@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / b0651021e5
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 7, 05:15 UTC

[Kernel] [Helion] [11/N] Retune configs for silu_mul_fp8 (#36062)

Yanan Cao committed Mar 4, 2026 at 21:25 UTC b0651021e5c042e0893929e1b80cf367c6611708
1 file changed +4034 -4026
vllm/kernels/helion/configs/silu_mul_fp8.json
+4034 -4026
@@ -47,8 +47,8 @@
47 },
48 "intermediate_4096_numtokens_256": {
49 "block_sizes": [
50 - 32,
51 - 512
50 + 256,
51 + 32
52 ],
53 "loop_orders": [
54 [
@@ -60,7 +60,7 @@
60 true
61 ],
62 "l2_groupings": [
63 - 2
63 + 1
64 ],
65 "range_unroll_factors": [
66 0
@@ -80,11 +80,11 @@
80 "",
81 ""
82 ],
83 - "num_warps": 2,
84 - "num_stages": 2,
83 + "num_warps": 8,
84 + "num_stages": 1,
85 "indexing": [
86 "pointer",
87 - "tensor_descriptor",
87 + "pointer",
88 "pointer",
89 "pointer"
90 ],
@@ -137,8 +137,8 @@
137 },
138 "intermediate_8192_numtokens_256": {
139 "block_sizes": [
140 - 32,
141 - 8
140 + 256,
141 + 64
142 ],
143 "loop_orders": [
144 [
@@ -167,23 +167,23 @@
167 ],
168 "load_eviction_policies": [
169 "",
170 - "first",
170 + "",
171 ""
172 ],
173 - "num_warps": 4,
173 + "num_warps": 8,
174 "num_stages": 1,
175 "indexing": [
176 - "tensor_descriptor",
176 "pointer",
177 "pointer",
179 - "tensor_descriptor"
178 + "pointer",
179 + "pointer"
180 ],
181 "pid_type": "flat"
182 },
183 "intermediate_14336_numtokens_256": {
184 "block_sizes": [
185 - 16,
186 - 32
185 + 8,
186 + 4096
187 ],
188 "loop_orders": [
189 [
@@ -215,13 +215,13 @@
215 "",
216 ""
217 ],
218 - "num_warps": 1,
218 + "num_warps": 8,
219 "num_stages": 1,
220 "indexing": [
221 "pointer",
222 "pointer",
223 "pointer",
224 - "tensor_descriptor"
224 + "pointer"
225 ],
226 "pid_type": "flat"
227 },
@@ -272,8 +272,8 @@
272 },
273 "intermediate_7688_numtokens_256": {
274 "block_sizes": [
275 - 8,
276 - 16
275 + 32,
276 + 512
277 ],
278 "loop_orders": [
279 [
@@ -302,10 +302,10 @@
302 ],
303 "load_eviction_policies": [
304 "",
305 - "last",
305 + "",
306 ""
307 ],
308 - "num_warps": 1,
308 + "num_warps": 8,
309 "num_stages": 1,
310 "indexing": [
311 "pointer",
@@ -363,7 +363,7 @@
363 "intermediate_2048_numtokens_1": {
364 "block_sizes": [
365 1,
366 - 16
366 + 256
367 ],
368 "loop_orders": [
369 [
@@ -372,7 +372,7 @@
372 ]
373 ],
374 "flatten_loops": [
375 - false
375 + true
376 ],
377 "l2_groupings": [
378 1
@@ -395,20 +395,20 @@
395 "",
396 ""
397 ],
398 - "num_warps": 16,
399 - "num_stages": 2,
398 + "num_warps": 2,
399 + "num_stages": 1,
400 "indexing": [
401 "pointer",
402 "pointer",
403 "pointer",
404 - "tensor_descriptor"
404 + "pointer"
405 ],
406 "pid_type": "flat"
407 },
