Files
aituner/runs/frontier-qwen30-vllm020-profile-v1/frozen/profile-v2/attention.csv

81 KiB

1time_stats.attn_input_reshape.mintime_stats.attn_input_reshape.maxtime_stats.attn_input_reshape.meantime_stats.attn_input_reshape.mediantime_stats.attn_input_reshape.stdtime_stats.attn_kv_cache_save.mintime_stats.attn_kv_cache_save.maxtime_stats.attn_kv_cache_save.meantime_stats.attn_kv_cache_save.mediantime_stats.attn_kv_cache_save.stdtime_stats.attn_prefill.mintime_stats.attn_prefill.maxtime_stats.attn_prefill.meantime_stats.attn_prefill.mediantime_stats.attn_prefill.stdtime_stats.attn_decode.mintime_stats.attn_decode.maxtime_stats.attn_decode.meantime_stats.attn_decode.mediantime_stats.attn_decode.stdtime_stats.attn_output_reshape.mintime_stats.attn_output_reshape.maxtime_stats.attn_output_reshape.meantime_stats.attn_output_reshape.mediantime_stats.attn_output_reshape.stdn_embdn_q_headn_kv_headblock_sizenum_tensor_parallel_workersmax_model_lenbatch_sizeprefill_chunk_sizekv_cache_sizeis_prefillattention_backendis_mixed_batchmodeseq_lenstotal_tokensmax_seq_lenmin_seq_lenavg_seq_lenequal_seq_lenseq_len_varianceseq_len_stdseq_len_cvis_chunked_prefill_samplechunk_start_tokenchunk_end_tokentotal_prefill_tokensprofiling_precisionmodel_archquant_signaturemeasurement_typeis_true_mixed_batchprefill_seq_lensprefill_kv_cache_sizesdecode_kv_cache_sizesnum_prefill_seqsnum_decode_seqsdecode_batch_sizetotal_batch_sizetotal_decode_tokensdecode_avg_kv_cache_sizebatch_composition_ratiobatch_specprojection_policy
20.00.00.00.00.00.014143999665975570.0288639999926090240.0197055999189615260.017711999826133250.0051572008498366810.0479680001735687260.070464000105857850.058102400973439220.058102400973439220.0074774634860415610.00.00.00.00.00.00.00.00.00.02048324161409601640.0TrueFLASH_ATTNFalsevllm020_batch_spec[64]64646464.0True0.00.00.0False0.064.064BF16genericnoneCUDA_EVENTFalseq64measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
30.00.00.00.00.00.0143999997526407240.049472000449895860.0204127999022603030.016351999714970590.0101074973797224170.0465600006282329560.083232000470161440.055875200033187870.055875200033187870.0111267587395034280.00.00.00.00.00.00.00.00.00.020483241614096011280.0TrueFLASH_ATTNFalsevllm020_batch_spec[128]128128128128.0True0.00.00.0False0.0128.0128BF16genericnoneCUDA_EVENTFalseq128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
40.00.00.00.00.00.014848000369966030.0222079996019601820.0170336001552641380.0153120001778006550.0028192359919702410.051040001213550570.076927997171878810.0563968002796173050.0563968002796173050.0074819821786375390.00.00.00.00.00.00.00.00.00.020483241614096012560.0TrueFLASH_ATTNFalsevllm020_batch_spec[256]256256256256.0True0.00.00.0False0.0256.0256BF16genericnoneCUDA_EVENTFalseq256measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
50.00.00.00.00.00.0150720002129673960.0222720000892877580.017068800237029790.016160000115633010.0024608895793191970.069311998784542080.08387199789285660.074320000410079950.074320000410079950.0047777664331758660.00.00.00.00.00.00.00.00.00.020483241614096015120.0TrueFLASH_ATTNFalsevllm020_batch_spec[512]512512512512.0True0.00.00.0False0.0512.0512BF16genericnoneCUDA_EVENTFalseq512measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
60.00.00.00.00.00.0185920000076293950.0285439994186162950.020959999784827230.0191839998587965970.0031981754960534940.121792003512382510.154080003499984740.13077120110392570.13077120110392570.008588077975382980.00.00.00.00.00.00.00.00.00.0204832416140960110240.0TrueFLASH_ATTNFalsevllm020_batch_spec[1024]1024102410241024.0True0.00.00.0False0.01024.01024BF16genericnoneCUDA_EVENTFalseq1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
70.00.00.00.00.00.0277759991586208340.033856000751256940.0301312003284692760.0296800006181001660.00211525581035752150.326784014701843260.34502398967742920.333964803814887970.333964803814887970.00458724249176063750.00.00.00.00.00.00.00.00.00.0204832416140960120480.0TrueFLASH_ATTNFalsevllm020_batch_spec[2048]2048204820482048.0True0.00.00.0False0.02048.02048BF16genericnoneCUDA_EVENTFalseq2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
80.00.00.00.00.00.044383998960256580.050847999751567840.0465407997369766260.045311998575925830.0022779058112372231.09596800804138181.11513602733612061.09997758865356451.09997758865356450.0056944032461204850.00.00.00.00.00.00.00.00.00.0204832416140960140960.0TrueFLASH_ATTNFalsevllm020_batch_spec[4096]4096409640964096.0True0.00.00.0False0.04096.04096BF16genericnoneCUDA_EVENTFalseq4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
90.00.00.00.00.00.0780159980058670.086911998689174650.081142399460077290.080191999673843380.002927957063344754.0704002380371094.1131520271301274.0870880126953124.0870880126953120.0136605670125095540.00.00.00.00.00.00.00.00.00.0204832416140960181920.0TrueFLASH_ATTNFalsevllm020_batch_spec[8192]8192819281928192.0True0.00.00.0False0.08192.08192BF16genericnoneCUDA_EVENTFalseq8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
100.00.00.00.00.00.0160639993846416470.051679998636245730.0221152000129222860.0175839997828006740.0103400948223408180.051968000829219820.090112000703811650.063283200934529330.063283200934529330.0125573412554674520.00.00.00.00.00.00.00.00.00.0204832416140960164448.0TrueFLASH_ATTNFalsevllm020_batch_spec[64]64646464.0True0.00.00.0True448.0512.064BF16genericnoneCUDA_EVENTFalseq64s512measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
110.00.00.00.00.00.015839999541640280.0266239996999502180.0189279997721314430.0173760000616312030.0035146503162606190.066816002130508420.079935997724533080.07252800017595290.07252800017595290.0043435025586137160.00.00.00.00.00.00.00.00.00.02048324161409601128896.0TrueFLASH_ATTNFalsevllm020_batch_spec[128]128128128128.0True0.00.00.0True896.01024.0128BF16genericnoneCUDA_EVENTFalseq128s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
120.00.00.00.00.00.016480000689625740.0308800004422664640.019452800229191780.0179679999127984050.0040962111830074850.13113600015640260.15468800067901610.139081600308418260.139081600308418260.0075119068743661780.00.00.00.00.00.00.00.00.00.020483241614096012561792.0TrueFLASH_ATTNFalsevllm020_batch_spec[256]256256256256.0True0.00.00.0True1792.02048.0256BF16genericnoneCUDA_EVENTFalseq256s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
130.00.00.00.00.00.0178559999912977220.035583999007940290.0208511998876929270.0185599997639656070.0052359115941307160.329504013061523440.3502719998359680.339129605889320340.339129605889320340.0060274009866636480.00.00.00.00.00.00.00.00.00.020483241614096015123584.0TrueFLASH_ATTNFalsevllm020_batch_spec[512]512512512512.0True0.00.00.0True3584.04096.0512BF16genericnoneCUDA_EVENTFalseq512s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
140.00.00.00.00.00.0193280000239610670.0406080000102520.0227904003113508220.0207040002569556240.0061130519657783371.14156794548034671.15187203884124761.1444832086563111.1444832086563110.00323323113743891270.00.00.00.00.00.00.00.00.00.0204832416140960110247168.0TrueFLASH_ATTNFalsevllm020_batch_spec[1024]1024102410241024.0True0.00.00.0True7168.08192.01024BF16genericnoneCUDA_EVENTFalseq1ks8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
