8 1024.0 0.07362559959292413 0.07362559959292413 0.014431999996304512 0.05882879942655563 0.05778240002691746 0.03714313592014963 0.036482463672774496 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
14 2 q128_16q1s1k 1 16 128 16 1024.0 0.0736224003136158 0.0736224003136158 0.014479999896138906 0.05882879942655563 0.06270079985260964 0.03563837490653034 0.037984025407085474 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
15 2 q256_16q1s2k 1 16 256 16 2048.0 0.10047360062599181 0.10047360062599181 0.01425600005313754 0.05459520071744919 0.06628479920327664 0.045378692890289625 0.05509490773570219 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
16 2 q256_32q1s2k 1 32 256 32 2048.0 0.14032640159130094 0.14032640159130094 0.01566399959847331 0.05459520071744919 0.09493440166115762 0.05123499252968345 0.0890914090616175 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
17 2 q512_32q1s4k 1 32 512 32 4096.0 0.22105600088834762 0.22105600088834762 0.016032000072300434 0.06650560013949873 0.13783999979496003 0.07194410845270183 0.14911189243564577 -2.7755575615628914e-17 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
18 2 q512_64q1s4k 1 64 512 64 4096.0 0.36867519915103913 0.36867519915103913 0.015263999812304974 0.06650560013949873 0.22715839892625808 0.0834932625520734 0.2851819365989658 5.551115123125783e-17 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
19 2 q1k_64q1s8k 1 64 1024 64 8192.0 0.6853824079036711 0.6853824079036711 0.015856000129133463 0.08607039973139763 0.40332479774951924 0.12053885718696096 0.5648435507167102 1.1102230246251565e-16 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
20 2 q2k_32q1s4k 1 32 2048 32 4096.0 0.3503839999437332 0.3503839999437332 0.018432000651955605 0.19546559900045393 0.13783999979496003 0.2054811520619412 0.14490284788179197 -5.551115123125783e-17 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
21 2 q4k_16q1s4k 1 16 4096 16 4096.0 0.7185311973094941 0.7185311973094941 0.026575999334454536 0.5800191938877105 0.10921279862523078 0.6046757700946376 0.11385542721485654 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
22 4 q64_8q1s512 1 8 64 8 512.0 0.0562208004295826 0.0562208004295826 0.015520000364631414 0.06054079942405224 0.0591108573866742 0.028446344104128624 0.027774456325453972 -6.938893903907228e-18 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
23 4 q128_8q1s1k 1 8 128 8 1024.0 0.061199999228119854 0.061199999228119854 0.018240000121295452 0.05626560002565384 0.0588383998721838 0.029916029687899703 0.03128396954022015 -6.938893903907228e-18 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
24 4 q128_16q1s1k 1 16 128 16 1024.0 0.06104319989681244 0.06104319989681244 0.016207999549806118 0.05626560002565384 0.05755840018391609 0.030174939058162802 0.03086826083864964 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
25 4 q256_16q1s2k 1 16 256 16 2048.0 0.07356479987502099 0.07356479987502099 0.015584000386297703 0.0556256003677845 0.06238719932734965 0.0346749350032807 0.038889864871740294 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
26 4 q256_32q1s2k 1 32 256 32 2048.0 0.09806400015950203 0.09806400015950203 0.015008000191301107 0.0556256003677845 0.06935679838061332 0.04364509673334097 0.05441890342616107 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
27 4 q512_32q1s4k 1 32 512 32 4096.0 0.1509471982717514 0.1509471982717514 0.014864000026136637 0.0640383992344141 0.10621120035648347 0.05677791295527661 0.09416928531647478 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
28 4 q512_64q1s4k 1 64 512 64 4096.0 0.21687040030956264 0.21687040030956264 0.015343999955803156 0.0640383992344141 0.14040640145540234 0.06793047918211377 0.1489399211274489 2.7755575615628914e-17 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
29 4 q1k_64q1s8k 1 64 1024 64 8192.0 0.3848575979471207 0.3848575979471207 0.015647999942302704 0.08209280073642732 0.22809920012950893 0.10185316839884552 0.2830044295482752 0.0 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
30 4 q2k_32q1s4k 1 32 2048 32 4096.0 0.2785151988267898 0.2785151988267898 0.01726400014013052 0.12431039959192276 0.10621120035648347 0.1501912864839173 0.1283239123428725 -5.551115123125783e-17 one_fused_FA3_call_projected_for_Frontier_with_total_conservation
31 4 q4k_16q1s4k 1 16 4096 16 4096.0 0.456512001156807 0.456512001156807 0.017680000513792038 0.3261695951223373 0.06970879957079888 0.3761264483787333 0.08038555277807365 -5.551115123125783e-17 one_fused_FA3_call_projected_for_Frontier_with_total_conservation