submission 754149
skylers28_13449 · python · License unknown
Use it
Vendorable · source mirrored · license unknownView source →
No package. Vendor the mirrored source: 607 lines, June 9 Researcher Reciprocity License v1.0.
submission.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-mxfp4-mm-754149?include=source"interfacepython
Compatibility
measured onAMD Instinct MI355X
declared hardwareAMD Instinct MI355X
architecturesgfx950
dtypesbf16, mxfp4
Benchmark evidence
1 measurement across 1 GPU, fastest first.
Operation / workload
Hardware
Latency
Rank
Observed
Reported · How evidence levels are derived →
Source and license
sourceavailable
revision digestsha256:355a411bd2360d68c3e1a12f2b291746c46ea21110179413027f6be068d9e91f
license declaredunknown
license concludedunknown
authorsskylers28_13449
imported2026-08-26
Techniques
Extracted from the mirrored source by pattern, never inferred. Each row cites its line.
fp4
FP4 quant + FP4 GEMM single-file submission.Kernel source
submission.py607 lines
#!POPCORN leaderboard amd-mxfp4-mm
#!POPCORN gpu MI355X
"""
FP4 quant + FP4 GEMM single-file submission.
bf16 A, MXFP4 B -> MXFP4 per-1x32 quant A -> custom GEMM kernels -> bf16 C.
Kernel dispatch (6 separately-compiled kernels):
- v25_2x16 : M=4, N=2880, K=512
- v25_8x64 : M=32, N=2880, K=512
- v25_8x64 : M=32, N=4096, K=512
- v31_32x64 : M=64, N=7168, K=2048
- v24_64x16 : M=256, N=3072, K=1536
- v31_8x32 : M=16, N=2112, K=7168
- aiter fallback for all other shapes
Each kernel is gzip-compressed and base64-encoded, then extracted and
compiled separately via torch.utils.cpp_extension.load at first use.
"""
from task import input_t, output_t
import os
import gzip
import base64
import torch
import torch.utils.cpp_extension as cpp_ext
DIR = os.path.dirname(os.path.abspath(__file__))
BUILD_DIR = os.path.join(DIR, "build")
# --- Kernel: v25_2x16_t64_o4_ks512_4_2880_512 (from custom_kernels_B_register/v25_2x16_t64_o4_ks512_4_2880_512.cu) ---
_B64_V25_2X16_4_2880_512 = (
"H4sIAP1a1GkC/+1ce1PbSBL/n0+hylYoyTjBI8mPxcAVEMhSCSQXyO3tUdSUbMugxbK9spw1l+W7X/c8pBlpZBvY262tjSsBM9Ov"
"+U2rZ6an7e+icX80H4TW7m003Yb/NJmP0ygOX9/ub3xX6uwNR5MgJS29d3rfi8YDQrblG727P0sH0Tjd3+hPxrM0XEwTC/60zqw9"
"y+8WGs+h0e10GsX2d9DeJG6x+fD9h6N3FCVVdKE80ipJoxeXxx8ts0zeSQ9/ujy+QMHNEvvF0cH7Y/r204fPTIZXEnH++YwqRBdG"
"I85Ozg7oOxRA3E7ZiNPL408XJogYX0F6iejHg08fLxgwxNx1XtVFL07/cwydLd80KsZsmXix9/KHT8cHby4quf99xDpLoz0+OTk+"
"ujz917F5KvNu83R+On5rZsSO82rwL0/fc43L+s38F2fHZ/SAnnz0OV4VviTI+GwJYD0z1cfjT/Tw8wnK8Y3gQecJuoRpoEzC5YfL"
"g/fMEtLolDA+oP/8fHB+KX3GaAin4J5nWcZpPqRvDi4PKCB7wUxGYsvkgYdi1Dolk5jeT8NBOLQoDdI0iXrzNKTUtin9EvbTSUJn"
"0X+xwbdqFr6dDG0Q6MCLCY48d+GvLaRTFmIxMeznokPTp9jDwiATxt7xnwsfhVE6CL9EfWAAecNJ0g+j8Sgaw98bc9AICjk1TSd0"
"OPVp6MaEy7MWjvV1A3FHQs8Fyl6UzgAzEMRYghnFLnvhdHW6GQVKILQZw/4+xFDH2rQai868QBlqVB6nOjkpksVAxqiwu32CryJJ"
"MPh5DvM9oDDlLIi15tYrKyxadhsNwfB4AssKqob1AEjbc8f6h2XbmoxdyyOsXWvd4a07VqNowCAcT5KYoql2Y+E38DW3frPsGMdG"
"YHLgl6qf88dFMzI5O1bMSRhIQLPPRoU0IXTy96/Y74Ip/UkMK184YM8MCt+1XCczxSXF+UrAnAGbM6DPmDctUhQ8S6P+3b14GGM2"
"G8Q4G0k4m49SQZhJRByMqudTRpjbsQkDFsrAbCENDXJw+GTOJqA4DPRdmsZCFufZyjQILCUNAiPf71oc+qxhBxo4fRKm82Rs2eJh"
"cWzu279ltGDEw7Kn7MskGoCUmwhcKKG/zAPYz8Dze4AP3CwOY5upYcHKUvc0NZAEY4AA0E9B7AF0kFadEfMVOZ2kwShvSCa/0lka"
"JGnedEfDURirrWIUBdkBe/LRmOVUs34wChU61JFGgw0ZJ/SIy+I//eHg/Um+MShE/RouHZZ4cQiQ8wambEoBtj0Ur3iNTnYbjIbC"
"w5Bs00w1moDVFNABskzwdmlPZGKd3cg2hfXlWqw3o0kvU5tNDXhjZo6J606QMY3q7AEjGFPTtnpbHIAa2w7osnACSQtncJb0SYs9"
"DdAUJlNwZ9oPZulugXDfzkTga5O725UyjJp0OVAs7LzOeJzcAr52fAlGsyvSus7b81geLGjQg6XDc8GwRk7xVTMhNxC52EgYw5KR"
"MMJ9m41ZsUjTPg2ipAFimLirxnUF2WiCRDan5msSBDkHI6kKt8ZzGyGPzsIWAZ2eQdO4Zuso8uIyykCzQaljoCUmWlBWoEWTN/ey"
"9TGLyfKF5i3r1+cF/rKVljprABV1lOMsQ5dk6JJqdIlEl6yPLhHokpXoumZ0iQldz4wuKaFLVqBLno8uQXTJUnTdDF23Gl1Xouuu"
"j64r0HVXouub0XVN6DbN6LoldN0V6LrPR9dFdN2l6HoZul41up5E11sfXU+g661Et2VG1zOh2zaj65XQ9Vag6z0fXQ/R9Zai62fo"
"+tXo+hJdf310fYGuvxLdjhld34Tu92Z0/RK6/gp0/eej6yO6/lJ0mxm6zWp0mxLd5vroNgW6zZXokoplrWlc1irWtWYJ3+YKfJvP"
"x7eJ+DaX4tvK8G1V49uS+LbWx7cl8G2txrdiYWsZ8a1Y2VolfFsr8G09H98W4ttaim87w7ddjW9b4tteH9+2wLe9Gt+Kpa1txLdi"
"bWuX8G2vwLf9fHzbiG9bxfdhw4Byko7hEAoszPLZ7XBEF5NEF6jufFboVgQ6hp1+gMQRY1X54ADRWLg8P+KIjE6nUZgTPA71omCW"
"JXRYvszORPJ0kU4fduIGzyflfAAsnuwbmE0B85WuVxaMoFEQ8Qse7xWtbtPHLBIILp5w2EmeH4kNbmBLEBxdImZivAJWHTY1/btw"
"AMtT3iWaWEgtp+jkKaKmGsJyPBW0pEQLtqhrmdBHluhzH6HPW1Ofu0Sf/wh9zTX1eUv0tR6hr72mPn+Jvs4j9H2/pr7mEn3kUQ6z"
"rse0lml8jMuQdX2mvUzjY5yGrPQaDArIMxkOQWeeTKrpF3MiP0NaeVqmownhSYzBLDUnMfLsxWaejbsSiq8dQ7JoraxILpYjpwhi"
"pvDoUs6F8E5yXXmUV5OBVyoqpbtGRAblZBlUHkxX50xhxgb0kA6CNMCJS8KbmZorNScpBf10FETjPEnZo3kiK2+Nh3FA71hiT82r"
"Iui9+zSc5Y2HNIHnGP7pqc7cBszwTdMEhqkaAGO32WVN6qgWMN/JtHBcv5smwU0cWPNxMhmNWBNAwhY96y5ieTL8vStvYvGvrS1H"
"yZmxcYJA4alAXLNscam7jen/LXXAxrSh6lXr+JQc9VamWPEvxOwuus48TPMv2Sk9TPMv2enKTtfQ6clOeVB/WN+luOs+wqeEr9/O"
"h8NROHiiXzEhKu+sjwaoXoXtkyS6EXljVXw3IxjDAw+dGR3shJpKr8svfdRu38mT44wGwpQqAToVCT3YrNBgMEBvRl3gRuynGIGY"
"Yt3xYCj9yQg3X8rg9SB610GdkhCs8gr9bFS2QuA6ek6fUfmqlM2ikOEoSNGKO7xndZsYje/wnrTlwzscds3Cd3c4HtyHjpX7h2xC"
"mMey7WZx5q9ybLa4ruuuAQdMVJbLIrbM2Ki4kAIuGibEhImKBzHjQZ6LB1kfD2LEAxco1PAkTNxlmLirMHHNmLjPxcRdHxPXiAkm"
"BL2nYuItw8RbhYlnxsR7Libe+pjoQVtEOQzao2A+7t/SHt7Zzii1lTKeOsQxHsZx7bwJ45jehck4HD3hBrUQ7tlrxT5iDRbTMqEb"
"ZGA7opO5cmd7VucVNvzXu2xpCBewFI/ZoT1I4DBJs6Mj24Jdd/WbW2UTqRTrlKmUXVyZ7mvxmFpjiwO7PgMOPMfjnaRSLpR7hWIA"
"xFN0PsaqEii6WcjlsreKxUxdsxUkt4KsYwXJrSAVVhBpBamy4iGbp19v2P10D/a+d6eDxetFd0O/y2XlaW8v2NVxNHgTxZIkGlq2"
"7Hyp1KFhpiKzS4i32e+cysHVWDJvK81bglRp6xru5H98K8octY2fVtf18fQNK0Kzd3bOQKwsaXxlEQekiz9XCTjnAs4zAee6gPNl"
"An58CxJOz8U19x4zupaLrmIdz2OKGNBoTPked68sLt/mZBf6/LZdoFcSkjMMo2SW0in0xVpBQA0N1MMp72NbfyzviaOxnUP7SpVU"
"R2YnV8LF72nKYG6lH5TMc6BR1VYQNVZ8yMC7rfEWPTgV4LACi9skDAZGR/81SASOvA5jO6/bLJKOgnFoKSUbL6tJmVTcnOLcCA3b"
"sljUSAwrm0r8soqYra6DKOZmMJNeZoehAtmddCRGtl0+MykVsjIVCH7/DqFVFnVVdm8CdgYUBdKgn0ZfGE9m1K5afeqUGHtrMZ47"
"xdEEiCXOfBqNQivDiZ9G1XpXcDbduH/kgO1k6VJVMA+gKJ6fPHVVtYpyGZ5EPRInIGD72njdGNat4s8H80h4VcyeeF5qelzKiXsl"
"4nFGfF4m1qp2VN6t3MVUuM61nEjhBUj2ViOppRe6ptMhbzRlQ7RNin4Y1Q6hFssA1FGPo0nTD8KlDcxSkeI8m1nqyHq56ko3sQOz"