150.00.00.00.00.00.0158079992979764940.0306880008429288860.0195967998355627070.017744000069797040.0043433847710334620.00.00.00.00.00.0490560010075569150.075808003544807430.059481600672006610.059481600672006610.0090315414714469550.00.00.00.00.020483241614096010128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1]1111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalseq1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
160.00.00.00.00.00.016031999140977860.024863999336957930.019238399714231490.0180799998342990880.00322825285372664240.00.00.00.00.00.051424000412225720.073536001145839690.0593280006200075160.0593280006200075160.00733078077351430840.00.00.00.00.020483241614096080128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
170.00.00.00.00.00.0165439993143081670.039776001125574110.0213791996240615850.0185119993984699250.0065935761762461710.00.00.00.00.00.04883199930191040.064351998269557950.054790399968624110.054790399968624110.0056725229984918640.00.00.00.00.0204832416140960160128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
180.00.00.00.00.00.016351999714970590.028448000550270080.0192672001197934160.0179520007222890850.00350686876669495770.00.00.00.00.00.0498559996485710140.077983997762203220.059868799895048150.059868799895048150.010439147548788280.00.00.00.00.0204832416140960320128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
190.00.00.00.00.00.0163839999586343770.0260799992829561230.019235199689865110.0177919995039701460.00321619743312845680.00.00.00.00.00.0581119991838932040.10457599908113480.067084800079464920.067084800079464920.0134790224626464940.00.00.00.00.0204832416140960640128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
200.00.00.00.00.00.0147200003266334530.040575999766588210.0191008001565933230.0154559998773038390.0075122812430115770.00.00.00.00.00.053631998598575590.077823996543884280.060905599594116220.060905599594116220.0075446201763480910.00.00.00.00.0204832416140960801024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
210.00.00.00.00.00.0144319999963045120.021919999271631240.0166847999207675460.015631999820470810.00242931186218112160.00.00.00.00.00.06291200220584870.078911997377872470.068915200978517530.068915200978517530.0054726956656954250.00.00.00.00.02048324161409601601024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
220.00.00.00.00.00.0145600000396370890.0389439985156059270.0183136000297963630.015616000629961490.0071272702601157690.00.00.00.00.00.086751997470855710.106623999774456020.093919999152421940.093919999152421940.0069880995890866350.00.00.00.00.02048324161409603201024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
230.00.00.00.00.00.0145279997959733010.0546879991888999940.0214399999007582680.015391999855637550.0120527648495977750.00.00.00.00.00.134880006313323970.15286399424076080.14313599914312360.14313599914312360.0054362714640335990.00.00.00.00.02048324161409606401024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
240.00.00.00.00.00.0143999997526407240.0419199988245964050.018998399749398230.0153439999558031560.0079898435266232870.00.00.00.00.00.061760000884532930.083743996918201450.067475199699401860.067475199699401860.00660677471287780.00.00.00.00.0204832416140960802048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
250.00.00.00.00.00.0162560008466243740.113536000251770020.0427616000175476060.0289600007236003880.0291043015380207620.00.00.00.00.00.097344003617763520.144224002957344060.113929601758718480.113929601758718480.0131985946004178670.00.00.00.00.02048324161409601602048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
260.00.00.00.00.00.0146880000829696660.0341439992189407350.0189184000715613350.016432000324130060.0055009439930806840.00.00.00.00.00.129183992743492130.150879994034767150.138079997897148140.138079997897148140.0076583305386775870.00.00.00.00.02048324161409603202048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
270.00.00.00.00.00.0160639993846416470.036416001617908480.01982080005109310.017807999625802040.00572641281698457650.00.00.00.00.00.220991998910903930.239040002226829530.22935039848089220.22935039848089220.0048613429070060280.00.00.00.00.02048324161409606402048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
280.00.00.00.00.00.0147519996389746670.0358400009572505950.0189088000915944580.0157920001074671750.00643748179247574750.00.00.00.00.00.101343996822834010.122015997767448430.108963198959827420.108963198959827420.0063363308091651790.00.00.00.00.0204832416140960804096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
290.00.00.00.00.00.0138560002669692040.034175999462604520.0178463999181985860.0148000000044703480.0064950075396352550.00.00.00.00.00.131264001131057740.155616000294685360.13892800062894820.13892800062894820.0083814728110750220.00.00.00.00.02048324161409601604096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
300.00.00.00.00.00.0144319999963045120.035199999809265140.0191680003888905040.016000000759959220.0059955224776546950.00.00.00.00.00.217280000448226930.23436799645423890.22314559817314150.22314559817314150.0047207306467391230.00.00.00.00.02048324161409603204096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
310.00.00.00.00.00.0140479998663067820.036704000085592270.0184415998868644250.0150239998474717140.0065965351627931270.00.00.00.00.00.393216013908386230.45247998833656310.40580800473690030.40580800473690030.015783497550889410.00.00.00.00.02048324161409606404096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
320.00.00.00.00.00.0143360001966357230.0221439991146326070.0169120000675320630.0151680000126361850.00281560892951363470.00.00.00.00.00.155872002243995670.30790400505065920.178380799293518050.178380799293518050.043558655752659270.00.00.00.00.0204832416140960808192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
330.00.00.00.00.00.0142719997093081470.021120000630617140.0155168000608682630.014880000147968530.0019408707315939040.00.00.00.00.00.215872004628181460.235615998506546020.222508800029754680.222508800029754680.0061819511706466660.00.00.00.00.02048324161409601608192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
340.00.00.00.00.00.0155520001426339150.0392640009522438050.0236096001230180280.021312000229954720.0072369796257115480.00.00.00.00.00.4087359905242920.4703359901905060.43368639945983880.43368639945983880.018446623831600740.00.00.00.00.02048324161409603208192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
350.00.00.00.00.00.0143999997526407240.0254079997539520260.0162271999754011640.0149600002914667130.00316173754417361850.00.00.00.00.00.74128001928329470.76278400421142580.74640000462532030.74640000462532030.0061121674488375470.00.00.00.00.02048324161409606408192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
360.00.00.00.00.00.014623999595642090.028127999976277350.0206527997739613040.0213599996641278270.0043087069576131020.0283834985654506270.0398596879706466440.0324922580745924260.0324922580745924260.004535972662085970.0293125011761036330.041164310587874630.033555741933275390.033555741933275390.0046844367577016480.00.00.00.00.0204832416140960900TrueFLASH_ATTNFalsetrue_mixed_fused_projected72False64BF16genericnoneCUDA_EVENTTrue[64][0][512, 512, 512, 512, 512, 512, 512, 512]18898512.00.1111111111111111q64_8q1s512fused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