408 "intermediate_2880_numtokens_1": {
409 "block_sizes": [
410 1,
411 - 1
411 + 128
412 ],
413 "loop_orders": [
414 [
@@ -417,7 +417,7 @@
417 ]
418 ],
419 "flatten_loops": [
420 - false
420 + true
421 ],
422 "l2_groupings": [
423 1
@@ -440,8 +440,8 @@
440 "",
441 ""
442 ],
443 - "num_warps": 16,
444 - "num_stages": 2,
443 + "num_warps": 4,
444 + "num_stages": 1,
445 "indexing": [
446 "pointer",
447 "pointer",
@@ -453,16 +453,16 @@
453 "intermediate_4096_numtokens_1": {
454 "block_sizes": [
455 1,
456 - 32
456 + 256
457 ],
458 "loop_orders": [
459 [
460 - 1,
461 - 0
460 + 0,
461 + 1
462 ]
463 ],
464 "flatten_loops": [
465 - false
465 + true
466 ],
467 "l2_groupings": [
468 1
@@ -485,8 +485,8 @@
485 "",
486 ""
487 ],
488 - "num_warps": 4,
489 - "num_stages": 2,
488 + "num_warps": 2,
489 + "num_stages": 1,
490 "indexing": [
491 "pointer",
492 "pointer",
@@ -498,7 +498,7 @@
498 "intermediate_8192_numtokens_1": {
499 "block_sizes": [
500 1,
501 - 32
501 + 256
502 ],
503 "loop_orders": [
504 [
@@ -507,7 +507,7 @@
507 ]
508 ],
509 "flatten_loops": [
510 - false
510 + true
511 ],
512 "l2_groupings": [
513 1
@@ -530,8 +530,8 @@
530 "",
531 ""
532 ],
533 - "num_warps": 1,
534 - "num_stages": 2,
533 + "num_warps": 2,
534 + "num_stages": 1,
535 "indexing": [
536 "pointer",
537 "pointer",
@@ -543,7 +543,7 @@
543 "intermediate_11008_numtokens_1": {
544 "block_sizes": [
545 1,
546 - 32
546 + 256
547 ],
548 "loop_orders": [
549 [
@@ -571,11 +571,11 @@
571 null
572 ],
573 "load_eviction_policies": [
574 - "first",
574 + "",
575 "",
576 ""
577 ],
578 - "num_warps": 4,
578 + "num_warps": 2,
579 "num_stages": 1,
580 "indexing": [
581 "pointer",
@@ -588,7 +588,7 @@
588 "intermediate_14336_numtokens_1": {
589 "block_sizes": [
590 1,
591 - 32
591 + 256
592 ],
593 "loop_orders": [
594 [
@@ -597,10 +597,10 @@
597 ]
598 ],
599 "flatten_loops": [
600 - false
600 + true
601 ],
602 "l2_groupings": [
603 - 2
603 + 1
604 ],
605 "range_unroll_factors": [
606 0
@@ -618,9 +618,9 @@
618 "load_eviction_policies": [
619 "",
620 "",
621 - "first"
621 + ""
622 ],
623 - "num_warps": 8,
623 + "num_warps": 2,
624 "num_stages": 1,
625 "indexing": [
626 "pointer",
@@ -633,7 +633,7 @@
633 "intermediate_2048_numtokens_2": {
634 "block_sizes": [
635 2,
636 - 16
636 + 128
637 ],
638 "loop_orders": [
639 [
@@ -645,7 +645,7 @@
645 true
646 ],
647 "l2_groupings": [
648 - 2
648 + 1
649 ],
650 "range_unroll_factors": [
651 0
@@ -670,15 +670,15 @@
670 "indexing": [
671 "pointer",
672 "pointer",
673 - "tensor_descriptor",
673 + "pointer",
674 "pointer"
675 ],
676 "pid_type": "flat"
677 },
678 "intermediate_2880_numtokens_2": {
679 "block_sizes": [
680 - 1,
681 - 4
680 + 2,
681 + 256