"3tXVWYR5ri/D0bCfqBd3D3UMJ8IC2CLdj/s8YM3s0nHVVGXEa69YEgOXf0y3FLInevaLFzAnWrq+LFtSBhUPRiGBqBqwLFdTkbMJ"
"RAJqZdamJGhTw/YqKKYc5QtHzDZqTBXmc1gTyZuKWULJ5eYkruDy8iavgouFhQYnbyrvW8r79rVeQ6bffOUzMOsvmyle7AerxkJM"
"/dqnMrmHlCefgl9eFcPjltBzvcpkeFYpaiz5WNbBpuKQzQmbi6yDKB2kittViFyV21M6vCpuOTVZQ7PY0Co2mGZKRjoWbbPsy6pn"
"Kqtuh+Uj6PfLzsOaOT5HUqZoJHmjyVk5kZsTuRmnlzd6VZzsIrA3j0YpPOVBPLjpj9nghRMM4QkkrQX+czt02Bm2hr5tDH01W3wa"
"oOZIAOtqI7h+nas0h04fDsvmngaP2JWdaKtjGl6WGZPodjMcs0YzpFm+RKLbzXDMGr3KB+LhkZGb/A6R+/8Utf/QiP34aP34SP0n"
"RmnyB0fpZ0VoUhWhyToRmlRFaPLnRGiy9NlZGp2fFpmfFpV/v4j8nGhsjsSVUdgcgZ8Wff/oyOt+i7x/h8jr/tX2x25V9HXXib5u"
"VfR1/5zo636Lvt+iryH6et+i798h+np/tejrVUVfb53o61VFX+/Pib7et+j7N4q+D9mFbcX9jDL/eVyezFNxhVC4nZDXCKZbF6mg"
"WyEQrx5KNxiV9xKlkKxc+aoVCVYaT8X1S+EzCTmNLWfAqZvaCbSv4nUreL3r/Mb7oZw6qpVWCKwAaPm8TpYVTlxJsGsWDlogde1Y"
"e+a82pJPDwu5/CP9sjiF1ZsMotk0SPu3VBaeVFecyCuOQrXIynKS8t2LLnfNKhF5D87KEmbsayOqL/FZMYFdfUcvYOAVrKzWgl0U"
"g9T8+126ssjlZD7uX4TpgfzGEttG6GpOoVan+qZHyMgEnAWLN/fjII76F6zi5SyMJ8n9BRhQz20RJhaU7O7uwpPk2TkS4H6sRSkn"
"chQx+/vKZkNOoX67Vp4gPiU4FTAN79jXQIC5IbhKP7Sm9/i0iu+f6m58/Onw9PwNIfTsw5vP74/tL26Tugv85D143cSnd7MmcalP"
"8bumKLytW7Gc0vj1YNK3waGtF8BlnXz0rbfHZ2cWslvAbk18i7Fb9uGraPwKr/L4J3OiYMS2CtZsPp1OktR50ZUSw6H9IpmPX9St"
"q2t7er+zM+n9HPZxiwgjUv6WF47madMIxV2ixn5UxWnwYTWWBvN0YvFNom5fClTyKznS1/j1OPYLPktp8sKxnddalNi3na4aVyqe"
"Y/mqiA3ag7hvgyYbYHIKUW9JYOnwAnzgk3A+kVlAvJLbYPFRmS13XBbsmAP/D8MCWRSATQAA"
)
# --- Kernel: v25_8x64_t256_o2_ks512_32_2880_512 (from custom_kernels_B_register/v25_8x64_t256_o2_ks512_32_2880_512.cu) ---
_B64_V25_8X64_32_2880_512 = (
"H4sIAP1a1GkC/+1cbVPbuBb+zq/wdKeMHdISyY6TJZSdQKHLtNDeQu/evQyjcRIHvORtHacbbpf/fs+RbFmy5SSUDzs7S6aFIJ3X"
"R/KRdHSSH6JJf7QYhNb+bTTbhf8sXkySaBy+vj3Y+qHU2RuOpkFCfL13dt+LJgNCdrM3end/ngyiSXKw1Z9O5km4nMUW/GmdWW8s"
"l3YKrefQStvtRrH9PbQ3SYn88MPHo/cMRbXNXSjP90rS2MXl8SfLLFN0ssNfL48v0JqmX6S4OOp+OGbvPn/8wmUYvPhyxhQiFENK"
"Us5OzrrsPQogtF024vTy+DMyeka+gvQS0S/dz58uODDE3HVexcUuTv97bEQNveLMlokXey9//nzcfVuBGhL854j3ltw9Pjk5Pro8"
"/fexeSzz7nMjlJ+P33FGRYyJ5FwjOTcCe3n6QRhRMWC832zFxdnxGeuyk0+ewBBhog2vXUEnhjBF2zAFONmn48/s8MsJSiItI6TQ"
"e4IzBUmMIi4/XnY/cGM8t1ki6bJ/femeX2ZTyWyKIBEz0uKiSrPqkL3tXnYZwHzBjUZi4zw5TB3XKbnE5H4WDsKhxViQJHHUWyQh"
"Y7bN2Newn0xjNo/+hw2eVbPw7XRog0AHXlxw5NKlt7GQdlmIxcXwn8s2S77HHh4fuTD+TvxceiiMsUH4NeoDA8gbTuN+GE1G0QT+"
"3lqARlAoqFkyZcOZx0I6JkKetXSsb1uIOxK6FCh7UTIHzEAQZwnmDLvspdPR6eYMKIHQ5gwHBxBbHWvbaizbiwJlqFG5gurkpEg2"
"BjJOhd2tE3wVSYLBbwsY7wGDIefTyV9Yr6ywaNltNATDx1NYb1A1LBRA2lo41k+WbWsy9i2X8HatdU+07lmNogGDcDKNxwxNtRtL"
"r4GvhfWnZY/RNwKDA79U/YJ/XDRDytmzxoKEgwQ0B9wrpAmhU7x/xX8XTOlPx7AkhgP+zKDwfYs60hRKiuMVgzkDPmZAL5m3LVIU"
"PE+i/t19+jCO+WgQ42jE4XwxSlJCKRFxMKpezDhhbsc2OJwqA7NTaWiQg+6TBR+Aohs4d1kyTmUJnh2pIcUyo0Fgsvf7loBeNuxB"
"g6CPw2QRTyw7fVgcW8ztPyUtGPGw6in7Oo0GIOUmgikUs98XAWx04Pnt4gM3H4djm6vhwcpSNzs1kAQ+QADoJyC2Cx3Er3NisVIn"
"0yQY5Q3x9A82T4I4yZvuWDgKx2pr6kVBdsCffDRmNdW8H4xChQ51JNFgK4sTesTlCwD7ufvhJN8wFMJ+DdcOK30JCJDzBoZsxgC2"
"NyhemTU62W0wGqYzDMm2zVSjKVjNAB0gk4J3S3slE+v8JmtTWF9uxHozmvakWjk0MBulOSauu5SMa1RHDxjBmJq2BdwRANT4lkCX"
"hQNIfBzBedwnPn8aoCmMZzCdWT+YJ/sFwgNbisDXtphuV4obtWzKgeLUzmvJ4+QWiLXjazCaXxH/Om/PY3mwZEEPlg6XgmGNnOKb"
"ZkJuIHJxTzjDCk844YHNfVYs0rTPgihugBgu7qpxXUE2miKRLajFmgRBzsFIqsKt8dxGyKOz8EVAp+fQNK75Ooq8uIxy0GxQ6hho"
"iYkWlBVo0eTtN3J9lDE5e6F5q/r1cYG/bKWlzhtARR3lOKvQJRJdUo0uydAlm6NLUnTJWnSpGV1iQtc1o0tK6JI16JKno0sQXbIS"
"XSrRpdXo0gxdujm6NEWXrkXXM6NLTeg2zejSErp0Dbr06ehSRJeuRNeV6LrV6LoZuu7m6Lopuu5adH0zuq4J3ZYZXbeErrsGXffp"
"6LqIrrsSXU+i61Wj62Xoepuj66XoemvRbZvR9Uzo/mhG1yuh661B13s6uh6i661EtynRbVaj28zQbW6ObjNFt7kWXVKxrDWNy1rF"
"utYs4dtcg2/z6fg2Ed/mSnx9ia9fja+f4etvjq+f4uuvx7diYfON+FasbH4JX38Nvv7T8fURX38lvi2Jb6sa31aGb2tzfFspvq31"
"+FYsbS0jvhVrW6uEb2sNvq2n49tCfFsqvg9bBpTjZAKHUGDhls9vhyO2nMa6QHXns0a3ItAx7PQDJI44q8oHB4jGkor8iJNmdNqN"
"wpjgcagXBXOZ0OH5MluKFOkinT5sjxsin5TzAbB4sm9gNgXMV7peWeBBoyDidzzeK1pp08MsEggunnD4SV4ciQ3TwM5AcHSJmIlx"
"C1i1+dD078IBLE95V9rEQ2o5RZedImqqITzHU0FLSrRgi7qWpfrICn30EfrcDfXRFfq8R+hrbqjPXaHPf4S+1ob6vBX62o/Q9+OG"
"+por9JFHTZhNZ4y/SuNjpgzZdM60Vml8zKQha2cNBgXkmQ6HoDNPJtX0C7s0P0P8PC3T1oSIJMZgnpiTGHn2YjvPxl2liq8dQ7Jo"
"o6xILlYgpwjipojoUs6FiE5yXXmUV5OBVyoqpTtIRAblyAyqCKbrc6YwYgN2yAZBEuDAxeHNXM2VmpOUKf1sFESTPEnZY3kiK28d"
"D8cBu+OJPTWviqD37pNwnjceshieY/inpzpzGzDDN0ticFM1AHy3+WVN4qgW8LkjtQhcv60Y4MYm46uNTmYO6G+ANju9893FW4Ad"
"1W8+ZdVdA3rauJbTooGzImskspHkjVQ20rzRlY3qMbrsGXmKZ+TxnpHcMyI9I7lnRHpGcs+I9IzknpE1ntGneEYf7xnNPaPSM5p7"
"RqVnNPeMSs9o7hld45n7FM/cx3vm5p650jM398yVnrm5Z670zM09c6VnD5uHHxHmHhF/0rh4uxgOR+HgO2MQF6LyzvtogBqBsH0a"
"RzfpHYMqviMJJrA4QKekg11zU+ml4oJQ7fac/CKF08CSpkqATkVCDza2LBgMMPKhLhhd/jP1wCkGN36r0mf96QiDmuK8vuDetVFn"
"RghWuYV+7pWtEFBHv//hVJ4qZbsoZDgKErTiDu/kaRNX7ju8U/c9eIdu1yx8d4f+4JllotxVyQHhU5MfTYojf5VjsyN0XXcMOGAI"
"LJfW7JixUXEhBVw0TIgJExUPYsaDPBUPsjkexIgHBk7U8F2Y0FWY0HWYUDMm9KmY0M0xoUZMMOS634uJuwoTdx0mrhkT96mYuJtj"
"ogftNMph0B4Fi0n/lvXwfn/OmK2UgtXBKRHGcZ91E47H7C6MJ+HoO27bC+Gev9bsOTdgMS0TukEGtiM2XSj3+2d1UY4lfr2XS0O4"
"hMV5whM8QRwOgDFLM/Dt+nVHv+VXDhxKZVeZStnxl+m+FVMaNb448KtW4MCcD95fK8Vl+axQDIB4ipOPs6oEim4ecoXsnWLxW8ds"
"BcmtIJtYQXIrSIUVJLOCVFnxIMfpjxtey9CDc9Ld6WD5etnZ0u/9eYnjuwteZhAN3kbjjCQaWnbW+VIpZcSslrQrFW/z3zmVg6tx"
"xryrNO+kpEpbx1C/8cu7M1FIp+0FtSLAT6dvedGivbd3BmKzsthXFnFAevrnOgHnQsC5FHCuCzhfJeCXdyDh9DwtiXjDja7loqtY"
"J4sxQwxYNGFi6/mmLE4Pe7IARFRnpAiWBBViZRTPEzaD/rFWRFJDQ03y+XERS8LG0cTOIX6lSqojcyFFIVS80RTCOGdzomSmA42q"
"RoO4iTKnDPy7Gr+pmiVJweIFOrdxGAyUya+T/hHEKbailmc3rwk21vQEk9BSSn9eribn0nHziuOWatrNCpIrGWD1UxlermLgq/Ag"
"GguTuHkvDXU5hR1/RrpbJlUqsrMUMzwj7xF2ZQOgxpHeFOwNGApkQT+JvnIeadi+WqXslBh7GzGeO8XIFSCuOCuSaBRaEi+R5VDL"
"q2Ey6sb9lIO2J9PwqmARbFG8yL3pqmoVZVgiOX+UnpaA7VvjdWNYt4o/H8yeiGqrN+nzVNNjWE7cKxFPJPF5mVirBlN5d/KppsJ1"
"rp2fCy9AsrceSS1t1TGdJEWjKcumbWj0g6t2YLX4Ib6OehxNmn5oLm12VopMz77SUierw6yuoEx3a9b7ujqKMM71VTga9h714k6j"
"juEltQC2U/eTvghkc7t0tDVVr4maPkzzMdwqYLlbIQGiZ1VFZXysXQOVZYvZan4oCklpVXk51dIxmyxSPUGa0HxcrocXD2qYXgXF"
"FDa+ujxRyB0ANTyv0+UZNtlUzDh3eZpKdtOUw82bXAMHDwENQdpU3vvK+9a1Xoeo357miM/760ZlfgMrxTId5o1Pa9neMjsRFebg"
"VTEU7qR6rteZDM8lQ42l+SQ7+BDInG8n7yBKB6nipgoRVbldpcOt4s6GRjY0iw1+scE0UllU45FVZmVWjZT8dAQsE0G/MKi8SeBy"
"lMlKG0neWIREENCcgEouN290TVz88ri3iEYJPMXBeHDTn3Bn00EfwlNG/CX+o202bA/9oVd+5Gp2+umRmpOBVVcbYZrXhbpySPTq"