370.00.00.00.00.00.0155520001426339150.0341439992189407350.0231711997650563720.0242559993639588360.0059086149180969710.033331592431144380.0389353417824780950.035804284022418540.035804284022418540.0020822702970440950.036332409033699370.042440659456344840.039027715070875620.039027715070875620.00226973542614901470.00.00.00.00.0204832416140960900TrueFLASH_ATTNFalsetrue_mixed_fused_projected136False128BF16genericnoneCUDA_EVENTTrue[128][0][1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]188981024.00.1111111111111111q128_8q1s1kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
380.00.00.00.00.00.0146559998393058780.026111999526619910.0196351999416947350.0181120000779628750.00382980507492577950.041895294179912160.0574842399205263840.047448857184210940.047448857184210940.0047798307484557430.0516727030372661840.070899754181951640.058522341344794120.058522341344794120.0058953345397863780.00.00.00.00.02048324161409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected144False128BF16genericnoneCUDA_EVENTTrue[128][0][1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]1161617161024.00.058823529411764705q128_16q1s1kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
390.00.00.00.00.00.0147200003266334530.0293119996786117550.0186623998917639260.016735999844968320.00440171121627253550.043229739593259010.050498270647053930.045354143434084480.045354143434084480.00229441560002406970.087330256177195380.102013728363985780.091621855742417750.091621855742417750.0046350476318460230.00.00.00.00.02048324161409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected272False256BF16genericnoneCUDA_EVENTTrue[256][0][2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048]1161617162048.00.058823529411764705q256_16q1s2kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
400.00.00.00.00.00.0145920002833008770.0263679996132850650.0163360000587999820.015151999890804290.0034153534559464020.060318427751607650.066183231883751980.062744155498172470.062744155498172470.0019856472896442550.147681576539926780.162040762490523240.153620645431850720.153620645431850720.0048615909452162470.00.00.00.00.02048324161409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected288False256BF16genericnoneCUDA_EVENTTrue[256][0][2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048]1323233322048.00.030303030303030304q256_32q1s2kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
410.00.00.00.00.00.0147519996389746670.024544000625610350.0171679999679327020.0161919994279742240.00286850164544984360.091287006883597120.096891507759963290.093603960514757760.093603960514757760.00134775427449167640.27408898875238920.29091648732098460.28104563569953660.28104563569953660.0040466285268085610.00.00.00.00.02048324161409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected544False512BF16genericnoneCUDA_EVENTTrue[512][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1323233324096.00.030303030303030304q512_32q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
420.00.00.00.00.00.018815999850630760.030975999310612680.0245984001085162160.0248480001464486120.00375394373915659750.10352492558669320.106031321474374120.103997972208409730.103997972208409730.00070251094290568140.56527507078934440.57896069388741170.56785803775176480.56785803775176480.0038359063841948970.00.00.00.00.02048324161409606500TrueFLASH_ATTNFalsetrue_mixed_fused_projected576False512BF16genericnoneCUDA_EVENTTrue[512][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1646465644096.00.015384615384615385q512_64q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
430.00.00.00.00.00.0186240002512931820.0434879995882511140.0244607999920845030.0196000002324581150.0089220816297813940.1961692049453070.20761400479457990.20084444292509310.20084444292509310.003943509532928051.11967080092932681.18499404173709741.14635555736100911.14635555736100910.0225082855305297830.00.00.00.00.02048324161409606500TrueFLASH_ATTNFalsetrue_mixed_fused_projected1088False1024BF16genericnoneCUDA_EVENTTrue[1024][0][8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192]1646465648192.00.015384615384615385q1k_64q1s8kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
440.00.00.00.00.00.0271359998732805250.036671999841928480.029459200054407120.0282560000196099280.00294466784081695530.377572796196646130.38981136244763720.380755204309421840.380755204309421840.00366002038310422540.25228319421320490.26046064930925970.254409587036174330.254409587036174330.00244551949302531260.00.00.00.00.02048324161409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected2080False2048BF16genericnoneCUDA_EVENTTrue[2048][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1323233324096.00.030303030303030304q2k_32q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
450.00.00.00.00.00.04351999983191490.0496959984302520750.045027199387550360.043951999396085740.0020357856601693081.10489843882454971.11890538864761081.11123712367541281.11123712367541280.00502206967246249850.13954950772391220.141318596070711930.140350088410310850.140350088410310850.00063429119448562930.00.00.00.00.02048324161409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected4112False4096BF16genericnoneCUDA_EVENTTrue[4096][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1161617164096.00.058823529411764705q4k_16q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
460.00.00.00.00.00.0147200003266334530.023647999390959740.018095999956130980.0163520006462931630.00354811270589590380.048223998397588730.085663996636867520.059612799808382990.059612799808382990.0114456654139688770.00.00.00.00.00.00.00.00.00.02048324162409601640.0TrueFLASH_ATTNFalsevllm020_batch_spec[64]64646464.0True0.00.00.0False0.064.064BF16genericnoneCUDA_EVENTFalseq64measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
470.00.00.00.00.00.0141759999096393590.0208640005439519880.016681600082665680.0156640000641345980.00257690638330975840.0496959984302520750.080576002597808840.058828799426555630.058828799426555630.0095151261085193310.00.00.00.00.00.00.00.00.00.020483241624096011280.0TrueFLASH_ATTNFalsevllm020_batch_spec[128]128128128128.0True0.00.00.0False0.0128.0128BF16genericnoneCUDA_EVENTFalseq128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
480.00.00.00.00.00.0143999997526407240.0287040006369352340.0174304000101983550.0150560000911355020.0043492946133355550.0498559996485710140.073664002120494840.054595200717449190.054595200717449190.00696109207579255740.00.00.00.00.00.00.00.00.00.020483241624096012560.0TrueFLASH_ATTNFalsevllm020_batch_spec[256]256256256256.0True0.00.00.0False0.0256.0256BF16genericnoneCUDA_EVENTFalseq256measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
490.00.00.00.00.00.0143360001966357230.031615998595952990.0177887998521327960.0157119999639689920.0050053857233186590.061023999005556110.081791996955871580.066505600139498730.066505600139498730.0069479955958011050.00.00.00.00.00.00.00.00.00.020483241624096015120.0TrueFLASH_ATTNFalsevllm020_batch_spec[512]512512512512.0True0.00.00.0False0.0512.0512BF16genericnoneCUDA_EVENTFalseq512measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