682 ],
683 "loop_orders": [
684 [
@@ -687,7 +687,7 @@
687 ]
688 ],
689 "flatten_loops": [
690 - false
690 + true
691 ],
692 "l2_groupings": [
693 1
@@ -706,11 +706,11 @@
706 null
707 ],
708 "load_eviction_policies": [
709 - "first",
709 + "",
710 "",
711 ""
712 ],
713 - "num_warps": 32,
713 + "num_warps": 4,
714 "num_stages": 1,
715 "indexing": [
716 "pointer",
@@ -723,7 +723,7 @@
723 "intermediate_4096_numtokens_2": {
724 "block_sizes": [
725 2,
726 - 32
726 + 128
727 ],
728 "loop_orders": [
729 [
@@ -755,8 +755,8 @@
755 "",
756 ""
757 ],
758 - "num_warps": 8,
759 - "num_stages": 2,
758 + "num_warps": 2,
759 + "num_stages": 1,
760 "indexing": [
761 "pointer",
762 "pointer",
@@ -767,8 +767,8 @@
767 },
768 "intermediate_8192_numtokens_2": {
769 "block_sizes": [
770 - 1,
771 - 32
770 + 2,
771 + 128
772 ],
773 "loop_orders": [
774 [
@@ -777,7 +777,7 @@
777 ]
778 ],
779 "flatten_loops": [
780 - false
780 + true
781 ],
782 "l2_groupings": [
783 1
@@ -796,14 +796,14 @@
796 null
797 ],
798 "load_eviction_policies": [
799 - "first",
799 + "",
800 "",
801 ""
802 ],
803 - "num_warps": 8,
804 - "num_stages": 2,
803 + "num_warps": 2,
804 + "num_stages": 1,
805 "indexing": [
806 - "tensor_descriptor",
806 + "pointer",
807 "pointer",
808 "pointer",
809 "pointer"
@@ -813,19 +813,19 @@
813 "intermediate_11008_numtokens_2": {
814 "block_sizes": [
815 1,
816 - 256
816 + 16384
817 ],
818 "loop_orders": [
819 [
820 - 0,
821 - 1
820 + 1,
821 + 0
822 ]
823 ],
824 "flatten_loops": [
825 - false
825 + true
826 ],
827 "l2_groupings": [
828 - 1
828 + 16
829 ],
830 "range_unroll_factors": [
831 0
@@ -841,24 +841,24 @@
841 null
842 ],
843 "load_eviction_policies": [
844 - "",
844 + "last",
845 "first",
846 - ""
846 + "last"
847 ],
848 - "num_warps": 4,
849 - "num_stages": 2,
848 + "num_warps": 32,
849 + "num_stages": 3,
850 "indexing": [
851 - "pointer",
851 + "tensor_descriptor",
852 "pointer",
853 "pointer",
854 "tensor_descriptor"
855 ],
856 - "pid_type": "flat"
856 + "pid_type": "xyz"
857 },
858 "intermediate_14336_numtokens_2": {
859 "block_sizes": [
860 - 1,
861 - 64
860 + 2,
861 + 128
862 ],
863 "loop_orders": [
864 [
@@ -867,7 +867,7 @@
867 ]
868 ],
869 "flatten_loops": [
870 - false
870 + true
871 ],
872 "l2_groupings": [
873 1
@@ -890,8 +890,8 @@
890 "",
891 ""
892 ],
893 - "num_warps": 8,
894 - "num_stages": 2,
893 + "num_warps": 2,
894 + "num_stages": 1,
895 "indexing": [
896 "pointer",
897 "pointer",
@@ -902,8 +902,8 @@
902 },
903 "intermediate_2048_numtokens_4": {
904 "block_sizes": [
905 - 1,
906 - 256
905 + 4,
906 + 64
907 ],
908 "loop_orders": [
909 [
@@ -935,8 +935,8 @@
935 "",
936 ""
937 ],
938 - "num_warps": 32,