"8K/U2hAR2NiBtjlFV2Q2LEOwI7GSjWXYZH4kQ7AjsZKNbuVEf3hk9CXP0fefEH3J3y36kqroSzaJvqQq+pK/JvqS5+j7HH0N0Zc+"
"R99/QvSlf7foS6uiL90k+tKq6Ev/muhLn6Pvc/Q1RF/3Ofr+E6Kv+3eLvm5V9HU3ib5uVfR1/5ro6z5H339Q9H2oTgzjzW3F5Yup"
"2NOaLpL0fqBw9ZDdEZiuVDIFnQqBeK9Qup6ovHQoxWnl7lctTbCS8Sy9Wyl8kCWnsbNhceqmdgLt63hpBa97nV99P+QW1krLBZYA"
"+J4oquaVE1cZyDULnU0RunasN5ottRUfMU/lie99yKpSeKHJIJrPgqR/y7KKk+pSk+y+olAmsraOpHyRosvdsDwkuwDn9Qhz/t0i"
"1bf3vIrArr6cT2EQZc68yILfCIPU/FuAOll1y8li0r8Ik272tTa2jdDVnEKRTvW1TSpDCjgLlm/vJ8E46l/wUpezcDyN7y/AgHpu"
"S2piQcn+/j48Oa6dIwHTjbcodUSOIubgQNlxZEOoX5WVB0gMCQ4FDMN7/l0hYG4IU6UfWrN7fDrTby/rbH369fD0/C0h7Ozj2y8f"
"ju2vtMnaS5xxtOmzKWV38yahDMIyflUZg/d1a5wN6vj1YNq3YSpbL4DPOvnkWe+Oz84sFGChAGtKLS7Asg9fRZNXeDUnPsEVBSO+"
"ZbDmi9lsGifOi04mMhzaL+LF5EXdurq2Z/d7e9Peb2Eft4rglPJ3doFoHjmNML0b1NiPqjgN01gNn8EimVpis6jblwBV9tUtyWv8"
"GiX7hRioJH7h2M5rLUAcQNBWQ0nFoyzrmc3hQXsWD2zQZANMTiHQrYgtbfFBDeDL4PxO5hTitdwGi4/KbPnc5fGOz+H/A1K/KyrB"
"TwAA"
)
# --- Kernel: v25_8x64_t256_o2_ks512_32_4096_512 (from custom_kernels_B_register/v25_8x64_t256_o2_ks512_32_4096_512.cu) ---
_B64_V25_8X64_32_4096_512 = (
"H4sIAP1a1GkC/+1cbVPbuBb+zq/wdKeMHdISyc7LEspOoNBlWmhvoXfvXobROIkDXuIk6zjdcLv893uOZMuSLSehfNjZWTItBOm8"
"PpKPpKOT/BBOBuPFMLD2b8PZLvxn8WKShFHw+vZg64dSZ380nvoJaem9s/t+OBkSspu90bsH82QYTpKDrcF0Mk+C5Sy24E/rzHpj"
"ubRbaD2HVq/xY6vY/h7am6REfvjh49F7hqI65i6U1/JK0tjF5fEnyyxTdLLDXy+PL4CCNkvGXBz1Phyzd58/fuEyDF58OWMKEYoh"
"JSlnJ2c99h4FENopG3F6efwZGT0jX0F6ieiX3udPFxwYYu46r+JiF6f/PTaihl5xZsvEi72XP38+7r2tQA0J/nPEe0vuHp+cHB9d"
"nv772DyWefe5EcrPx+84oyLGRHKukZwbgb08/SCMqBgw3m+24uLs+Iz12MknT2CIMNGG16mgE0OYom2YApzs0/FndvjlBCWRthFS"
"6D3BmYIkRhGXHy97H7gxntsskfTYv770zi+zqWQ2RZCIGWlxUaVZdcje9i57DGC+4EYjsXGeHKaO65RcYnI/C4bByGLMT5I47C+S"
"gDHbZuxrMEimMZuH/8MGz6pZ+HY6skGgAy8uOHTp0ttYSKcsxOJi+M9lhyXfYw+Pj1wYfyd+Lj0Uxtgw+BoOgAHkjabxIAgn43AC"
"f28tQCMoFNQsmbLRzGMBjYiQZy0d69sW4o6ELgXKfpjMATMQxFn8OcMue+l0dbo5A0ogtDnDwYFFO461bTWWnUWBMtCoXEF1clIk"
"i4CMU2F3+wRfRRJ/+NsCxnvIYMj5dGotrFdWULTsNhyB4dEU1htUDQsFkLYXjvWTZduajH3LJbxda90TrXtWo2jAMJhM44ihqXZj"
"6TXwtbD+tOwIfSMwOPBL1S/4o6IZUs6eFQkSDhLQHHCvkCaATvH+Ff9dMGUwjWBJDIb8mUHh+xZ1pCmUFMcrBnOGfMyAXjJvW6Qo"
"eJ6Eg7v79GGM+GgQ42jEwXwxTlJCKRFxMKpezDhhbsc2OJwqA7NTaWiQg+6TBR+Aohs4d1kSpbIEz47UkGKZ0SAw2ft9S0AvG/ag"
"QdDHQbKIJ5adPiyOLeb2n5IWjHhY9ZR9nYZDkHITwhSK2e8LHzY68Pz28IGbR0FkczU8WFnqZqcGksAHCACDBMT2oIO06pxYrNTJ"
"NPHHeUM8/YPNEz9O8qY7FoyDSG1NvSjI9vmTj8asppoP/HGg0KGOJBxuZXFCj7h8AWA/9z6c5BuGQtiv4dphpS8BAXLewJDNGMD2"
"BsUrs0Ynu/XHo3SGIdm2mWo8BasZoANkUvBuaa9kYp3fZG0K68uNWG/G075UK4cGZqM0x8R1l5JxjeroASMYU9O2gDsCgBrfEuiy"
"cABJC0dwHg9Iiz8N0BTEM5jObODPk/0C4YEtReBrW0y3K8WNWjblQHFq57XkcXILxNrx1R/Pr0jrOm/PY7m/ZH4flg6XgmGNnOKb"
"ZkJuIHJxTzjDCk844YHNfVYs0rTP/DBugBgu7qpxXUE2niKRLajFmgRBzsFIqsKt8dyGyKOz8EVAp+fQNK75Ooq8uIxy0GxQ6hho"
"iYkWlBVo0eTtN3J9lDE5e6F5q/r1cYG/bKWlzhtARR3lOKvQJRJdUo0uydAlm6NLUnTJWnSpGV1iQtc1o0tK6JI16JKno0sQXbIS"
"XSrRpdXo0gxdujm6NEWXrkXXM6NLTeg2zejSErp0Dbr06ehSRJeuRNeV6LrV6LoZuu7m6Lopuu5adFtmdF0Tum0zum4JXXcNuu7T"
"0XURXXclup5E16tG18vQ9TZH10vR9dai2zGj65nQ/dGMrldC11uDrvd0dD1E11uJblOi26xGt5mh29wc3WaKbnMtuqRiWWsal7WK"
"da1Zwre5Bt/m0/FtIr7Nlfi2JL6tanxbGb6tzfFtpfi21uNbsbC1jPhWrGytEr6tNfi2no5vC/FtrcS3LfFtV+PbzvBtb45vO8W3"
"vR7fiqWtbcS3Ym1rl/Btr8G3/XR824hvW8X3YcuAcpxM4BAKLNzy+e1ozJbTWBeo7nzW6FYEOoadvo/EIWdV+eAA0VhSkR9x0oxO"
"p1EYEzwO9UN/LhM6PF9mS5EiXaTTB52oIfJJOR8Aiyf7BmZTwHyl65UFHjQKIn7H472ilTY9zCKB4OIJh5/kxZHYMA3sDARHl4iZ"
"GLeAVYcPzeAuGMLylHelTTykllN02SmiphrCczwVtKREC7aoa1mqj6zQRx+hz91QH12hz3uEvuaG+twV+lqP0NfeUJ+3Ql/nEfp+"
"3FBfc4U+8qgJs+mMaa3S+JgpQzadM+1VGh8zacjaWYNBAXmmoxHozJNJNf3CLs3PkFaeluloQkQSYzhPzEmMPHuxnWfjrlLF144h"
"WbRRViQXK5BTBHFTRHQp50JEJ7muPMqrycArFZXSHSQig3JkBlUE0/U5UxixITtkQz/xceDi4Gau5krNScqUfjb2w0mepOyzPJGV"
"t0ajyGd3PLGn5lUR9P59EszzxkMWw3MM//RUZ24DZvhmSQxuqgaA7za/rEkc1QI+d6QWgeu3FQPc2GR8tdHJzAH9DdBmp3e+u3gL"
"sKP6zaesumtATxvXclo0cFZkjUQ2kryRykaaN7qyUT1Glz0jT/GMPN4zkntGpGck94xIz0juGZGekdwzssYz+hTP6OM9o7lnVHpG"
"c8+o9IzmnlHpGc09o2s8c5/imft4z9zcM1d65uaeudIzN/fMlZ65uWeu9Oxh8/Ajwtwj4k8aF28Xo9E4GH5nDOJCVN75AA1QIxC2"
"T+PwJr1jUMV3JcEEFgfolHSwa24qvVRcEKrdnpNfpHAaWNJUCdCpSOjDxpb5wyFGPtQFo8t/ph44xeDGb1UGbDAdY1BTnNcX3LsO"
"6swIwSq30M+9shUC6uj3P5zKU6VsF4WMxn6CVtzhnTxt4sp9h3fqLQ/eods1C9/doT94Zpkod1VyQPjU5EeT4shf5djsCF3XXQMO"
"GALLpTU7ZmxUXEgBFw0TYsJExYOY8SBPxYNsjgcx4oGBEzV8FyZ0FSZ0HSbUjAl9KiZ0c0yoERMMue73YuKuwsRdh4lrxsR9Kibu"
"5pjoQTuNchi0x/5iMrhlfbzfnzNmK6VgdXBKhHHcZ90EUcTugngSjL/jtr0Q7vlrzZ5zAxbTMqEbZGA7YtNFIleAYAlr8ITncfw4"
"GEJ/lk3gu/Lrrn6Zr5wrlAKuMpWysS/TfStmLmp8DeA3qsCBqR28plZqyPLBVwyAsIlzjLOqBIpuHlmF7J1ijVvXbAXJrSCbWEFy"
"K0iFFSSzglRZ8SAXwz9ueMlCH45Dd6fD5etld0u/3ueVjO8ueDVBOHwbRhlJOLLsrPOlUrGIyStpVyre5r9zKgcX3Yx5V2neSUmV"
"tq6hTOOXd2eiXk7b8mm1fp9O3/LaRHtv7wzEZtWvryzigPT0z3UCzoWAcyngXBdwvkrAL+9Awul5Wvnwhhtdy0VXsU4WEUMMWDhh"
"Yof5pixOj26yzkMUYaQIlgQVQmIYzxM2g/5IqxWpoaEm+fxUiJVfUTixc4hfqZLqyFzIRAgVbzSFMM7ZnCiZ6UCjqtEgbqLMKQP/"
"rsZvKlpJUrB4Hc5tHPhDZfLrpH/4cYqtKNnZzUt/jaU7/iSwlAqfl6vJuXTco+K4pZp2s7rjSgZY5FSGl6sY+GI7DCNhEjfvpaH8"
"prCxz0h3y6RK4XWWSYZn5D3CrqzzahzpT8Fen6FA5g+S8CvnkYbtq8XITomxvxHjuVOMXD7iirMiCceBJfESyQy1ihomo27cTzlo"
"ezLbrgoWwRbFixSbrqpWUW0lcvBH6aEI2L41XjdGdav488HsiSiqepM+TzU9huXE/RLxRBKfl4m1oi+Vdyefaipc59oxufACJPvr"
"kdSyU13TgVE0mpJp2r5FP59q51KLn9XrqMfRpOln49KeZqXI9IgrLXWycsvqQsl0U2a9r6ujCONcX4WjYe9RL+406hheUgtgO3U/"
"GYhANrdLJ1hTkZoo3cNsHsOtAla1FfIcevJUFMDH2m1PWbaYreaHopB7VpWXMypds8kio+OnecvHpXR4jaCG6ZVfzFTjq8fzgdwB"
"UMPTNz2eSJNNxcRyj2ejZDdNOdy8yTVw8BDQEKRN5X1Led++1ssN9UvSHPH5YN2ozG9gpVimw7zxoSzbW2YHn8IcvCqGwp1Uz/U6"
"k+G5ZKixNJ9kBx8Cmdrt5h1E6SBV3FQhoiq3q3S4VdzZ0MiGZrGhVWwwjVQW1XhklcmXVSMlPwQBy4Q/KAwqbxK4HGWy0kaSNxYh"
"EQQ0J6CSy80bXRMXvyPuL8JxAk+xHw1vBhPubDroI3jKSGuJ/2iHjTqj1sgrP3I1O/2QSM3JwKqrjTDN60JdOSR6dfhXam2ICGzs"
"QNucoisy6ZUh2JVYycYybDINkiHYlVjJRrdyoj88MvqS5+j7T4i+5O8WfUlV9CWbRF9SFX3JXxN9yXP0fY6+huhLn6PvPyH60r9b"
"9KVV0ZduEn1pVfSlf030pc/R9zn6GqKv+xx9/wnR1/27RV+3Kvq6m0Rftyr6un9N9HWfo+8/KPo+VCeG8ea24vLFVNNpTRdJej9Q"
"uHrI7ghMVyqZgm6FQLxXKF1PVF46lOK0cverViBYSTRL71YKn1fJaexsWJy6qZ1A+zpeWsHrXudX3w+5hbXScoElAC1P1E7zAomr"
"DOSahc6mCF071hvNltqKT5Kn8sTXO2TFJ7yeZBjOZ34yuGVZYUl1RUl2X1GoBllbLlK+SNHl6lUg6pdaZX/nX2clSwqzL52SdX2L"