500.00.00.00.00.00.0146559998393058780.044160000979900360.0192000001668930050.0154880001209676270.0085612413233640380.080544002354145050.091967999935150150.086070399731397630.086070399731397630.0041453290354837540.00.00.00.00.00.00.00.00.00.0204832416240960110240.0TrueFLASH_ATTNFalsevllm020_batch_spec[1024]1024102410241024.0True0.00.00.0False0.01024.01024BF16genericnoneCUDA_EVENTFalseq1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
510.00.00.00.00.00.0178559999912977220.0233919993042945860.0196671998128294940.0184639999642968180.00225776696878325850.187296003103256230.205855995416641240.195465599000453930.195465599000453930.0063390688243036630.00.00.00.00.00.00.00.00.00.0204832416240960120480.0TrueFLASH_ATTNFalsevllm020_batch_spec[2048]2048204820482048.0True0.00.00.0False0.02048.02048BF16genericnoneCUDA_EVENTFalseq2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
520.00.00.00.00.00.0262400005012750630.034464001655578610.0282976005226373670.0270880004391074180.00251484415229223740.57542401552200320.58899199962615970.58001919388771050.58001919388771050.0038588692738295960.00.00.00.00.00.00.00.00.00.0204832416240960140960.0TrueFLASH_ATTNFalsevllm020_batch_spec[4096]4096409640964096.0True0.00.00.0False0.04096.04096BF16genericnoneCUDA_EVENTFalseq4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
530.00.00.00.00.00.041824001818895340.0472000017762184140.0430367998778820040.04233600012958050.00170734057720767282.0631999969482422.07872009277343752.0671519994735722.0671519994735720.0049596512711278350.00.00.00.00.00.00.00.00.00.0204832416240960181920.0TrueFLASH_ATTNFalsevllm020_batch_spec[8192]8192819281928192.0True0.00.00.0False0.08192.08192BF16genericnoneCUDA_EVENTFalseq8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
540.00.00.00.00.00.0143999997526407240.056480001658201220.022703999932855370.0179359996691346170.0123779151507276890.0495360009372234340.071967996656894680.0560159996151924150.0560159996151924150.00704766375527428840.00.00.00.00.00.00.00.00.00.0204832416240960164448.0TrueFLASH_ATTNFalsevllm020_batch_spec[64]64646464.0True0.00.00.0True448.0512.064BF16genericnoneCUDA_EVENTFalseq64s512measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
550.00.00.00.00.00.014080000109970570.0213440004736185070.016115200053900480.0149280000478029250.00254998849939617020.070720002055168150.26428800821304320.15882560089230540.15882560089230540.0532203470861023760.00.00.00.00.00.00.00.00.00.02048324162409601128896.0TrueFLASH_ATTNFalsevllm020_batch_spec[128]128128128128.0True0.00.00.0True896.01024.0128BF16genericnoneCUDA_EVENTFalseq128s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
560.00.00.00.00.00.013600000180304050.030144000425934790.0169759999029338370.0150560000911355020.0047151596973641110.083935998380184170.110367998480796810.091600000113248810.091600000113248810.0079116694344727920.00.00.00.00.00.00.00.00.00.020483241624096012561792.0TrueFLASH_ATTNFalsevllm020_batch_spec[256]256256256256.0True0.00.00.0True1792.02048.0256BF16genericnoneCUDA_EVENTFalseq256s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
570.00.00.00.00.00.0144959995523095130.0207679998129606250.0162335999310016630.0153920003212988380.0022026159079954270.19983999431133270.220704004168510440.208553601801395430.208553601801395430.006893072002306670.00.00.00.00.00.00.00.00.00.020483241624096015123584.0TrueFLASH_ATTNFalsevllm020_batch_spec[512]512512512512.0True0.00.00.0True3584.04096.0512BF16genericnoneCUDA_EVENTFalseq512s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
580.00.00.00.00.00.014848000369966030.0334400013089180.0186848001554608330.0160480001941323280.005539910056391740.60431998968124390.6358079910278320.61261439919471730.61261439919471730.0087459339534080960.00.00.00.00.00.00.00.00.00.0204832416240960110247168.0TrueFLASH_ATTNFalsevllm020_batch_spec[1024]1024102410241024.0True0.00.00.0True7168.08192.01024BF16genericnoneCUDA_EVENTFalseq1ks8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
590.00.00.00.00.00.0138240000233054160.033599998801946640.0178111997433006780.0149440001696348190.0059261404835654720.00.00.00.00.00.0454719997942447660.075392000377178190.0547584000974893560.0547584000974893560.0102535485061015490.00.00.00.00.020483241624096010128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1]1111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalseq1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
600.00.00.00.00.00.0140479998663067820.024095999076962470.016412799991667270.014735999982804060.0033471586387139870.00.00.00.00.00.04617600142955780.075295999646186830.054608000442385680.054608000442385680.0097489377983401350.00.00.00.00.020483241624096080128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
610.00.00.00.00.00.0142719997093081470.0284160003066062930.0186111998744308940.015984000172466040.0052098750938636470.00.00.00.00.00.0481280013918876650.078975997865200040.0613536003977060360.0613536003977060360.0104881536551578450.00.00.00.00.0204832416240960160128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
620.00.00.00.00.00.0141120003536343570.0257280003279447560.0160928001627326030.015120000112801790.0033008881230526050.00.00.00.00.00.048640001565217970.076480001211166380.058105600625276560.058105600625276560.0094735442184044080.00.00.00.00.0204832416240960320128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
630.00.00.00.00.00.0146559998393058780.035519998520612720.0185887998901307570.0160640003159642220.0060710574379924470.00.00.00.00.00.048223998397588730.078624002635478970.056604800373315820.056604800373315820.0095653942137304010.00.00.00.00.0204832416240960640128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
640.00.00.00.00.00.0143039999529719350.0288319997489452360.016995199955999850.0150240003131330010.0042850007448681880.00.00.00.00.00.048544000834226610.067199997603893280.057782400026917460.057782400026917460.0068521255546798050.00.00.00.00.0204832416240960801024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
650.00.00.00.00.00.01446400023996830.0249279998242855070.01611839998513460.015215999912470580.00297748085246739070.00.00.00.00.00.053791999816894530.089663997292518620.062700799852609640.062700799852609640.0099835912770926960.00.00.00.00.02048324162409601601024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
660.00.00.00.00.00.0144319999963045120.030015999451279640.0176480000838637360.015472000464797020.0045851263094848480.00.00.00.00.00.0616640001535415650.076927997171878810.067043200135231030.067043200135231030.0055001917984906290.00.00.00.00.02048324162409603201024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
670.00.00.00.00.00.0141759999096393590.0263679996132850650.0169471997767686840.0150399999693036080.00379515661035502050.00.00.00.00.00.087999999523162840.1114559993147850.09640319943428040.09640319943428040.0073975416155580880.00.00.00.00.02048324162409606401024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
680.00.00.00.00.00.013695999979972840.0212479997426271440.0153599999845027930.0146400001831352710.00209427709508143170.00.00.00.00.00.0517119988799095150.070656001567840580.0580544002354145060.0580544002354145060.0066330348159100250.00.00.00.00.0204832416240960802048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