939 - "num_stages": 2,
938 + "num_warps": 2,
939 + "num_stages": 1,
940 "indexing": [
941 "pointer",
942 "pointer",
@@ -947,8 +947,8 @@
947 },
948 "intermediate_2880_numtokens_4": {
949 "block_sizes": [
950 - 1,
951 - 8
950 + 4,
951 + 64
952 ],
953 "loop_orders": [
954 [
@@ -960,7 +960,7 @@
960 true
961 ],
962 "l2_groupings": [
963 - 1
963 + 64
964 ],
965 "range_unroll_factors": [
966 0
@@ -977,23 +977,23 @@
977 ],
978 "load_eviction_policies": [
979 "",
980 - "",
981 - ""
980 + "last",
981 + "first"
982 ],
983 - "num_warps": 1,
984 - "num_stages": 2,
983 + "num_warps": 8,
984 + "num_stages": 7,
985 "indexing": [
986 "pointer",
987 - "pointer",
988 - "pointer",
987 + "tensor_descriptor",
988 + "tensor_descriptor",
989 "pointer"
990 ],
991 - "pid_type": "flat"
991 + "pid_type": "xyz"
992 },
993 "intermediate_4096_numtokens_4": {
994 "block_sizes": [
995 4,
996 - 16
996 + 64
997 ],
998 "loop_orders": [
999 [
@@ -1005,7 +1005,7 @@
1005 true
1006 ],
1007 "l2_groupings": [
1008 - 2
1008 + 1
1009 ],
1010 "range_unroll_factors": [
1011 0
@@ -1021,11 +1021,11 @@
1021 null
1022 ],
1023 "load_eviction_policies": [
1024 - "last",
1024 + "",
1025 "",
1026 ""
1027 ],
1028 - "num_warps": 4,
1028 + "num_warps": 2,
1029 "num_stages": 1,
1030 "indexing": [
1031 "pointer",
@@ -1038,7 +1038,7 @@
1038 "intermediate_8192_numtokens_4": {
1039 "block_sizes": [
1040 1,
1041 - 16
1041 + 2048
1042 ],
1043 "loop_orders": [
1044 [
@@ -1047,7 +1047,7 @@
1047 ]
1048 ],
1049 "flatten_loops": [
1050 - true
1050 + false
1051 ],
1052 "l2_groupings": [
1053 1
@@ -1070,11 +1070,11 @@
1070 "",
1071 ""
1072 ],
1073 - "num_warps": 4,
1074 - "num_stages": 1,
1073 + "num_warps": 8,
1074 + "num_stages": 8,
1075 "indexing": [
1076 "pointer",
1077 - "pointer",
1077 + "tensor_descriptor",
1078 "pointer",
1079 "pointer"
1080 ],
@@ -1082,20 +1082,20 @@
1082 },
1083 "intermediate_11008_numtokens_4": {
1084 "block_sizes": [
1085 - 1,
1086 - 32
1085 + 4,
1086 + 2048
1087 ],
1088 "loop_orders": [
1089 [
1090 - 0,
1091 - 1
1090 + 1,
1091 + 0
1092 ]
1093 ],
1094 "flatten_loops": [
1095 - false
1095 + true
1096 ],
1097 "l2_groupings": [
1098 - 1
1098 + 4
1099 ],
1100 "range_unroll_factors": [
1101 0
@@ -1111,36 +1111,36 @@
1111 null
1112 ],
1113 "load_eviction_policies": [
1114 - "",
1114 "first",
1116 - ""
1115 + "last",
1116 + "last"
1117 ],
1118 - "num_warps": 16,
1119 - "num_stages": 1,
1118 + "num_warps": 8,
1119 + "num_stages": 6,
1120 "indexing": [
1121 - "pointer",
1121 "pointer",
1122 "tensor_descriptor",
1124 - "pointer"
1123 + "tensor_descriptor",
1124 + "tensor_descriptor"
1125 ],
1126 - "pid_type": "flat"
1126 + "pid_type": "xyz"
1127 },
1128 "intermediate_14336_numtokens_4": {