"iPFKhTn/cpHqe31eX2BXX9unAIk6Z15+we+KQWr+NUDdrLzlZDEZXARJL/teG9tGUGtOoUqn+kInlSEFnPnLt/cTPwoHF7wI5iyI"
"pvH9BRhQz21JTSwo2d/fh2fKtXMkYCLyFqWQyFHEHBwoe5FscPVLtPLQ8cHi3xECVgYwdwaBNbvHxzX91rLu1qdfD0/P3xLCzj6+"
"/fLh2P5Km6yzxClImy02pexuDkPHIE7jmDJ4X7eibPij18PpwIa5bb0APuvkk2e9Oz47s1CAhQKsKbW4AMs+fBVOXuFdnfjkVuiP"
"+R7Cmi9ms2mcOC+6mchgZL+IF5MXdevq2p7d7+1N+78FA9w7JsCb/53dKJoHTCNMLws19iOWqGHSXyRTS2wKdbVIlX0TS/IavxXJ"
"fiFgT+IXju281gLBAQRnNWRUPLKyPNkcBrRn7sAGTTZ47xQC2ooY0hGfuwC+DKXvZE6RW8ttsBgBdorf0vHAp+P/AS1BGjSETwAA"
)
# --- Kernel: v31_32x64_t512_o2_ks1024 (from custom_kernels_larger_k/v31_32x64_t512_o2_ks1024.cu) ---
_B64_V31_32X64_64_7168_2048 = (
"H4sIAP1a1GkC/+09/VPbSLK/81eo2Aol2V5AskMoDLwiCdnbIpBUyNbeHeWnkm0ZtLZsny0T8nL53193z4dmRiNZDuyrvL1Qu8GW"
"unt6evprej74KZkOJqth7BzfJfM9+D9crKZZksa7d6dbPxVe9keTWZT5B/rb+ed+Mh36/p74oL8eLLNhMs3UR9HkdrZIsrv0dGsw"
"my6z+GG+cADGefn23auL8NI5cdpB1/ruCt4ddMx3F+H1x/P3Drzz94OSt+HLf3w8vwYQDrznFJq4fnX29jz85cO7395bWbj67TJU"
"YDRiynMT7fLN5Vl4gcwFh0Xefv14/kGjxMBNwN/PPry/JtEE9lcomY7tVXj96z/PrWLD/nz824fzs9fIwHPf2uG/v6K3Bc4/nP9C"
"7Igx2xMs2gCvJOCVBLyyyglJ+gfmq7evr8Oz8MO735EVIYoGxygM4+X5JUC/ed9hXUe9UAg0NIUoQWbDySWnIZtKUELg/fmH8OVv"
"bwDbZKdZbMMmd0B+wzSjaAoc9aod5JroCvk2nbbvgZSL+kviMhSYK2elAp+Fr/7229XFNXUKxCC1ldvUnmXEzsKP7z6eveWYxTEw"
"aXa3ss/zeBiPnDCMsmyR9FdZHIauG4b38SCbLcJl8j/4oAPI+HE2cqEdD36ovaQdPHRqEzksITLNHg7D7Ft4IddIhOgT+/ehg8TC"
"cBjfJwNAAHqj2WIQJ9NJMoXvW7xBB4CHYTpKo3C5GHRcEqWzgrfwsgFYi3gJbAwyINGPlnGLuJ2NRp7zZQtk6zAEeAgaAQhzkPYC"
"WsnixXwRZ+EgWmbHOsypi4RAWZBKl4gIZu4B+8v8Zr/XcuY3Pv0b0L9t+Hdf/veVYQH91WLq3He3vm6tlsn01omW7XDF2plnCyDG"
"v2SmOKPhEDq2DJfzaBC7bRBeo8tpEAaKD9BLsOOHTAwHDpfbAfzuFshpNcic/mo0ihcot9kKBO58ISYOkCDw1M1ZWkTT21j5DlIa"
"Jbdd52v1uCFzThqNYxgwGDI+YvezZNjABlpGA2H/cxYvxWhBRDw6QgCABIhoGeJwWoZMBzx14RcfK7N/yAQO2zKLsmRgYGO3T13W"
"jNdSOYJhfPD9ffjRR7NRpj3Yb9CdHWzPwyGHUYgBYfvVtqPICwWxNZncp2GUDm8H03ARfQo5z6Trk+HSzccYqbVMvQEQ/N0ivqp+"
"UHhohcwq7kGhl3HGviz5l1o0VMRo9eAREijdMnW3sS+7rC+70Jdd1pdd7MsuMLrtVaoLisNRu+8fvAyzGZMCqRKTgN3iURBk9ZW9"
"kAp3O5n1owmNbyi6JF8SrfyN0MhcGREA7dKmjoomCp488CAGTaZI6wa/kQ9/wyN1yrvn6prguYIpUF7/oGXtIXdKpJMVI8ElzNwz"
"DsFo3gnjIPWZA3ceVImQxPpJtgRhACFCASvCV+4DN0QJBzJI0IhdQjg9dYJDz9kB+zpcGZCxBtVmUG/emGApgBEUvn7xBn9MkGj4"
"xwpi7jCEsEv55cHK+dmJTc7ukhEwns5W5GVcyMAB9MXKc/7LQSes0DiG3IGea0+P2NMjZ99kYBhPZ4s0RFbd/YcOOpL9lfNvx02x"
"bz5EQ/ilts/wU5MNSefISRkICQlgTqlXCBPDS/b5Z/ptsDKYpTD9iIec9LETeJKRwDdHawHMDPmIScwdxzepLsGbjj8DUErD4FuH"
"AWx1NdEoYeetLa4oNMvGd6CPvAngldNBNjzssb8imZu8o7qGWcqCPGE0JXUuvBzClZ+PHSZr+eAIHmhe3+XW4blMmf8tYblZccNj"
"/t351yqCySIkQmfhGGJAPHGVXESdLxr+7AxegBnLThU93hmaZTXAchBNuEek1Ja57YutQkYEqSrBYtJqprk6YDJ8QJObzAbjX4cP"
"uw+Q3NGX10kKX5pOdreIoyG9MlGzWQZy+TRbjDGlpikGNcoTq5HjIvHTEwXQ41I3SS1mnzDnAfA9nUoOsrx1OMizMpAxeWeAAthG"
"Waf1MWIZxA4bnRtkA7oB3WakeqZ3ih7CqI+ecZ+9+Wm+iG7TyFlNF7PJhB6B63VckiyBwa9jnJU4SbMphonAyPNi1qlmL/T01AWu"
"bpIeNwHyHbJh+OTyb62Cf773FL8JJsspfDUcKOCjkQiiTUAJmBcT+Ifsq0yRwSeDOITLpemDS2SYM8/h4sN0n3n6HB7EgCa4j34O"
"uFde/ewELHzlEvmX0NuQuoRdoxeuK/j33OB5Bz0+NOWRzxMMcAMRw8hNoAnq0EOehKEjYs4xxkKuNgzPRYsJPMKDrzjL22CoAbxk"
"qPfLxzqAdmC8CbrBRdA18f01+MCwDzSK+ORYbmQ/m9AUkComA/f7FDwsL3ySc0eq09e1eZ/iJnnK91RuMiK+lmmcVkOREBQ4SnBD"
"GOIQhLjgGe84zJIJz6EvWqpTzSCX19yqnOSzKf7v7z5crK+SaFrCPSVQ7rLPxwot/qh5ohaoVC3K3Ry4Z3CnzGUS0l5Jy0UHyn44"
"2rNaaDz8seYUAWIOLBix4UlnzGQsK1DCrgruWUcfTKJ0Dm6CtZsmUzdnpAXRBtIhz8Srcu0qPYuLr3bzJfa/gbvf3OU/3u3nrv9R"
"7n+TEPAEYeBpQgH+qD7gJjeaop1WhQhLmFBJaSVBI2JsqDWWyPEU0eMpIgiTpvC7TxhKchWtE1RoHv0yHEZZhEQX8a0WVuyRgMPP"
"J1EyzSNBP1TciXxKxcjxLcwr5vnDC+KeVY7kQx445PeX4eKGL2v0bjo9IyOXnKH3ZOVBlS2Qo0vF1cxT+SL1km1X1UGappMVZf51"
"icuY698YFZDzj990NSSB8boDOvQEsyRZQ6dUSZGcYQFGlZZ55bV1WiGopmxY0RgU9jjp3eyj8qH27veKL33x0re8DMTLwPKyLV62"
"e91NdZN5HJbyhKPVhE8PWb2rz1/nhZ+Sarc1eembyQtPcEvyFRa8cUCY/tDinjlJMydzoqq1JA0tLLg0VIImBb7ocVIgtSf1Qerd"
"4G41lbkQ+3LMCYjv5ZmQJLAM5/EiZEp3ovV1T1PBXBJstssa2CsQ6RabCJOpbIChPVuDNgVx8UDhKkOGYfI5WopkxJJ43DIDU+Kb"
"pKaJHsgY3KHNeRaKOAoForksNqNaUszV1FpXXupRS3BRb/7AUlzFmkaLWWpOITYzHBmxLYajTgHGScuMA9KoTCVSqNLY5jNKHHNF"
"bVTAHaftdw1y5dbJsnaeAUoPb8leyCMXFzo1t5yTnAY4BXA1NqEHHcz+FO6m/gGbE+dQ8F7p5/iQSi3EH+C3lTcBNpC/CgzS406O"
"uaMijiZRJk3ZpqBjXMQMnh/gR1yIPOjAJ+S04eCnMcowwEeBXuTDdFXVjhtsqdet69jPWLQmS1Y9ehRWrWFgQ2FUnH4ukiFXuIrp"
"qGUeymiMSRjFiZUa86sdrb5MXc/hqpNO4T+Ly9ma/5xNQtG8cJ3VGNpUc81Ec/M5YoWX1SeH6jCtXUNj6ZcclabS7XreU6iQqiwV"
"P9yhlk9BrO3XyWA+RUkW3qcDmFCKwY+WqXM/A0sBPXO3lyGCDHC5kaD2vW3n6OjI2QbOZ4vP294G5sQskbJfS4GoRv376co+lVlU"
"vsPlF9LZ3999eM32HBVB8J26XaTgqPdUKt2q0hLtPhKbjBo65WJVaahUlYZqVen1NX9W37YJfM/WohKRKCOhIELgz3RwnEEWOmq1"
"8bXlJAUD9HpNFOSM6dh8WdeYVYnKcE7b4ifm0WBMS2tlGwNW+XxlR9Sesb2eUQ3Jk7C1VQijBw3rcrRoUy9wsEZg/n7COd9w8iKM"
"sizSqUleVbwzV6kk5qNin24oha1ZrkWa/nNPmQjYLE1OWbQyri1EVQfUpwmnlcFUJOYqyjPn/0U0NYd/q1ZIrbJ0TRp1w6yqv0VN"
"3SjkbsBSbfPD3oYz9pxWT767FZO8Iias1CyHFay2uAzE2VWZ0pvWo7UgDE2CUwDS9ZMMEig4TuKBUhdb5m7xbvVE+QRiRGIPgZaG"
"hFQ80bxgmWyLExPRJ5VL2YhVrFKmCsFCFNBNRx8sg+31jahZ5jcO5XT21x3NP1f4T6oy2t6YMJxEq+ngLuzjnpxlGLpK8Gs5gcdG"
"Exu5jdOU750JZ/fxYhLNv8HVGRMFh21brKz710Dh0r2DGDKJhwxNZ8iC9iqcrbgiEFhDDx8a7BxkmYDEJIaiGldiHTov/swnYC1j"
"oS2TmFUd6bHUHL4ZNQS2owWE3VB0kaLcjbKHhbqtrOMoO+6LUEpqWYST6VCfrR/0IQlSoOCBvoIg6YoNOkAWa/s4U9CODqird5LP"
"fo8mG3pqrHHY7/H3+UEDcfTA3ALD2FCQc3aMUwwNrUtGtXu6SsN5MmS7AC8BWczWME2CZI5/LUObUs4q0a50tCsTDbsWIu6nW6wF"
"5o03cor2in6OpNFoCN3Ka2+fbpOhthUs38elk3qmHJHBRV05JJyCS79zKJwPGiT2lJdNjqA8K+z9Il5Z3ZXDqr0p9BzDBWeG81IF"
"/vsvl/mUXgwS4qlVXAQqCFutCdOcQLS7V6ShFDmTxTKTyiPxG9iGSZeWBVOej+fD/rNKpIWIyh4sQVltB4Ts5uw9K7LnwUO1SYMe"
"6esaAnsageKgDMXuk4o9hZNoGgsoakce5jIhP0ULEhuH3FsDyasbHEk5jmWBZaUGAfusDJaK68MkZewS58/06aa5mCDA9ixgPOry"
"HZtB176hk8X6sm2d8vwNxKRFfItnI/Z3Ye6i/fPV7ihYfF9y0qyoaDUsNo08UYyyYcYmK148HZKgUJO1Jlsq3aZJy7O50JxZ3Eoi"
"yf+sUMp3lKAH05BoJ4p9I6pakTpxhI8tceb9AvBUAl/ZKWO5M6FNLVIrG/xUYVMqVBE1n0Kz2bZOrGznmcqpVg3oa3U3qfMNfmyx"
"nBN1LetkPWJx2ZtGTB4gsidca3+UWoN7RYf+MMSo62fMTDzFifVXIzprlm/qKl2xr1jPrOZPW6E3knMjCS/O9tXktuRHSYnwnJqW"
"AcGDalRjylLyQ/PsKgDVwPI+mSaGxyAklW/vsNFp3+y036uBXq/jtTqvCwBLjKoQIAP/PB2w4LZ0PSNTnoxZqjzBgqEiK3xS3G4z"
"Fm5Ca24y1kt2xq6jri5yY6eUNiHSN0BpG5/y+Z+Y8kEzXv0tbfX2FBU2i+I2H7Fj1NxgpO9EYy7R7giNxTCVbtUmJfUk6Fm4oEq9"
"djhVnY1wT0LKqG9MUkve+tMvYumEJbH1F+xVTV4OxM5LffIjGLoxA0WTt9jrKhtB1b7CyBKqweyW5aRgyQFbfHPq7ghCXreAO2f7"
