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-v1/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,representation
1,q64_8q1s512,1,8,64,8,511.0,0.06604800000786781,0.06604800000786781,0.021359999664127827,one_fused_FA3_call_not_split_into_prefill_decode
1,q128_8q1s1k,1,8,128,8,1023.0,0.07483199909329416,0.07483199909329416,0.024255999363958836,one_fused_FA3_call_not_split_into_prefill_decode
1,q128_16q1s1k,1,16,128,16,1023.0,0.10597119852900506,0.10597119852900506,0.018112000077962875,one_fused_FA3_call_not_split_into_prefill_decode
1,q256_16q1s2k,1,16,256,16,2047.0,0.13697599917650224,0.13697599917650224,0.01673599984496832,one_fused_FA3_call_not_split_into_prefill_decode
1,q256_32q1s2k,1,32,256,32,2047.0,0.2163648009300232,0.2163648009300232,0.01515199989080429,one_fused_FA3_call_not_split_into_prefill_decode
1,q512_32q1s4k,1,32,512,32,4095.0,0.3746495962142944,0.3746495962142944,0.016191999427974224,one_fused_FA3_call_not_split_into_prefill_decode
1,q512_64q1s4k,1,64,512,64,4095.0,0.6718560099601746,0.6718560099601746,0.024848000146448612,one_fused_FA3_call_not_split_into_prefill_decode
1,q1k_64q1s8k,1,64,1024,64,8191.0,1.3472000002861022,1.3472000002861022,0.019600000232458115,one_fused_FA3_call_not_split_into_prefill_decode
1,q2k_32q1s4k,1,32,2048,32,4095.0,0.6351647913455962,0.6351647913455962,0.028256000019609928,one_fused_FA3_call_not_split_into_prefill_decode
1,q4k_16q1s4k,1,16,4096,16,4095.0,1.2515872120857237,1.2515872120857237,0.04395199939608574,one_fused_FA3_call_not_split_into_prefill_decode
2,q64_8q1s512,1,8,64,8,511.0,0.05652800053358078,0.05652800053358078,0.019600000232458115,one_fused_FA3_call_not_split_into_prefill_decode
2,q128_8q1s1k,1,8,128,8,1023.0,0.07362559959292413,0.07362559959292413,0.014431999996304512,one_fused_FA3_call_not_split_into_prefill_decode
2,q128_16q1s1k,1,16,128,16,1023.0,0.0736224003136158,0.0736224003136158,0.014479999896138906,one_fused_FA3_call_not_split_into_prefill_decode
2,q256_16q1s2k,1,16,256,16,2047.0,0.10047360062599181,0.10047360062599181,0.01425600005313754,one_fused_FA3_call_not_split_into_prefill_decode
2,q256_32q1s2k,1,32,256,32,2047.0,0.14032640159130094,0.14032640159130094,0.01566399959847331,one_fused_FA3_call_not_split_into_prefill_decode
2,q512_32q1s4k,1,32,512,32,4095.0,0.22105600088834762,0.22105600088834762,0.016032000072300434,one_fused_FA3_call_not_split_into_prefill_decode
2,q512_64q1s4k,1,64,512,64,4095.0,0.36867519915103913,0.36867519915103913,0.015263999812304974,one_fused_FA3_call_not_split_into_prefill_decode
2,q1k_64q1s8k,1,64,1024,64,8191.0,0.6853824079036711,0.6853824079036711,0.015856000129133463,one_fused_FA3_call_not_split_into_prefill_decode
2,q2k_32q1s4k,1,32,2048,32,4095.0,0.3503839999437332,0.3503839999437332,0.018432000651955605,one_fused_FA3_call_not_split_into_prefill_decode
2,q4k_16q1s4k,1,16,4096,16,4095.0,0.7185311973094941,0.7185311973094941,0.026575999334454536,one_fused_FA3_call_not_split_into_prefill_decode