690.00.00.00.00.00.0142080001533031460.03078399971127510.0171487999148666850.0150399999693036080.0048823746863122840.00.00.00.00.00.0619199983775615760.078432001173496250.066284799203276640.066284799203276640.0049628526898011920.00.00.00.00.02048324162409601602048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
700.00.00.00.00.00.0138879995793104170.029696000739932060.0175007999874651420.0152320000343024730.004672733197053140.00.00.00.00.00.088192000985145570.110976003110408780.094934401661157620.094934401661157620.0075092350429855770.00.00.00.00.02048324162409603202048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
710.00.00.00.00.00.0143999997526407240.0221120007336139680.017516800016164780.0160479992628097530.00303066217929157850.00.00.00.00.00.130656003952026370.151104003190994260.138572798669338220.138572798669338220.007501378412497710.00.00.00.00.02048324162409606402048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
720.00.00.00.00.00.0139199998229742050.040128000080585480.018924799934029580.015504000242799520.0074595452678169610.00.00.00.00.00.062784001231193540.082592003047466280.07045120075345040.07045120075345040.0059790559843827440.00.00.00.00.0204832416240960804096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
730.00.00.00.00.00.0143360001966357230.022368000820279120.0163328002206981180.0149280000478029250.0027840443631867310.00.00.00.00.00.10031999647617340.12793600559234620.109212798625230780.109212798625230780.008627696170467160.00.00.00.00.02048324162409601604096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
740.00.00.00.00.00.0139199998229742050.02256000041961670.0161280000582337370.0148000000044703480.0029589533325477080.00.00.00.00.00.131167992949485780.148128002882003780.137839999794960030.137839999794960030.0056963843611480530.00.00.00.00.02048324162409603204096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
750.00.00.00.00.00.0142080001533031460.0299839992076158520.0174688001163303860.015088000334799290.00473795509624830650.00.00.00.00.00.2176319956779480.24473600089550020.227158398926258080.227158398926258080.0088318282218471380.00.00.00.00.02048324162409606404096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
760.00.00.00.00.00.0140479998663067820.023040000349283220.0165120000950992120.0145120001398026940.00350267748286242540.00.00.00.00.00.110207997262477870.123071998357772830.116009599715471260.116009599715471260.0046676372669509020.00.00.00.00.0204832416240960808192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
770.00.00.00.00.00.0141120003536343570.0420800000429153440.0175103999674320230.0146079999394714830.008231178835301670.00.00.00.00.00.157023996114730830.175871998071670530.163993598520755760.163993598520755760.0065882990746763930.00.00.00.00.02048324162409601608192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
780.00.00.00.00.00.0139199998229742050.02083200030028820.0152991999872028830.014623999595642090.0019167680865053860.00.00.00.00.00.217792004346847530.24153600633144380.22607680112123490.22607680112123490.0072513520806852360.00.00.00.00.02048324162409603208192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
790.00.00.00.00.00.0140159996226429940.0210880003869533540.0156192001886665820.014735999982804060.0020132274523021410.00.00.00.00.00.39599999785423280.41526401042938230.403324797749519240.403324797749519240.0069424014319140520.00.00.00.00.02048324162409606408192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
800.00.00.00.00.00.0141759999096393590.033792000263929370.0205952000804245470.0196000002324581150.0065487542488813440.0261926237396945120.0400068825071684260.0291551801784920360.0291551801784920360.004088287484382410.0245913755245457560.0375611195379861460.0273728203550887460.0273728203550887460.00383835593485759440.00.00.00.00.0204832416240960900TrueFLASH_ATTNFalsetrue_mixed_fused_projected72False64BF16genericnoneCUDA_EVENTTrue[64][0][512, 512, 512, 512, 512, 512, 512, 512]18898512.00.1111111111111111q64_8q1s512fused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
810.00.00.00.00.00.0138560002669692040.0208960007876157760.0153184000402688990.0144319999963045120.00196403129609269660.0293490182086938620.062362629413566790.037143135920149630.037143135920149630.0096183137918725690.0288269814625269140.061253373086490410.0364824636727744960.0364824636727744960.0094472309570118630.00.00.00.00.0204832416240960900TrueFLASH_ATTNFalsetrue_mixed_fused_projected136False128BF16genericnoneCUDA_EVENTTrue[128][0][1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]188981024.00.1111111111111111q128_8q1s1kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
820.00.00.00.00.00.0138560002669692040.0321279987692832950.0172863999381661430.0144799998961389060.0055364954466361930.032730858745583540.043945784919370820.035638374906530340.035638374906530340.0034296017302565670.034885142025938990.046838213450799780.0379840254070854740.0379840254070854740.00365533163619026840.00.00.00.00.02048324162409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected144False128BF16genericnoneCUDA_EVENTTrue[128][0][1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]1161617161024.00.058823529411764705q128_16q1s1kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
830.00.00.00.00.00.0138560002669692040.0215039998292922970.0150784000754356390.014256000053137540.0022384359754788490.0421008135499741850.0517841446901477560.0453786928902896250.0453786928902896250.0031841099406505550.051115188437483090.062871856639274250.055094907735702190.055094907735702190.00386587255442991320.00.00.00.00.02048324162409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected272False256BF16genericnoneCUDA_EVENTTrue[256][0][2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048]1161617162048.00.058823529411764705q256_16q1s2kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
840.00.00.00.00.00.0143999997526407240.026879999786615370.0176671998575329770.015663999598473310.0038648118463871730.0484753237738213060.055999569787237330.051234992529683450.051234992529683450.0024273798204239410.084292681358853870.097376428352144080.08909140906161750.08909140906161750.00422091773320770050.00.00.00.00.02048324162409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected288False256BF16genericnoneCUDA_EVENTTrue[256][0][2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048]1323233322048.00.030303030303030304q256_32q1s2kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
850.00.00.00.00.00.014623999595642090.033599998801946640.0198047997429966930.0160320000723004340.0067505199081454740.071215084745799850.072923083671363960.071944108452701830.071944108452701830.00055000659159438440.147600912497127670.151140923730102380.149111892435645770.149111892435645770.00113994773843970050.00.00.00.00.02048324162409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected544False512BF16genericnoneCUDA_EVENTTrue[512][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1323233324096.00.030303030303030304q512_32q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