"tPJNW106Yas8Undo5ViBChIQVlt91LZidXpM++Y3z+WnA/npRU/f1Ktvkc03jJCk7ZtQNogHlgCh7EQxAoQwNtyDomiBIVCR14Vh"
"f5VMMlB8ftyQqSpjFUKjf/CA/wWH4ehwdDDquAVOxYi1UOFbjHDL6dB/+yzs0Qekat1QbHV3IhiAl2s6ge7pTPN/XFwwawSqVdqM"
"oybB+tGidsQoRA0mHDV06JKVucp/n4hdM5pcfVOu1qH4auRus1XG8zYjrRUpYL1kFMmwCZKR8NozSVdOeDBH6+Tpgpw9nZ6wNvWU"
"HEGUYpkWH+GV6Myxc2mqFdtJKypNYmOtOlvBmfyVsulWEGNPeQd1uysJelrgGzEvMwKmOvAL4l3Rj3LWUUZNgrzy1JrgDY9UI/RT"
"ZJI3o15hq7QTTwCqVCDOzo6TN9JmjegyUmuaTpbOteSB9DXlbluFdBlH+z2vSwB+CYBvnmNA4KAEOBDU2iUAbZOafVEcz/tjJKJq"
"7I11SHFtvDAkVYv7nCY7CyoHgYl/jRY+Wisq+iEUxCKsUa9kZ72o1Nvr8GJtxa1XKFeOqtaAVVZZv8+qvaUG/6O6/h9UXTcr5RXl"
"9e+8UP5kpezvt4r9mAL2j9r1f27t+kdl9+kru5GlpstyA+HpLsUhtgdRxSlu0TDx+X7cy0I590dRuUZRWd/xFBkbR/Q7h56mvlxn"
"0ixrikYPLqnGoBeSdZsuKSU/opuPrCrXLqWXd/lby8aaZCyFY1YggAl1oE8y9E1X/4cV5Mn4R/X4R/X4R/X4T68eo6H9xSrHmrPz"
"SirHhU2hG0eCJ68jb7TiWB4j7EXiOqVhXXB/xeJwafnyacqWT1OufJoyZe3yZPGqukU8XA1itotmLCp99/Ggwx69UIt9fOOqba+q"
"rZ5mqaRpVTTtyCoLWm6tO+ZUPRDTO3673JX9XjldM3jxf06KlVf/LfisMs5JmNcscXHM6SwVSQHvOlHP7LMrVaL9E3ZncOSfsGuD"
"o+CE3RwctU/mWpRZl/ng9NXvwq9j5wX8ooQHOGiesN50obEmtQaf/CY1B5+CJrUHn9pN1qDiaHU7ASs5+aJpbLTvtfQHvvkgMB+0"
"va/rzqfpOo4ZhLfmFJ3QZOGdzBUOKaQ/2B2AXfjABw8+ll2aQ/Oy3VHXXphn9ztp8o5I2mzAb+TKEegO06em80fPDKrYwz8KriBa"
"U4UvN9A/0Shb6p5yi4VucAnkN9mmZVDWm6ttuMTW5ZqDRinsVj5c+N02YF/FPdr8RMogGtwhd1/0rf18uIRfwHLxZIL3a6s9MsHC"
"eBKnyo1i+oKBU0olyi8AcqzY4o6sMnSWVog7JhR8fvMx2GEY8TuJlFsReZ6Fb9PSN+P8TX8GgZ2nPywhu48mVEkeRRNcWPiKt5Tj"
"1VOmdG/ZJ/maLCOeLlcLlo+y83zsKqkGv2ec9a/hDFbyizON46HUadDDBrxERcTnuCLawEuVC0vM7CmowptFHLNvMsW4jCZgBG7B"
"bdGF5w12MXmL0TdvxOcZALLAOGDaVejg6lD0Twzpd9rDyh6xA6qz2ZL7snDsWpbUVNYvcUEEYhJwfYUfA98P8MsFfnnhHxyKfoBr"
"Vu/x8JVWSeWWd7PVZJivX1a1u+kC0cYLUeKykbyVU0ptoWPKqgeVZzTp4eYX1p2Im8UdzGRsx6HE6Un19l/RP946t6dd3Q53dtS4"
"JWBU66+++pw17FXRSZHGZRXEGCEuil0ni0imxGmUxSGXgryGwt4l6VrKeiR8k8ml9Tn3ZVbWuLHyDBr/HsOSxxl2kBg/C17R0FbT"
"wXWcnYm/2+C6ZE9e2aJ7S83XEFliXkYPrz9PozQZXNPq8yVdtHFN1/7nLXc3bZifunt8u1ZppdHnPt/LxU7+kpSH5Wf8bEq9pXoL"
"1SoEMABeeNbcn4VMcoqOkVO4akFQgrMQSQgWcKXExhELvnsnV1C+9qE/EX92wlF4q0GLr4OYTwQ1lXFlTUK5j7uEf0re7tBAgucH"
"ymIGlZpddg0tAXAfBx+75mlupknHx8d9rDwBxOnpqRwbQxyFHtHYVdhtDUdUYtmXJZZ90a10JJD2YYrC9sYMk+U8ygZ3oTCbtZpr"
"26eydiNLnW0rNfeobLjSysBNB6boftVODWXbnPWCwOKCmaboug81XBjauhnQoc9XzNS/KLeh2z2M5huMW/k2iePflC3IHQ/oaGn2"
"tGYrCU0qSkICGJZCp+Voh6ml2NDmtOnoGsPb0qu46mJO9aIpn0yKsSiZpNsiuPUiapYj0jUmWtKo03/EjqBH7ArabHtB8ZibcnKN"
"jkxqBJqy77xp7RCucXXnej0yT/FKGxICPsFWjMqIPRf4RoVj18szq9tEoRSl4hNXqV18ddR2BtCqdWoZwZhr63t/x2Lnr17CKU41"
"ZcDVp/DrivklaIybls6csTbz3Q1KXQnIMZM3FZSPXEXdSHutZyYlev3Cs68ziiyGGR4WedHMOiUJjfqzrlpu5DpWZSgRm+o/DYFY"
"1dnoi+jKt/TgibiWA+yVrX7mVU6YN8T0x+2cOf4NIfEHUrtb7//x8ter174fXr57/dvbc/e+7Yft4AFrv8/9IJwF4XiJf8q05aRi"
"dNPd4WzgYvV4G6CdN+87zi/nl5cOoTmI5swCh6E57ksHlR1EB6K4TZaQQoLRvHRI55f4+C0azpn8xDz5CSJ7213RYDxytxer6XbL"
"uem5889HR7P+H/EAV/ZBGMp3sTnC7hc0QJ4NaeivyjArShhUX1tlM4flyTp/eEZCFAGyXcyw3G1m/Nli23O9Xa0+fwqx2VEq+SVZ"
"b65b1tq9lq5SvQiSsMzzWnWQRSrM8IQ4vxGZi3gttoXjV0U0I9Ohqe7/ArT9mHM4eAAA"
)
# --- Kernel: v24_64x16_t256_o1_ks512 (from custom_kernels_larger_k/v24_64x16_t256_o1_ks512.cu) ---
_B64_V24_64X16_256_3072_1536 = (
"H4sIAP1a1GkC/+09a1PbyLLf+RVz2QolPwLIOITCwC1ISM4WC0nlUXvOobhTki2D4ueRZEJulv9+u3tmpJnRyDYO2ZM9N9RusKXu"
"np5+TU/Pg1/icXc460Xs4CaebsH/PJmNs3gUbd4crf1Sehn2h5Mg83fNt9MvYTzu+f6W+mC+7qZZLx5n+qNgeD1J4uxmdLTWnYzT"
"LLqbJgxg2Mlvb16c8XN2yHbbHee7C3jn79rvzvj7D6dvGbx75rfcL/nJPz6cvgcICbvFSoDvXxz/dspfv3vz8S3A7ZTeX3w85xqM"
"QUx7bqOdvzo/5mfId2uvzNuvH07fGZQEuA34+/G7t+9JMm33KxKM6xV//+s/T50Sxf58+Nu70+OXyEDr2a4L4O8v6G2J83enr4kd"
"pbItxaIL8CIHvMgBL0riPz8958f81du2YJlp1OuGFiswhQ5kdwtMW21O7BNXuxfLtHviavdi2Xbfnr7jJx9fAaLd/Ua5W2us+qdR"
"6kejzJ9LwdD6K2GCLSeHH958OP5N8afYreuoJU/lL/728eLsPUG/e/N7bt3SB7ccHnwi2pGYlg5sgjby8QoNHlc1iCZzXGow+zKN"
"elGfcR5kWRKHsyzi3PM4v4262SThafy/+KANyPhx0vegkRr8UGPxTuuuvTSRvQoi4+xuj2er8EJxmwjRJ/HvXRuJcd6LbuMuIAC9"
"/iTpRvF4GI/h+5pskAFwj4/6o4CnSbftkRzZDN7CyzpgJVEKbHQzIBEGadQkbif9fo19JYsVCPBwp4UIUxB1Aq1kUTJNoox3gzQ7"
"MGGOPCQEFoxUOkREMXML2F+nl9tXTTa99OnfFv27A/9u5//dCyygP0vG7Lazdr82S+PxNQvSHT4T7UyzBIjJL5ktzqDXg46lPJ0G"
"3cjbAeHVO5IGYaD4AL0CO7rLlDpQXV4b8DtrIKdZN2PhrN+PEpTbZAYCZ1+JiV0kCDx1CpaSYHwdad9BSv34usPu5+sNmWOjYBCB"
"wkBlUmO3k7hXxwaaVgM8/JJFqdIWDNf7+wgAkAARpBzV6VCZCXjkwS+pK7t/yASqLc2CLO5a2NjtI080U2vqHIEa73x/G35Mbdar"
"rAf7Dbazge3VUOWghQgQ1l+sM01eKIi14fB2xINR77o75knwmUueydaHvdQrdIzUmrbdAAj+bs4LyUzaLXmh8IpbMOg0ysSXVH5Z"
"ioaOGMzuaoQERpeOvHXsy6boyyb0ZVP0ZRP7sgmMrtfmmguKg+nd93dPeDYRUiBTEhJwezwKgrx+bi9yg7seTsJgSPrlqkv5S6JV"
"vFEWWRgjAqBfusxRs0TFUw0iiEVTGNIi5dcL9ddrZE5F9zzTEmqeYgqM199tOnsogxLZ5BxNSAmL8Iwq6E/bPGqNfBHA2Z0uEZJY"
"GGcpCAMIEQp4Eb7y7qQj5nAggxid2COEoyPW2quxDfCvvZkFGRlQOwLq1SsbbARgBIWvn7/CHxsk6H2awYDb4zDmUva7O2NPWWRz"
"dhP3gfHRZEZRxoPpAYA+n9XYfzMMwhqNA7bj03Pj6b54us+2bQZ60XiSjDiy6m3ftTGQbM/YH8wbYd98GA3hl96+wB/ZbOR09tlI"
"gJCQAOaIeoUwEbwUn5/Sb4uV7mQEc6OoJ0kfsFYtZ6Tl29pKgJme1FiOucF8m2oK0XTwBYBGpAbfqQbw1dnQoISdd7Y4o6E5b3wD"
"+iibAF4lHWSjhj32ZyRzm3c0V56NxCBPGI2cuhReAeHlnw+YkHX+YB8eGFHfk95R84Qx/5HDSreSjifiO/vXLICZLCRCx3wAY0A0"
"9LRcRJ/MWvHsGF6AG+edKke8Y3TL+QBpNxjKiEgzQBG2z9ZKGRHkqQSLGas9izQB494dutxw0h382rvbvIPkjr68jEfwpcGymyQK"
"evTKRs0mGcjl8yQZAAUxl6JGZWLVZx4SPzrUAGtS6japZPIZcx4A3zKpFCDpNZMgT6pABhSdAQpg61WdNnUkMogNoZ1LZAO6Ad0W"
"pK7s6BTc8SDEyLgt3vwyTYLrUcBm42QyHNIjCL3MI8kSGPw6wDk/ixsNpSYCo8iLWaeevdDTIw+4uoyvpAtQ7Mgbhk+e/NYsxefb"
"mhY3wWUlhXsrgAI+Ooki2gCUlohiCn9PfM1TZIjJIA4Vcmn64BEZEcwLuGhvtC0ifQEPYkAX3MY4B9xrr56ylhi+Con8S9ktpy5h"
"1+iF5yn+a17rWRsjPjRVo5inGJAOotQoXaAB5nCFPClHR8SCYxwLpdkIPA89plUjPPiKU7wHqBrAK1S9Xa3rFrQD+ibouhRBx8b3"
"F+ADwz7QKONTYLnM+9mApoBUORm43abBw/HCJzm3c3O6X5j3aWFSpnyPFSYD4isdRaP5UCQEDY4SXA4q5iDERGa8A57FQ5lDn4lf"
"GSTx5XgKIRFCmAhTAAEWUi7DlOOV+BEYTxZiyIFGNKKxitmmar4y5Ime5IUlZb0rBEGNjZ+x8C8bC3UHuCyst1w+fEB81MkYhbD/"
"mFBZBJc/OV7SFPGE94IsEE4v4qWYJ4d83kwZueVhEeRCGTmSuCcj25ygp0c7gTygKWY5oqhStaHN7s1sPBABriO/HFg1V/W8cagv"
"DOhKt+OrgN9ylGd1jO5kyFXz4veTORhFoQDmzsrYhQd5zkCL1qFLcmExRRTLc/k1NAbRM7QAVlETUVrW9TnnBzvSnOeSzvaXtsXj"
"ClsMFttiUB5wV7LF4BFs8fiRbPH4wbZ4/Li2GDzcFoPVbTHQbTH4N9ri5yDO+O2oCwO6UlOQjmAqPgzQIrz1lCNIF2ugBLVdW2f7"
"+/tsHTifJF/WF9XG9OgrRsxyvuq2dQWf3oAMh1FvXjoaOtPRcG46Kmf77qS0WEt7TQb2+5t3L8UibhkE3+lrWKUMYEunUoz85ZzX"