1129 "block_sizes": [
1130 4,
1131 - 16
1131 + 256
1132 ],
1133 "loop_orders": [
1134 [
1135 - 1,
1136 - 0
1135 + 0,
1136 + 1
1137 ]
1138 ],
1139 "flatten_loops": [
1140 true
1141 ],
1142 "l2_groupings": [
1143 - 1
1143 + 64
1144 ],
1145 "range_unroll_factors": [
1146 0
@@ -1160,25 +1160,25 @@
1160 "",
1161 ""
1162 ],
1163 - "num_warps": 16,
1164 - "num_stages": 1,
1163 + "num_warps": 2,
1164 + "num_stages": 2,
1165 "indexing": [
1166 "pointer",
1167 "pointer",
1168 "pointer",
1169 "pointer"
1170 ],
1171 - "pid_type": "flat"
1171 + "pid_type": "xyz"
1172 },
1173 "intermediate_2048_numtokens_8": {
1174 "block_sizes": [
1175 8,
1176 - 256
1176 + 32
1177 ],
1178 "loop_orders": [
1179 [
1180 - 1,
1181 - 0
1180 + 0,
1181 + 1
1182 ]
1183 ],
1184 "flatten_loops": [
@@ -1201,14 +1201,14 @@
1201 null
1202 ],
1203 "load_eviction_policies": [
1204 - "last",
1204 + "",
1205 "",
1206 ""
1207 ],
1208 - "num_warps": 32,
1208 + "num_warps": 2,
1209 "num_stages": 1,
1210 "indexing": [
1211 - "tensor_descriptor",
1211 + "pointer",
1212 "pointer",
1213 "pointer",
1214 "pointer"
@@ -1217,8 +1217,8 @@
1217 },
1218 "intermediate_2880_numtokens_8": {
1219 "block_sizes": [
1220 - 8,
1221 - 32
1220 + 4,
1221 + 2048
1222 ],
1223 "loop_orders": [
1224 [
@@ -1230,7 +1230,7 @@
1230 true
1231 ],
1232 "l2_groupings": [
1233 - 1
1233 + 4
1234 ],
1235 "range_unroll_factors": [
1236 0
@@ -1246,23 +1246,23 @@
1246 null
1247 ],
1248 "load_eviction_policies": [
1249 - "",
1250 - "",
1249 + "last",
1250 + "last",
1251 ""
1252 ],
1253 - "num_warps": 4,
1254 - "num_stages": 1,
1253 + "num_warps": 16,
1254 + "num_stages": 5,
1255 "indexing": [
1256 "pointer",
1257 - "pointer",
1258 - "pointer",
1257 + "tensor_descriptor",
1258 + "tensor_descriptor",
1259 "pointer"
1260 ],
1261 - "pid_type": "flat"
1261 + "pid_type": "xyz"
1262 },
1263 "intermediate_4096_numtokens_8": {
1264 "block_sizes": [
1265 - 2,
1265 + 8,
1266 32
1267 ],
1268 "loop_orders": [
@@ -1275,7 +1275,7 @@
1275 true
1276 ],
1277 "l2_groupings": [
1278 - 2
1278 + 1
1279 ],
1280 "range_unroll_factors": [
1281 0
@@ -1295,7 +1295,7 @@
1295 "",
1296 ""
1297 ],
1298 - "num_warps": 32,
1298 + "num_warps": 2,
1299 "num_stages": 1,
1300 "indexing": [
1301 "pointer",
@@ -1307,8 +1307,8 @@
1307 },
1308 "intermediate_8192_numtokens_8": {
1309 "block_sizes": [
1310 - 4,
1311 - 64
1310 + 2,
1311 + 1024
1312 ],
1313 "loop_orders": [
1314 [
@@ -1317,10 +1317,10 @@
1317 ]
1318 ],
1319 "flatten_loops": [
1320 - false
1320 + true
1321 ],
1322 "l2_groupings": [
1323 - 1
1323 + 16
1324 ],
1325 "range_unroll_factors": [
1326 0
@@ -1336,36 +1336,36 @@
1336 null
1337 ],
1338 "load_eviction_policies": [