4,q64_8q1s512,1,8,64,8,511.0,0.0562208004295826,0.0562208004295826,0.015520000364631414,one_fused_FA3_call_not_split_into_prefill_decode
4,q128_8q1s1k,1,8,128,8,1023.0,0.061199999228119854,0.061199999228119854,0.018240000121295452,one_fused_FA3_call_not_split_into_prefill_decode
4,q128_16q1s1k,1,16,128,16,1023.0,0.06104319989681244,0.06104319989681244,0.016207999549806118,one_fused_FA3_call_not_split_into_prefill_decode
4,q256_16q1s2k,1,16,256,16,2047.0,0.07356479987502099,0.07356479987502099,0.015584000386297703,one_fused_FA3_call_not_split_into_prefill_decode
4,q256_32q1s2k,1,32,256,32,2047.0,0.09806400015950203,0.09806400015950203,0.015008000191301107,one_fused_FA3_call_not_split_into_prefill_decode
4,q512_32q1s4k,1,32,512,32,4095.0,0.1509471982717514,0.1509471982717514,0.014864000026136637,one_fused_FA3_call_not_split_into_prefill_decode
4,q512_64q1s4k,1,64,512,64,4095.0,0.21687040030956264,0.21687040030956264,0.015343999955803156,one_fused_FA3_call_not_split_into_prefill_decode
4,q1k_64q1s8k,1,64,1024,64,8191.0,0.3848575979471207,0.3848575979471207,0.015647999942302704,one_fused_FA3_call_not_split_into_prefill_decode
4,q2k_32q1s4k,1,32,2048,32,4095.0,0.2785151988267898,0.2785151988267898,0.01726400014013052,one_fused_FA3_call_not_split_into_prefill_decode
4,q4k_16q1s4k,1,16,4096,16,4095.0,0.456512001156807,0.456512001156807,0.017680000513792038,one_fused_FA3_call_not_split_into_prefill_decode
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 representation
2 1 q64_8q1s512 1 8 64 8 511.0 0.06604800000786781 0.06604800000786781 0.021359999664127827 one_fused_FA3_call_not_split_into_prefill_decode
3 1 q128_8q1s1k 1 8 128 8 1023.0 0.07483199909329416 0.07483199909329416 0.024255999363958836 one_fused_FA3_call_not_split_into_prefill_decode
4 1 q128_16q1s1k 1 16 128 16 1023.0 0.10597119852900506 0.10597119852900506 0.018112000077962875 one_fused_FA3_call_not_split_into_prefill_decode
5 1 q256_16q1s2k 1 16 256 16 2047.0 0.13697599917650224 0.13697599917650224 0.01673599984496832 one_fused_FA3_call_not_split_into_prefill_decode
6 1 q256_32q1s2k 1 32 256 32 2047.0 0.2163648009300232 0.2163648009300232 0.01515199989080429 one_fused_FA3_call_not_split_into_prefill_decode
7 1 q512_32q1s4k 1 32 512 32 4095.0 0.3746495962142944 0.3746495962142944 0.016191999427974224 one_fused_FA3_call_not_split_into_prefill_decode
8 1 q512_64q1s4k 1 64 512 64 4095.0 0.6718560099601746 0.6718560099601746 0.024848000146448612 one_fused_FA3_call_not_split_into_prefill_decode
9 1 q1k_64q1s8k 1 64 1024 64 8191.0 1.3472000002861022 1.3472000002861022 0.019600000232458115 one_fused_FA3_call_not_split_into_prefill_decode
10 1 q2k_32q1s4k 1 32 2048 32 4095.0 0.6351647913455962 0.6351647913455962 0.028256000019609928 one_fused_FA3_call_not_split_into_prefill_decode
11 1 q4k_16q1s4k 1 16 4096 16 4095.0 1.2515872120857237 1.2515872120857237 0.04395199939608574 one_fused_FA3_call_not_split_into_prefill_decode