"oFQAptdipMRU1xO5rgGJ9ltuoc+8ooWjfF3OLBFQiTqJryUb4bysmFghMaO52XmxM+8hqAJ7vNMClLw9SOieaS9b1D/9bRtzRF8D"
"8Xd1fHin4WNTqsqMDcFMk/6VirYLXtOgO6DC3sPS63YppxJiQZlgVmTKB3IbM5IN9hhBETj0cMd6LWRQvG9pEsiB2hqNDZtEHzwX"
"9YKrka1nuzi/wBXF3TZ8QvnVGX4aoGRa+KhVoEuR/KGHTNsLLwsxN6itK0zFaO0zNqYAZOG9lIsYvDBD1iyl7lwjkAuMG6GReIsG"
"INc8lNx3lh+ABZ3pMIjHSwclo0b4eFPjf2ssMimI8ZzoGGvrrtjUF+HoQEeaN9pXRrlvinQKefn5vYUhXHfpSJarVMQacz6fF8UK"
"2o7MJA8+VWuiusWrshu2p0/UV/ewxV4WLPKyBya7glpVyiveLk587QJ9jrliEpzbrzRz3XaXzSKBgMgfbVbWlkoiK61umSRSl1tZ"
"QgsTSux7ZcNLKxf55xPxnMqSK5QijXn4o2aGq5Yzq/LMZaiFy40AenqaQ52pEQEaumsZwwOndZ4CtMqsy7UONfXV2C9aMLJkRROt"
"WthBudIsFaeL1hSg2VVF8kynqs87OjrXxhyhpGhTtG7WTWnJNh9iyxARqdM083HNzR1hy1FN+g4W/HjW+z0st9pk/0RzLRdXlNYC"
"F7VgKWql0csMvabxW11fpoXvZvZlV1vJEcaTn77w5/nCX9zeflR/NvaacD4MZuPuDQ9xj0vKuacVLJvMrwlfwBauo9FI7kXhk9so"
"GQbTFTIcywOY2AZouAH1kaajS6O4PMdkyIH2gk9m0lYJrG7mhgbsFAQZg8RyDG1fzIVa1y2qQ9MhxJqBsuthlPIpJK70OLdxubmT"
"A9tBEvWgEbV/jiYb2joodVtbMtL2rpehtPnKHLhwKWrhfGp5XSgUdaEQZhD6vnoWmhWigq7cPgNkcS0DJ7Pa7nznMll4RZNhc+5m"
"9Da8ku+LowfqLIE2hXBSLGO4ji/oVJZr10WlfNJBrxYVK+7j2YhP457Y/ncOWKoE8RS8EletxNdOBdoY0S5ytAsT7cJGo2oW4n6+"
"xhXvovF6QdG9L6tAMmjUlRMUhcnP13HP2AOmVU8MUk+0kzu4kJ3LXVLw6HcBhfNOi8SW9rIhEbRnpU0UxKvYoyZh9d6Ueo5ZgWRG"
"8jIP/PfX50WdSikJ8Xg85tfJhPYtIlBJ2FScQQBObRXtbpVpaEvpcZJmufHk+HVsw6ZLhy0QbBSPvULtT3UiTUTUNhwoyno7IGSv"
"YO9Jmb0aPNSbtOiNRaVrLoEtg0BZKb18H0z1ZsJhMI6M3TL5GTMb8nOQkNiKet1cSFFuU0jaKTEHrCi0KdgnVbB0ZqUXjwS7xPmT"
"fPODBTZQdkRgWw4wmRvIrZqtjnsnp8hjqvZz5gdvYPBMoms8FLG92W8y4597d6AQWUgqSYuFO6djiYrSoeaUdXsQdeJF4x4JCi3Z"
"aLKp023YtGquEFowi9tncvJPNUrFLhqMYAYS7b5x70DV67GHTMXYimAeloDHOfCFmzLW8GPazJNbZV0edmzkBuVux0JFI63L44/V"
"qHq9TxQITT7qFdvR9HkRSTk/7WMkgEycYSTL1SJGOOvTia5i2aqiBKdqNCr1dqeKFT9a+oHntYxsAx7MQw1N1PABqEbqXppVVfyc"
"abm/o0xW8aMbczHVsM0ZzxrkVL6HnC1Z+7as/auF6KGJHj4QfSWZry53U/a4S0yXP8wIvoy7YgxLVVkwz7WHA5FsD3HbiaYmfGKm"
"246VXIPSQC7nDnA9V57kxm8mldI+UDwYpBa9Y8y85AFx3MzcWbPPfQUVUcHauKHTbRgjWrHdTyNLRymPeULrPcbpTn3KIIMEmVNx"
"EtPkLu2aT7+q9TeRDFo91OJYzeSzs2ZbVNpV2zbNaYri6tIOnQ3Z7FVHTgZ0TkMpx/Dx5XjikmPolGNYIcfwu8kx1OQYuuUYLhhw"
"qgWrkhjOw1k8zICEPFQn5ECN9Xda3N+9w/9ae7y/19/tt72Sc4MMm2iQTUGzydr037YIRfQBCWqiKzhx+nu+dWUA7LdMV7fdk5KW"
"6C5TM4R8tZdwTYl+r/htT9d133NpbSmCVUa4MsGV4/y3xXr8QQXpUd60gTyh+Z9DY5uLsgDftgCn0dg1hMksk/MSK9EEig9ID5GM"
"mLJYKagiY6aKXj4FwYl5u0gqtCKasTsCXilOD9i5bd0028tUzcgTX2v65IBuycCG5TtFTDyV3JtuUDE2GuNjXwyPfbHbqQ/DYknP"
"inUUQIMgL2p6rfBSBuI+1ogoMlz2r2z937NoCFCVAmEbG6xoZEc0YspIr3WybDS9bF+Z/cVn28iEDukJjravah0C8CsA/KtamVqr"
"ArilqO1UAOzY1Nw7MPBcPe7AoCrtpVOluBGjpJJ5O0kkTXHmMleCEP8CK/xmq5jTD2UgDmH1r0rh4t5xWtSuz6sVK2+5Arp2JHQJ"
"WG1rxY9ZzXfU5n9W3X9W3X+8qrtdQZ9Tdv/BC+iPVuL+cavb31LY/lnT/v9b0/5Z8e2U9suYFV+RfqgYdV6u+Jb21tj4ctu4djmE"
"0HntexebzT1ygbUxRW33eOCcesW684o15wfPiSsmwVjjMIvIpo9UlJG/iwS/saL8DdXk1ar3FRJdtSxsCL5cGJZVhaND1jInL+Xd"
"qD/ryD/ryD/ryP9JdWQjNtQqasOlHa+PGJcfvVL8qFXi1VcCK2K4ELxd+V2m3mtq6i9X8a2sST5OLfJxapCPU3tcuuZYrtwlUW/W"
"jcROlIEq391G3bZ49Fyv4Mldqq6Nqa4imaM8ZpTGjEsQRHj1lrqgTS/rq6mQvJrtwn0pm7kbRlb0p2Q1RUnfgS/K3ZJEkUsY4pjS"
"aTySAp4z7t3Z17QE24fiwt3APxR37gatQ3HtbrBziEayfK6T0jXl8OuAPYdflOIAB41D0ZsONNag1uCT36Dm4FOrQe3Bp52GaFAL"
"26afgJccfjUsNtiuNc0Hvv2gZT/Yqd0vOuFo2jiOarUF5zCVJavQYy9b5EL6JC7Q68AHqTz4WM4GpXYwedzsd9zVdnHvkyHvgKQt"
"FH6ZLweB7Qh7arBP1tqH6OGnUigIliqtuxz0OzplU99A7vDQB9yguJJvOpSy2F1d6lLbf5dUGqVVa4W68LtLYffqEmp5eKcbdG+Q"
"u6/mPn6pLhUXsLQ6HOLl1HqPbDAeDaORdtOYuQrAKqmIbIbufGZObHW/VhW6SBgEAQNfXhsMfsgDed+3dmWWzNrw7ajyzaB4E05g"
"1JZZjkjvboMhVV37wVBV+AugPAnEm8F5GmVVgPI0SAFXAN7jpeF4QZatr2vxKX9NvhaN01kiUmVx1ldcelWX134LidVZd5Z/YeMo"
"6unnmOvwEk0bn+PCaR3vODav+kAoegrG9SqJIvEtT1rOgyG4lVcKhHT/eF3cE94U9O0L6mVOgSwIDoS9ljo421P9U0byg/Zwbo/E"
"ZUiTSSqjIx94jpU3nfVzXI6AUQ64vsCPLd9v4Zcz/PLc391T/YBgr9+d62utktmlN5PZsFcsc85r96HLMw9eBlIX/BatHFEmDB3T"
"1hyo2mJID7ffiO4E0i1uYKblOk2ldufol/Gq/snWpT9tmp69saGPhApGjyfzbyIXDdfm0RkhjfN5EAOEOCt3nTwiHhOnQRZxKYX8"
"AiZ3l4wY5OqRinY2l87nMjo6WZPOKnNyDG6pHLnEJQP4Wbfu/1KUS4FTz3zQJWfj7vsoO1Z/cMHzyPNqVav4TXuvBxLIsc+Du5df"
"xsEo7r6nJe1zupDqPd3ZX/BZpDqVXOKiUDKLjMtr9G7ZYX7FXkky36dTpaHI7JNTz6PgSyj3wYnz/USqV3240eWOxmUauj8rYAA8"
"qznnQSJ9oHDOrPzK08uhObhIFwjBAa7XxOSN5faos1G4llx5MZ+ov1/BNN6WoCVXYewniprOuLa8ol3sXcE/JbI3ifojT3mxkGre"
"HuFT+quiM3ws3dkgjO/g4ACwmghxdHSU68YSR6lHpLs5EWeJEFoRk84rYtJZZ24IFDZ9Lw7n9uJ0GmTdG648baHlujbiLNyps8y+"
"nCU34QjKdixVf6SJLlEyrM0MwZb7o8PZ+QA0fCH87at2n6/bzQ0HNW8VelAasFKykW9XwABJ07kF+0BollMxToB1a3SazDjKnYsN"
"Dd+IvQusf82sT+vrpfPrvnJ2q3RRUTVwJQCdtfKKkUwx6eZLI+c06X/Ddp5v2NLzsL0B5bNr2nE0OgdpEGjkfZdNGydrrbu9F9uR"
"fTQ39yEl4ENsxSrVuMfwFQ1OXBYvvO4hBqUZlZxJ59Yl10pdB/ucVqfXNazJv7nDeKD2F5s1pfJMNR/1zJrConWDCjTBTdNkzlow"
"++GUsqwEcp3l9yRUa25OIct4baYHFXb93LZrK5UQjodVZ3SzdkVWof8sKt9bCYfTGCrEpsdPSyBOc7b6orqySg8eietcwTb7jrIr"
"5PoR/ak6NsW/CKT+Fmtn7e0/Tn69eOn7/PzNy4+/nXq3rTbfbd/5uzwDdfOJzwfpM7/VZCOl3NFmb9L1sJq9DsDs1ds2e316fs4I"
"iyEWm/iMsJg3mg2z+Cmt+IJBqpG0IX2ptt5RJKO+t57MxutNdnnlTb/s70/CT1EXdw9Ab7Xv0ncqFgwNQLlOaKC/qMKcU+Kgit4s"
"mzCRjZr84TRNFQmyTUyhvHXh3VmyXvNqm8aKwBEMvkxbO6jILQvjca4WGEkh1ZMgy8pqteYyyCrhFHhKnCsiSxEvxHZw/KKMZqUy"
"VM75P++T2+aEeAAA"
)
# --- Kernel: v31_8x32_t128_o4_ks1024 (from custom_kernels_larger_k/v31_8x32_t128_o4_ks1024.cu) ---
_B64_V31_8X32_16_2112_7168 = (
"H4sIAP1a1GkC/+09a1PbyLLf+RVTbIWSbC/gRwiFgVskIXu2CCQVsrXnHMpXJdsyaG3ZPrZMyM3Jf7/dPQ/NjEayHNhbuXtC7QZb"
"6u7p6enX9Dz4KZ4OJqthxI7v4vke/B8sVtM0TqLdu9Otn3Iv+6PJLEybB+bb+ed+PB02m3vyg/l6sEyH8TTVH4WT29kiTu+S063B"
"bLpMo4f5ggEMe/n23auL4JKdsMOu89UVvGq37HcXwfXH8/cM3jX3Wx332+DlPz6eXwOIAN5jOTLXr87enge/fHj323tnM1e/XQYa"
"jEFMe26jXb65PAsukLnWYZ63Xz+efzAocXAb8PezD++vSTJN9yuUTMv1Krj+9Z/n8PKg4+rPx799OD97fe3mDgH+/ore5t59OP+F"
"2JFDtidZdAFeKcArBXjllBP18MB+9fb1dXAWfHj3O7IiRVETGLlhvDy/BOg37zu866gXGoGaoRAFyHw4heQMZFsJCgi8P/8QvPzt"
"DWDb7NTzbbjkDshvuGbkRvWlQL1qtzJN9KR866zd9EHKef0lcVkKLJSzVIHPgld/++3q4po6BWJQ2ipsas8xYmfBx3cfz94KzPwY"
"2DS7W+nneTSMRiwIwjRdxP1VGgWB5wXBfTRIZ4tgGf8PPugAMn6cjTxox4cfai9utx46lYkcFhCZpg+HQfotvJBnJEL0if/70EFi"
"QTCM7uMBIAC90WwxiOLpJJ7C9y3RIAPgYZCMkjBYLgYdj0TJVvAWXtYAaxEtgY1BCiT64TJqELez0chnX7ZAtowjwEPQCECYg7QX"
"0EoaLeaLKA0G4TI9NmFOPSQEyoJUukREMnMP2F/mN/u9BpvfNOnfFv3bhn/31X9fORbQXy2m7L679XVrtYyntyxctoMVb2eeLoCY"