860.00.00.00.00.00.0143999997526407240.0218560006469488140.0168352002277970330.0152639998123049740.002838544489750650.081463382721429350.085608657143840960.08349326255207340.08349326255207340.00122596359053478740.27824862194013070.292407337881793850.28518193659896580.28518193659896580.0041874357314816650.00.00.00.00.02048324162409606500TrueFLASH_ATTNFalsetrue_mixed_fused_projected576False512BF16genericnoneCUDA_EVENTTrue[512][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1646465644096.00.015384615384615385q512_64q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
870.00.00.00.00.00.0143999997526407240.044959999620914460.021436799876391890.0158560001291334630.0097099878726833420.11944592087021780.123025247551814630.120538857186960960.120538857186960960.0010051181193413390.55972206504591990.57649474432288020.56484355071671020.56484355071671020.0047099707154007880.00.00.00.00.02048324162409606500TrueFLASH_ATTNFalsetrue_mixed_fused_projected1088False1024BF16genericnoneCUDA_EVENTTrue[1024][0][8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192]1646465648192.00.015384615384615385q1k_64q1s8kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
880.00.00.00.00.00.0181120000779628750.0264960005879402160.0193184001371264450.0184320006519556050.00242496051123599050.197702535746104020.2228117978683480.20548115206194120.20548115206194120.0080009442548680430.139417460801594920.1571242117771140.144902847881791970.144902847881791970.0056421700805159920.00.00.00.00.02048324162409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected2080False2048BF16genericnoneCUDA_EVENTTrue[2048][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1323233324096.00.030303030303030304q2k_32q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
890.00.00.00.00.00.025952000170946120.031552001833915710.027273599989712240.0265759993344545360.00162609487648431660.60146848811166590.60954729882120.60467577009463760.60467577009463760.00235374355631773030.113251522645780450.114772695628415450.113855427214856540.113855427214856540.00044319036980236280.00.00.00.00.02048324162409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected4112False4096BF16genericnoneCUDA_EVENTTrue[4096][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1161617164096.00.058823529411764705q4k_16q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
900.00.00.00.00.00.0145600000396370890.0227840002626180650.0164352000690996640.0155199998989701270.00236883014214695230.048799999058246610.091392002999782560.060540799424052240.060540799424052240.0121526087024487750.00.00.00.00.00.00.00.00.00.02048324164409601640.0TrueFLASH_ATTNFalsevllm020_batch_spec[64]64646464.0True0.00.00.0False0.064.064BF16genericnoneCUDA_EVENTFalseq64measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
910.00.00.00.00.00.0148160001263022420.028448000550270080.0170080001465976250.0156960003077983860.0039414662946626790.0478079989552497860.073568001389503480.056265600025653840.056265600025653840.008421791254122360.00.00.00.00.00.00.00.00.00.020483241644096011280.0TrueFLASH_ATTNFalsevllm020_batch_spec[128]128128128128.0True0.00.00.0False0.0128.0128BF16genericnoneCUDA_EVENTFalseq128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
920.00.00.00.00.00.0146880000829696660.0534080006182193760.0193536002188920970.015328000299632550.011383024008416260.0484480001032352450.07856000214815140.05562560036778450.05562560036778450.0093482745026169080.00.00.00.00.00.00.00.00.00.020483241644096012560.0TrueFLASH_ATTNFalsevllm020_batch_spec[256]256256256256.0True0.00.00.0False0.0256.0256BF16genericnoneCUDA_EVENTFalseq256measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
930.00.00.00.00.00.0150399999693036080.031392000615596770.018236799910664560.0161599996499717240.0048600712383025120.0554240010678768160.085055999457836150.06403839923441410.06403839923441410.009214486361786230.00.00.00.00.00.00.00.00.00.020483241644096015120.0TrueFLASH_ATTNFalsevllm020_batch_spec[512]512512512512.0True0.00.00.0False0.0512.0512BF16genericnoneCUDA_EVENTFalseq512measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
940.00.00.00.00.00.0147519996389746670.0265280008316040040.0173247999511659150.015360000077635050.00360048226864283050.076608002185821530.09420800209045410.082092800736427320.082092800736427320.005158955876697520.00.00.00.00.00.00.00.00.00.0204832416440960110240.0TrueFLASH_ATTNFalsevllm020_batch_spec[1024]1024102410241024.0True0.00.00.0False0.01024.01024BF16genericnoneCUDA_EVENTFalseq1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
950.00.00.00.00.00.0148799996823072430.0212479997426271440.0164031998254358760.015328000299632550.0019958970226479340.113952003419399260.151135995984077450.124310399591922760.124310399591922760.0111644311236837320.00.00.00.00.00.00.00.00.00.0204832416440960120480.0TrueFLASH_ATTNFalsevllm020_batch_spec[2048]2048204820482048.0True0.00.00.0False0.02048.02048BF16genericnoneCUDA_EVENTFalseq2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
960.00.00.00.00.00.0171840004622936250.0286720003932714460.0198656000196933760.0178239997476339340.00357984913159299770.31718400120735170.33411198854446410.32616959512233730.32616959512233730.0051110464528789180.00.00.00.00.00.00.00.00.00.0204832416440960140960.0TrueFLASH_ATTNFalsevllm020_batch_spec[4096]4096409640964096.0True0.00.00.0False0.04096.04096BF16genericnoneCUDA_EVENTFalseq4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
970.00.00.00.00.00.0246399994939565660.0304000005125999450.026566400006413460.025568000040948390.00201898466032373031.06486403942108151.08284795284271241.07153278589248661.07153278589248660.0055586392630498510.00.00.00.00.00.00.00.00.00.0204832416440960181920.0TrueFLASH_ATTNFalsevllm020_batch_spec[8192]8192819281928192.0True0.00.00.0False0.08192.08192BF16genericnoneCUDA_EVENTFalseq8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
980.00.00.00.00.00.0149760004132986070.0216959994286298750.0173055998980998980.015935999341309070.00257180465172680540.049568001180887220.079328000545501710.062284800410270690.062284800410270690.010271603497578270.00.00.00.00.00.00.00.00.00.0204832416440960164448.0TrueFLASH_ATTNFalsevllm020_batch_spec[64]64646464.0True0.00.00.0True448.0512.064BF16genericnoneCUDA_EVENTFalseq64s512measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