12 2 q64_8q1s512 1 8 64 8 511.0 0.05652800053358078 0.05652800053358078 0.019600000232458115 one_fused_FA3_call_not_split_into_prefill_decode
13 2 q128_8q1s1k 1 8 128 8 1023.0 0.07362559959292413 0.07362559959292413 0.014431999996304512 one_fused_FA3_call_not_split_into_prefill_decode
14 2 q128_16q1s1k 1 16 128 16 1023.0 0.0736224003136158 0.0736224003136158 0.014479999896138906 one_fused_FA3_call_not_split_into_prefill_decode
15 2 q256_16q1s2k 1 16 256 16 2047.0 0.10047360062599181 0.10047360062599181 0.01425600005313754 one_fused_FA3_call_not_split_into_prefill_decode
16 2 q256_32q1s2k 1 32 256 32 2047.0 0.14032640159130094 0.14032640159130094 0.01566399959847331 one_fused_FA3_call_not_split_into_prefill_decode
17 2 q512_32q1s4k 1 32 512 32 4095.0 0.22105600088834762 0.22105600088834762 0.016032000072300434 one_fused_FA3_call_not_split_into_prefill_decode
18 2 q512_64q1s4k 1 64 512 64 4095.0 0.36867519915103913 0.36867519915103913 0.015263999812304974 one_fused_FA3_call_not_split_into_prefill_decode
19 2 q1k_64q1s8k 1 64 1024 64 8191.0 0.6853824079036711 0.6853824079036711 0.015856000129133463 one_fused_FA3_call_not_split_into_prefill_decode
20 2 q2k_32q1s4k 1 32 2048 32 4095.0 0.3503839999437332 0.3503839999437332 0.018432000651955605 one_fused_FA3_call_not_split_into_prefill_decode
21 2 q4k_16q1s4k 1 16 4096 16 4095.0 0.7185311973094941 0.7185311973094941 0.026575999334454536 one_fused_FA3_call_not_split_into_prefill_decode
22 4 q64_8q1s512 1 8 64 8 511.0 0.0562208004295826 0.0562208004295826 0.015520000364631414 one_fused_FA3_call_not_split_into_prefill_decode
23 4 q128_8q1s1k 1 8 128 8 1023.0 0.061199999228119854 0.061199999228119854 0.018240000121295452 one_fused_FA3_call_not_split_into_prefill_decode
24 4 q128_16q1s1k 1 16 128 16 1023.0 0.06104319989681244 0.06104319989681244 0.016207999549806118 one_fused_FA3_call_not_split_into_prefill_decode
25 4 q256_16q1s2k 1 16 256 16 2047.0 0.07356479987502099 0.07356479987502099 0.015584000386297703 one_fused_FA3_call_not_split_into_prefill_decode
26 4 q256_32q1s2k 1 32 256 32 2047.0 0.09806400015950203 0.09806400015950203 0.015008000191301107 one_fused_FA3_call_not_split_into_prefill_decode
27 4 q512_32q1s4k 1 32 512 32 4095.0 0.1509471982717514 0.1509471982717514 0.014864000026136637 one_fused_FA3_call_not_split_into_prefill_decode
28 4 q512_64q1s4k 1 64 512 64 4095.0 0.21687040030956264 0.21687040030956264 0.015343999955803156 one_fused_FA3_call_not_split_into_prefill_decode
29 4 q1k_64q1s8k 1 64 1024 64 8191.0 0.3848575979471207 0.3848575979471207 0.015647999942302704 one_fused_FA3_call_not_split_into_prefill_decode
30 4 q2k_32q1s4k 1 32 2048 32 4095.0 0.2785151988267898 0.2785151988267898 0.01726400014013052 one_fused_FA3_call_not_split_into_prefill_decode
31 4 q4k_16q1s4k 1 16 4096 16 4095.0 0.456512001156807 0.456512001156807 0.017680000513792038 one_fused_FA3_call_not_split_into_prefill_decode