"+JLa4gyHQ+jYMljOw0HktUF4ta6gQRgoPkAvwI4eUjkcOFxeB/C7WyCn1SBl/dVoFC1QbrMVCJx9ISYOkCDw1M1YWoTT20j7DlIa"
"xbdd9rV83JA5loTjCAYMhkyM2P0sHtawgYbVQND/nEZLOVoQEI+OEAAgASJcBjicjiEzAU89+CXGyu4fMoHDtkzDNB5Y2NjtU483"
"4zd0jmAYH5rNffgxR7NWpD3Yb9CdHWzPxyGHUYgAYfvVNtPkhYLYmkzukyBMhreDabAIPwWCZ9L1yXDpZWOM1Bq23gAI/m4QX2U/"
"KDy0Qm4V96DQyyjlX5biSyUaOmK4evAJCZRumXjb2Jdd3pdd6Msu78su9mUXGN32S9UFxcH07jcPXgbpjEuBVIlLwG3xKAiy+tJe"
"KIW7ncz64YTGN5BdUi+JVvZGamSmjAiAdulSR00TJU8+eBCLJlekdYNfy4a/5pM6Zd3zTE3wPckUKG/zoOHsoXBKpJMlIyEkzN0z"
"DsFo3gmiVtLkDpw96BIhifXjdAnCAEKEAlaEr7wHYYgKDmQQoxF7hHB6ylqHPtsB+zpcWZCRAdXmUG/e2GAJgBEUvn7xBn9skHD4"
"xwpi7jCAsEsZ3MGK/cwim7O7eASMJ7MVeRkPEnAAfbHy2X8xdMIajWPIHei58fSIPz1i+zYDw2g6WyQBsurtP3TQkeyv2L+Zl2Df"
"mhAN4ZfePsdPbDYUnSOWcBASEsCcUq8QJoKX/PPP9NtiZTBLYPYRDQXpY9byFSOtpj1aC2BmKEZMYe6wpk11Cd50/BmAEhqGpnMY"
"wFZXE4MSdt7Z4opCs2p8B/oomgBeBR1kw8ceN1ckc5t3VNcgTXiQJ4y6oi6El0F46vMx47JWD47ggeH1PWEdvseV+d8KVpiVMDzu"
"39m/ViHMFSEROgvGEAOiiaflIvp00fJnZ/ACzFh1Ku/xztAsywGWg3AiPCKlttxtX2zlMiJIVQkWk1Y7zTUB4+EDmtxkNhj/OnzY"
"fYDkjr68jhP4Umfp3SIKh/TKRk1nKcjl02wxxpSaphjUqEisRsxD4qcnGqAvpG6TWsw+Yc4D4HsmlQxkecsEyLMikDF5Z4AC2FpR"
"p80x4hnEDh+dG2QDugHd5qR6tncKH4Kwj55xn7/5ab4Ib5OQraaL2WRCj8D1Mo8kS2Dw6xhnJSyu1+UwERh5Xsw69eyFnp56wNVN"
"3BMmQL5DNQyfPPGtkfPP977mN8FkBYWvlgMFfDQSSbQOKC3uxST+If+qUmTwySAO6XJp+uARGe7MM7joMNnnnj6DBzGgCe6jnwPu"
"tVc/sxYPX5lE/iX1NqAuYdfohedJ/n2v9byDHh+a8snnSQaEgchhFCZQB3XoIU/S0BEx4xhjoVAbjuehxbR8woOvOMvbYKgBvGCo"
"94vHugXtwHgTdE2IoGvjN9fgA8NNoJHHJ8dyo/pZh6aAVD4ZuN+n4OF40SQ5d5Q6fV2b92luUqR8T+UmQ+JrmURJORQJQYOjBDeA"
"IQ5AiAuR8Y6DNJ6IHPqioTvVFHJ5w62qST6f4v/+7sPF+iqJoSXCUwLlLv98rNESj+oneoFK16LMzYF7BnfKXSYh7RW0nHeg/Eeg"
"PauEJsIfb04TIObAkhEXnnLGXMaqAiXtKueeTfTBJEzm4CZ4u0k89TJGGhBtIB3ybbwy167Tc7j4cjdfYP8buPvNXf7j3X7m+h/l"
"/jcJAU8QBp4mFOCP7gNuMqPJ22lZiHCECZ2UURK0IsaGWuOIHE8RPZ4ignBpSr/7hKEkU9EqQYXm0S+DYZiGSHQR3RphxR0JBPx8"
"EsbTLBL0A82dqKdUjBzfwrxinj28IO555Ug9FIFDfX8ZLG7EskbvptOzMnLFGXpPXh7U2QI5elRcTX2dL1Iv1XZZHaRuO1lZ5l+X"
"uIyF/o1RAQX/+M1UQxKYqDugQ48xS1I1dEqVNMlZFmBVablXXlunlYKqq4Y1jUFhj+PezT4qH2rvfi//silfNh0vW/Jly/GyLV+2"
"e91NdZN7HJ7yBKPVREwPeb2rL15nhZ+CarczeenbyYtIcAvyFR68cUC4/tDinj1Jsydzsqq1JA3NLbjUdII2BbHocZIjtaf0Qend"
"4G41VbkQ/3IsCMjvxZmQIrAM5tEi4Ep3YvR1z1DBTBJ8tssb2MsR6eabCOKpaoCjPVuDNgVxiUDhaUOGYfI5WopixJF43HID0+Kb"
"omaIHshY3KHN+Q6KOAo5opksNqNaUMw11NpUXupRQ3JRbf7AU1zNmkaLWWJPITYzHBWxHYajTwHGccOOA8qobCXSqNLYZjNKHHNN"
"bXTAHdZudi1yxdbJs3aRASoP78heyCPnFzoNt5yRnLZwCuAZbEIPOpj9adxNmwd8TpxBwXutn+NDKrUQf4Df1t60sIHsVcsiPe5k"
"mDs64mgSpsqUXQo6xkXM1vMD/IgLkQcd+ISc1hh+GqMMW/ioZRb5MF3VteMGW+p1qzr2Mx6tyZJ1jx4GZWsY2FAQ5qefi3goFK5k"
"OuqYh3IaYxJGfmKlx/xyR2suU1dzuPqkU/rP/HK24T9nk0A2L11nOYYx1Vwz0dx8jljiZc3JoT5Ma9fQePqlRqWudbua95QqpCtL"
"yY9wqMVTEGf7VTKYT2GcBvfJACaUcvDDZcLuZ2ApoGfe9jJAkAEuNxLUvr/Njo6O2DZwPlt83vY3MCduiZT9OgpEFerfT1f2Kc2i"
"sh0uv5DO/v7uw2vQnU7XAYLv9O0iOUe9p1PplpWWaPeR3GRUMynnq0pDrao01KtKr6/Fs+q2TeB7rha1iEQZCQURAn9mguMMMtdR"
"p42vLSdpGKDXa6KgYMzEFsu61qxKVoYz2g4/MQ8HY1paK9oYsMrmKzuy9ozt9axqSJaEra1CWD2oOZejZZtmgYM3AvP3E8H5hpMX"
"aZRFkU5P8srinb1KpTAfFftMQ8ltzfIc0mw+97WJgMvS1JTFKOO6QlR5QH2acFoaTGVirqM8Y/8voqk9/FuVQmqZpRvSqBpmdf3N"
"a+pGIXcDliqbH/Y2mPHntHry3a2YZBUxaaV2OSxntfllIMGuzpTZtBmtJWFoEpwCkK6eZJBAwXESD5S6uDJ3h3erJsonECMSe2gZ"
"aUhAxRPDCxbJNj8xkX3SuVSNOMWqZKoRzEUB03TMwbLYXt+InmV+41BOZ3/d0fxzhf+kKmPsjQmCSbiaDu6CPu7JWQaBpwW/Buv4"
"fDSxkdsoScTemWB2Hy0m4fwbXJ01UWB822Jp3b8CipDuHcSQSTTkaCZDDrRXwWwlFIHAamb4MGDnIMsYJKYwNNW4kuvQWfFnPgFr"
"GUttmUS86kiPleaIzagBsB0uIOwGsosU5W60PSzUbW0dR9txn4fSUss8nEqH+nz9oA9JkAYFD8wVBEVXbtABsljbx5mCcXRAX71T"
"fPZ7NNkwU2ODw35PvM8OGsijB/YWGM6GhpyxY51iqBldsqrd01USzOMh3wV4CchytoZpEiRz4msR2pRyVoV2ZaJd2WjYtQBxP91i"
"LTBrvJZRdFf0MySDRk3qVlZ7+3QbD42tYNk+LpPUM+2IDC7qqiERFDz6nUHhfNAisae9rAsE7Vlu7xfxyuuuAlbvTa7nGC4EM4KX"
"MvDff7nMpvRykBBPr+IiUE7Yek2Y5gSy3b08Da3IGS+WqVIehV/DNmy6tCyYiHw8G/afdSINRNT2YEnKejsgZC9j71mePR8e6k1a"
"9Ehf1xDYMwjkB2Uod5+U7CmchNNIQlE76jCXDfkpXJDYBOTeGkhR3RBI2nEsBywvNUjYZ0WwVFwfxglnlzh/Zk437cUECbbnABNR"
"V+zYbHXdGzp5rC/a1qnO30BMWkS3eDZifxfmLsY/X92Ogsf3pSDNi4pOw+LTyBPNKGt2bHLiRdMhCQo12WiyodOt27R8lwvNmMWt"
"JIr8zxqlbEcJejADiXaiuDei6hWpEyZ9bIEz7+eApwr4yk0Zy50xbWpRWlkTpwrrSqHyqNkUms+2TWJFO890To1qQN+ouymdr4lj"
"i8Wc6GtZJ+sR88veNGLqAJE74Vr7o9UavCs69IchRl8/42bia06svxrRWbNsU1fhin3JemY5f8YKvZWcW0l4fravJ7cFP1pKhOfU"
"jAwIHpSjWlOWgh+aZ5cB6AaW9ck2MTwGoah8e4etTjftTjd7FdCrdbxS500BYIlRFwJk4J+nAx7clp5vZcqTMU+VJ1gw1GSFT/Lb"
"bcbSTRjNTcZmyc7addQ1RW7tlDImROYGKGPjUzb/k1M+aMavvqWt2p6i3GZR3OYjd4zaG4zMnWjcJbodobUYptMt26SknwQ9CxZU"
"qTcOp+qzEeFJSBnNjUl6ydt8+kUunfAktvqCva7Jy4HceWlOfiRDN3agqIsWe11tI6jeVxhZQrWY3XKcFCw4YItvTr0dScjv5nDn"
"fJ9WtmmrSydstUf6Dq0Mq6WDtAirrT9qO7E6Pa5985vn6tOB+vSiZ27qNbfIZhtGSNLuTSgbxANHgNB2olgBQhob7kHRtMASqMzr"
"gqC/iicpKL44bshVlbMKobF58ID/tQ6D0eHoYNTxcpzKEWugwjc44Qbr0H/7POzRB6Tq3FDsdHcyGICXq7OW6els839cXLBrBLpV"
"uoyjIsHq0aJyxMhFDS4cPXSYklW5yn+fyF0zhlybtlydQ/HVyt1mq1TkbVZaK1PAaskokuETJCvhdWeSnprwYI7WydIFNXs6PeFt"
"mik5gmjFMiM+wivZmWN2aasV30krK01yY60+W8GZ/JW26VYS409FB027Kwh6RuAbcS8zAqY68AviXd6PCtZRRnWCvPL1muCNiFQj"
"9FNkkjejXm6rNIsmAFUoELazw7JG2rwRU0Z6TZOlydxIHkhfE+G2dUiPc7Tf87sE0CwAaNrnGBC4VQDcktTaBQBtm5p7URzP+2Mk"
"omrsjXNIcW08NyRli/uCJj8LqgaBi3+NFj5aK0r6IRXEIaxRr2BnvazUu+vwcm3Fq1Yo146qVoDVVlm/z6q9owb/o7r+H1Rdtyvl"
"JeX177xQ/mSl7O+3iv2YAvaP2vV/bu36R2X36Su7oaOmy3MD6eku5SG2B1nFyW/RsPHFftzLXDn3R1G5QlHZ3PEUWhtHzDuHnqa+"
"XGXSrGqKVg8uqcZgFpJNmy4oJT+im4+sKlcupRd3+VvLxoZkHIVjXiCACXXLnGSYm67+DyvIk/GP6vGP6vGP6vGfXj1GQ/uLVY4N"
"Z+cXVI5zm0I3jgRPXkfeaMWxOEa4i8RVSsOm4P6KxeHC8uXTlC2fplz5NGXKyuXJ/FV1i2i4GkR8F81YVvruo0GHP3qhF/vExlXX"
"XlVXPc1RSTOqaMaRVR60vEp3zOl6IKd34na5K/e9cqZmiOL/nBQrq/478HllXJCwr1kS4pjTWSqSAt51op/Z51eqhPsn/M7gsHnC"
"rw0OWyf85uCwfTI3osy6zGdJ15zDr2P2An5RwgMc1E94b7rQWJ1ag0/NOjUHn1p1ag8+teu8Qc3RmnYCVnLyxdDYcN9vmA+a9oOW"
"/aDtf113Ps3Uccwg/DWn6KQmS+9kr3AoIf3B7wDswgcxePCx6NIcmpftjrruwjy/38mQd0jS5gN+o1aOQHe4PtXZHz07qGIP/8i5"
"gnBNFb7YQP9Eo2zoe8odFrrBJZDfZJuOQVlvrq7hkluXKw4apbBb2XDhd9eAfZX3aIsTKYNwcIfcfTG39ovhkn4By8WTCd6vrffI"
"BguiSZRoN4qZCwaskEqYXQDEnNjyjqwidJ5WyDsmNHxx8zHYYRCKO4m0WxFFnoVvk8I34+xNfwaBXaQ/PCG7DydUSR6FE1xY+Iq3"
"lOPVU7Z0b/kn9ZosI5ouVwuej/LzfPwqqZq4Z5z3r8YGK/WFTaNoqHQa9LAGL1ER8TmuiNbwUuXcEjN/CqrwZhFF/JtKMS7DCRiB"
