Add scripts for benchmark sparsity 0.9 cases with mattn256 & full256

This commit is contained in:
Qianfeng Zhang
2026-01-30 09:58:12 +00:00
parent 749e83f2fd
commit c360e0cbc4

View File

@@ -0,0 +1,71 @@
#!/bin/bash
set +x
BUILD=build
USE_SOFTMAX=0
if [ $# -ge 1 ]; then
USE_SOFTMAX=$1
fi
if [ $USE_SOFTMAX -eq 1 ]; then
EXE="$BUILD/bin/tile_example_hstu_attention -softmax=1"
else
EXE="$BUILD/bin/tile_example_hstu_attention"
fi
dtype="bf16"
set -x
seqlens8="16157,16136,14491,14708,14989,14845,15311,15958"
## sparsity 0.9 lengths from seqlen 16384
$EXE -v=0 -prec=$dtype -b=8 -jagged=1 -nhead=4 -hdim_qk=128 -hdim_v=128 -seqlens=$seqlens8 -causal=1 -local_len=256 -context_len=0 -minfull_len=256 -targets=200 -max_seqlen=16184 -perf=1 -alpha=0.0078125
echo -e ""
seqlens16="16157,16136,14491,14708,14989,14845,15311,15958,14005,12961,12953,13808,14951,15988,14919,14383"
## sparsity 0.9 lengths from seqlen 16384
$EXE -v=0 -prec=$dtype -b=16 -jagged=1 -nhead=4 -hdim_qk=128 -hdim_v=128 -seqlens=$seqlens16 -causal=1 -local_len=256 -context_len=0 -minfull_len=256 -targets=200 -max_seqlen=16184 -perf=1 -alpha=0.0078125
echo -e ""
seqlens32="16157,16136,14491,14708,14989,14845,15311,15958,14005,12961,12953,13808,14951,15988,14919,14383,15252,13338,15367,14524,15074,14557,13742,15494,14717,15922,13569,13225,15853,16117,13634,14023"
## sparsity 0.9 lengths from seqlen 16384
$EXE -v=0 -prec=$dtype -b=32 -jagged=1 -nhead=4 -hdim_qk=128 -hdim_v=128 -seqlens=$seqlens32 -causal=1 -local_len=256 -context_len=0 -minfull_len=256 -targets=200 -max_seqlen=16184 -perf=1 -alpha=0.0078125
echo -e ""
seqlens64="16157,16136,14491,14708,14989,14845,15311,15958,14005,12961,12953,13808,14951,15988,14919,14383,15252,13338,15367,14524,15074,14557,13742,15494,14717,15922,13569,13225,15853,16117,13634,14023,14817,14848,12948,13039,14400,13695,15897,13616,13083,15300,14298,16142,13397,16107,15078,13919,15399,14857,14394,16008,15287,13996,13685,15305,13769,15282,15837,15324,13282,14877,15657,13165"
## sparsity 0.9 lengths from seqlen 16384
$EXE -v=0 -prec=$dtype -b=64 -jagged=1 -nhead=4 -hdim_qk=128 -hdim_v=128 -seqlens=$seqlens64 -causal=1 -local_len=256 -context_len=0 -minfull_len=256 -targets=200 -max_seqlen=16184 -perf=1 -alpha=0.0078125
echo -e ""
seqlens128="16157,16136,14491,14708,14989,14845,15311,15958,14005,12961,12953,13808,14951,15988,14919,14383,15252,13338,15367,14524,15074,14557,13742,15494,14717,15922,13569,13225,15853,16117,13634,14023,14817,14848,12948,13039,14400,13695,15897,13616,13083,15300,14298,16142,13397,16107,15078,13919,15399,14857,14394,16008,15287,13996,13685,15305,13769,15282,15837,15324,13282,14877,15657,13165,13659,15666,13725,15422,15638,14874,15419,15059,13031,16127,13001,15973,15729,16128,16166,13913,15509,15757,13762,14886,16055,16101,15049,12922,14111,13801,15192,16107,13582,16121,15824,13118,13590,13330,14762,14463,15247,13830,14766,13383,14717,15513,14163,12972,15208,15048,15822,15544,16029,14336,13720,14320,14440,14295,14598,14142,14967,13086,13798,14621,13441,14688,14986,14044"