990.00.00.00.00.00.015391999855637550.035808000713586810.0196864003315567960.0171360000967979430.0059183878113040030.051584001630544660.107136003673076630.063641600683331480.063641600683331480.0158329578096967660.00.00.00.00.00.00.00.00.00.02048324164409601128896.0TrueFLASH_ATTNFalsevllm020_batch_spec[128]128128128128.0True0.00.00.0True896.01024.0128BF16genericnoneCUDA_EVENTFalseq128s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1000.00.00.00.00.00.0151680000126361850.022239999845623970.0166720000095665450.0158879999071359630.00209349459460345630.065215997397899630.089023999869823460.075209599733352660.075209599733352660.0079138406841029290.00.00.00.00.00.00.00.00.00.020483241644096012561792.0TrueFLASH_ATTNFalsevllm020_batch_spec[256]256256256256.0True0.00.00.0True1792.02048.0256BF16genericnoneCUDA_EVENTFalseq256s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1010.00.00.00.00.00.0152319995686411860.040288001298904420.0199712000787258160.016463999636471270.0073513058865772860.144672006368637080.168448001146316530.153635199368000050.153635199368000050.0081744358999562230.00.00.00.00.00.00.00.00.00.020483241644096015123584.0TrueFLASH_ATTNFalsevllm020_batch_spec[512]512512512512.0True0.00.00.0True3584.04096.0512BF16genericnoneCUDA_EVENTFalseq512s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1020.00.00.00.00.00.0155199998989701270.023040000349283220.017756799887865780.0167840002104640.00250772437120825840.337408006191253660.353439986705780030.34451200067996980.34451200067996980.004456484635903580.00.00.00.00.00.00.00.00.00.0204832416440960110247168.0TrueFLASH_ATTNFalsevllm020_batch_spec[1024]1024102410241024.0True0.00.00.0True7168.08192.01024BF16genericnoneCUDA_EVENTFalseq1ks8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1030.00.00.00.00.00.0156479999423027040.023935999721288680.018095999956130980.016544000245630740.0029983931024662540.00.00.00.00.00.05043200030922890.08438400179147720.0600832004100084260.0600832004100084260.009869593185722960.00.00.00.00.020483241644096010128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1]1111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalseq1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1040.00.00.00.00.00.016031999140977860.0300479996949434280.0195103999227285370.0176800005137920380.0040651852659633320.00.00.00.00.00.050048001110553740.070367999374866490.0593152005225420.0593152005225420.0065377688213296470.00.00.00.00.020483241644096080128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1050.00.00.00.00.00.014848000369966030.023935999721288680.0180351997725665580.0165120000019669530.00306676891167777240.00.00.00.00.00.051008000969886780.067359998822212220.0583872001618146940.0583872001618146940.0058327877392413810.00.00.00.00.0204832416440960160128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1060.00.00.00.00.00.015007999725639820.0262080002576112750.0175007999874651420.016463999636471270.00306661033061657140.00.00.00.00.00.050239998847246170.072544001042842860.0572544004768133150.0572544004768133150.0068908536530681510.00.00.00.00.0204832416440960320128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1070.00.00.00.00.00.0146880000829696660.0273279994726181030.01797760007902980.016480000689625740.0037834174475313920.00.00.00.00.00.048191998153924940.071007996797561650.056211199238896380.056211199238896380.0078243914841247250.00.00.00.00.0204832416440960640128.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s128measured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1080.00.00.00.00.00.0156479999423027040.0364480018615722660.0191360002383589750.0167040005326271060.0060765300956248030.00.00.00.00.00.0474880002439022060.084416002035140990.05883839987218380.05883839987218380.0112759407263325220.00.00.00.00.0204832416440960801024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1090.00.00.00.00.00.0151359997689723970.0339520014822483060.021193600166589020.0180000001564621930.0071366466143053040.00.00.00.00.00.049311999231576920.069920003414154050.057558400183916090.057558400183916090.0072975876943152260.00.00.00.00.02048324164409601601024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1100.00.00.00.00.00.0148799996823072430.0342720001935958860.02044159993529320.0173119995743036270.006954047018030130.00.00.00.00.00.052159998565912250.07350400090217590.061174400150775910.061174400150775910.0073811184368949220.00.00.00.00.02048324164409603201024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1110.00.00.00.00.00.0153599996119737630.037439998239278790.020124799665063620.017759999260306360.0061810455090790570.00.00.00.00.00.063551999628543850.085087999701499940.070192000269889840.070192000269889840.00622465804421178450.00.00.00.00.02048324164409606401024.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s1kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1120.00.00.00.00.00.0155520001426339150.0326079986989498140.019599999953061340.0174879999831318860.0048824750684607490.00.00.00.00.00.049183998256921770.08652800321578980.059737600386142740.059737600386142740.0115468479421139740.00.00.00.00.0204832416440960802048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1130.00.00.00.00.00.0151680000126361850.026752000674605370.0174304000101983550.0165600003674626350.0032431350942398190.00.00.00.00.00.0526719987392425540.083679996430873870.062387199327349650.062387199327349650.0112078247806301040.00.00.00.00.02048324164409601602048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1140.00.00.00.00.00.015007999725639820.036127999424934390.0193279998376965530.016880000010132790.0060582514251530850.00.00.00.00.00.063648000359535220.078783996403217320.069356798380613320.069356798380613320.0045155883286776430.00.00.00.00.02048324164409603202048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1150.00.00.00.00.00.0151040004566311840.037119999527931210.0196703999303281320.0171520002186298370.0061659946465222640.00.00.00.00.00.090719997882843020.115071997046470640.102524800598621360.102524800598621360.0087825440515356570.00.00.00.00.02048324164409606402048.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s2kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1160.00.00.00.00.00.0150399999693036080.0265280008316040040.0183968001045286650.016015999950468540.0042797880489862830.00.00.00.00.00.053311999887228010.079775996506214140.060764800012111670.060764800012111670.0087615791220696060.00.00.00.00.0204832416440960804096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1170.00.00.00.00.00.0153599996119737630.026432000100612640.017264000046998260.0158559996634721760.0032100084172420290.00.00.00.00.00.0624319985508918760.082464002072811130.069708799570798880.069708799570798880.0070163416760682330.00.00.00.00.02048324164409601604096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1180.00.00.00.00.00.0149119999259710310.022239999845623970.01660159993916750.0158879999071359630.002049139691293540.00.00.00.00.00.101279996335506440.11411199718713760.106211200356483470.106211200356483470.0040293183117546050.00.00.00.00.02048324164409603204096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1190.00.00.00.00.00.0153599996119737630.033631999045610430.0193055999465286750.0166719993576407430.00554455823079812340.00.00.00.00.00.13190400600433350.158399999141693120.140406401455402340.140406401455402340.0079980410735969420.00.00.00.00.02048324164409606404096.