"l3NbdOF5jV9M3uD07RvxRQaALHAOuHblOrg6lP2TQ/qd9rC0R/yA6my2FL4sGHuOJTWd9UtcEIGYBFxf4cdWs9nCLxf45UXz4FD2"
"A1yzfo9HU2uVVG55N1tNhtn6ZVm7my4QbbwQJS8byVo5pdQWOqatelB5xpAebn7h3QmFWdzBTMZ1HEqentRv/5X9E60Le9o17XBn"
"R49bEka3/vKrz3nDfhmdBGlclkGMEeIi33WyiHhKnIZpFAgpqGso3F1SrqWoR9I32Vw6nwtf5mRNGKvIoPHvMSxFnOEHifGz5BUN"
"bTUdXEfpmfy7DZ5H9uQXLbo39HwNkRXmZfjw+vM0TOLBNa0+X9JFG9d07X/WcnfThsWpu8e365RWEn7ui71c/OQvSXlYfMbPpdRb"
"urfQrUICA+CF78z9ecgkp8isnMLTC4IKnIdIQnCAayU2gZjz3TuZgoq1D/OJ/LMTTOOtAi2xDmI/kdR0xrU1Ce0+7gL+KXm7QwNp"
"PT/QFjOo1Ozxa2gJQPg4+Ni1T3NzTTo+Pu5j5QkgTk9P1dhY4sj1iMauxG4rOKICy74ssOyLbqkjgbQPUxS+N2YYL+dhOrgLpNms"
"1VzXPpW1G1mqbFupuEdlw5VWDm47ME33y3ZqaNvmnBcE5hfMDEU3fajlwtDW7YAOfb7ipv5Fuw3d7WEM32DdyrdJHP+mbEHteEBH"
"S7OnNVtJaFJREBLAsDQ6DWYcplZiQ5szpqNrDG/LrOLqiznli6ZiMinHomCS7orgzouoeY5I15gYSaNJ/xE7gh6xK2iz7QX5Y27a"
"yTU6MmkQqKu+i6aNQ7jW1Z3r9cg+xatsSAr4BFuxKiPuXOAbFY5fL8+tbhOF0pRKTFyVdonVUdcZQKfW6WUEa65t7v0dy52/Zgkn"
"P9VUAdecwq8r5hegcW4aJnPW2sx3NyhVJaDGTN1UUDxyJXUj47WZmRTo9Qvfvc4osxhueFjkRTPrFCQ0+s+6armV6ziVoUBsuv+0"
"BOJUZ6svsivf0oMn4loNsF+0+plVOWHeENEft2Nz/BtC8u+jdrfe/+Plr1evm83g8t3r396ee/ftZnD4gFtucLVx1gnGS/xLpg2W"
"yMFNdoezgYfF420AZm/ed9gv55eXDLEYYrFZh3Es5r1kqOogOBDEbbyEBBJM5iUjjV/i47doNmfqE/fjJ4jsb3dle9HI216sptsN"
"dtPz5p+Pjmb9P6IBruuDKLTvcmuE2ysYgCIXMtBfFWGWFDCourZKZ4xnySZ/eEJClgDSXcyvvG1u+uli2/f8XaM6fwqRmWl1/IKc"
"N9MsZ+XeSFapWgQpWOr7jSrIMhHmeFKc34gsRLwW28HxqzyalefQRPd/AS8Ud681eAAA"
)
# ---------------------------------------------------------------------------
# Module cache: one per kernel
# ---------------------------------------------------------------------------
_mod_cache: dict = {}
_CUDA_FLAGS = ["-std=c++20", "--offload-arch=gfx950", "-O3"]
def _load_kernel(name: str, b64_blob: str) -> object:
"""Decompress, write, and compile a single kernel .cu file. Cached."""
if name in _mod_cache:
return _mod_cache[name]
os.makedirs(BUILD_DIR, exist_ok=True)
os.environ["PYTORCH_ROCM_ARCH"] = "gfx950"
cu_source = gzip.decompress(base64.b64decode(b64_blob)).decode("utf-8")
cu_path = os.path.join(BUILD_DIR, f"{name}.cu")
need_write = True
if os.path.exists(cu_path):
with open(cu_path, "r") as f:
if f.read() == cu_source:
need_write = False
if need_write:
with open(cu_path, "w") as f:
f.write(cu_source)
mod = cpp_ext.load(
name=name,
sources=[cu_path],
extra_cuda_cflags=_CUDA_FLAGS,
build_directory=BUILD_DIR,
verbose=False,
)
_mod_cache[name] = mod
return mod
# ---------------------------------------------------------------------------
# Caches
# ---------------------------------------------------------------------------
_b_cache: dict = {}
_c_cache: dict = {}
def _prep_b_buffers(B_q, B_scale_sh):
"""Materialise static B buffers once and cache their raw int pointers."""
key = (B_q.data_ptr(), B_scale_sh.data_ptr())
if key in _b_cache:
return _b_cache[key]
B_data_u8 = B_q.view(torch.uint8).contiguous()
B_scale_u8 = B_scale_sh.view(torch.uint8).contiguous()
b_ptr = int(B_data_u8.data_ptr())
bs_ptr = int(B_scale_u8.data_ptr())
_b_cache[key] = (B_data_u8, B_scale_u8, b_ptr, bs_ptr)
return _b_cache[key]
def _get_c_buf(m, n, device):
"""Return a pre-allocated output buffer for shape (m, n), reusing across calls."""
key = (m, n)
if key not in _c_cache:
_c_cache[key] = torch.empty(m, n, dtype=torch.bfloat16, device=device)
return _c_cache[key]
def _quant_mxfp4(x, shuffle=True):
from aiter import dtypes
from aiter.ops.triton.quant import dynamic_mxfp4_quant
from aiter.utility.fp4_utils import e8m0_shuffle
x_fp4, bs_e8m0 = dynamic_mxfp4_quant(x.contiguous())
if shuffle:
bs_e8m0 = e8m0_shuffle(bs_e8m0)
return x_fp4.view(dtypes.fp4x2), bs_e8m0.view(dtypes.fp8_e8m0)
def custom_kernel(data: input_t) -> output_t:
"""
Uses optimized custom kernels for specific shapes,
otherwise falls back to aiter gemm_a4w4.
"""
import aiter
from aiter import dtypes
A, B, B_q, B_shuffle, B_scale_sh = data
A = A.contiguous()
B = B.contiguous()
m, k = A.shape
n, _ = B.shape
# v31_32x64: M=64, N=7168, K=2048
if m == 64 and n == 7168 and k == 2048:
B_data_u8, B_scale_u8, _, _ = _prep_b_buffers(B_q, B_scale_sh)
C = _get_c_buf(m, n, A.device)
mod = _load_kernel("v31_32x64_t512_o2_ks1024", _B64_V31_32X64_64_7168_2048)
mod.run(A, B_data_u8, B_scale_u8, C, m, n, k)
return C
# v24_64x16: M=256, N=3072, K=1536
if m == 256 and n == 3072 and k == 1536:
B_data_u8, B_scale_u8, _, _ = _prep_b_buffers(B_q, B_scale_sh)
C = _get_c_buf(m, n, A.device)
mod = _load_kernel("v24_64x16_t256_o1_ks512", _B64_V24_64X16_256_3072_1536)
mod.run(A, B_data_u8, B_scale_u8, C, m, n, k)
return C
# v31_8x32: M=16, N=2112, K=7168
if m == 16 and n == 2112 and k == 7168:
B_data_u8, B_scale_u8, _, _ = _prep_b_buffers(B_q, B_scale_sh)
C = _get_c_buf(m, n, A.device)
mod = _load_kernel("v31_8x32_t128_o4_ks1024", _B64_V31_8X32_16_2112_7168)
mod.run(A, B_data_u8, B_scale_u8, C, m, n, k)
return C
# v25_2x16: M=4, N=2880, K=512
if m == 4 and n == 2880 and k == 512:
B_data_u8, B_scale_u8, _, _ = _prep_b_buffers(B_q, B_scale_sh)
C = _get_c_buf(m, n, A.device)
mod = _load_kernel("v25_2x16_t64_o4_ks512_4_2880_512", _B64_V25_2X16_4_2880_512)
mod.run(A, B_data_u8, B_scale_u8, C, m, n, k)
return C
# v25_8x64: M=32, N=2880, K=512
if m == 32 and n == 2880 and k == 512:
B_data_u8, B_scale_u8, _, _ = _prep_b_buffers(B_q, B_scale_sh)
C = _get_c_buf(m, n, A.device)
mod = _load_kernel("v25_8x64_t256_o2_ks512_32_2880_512", _B64_V25_8X64_32_2880_512)
mod.run(A, B_data_u8, B_scale_u8, C, m, n, k)
return C
# v25_8x64: M=32, N=4096, K=512
if m == 32 and n == 4096 and k == 512:
B_data_u8, B_scale_u8, _, _ = _prep_b_buffers(B_q, B_scale_sh)
C = _get_c_buf(m, n, A.device)
mod = _load_kernel("v25_8x64_t256_o2_ks512_32_4096_512", _B64_V25_8X64_32_4096_512)
mod.run(A, B_data_u8, B_scale_u8, C)
return C
# Fallback to aiter reference
A_q, A_scale_sh = _quant_mxfp4(A, shuffle=True)
out_gemm = aiter.gemm_a4w4(
A_q,
B_shuffle,
A_scale_sh,
B_scale_sh,
dtype=dtypes.bf16,
bpreshuffle=True,
)
return out_gemm
scrolls · 607 lines total
Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0
Best evidence level for this revision: reported
JSON