## sparsity 0.9 lengths from seqlen 16384
$EXE -v=0 -prec=$dtype -b=128 -jagged=1 -nhead=4 -hdim_qk=128 -hdim_v=128 -seqlens=$seqlens128 -causal=1 -local_len=256 -context_len=0 -minfull_len=256 -targets=200 -max_seqlen=16184 -perf=1 -alpha=0.0078125
echo -e ""
seqlens256_0="3336,3214,3647,3481,3486,3088,3572,3465,3141,3298,3577,3808,3744,3220,3620,3266,3606,3805,3228,3247,3422,3293,3538,3729,3120,3536,3855,3494,3109,3777,3079,3688,3744,3320,3831,3229,3280,3351,3532,3218,3815,3365,3130,3173,3692,3863,3713,3104,3330,3113,3143,3252,3202,3825,3303,3391,3184,3529,3430,3498,3724,3741,3448,3502,3730,3481,3479,3118,3479,3093,3764,3156,3457,3811,3646,3644,3888,3290,3098,3076,3665,3796,3794,3570,3734,3722,3356,3476,3124,3638,3568,3674,3378,3410,3870,3124,3242,3768,3423,3318,3600,3499,3784,3312,3211,3486,3702,3136,3450,3635,3271,3617,3446,3166,3302,3518,3295,3580,3829,3876,3771,3692,3286,3155,3298,3431,3563,3846,3441,3181,3121,3456,3252,3609,3410,3448,3577,3537,3508,3362,3687,3491,3600,3146,3163,3609,3617,3553,3570,3422,3147,3457,3161,3231,3568,3238,3167,3084,3392,3878,3660,3724,3887,3430,3775,3507,3551,3534,3450,3145,3626,3694,3223,3098,3595,3364,3798,3153,3618,3208,3403,3856,3563,3564,3291,3521,3235,3127,3316,3814,3241,3639,3751,3724,3285,3504,3115,3811,3293,3185,3585,3160,3667,3537,3425,3728,3559,3644,3595,3329,3303,3488,3177,3547,3671,3455,3894,3698,3663,3497,3861,3507,3826,3765,3751,3680,3145,3209,3631,3630,3747,3604,3657,3739,3255,3119,3376,3153,3272,3783,3564,3845,3089,3694,3628,3382,3439,3674,3827,3501,3525,3257,3773,3511,3099,3801"
## sparsity 0.9 lengths from seqlen 4096
$EXE -v=0 -prec=$dtype -b=256 -jagged=1 -nhead=4 -hdim_qk=128 -hdim_v=128 -seqlens=$seqlens256_0 -causal=1 -local_len=256 -context_len=0 -minfull_len=256 -targets=200 -max_seqlen=3896 -perf=1 -alpha=0.0078125
echo -e ""
seqlens256_1="6676,7452,7636,7424,7539,6958,6583,6418,7893,7685,6511,6922,6806,7952,7766,6544,7952,6431,7901,7966,7573,7730,6759,7884,7136,7011,6949,7709,6353,6944,6460,7163,7654,6432,7855,7240,7598,7239,6837,7814,7929,7917,6874,7552,7248,7406,6828,6658,7057,7919,7399,7741,6742,7869,6501,7456,7290,7543,6995,7806,7771,7402,7348,7326,7617,7406,7281,6940,6537,7532,6601,6555,6692,7572,7405,7151,6937,7913,7091,7712,7638,6415,7903,7706,7722,6548,7565,6585,6774,6507,6542,6916,7344,7719,7019,6905,6729,7637,7418,6874,6819,6587,7283,6940,7649,6744,7464,7028,7612,7793,7552,7771,6888,7398,7124,6972,6576,7289,7529,7537,7675,6746,7271,7621