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s4kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1200.00.00.00.00.00.0143680004402995110.044479999691247940.0189088000915944580.01572799962013960.00868149374617210.00.00.00.00.00.064287997782230380.089823998510837550.073497600853443150.073497600853443150.0086052845121972580.00.00.00.00.0204832416440960808192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1]8111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse8q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1210.00.00.00.00.00.0151040004566311840.0286720003932714460.018905600067228080.0160800004377961160.0047975846278150210.00.00.00.00.00.111231997609138490.141151994466781620.119423998892307290.119423998892307290.0085133947067910270.00.00.00.00.02048324164409601608192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]16111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse16q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1220.00.00.00.00.00.0142719997093081470.0249279998242855070.0169152000918984420.0152160003781318660.0033741474994426350.00.00.00.00.00.158879995346069340.181119993329048160.169343997538089760.169343997538089760.0074151812607611230.00.00.00.00.02048324164409603208192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]32111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse32q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1230.00.00.00.00.00.0143360001966357230.0269440002739429470.0177376002073287960.0151999997906386850.004155987143752550.00.00.00.00.00.21923199295997620.234720006585121150.228099200129508930.228099200129508930.0047308413763273350.00.00.00.00.02048324164409606408192.0FalseFLASH_ATTNFalsevllm020_batch_spec[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]64111.0True0.00.00.0False000BF16genericnoneCUDA_EVENTFalse64q1s8kmeasured_FA3_core_plus_measured_KV;reshape_assumed_zero;mean_as_median
1240.00.00.00.00.00.0146880000829696660.050528001040220260.0203200003132224080.0155200003646314140.0106045221779246780.0242706298824989580.033402476259540760.0284463441041286240.0284463441041286240.00323303690997938340.0236973702910697680.032613527229953560.0277744563254539720.0277744563254539720.00315667426809233860.00.00.00.00.0204832416440960900TrueFLASH_ATTNFalsetrue_mixed_fused_projected72False64BF16genericnoneCUDA_EVENTTrue[64][0][512, 512, 512, 512, 512, 512, 512, 512]18898512.00.1111111111111111q64_8q1s512fused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1250.00.00.00.00.00.0153599996119737630.0336639992892742160.0199871998280286780.0182400001212954520.0055229291392537320.02591940556609470.037197551839907190.0299160296878997030.0299160296878997030.0039666665905543180.0271045958534495180.038898446447293740.031283969540220150.031283969540220150.0041480463179678810.00.00.00.00.0204832416440960900TrueFLASH_ATTNFalsetrue_mixed_fused_projected136False128BF16genericnoneCUDA_EVENTTrue[128][0][1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]188981024.00.1111111111111111q128_8q1s1kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1260.00.00.00.00.00.0145920002833008770.032992001622915270.0198400001041591150.0162079995498061180.0065463430087268140.025878695957699260.037631672654314820.0301749390581628020.0301749390581628020.00373035817822773640.0264733040701957130.038496325876549890.030868260838649640.030868260838649640.0038160696545292220.00.00.00.00.02048324164409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected144False128BF16genericnoneCUDA_EVENTTrue[128][0][1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024]1161617161024.00.058823529411764705q128_16q1s1kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1270.00.00.00.00.00.0138240000233054160.0210239998996257780.0163040001876652230.0155840003862977030.00232361658065454760.0318106215259669960.041569499368781060.03467493500328070.03467493500328070.0031302091811439850.0356773782709003740.046622501616364510.0388898648717402940.0388898648717402940.00351070339608287270.00.00.00.00.02048324164409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected272False256BF16genericnoneCUDA_EVENTTrue[256][0][2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048]1161617162048.00.058823529411764705q256_16q1s2kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1280.00.00.00.00.00.0145279997959733010.021759999915957450.015696000121533870.0150080001913011070.00208826661680368630.0382117121070628530.072122292565200570.043645096733340970.043645096733340970.0096719530740250850.04764428599178740.089925704551841980.054418903426161070.054418903426161070.012059477917840380.00.00.00.00.02048324164409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected288False256BF16genericnoneCUDA_EVENTTrue[256][0][2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048, 2048]1323233322048.00.030303030303030304q256_32q1s2kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1290.00.00.00.00.00.0143680004402995110.0338240005075931550.0176288002170622360.0148640000261366370.0057914954038577380.052142612500300330.062662615087066690.056777912955276610.056777912955276610.00302982230796515140.086481385068783820.103929386827911320.094169285316474780.094169285316474780.0050251266122044890.00.00.00.00.02048324164409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected544False512BF16genericnoneCUDA_EVENTTrue[512][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1323233324096.00.030303030303030304q512_32q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1300.00.00.00.00.00.0145279997959733010.0311999991536140440.0179007999598979960.0153439999558031560.0049744756613162490.064119588914211110.074052761232639560.067930479182113770.067930479182113770.00339186585567000060.140584412631694970.162363235914920580.14893992112744890.14893992112744890.007436783003753530.00.00.00.00.02048324164409606500TrueFLASH_ATTNFalsetrue_mixed_fused_projected576False512BF16genericnoneCUDA_EVENTTrue[512][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1646465644096.00.015384615384615385q512_64q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1310.00.00.00.00.00.0147200003266334530.038559999316930770.0184959999285638330.0156479999423027040.0069185300567616270.098729855494012770.106834545830437530.101853168398845520.101853168398845520.00208022540370420740.27432613901663790.29684545089845960.28300442954827520.28300442954827520.0057800165960650920.00.00.00.00.02048324164409606500TrueFLASH_ATTNFalsetrue_mixed_fused_projected1088False1024BF16genericnoneCUDA_EVENTTrue[1024][0][8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192, 8192]1646465648192.00.015384615384615385q1k_64q1s8kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1320.00.00.00.00.00.0149119999259710310.0304000005125999450.019779200013726950.017264000140130520.0053508834642061690.139188724722333650.162795228553352070.15019128648391730.15019128648391730.0089921855290115130.118923288617662660.139092760490837350.12832391234287250.12832391234287250.0076829518849569390.00.00.00.00.02048324164409603300TrueFLASH_ATTNFalsetrue_mixed_fused_projected2080False2048BF16genericnoneCUDA_EVENTTrue[2048][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1323233324096.00.030303030303030304q2k_32q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio
1330.00.00.00.00.00.017279999330639840.022239999845623970.018195199780166150.0176800005137920380.00143972982706208730.37280441816254430.382955043537016760.37612644837873330.37612644837873330.00336602723324909770.079675570174018830.081844956653718050.080385552778073650.080385552778073650.00071938562410884550.00.00.00.00.02048324164409601700TrueFLASH_ATTNFalsetrue_mixed_fused_projected4112False4096BF16genericnoneCUDA_EVENTTrue[4096][0][4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096, 4096]1161617164096.00.058823529411764705q4k_16q1s4kfused_total_conserving_projection_by_same_tp_pure_prefill_decode_reference_ratio