,6540,7529,7156,7893,6829,6462,7311,7394,7669,6525,6440,7207,7445,7155,7202,7614,6983,6415,7498,6551,7843,7788,7757,7072,7194,7621,6877,7668,6582,7511,6479,7521,6606,6812,7213,7893,6784,7330,6473,6738,6480,6766,7093,7012,7053,6734,7063,7093,7875,6831,7787,7818,7200,7352,6742,7389,7977,7653,6415,6539,7952,7782,7147,7872,7990,7357,7381,7951,7580,7428,7083,6720,7283,7832,7180,6879,7267,7358,7391,7736,6727,6962,7059,6592,6835,6762,7615,6591,7259,7302,6491,7620,7425,7687,7264,7207,6703,7593,6555,6901,6651,7705,6991,7854,6358,6475,7501,7427,7390,7232,6758,7281,7343,7349,7329,7314,7972,6908,7014,7091,7830,6807,7827,7817,7596,6828,6657,7760,6646,7233,6538,7498"
## sparsity 0.9 lengths from seqlen 8192
$EXE -v=0 -prec=$dtype -b=256 -jagged=1 -nhead=4 -hdim_qk=128 -hdim_v=128 -seqlens=$seqlens256_1 -causal=1 -local_len=256 -context_len=0 -minfull_len=256 -targets=200 -max_seqlen=7992 -perf=1 -alpha=0.0078125
echo -e ""
seqlens256_2="16157,16136,14491,14708,14989,14845,15311,15958,14005,12961,12953,13808,14951,15988,14919,14383,15252,13338,15367,14524,15074,14557,13742,15494,14717,15922,13569,13225,15853,16117,13634,14023,14817,14848,12948,13039,14400,13695,15897,13616,13083,15300,14298,16142,13397,16107,15078,13919,15399,14857,14394,16008,15287,13996,13685,15305,13769,15282,15837,15324,13282,14877,15657,13165,13659,15666,13725,15422,15638,14874,15419,15059,13031,16127,13001,15973,15729,16128,16166,13913,15509,15757,13762,14886,16055,16101,15049,12922,14111,13801,15192,16107,13582,16121,15824,13118,13590,13330,14762,14463,15247,13830,14766,13383,14717,15513,14163,12972,15208,15048,15822,15544,16029,14336,13720,14320,14440,14295,14598,14142,14967,13086,13798,14621,13441,14688,14986,14044,15040,14729,14553,14420,15312,14068,14518,15056,15390,15324,14413,13199,14778,15666,13888,15616,15987,14236,13413,15637,15398,13811,13525,14926,13120,12964,13448,15091,12935,15236,15586,15735,14259,16099,15261,13482,13275,13025,15219,12935,15179,15045,14351,13844,14752,13533,15802,14471,13115,16074,13241,13525,14175,14919,14843,13538,14964,15117,13524,14187,14044,16177,14522,13249,13537,13404,15015,16095,13439,14936,15221,14499,13621,13886,15300,14649,15778,14591,13971,14746,14469,15913,15136,15005,13302,16037,15211,15921,13614,13628,13434,14495,15959,12999,15581,15203,14061,15810,15096,14835,15662,14120,13315,13560,14204,13593,14151,13637,15058,13855,14515,13746,15481,16019,13957,14539,13350,13822,16167,14884,15475,15832,13441,15021,13609,15541,15287,15413"
## sparsity 0.9 lengths from seqlen 8192
$EXE -v=0 -prec=$dtype -b=256 -jagged=1 -nhead=4 -hdim_qk=128 -hdim_v=128 -seqlens=$seqlens256_2 -causal=1 -local_len=256 -context_len=0 -minfull_len=256 -targets=200 -max_seqlen=16184 -perf=1 -alpha=0.0078125
echo -e ""
set +x