1339 - "",
1340 - "",
1339 + "first",
1340 + "last",
1341 ""
1342 ],
1343 "num_warps": 1,
1344 - "num_stages": 1,
1344 + "num_stages": 8,
1345 "indexing": [
1346 + "tensor_descriptor",
1347 "pointer",
1347 - "pointer",
1348 - "pointer",
1348 + "tensor_descriptor",
1349 "pointer"
1350 ],
1351 "pid_type": "flat"
1352 },
1353 "intermediate_11008_numtokens_8": {
1354 "block_sizes": [
1355 - 8,
1356 - 128
1355 + 1,
1356 + 1024
1357 ],
1358 "loop_orders": [
1359 [
1360 - 0,
1361 - 1
1360 + 1,
1361 + 0
1362 ]
1363 ],
1364 "flatten_loops": [
1365 - false
1365 + true
1366 ],
1367 "l2_groupings": [
1368 - 1
1368 + 4
1369 ],
1370 "range_unroll_factors": [
1371 0
@@ -1383,22 +1383,22 @@
1383 "load_eviction_policies": [
1384 "last",
1385 "",
1386 - ""
1386 + "first"
1387 ],
1388 - "num_warps": 4,
1389 - "num_stages": 2,
1388 + "num_warps": 2,
1389 + "num_stages": 5,
1390 "indexing": [
1391 - "pointer",
1392 - "pointer",
1393 - "pointer",
1394 - "pointer"
1391 + "tensor_descriptor",
1392 + "tensor_descriptor",
1393 + "tensor_descriptor",
1394 + "tensor_descriptor"
1395 ],
1396 "pid_type": "flat"
1397 },
1398 "intermediate_14336_numtokens_8": {
1399 "block_sizes": [
1400 8,
1401 - 32
1401 + 128
1402 ],
1403 "loop_orders": [
1404 [
@@ -1407,7 +1407,7 @@
1407 ]
1408 ],
1409 "flatten_loops": [
1410 - false
1410 + true
1411 ],
1412 "l2_groupings": [
1413 1
@@ -1427,7 +1427,7 @@
1427 ],
1428 "load_eviction_policies": [
1429 "",
1430 - "first",
1430 + "",
1431 ""
1432 ],
1433 "num_warps": 4,
@@ -1442,8 +1442,8 @@
1442 },
1443 "intermediate_2048_numtokens_16": {
1444 "block_sizes": [
1445 - 16,
1446 - 64
1445 + 8,
1446 + 512
1447 ],
1448 "loop_orders": [
1449 [
@@ -1452,10 +1452,10 @@
1452 ]
1453 ],
1454 "flatten_loops": [
1455 - false
1455 + true
1456 ],
1457 "l2_groupings": [
1458 - 2
1458 + 64
1459 ],
1460 "range_unroll_factors": [
1461 0
@@ -1471,24 +1471,24 @@
1471 null
1472 ],
1473 "load_eviction_policies": [
1474 - "",
1475 - "",
1476 - ""
1474 + "last",
1475 + "last",
1476 + "first"
1477 ],
1478 - "num_warps": 4,
1478 + "num_warps": 16,
1479 "num_stages": 2,
1480 "indexing": [
1481 "pointer",
1482 "pointer",
1483 - "pointer",
1483 + "tensor_descriptor",
1484 "pointer"
1485 ],
1486 - "pid_type": "flat"
1486 + "pid_type": "xyz"
1487 },
1488 "intermediate_2880_numtokens_16": {
1489 "block_sizes": [
1490 - 16,
1491 - 32
1490 + 2,
1491 + 256
1492 ],
1493 "loop_orders": [
1494 [
@@ -1520,7 +1520,7 @@
1520 "",
1521 ""
1522 ],
1523 - "num_warps": 8,
1523 + "num_warps": 4,
1524 "num_stages": 1,
1525 "indexing": [
1526 "pointer",
@@ -1533,7 +1533,7 @@
1533 "intermediate_4096_numtokens_16": {
1534 "block_sizes": [
1535 16,
1536 - 32
1536 + 256
1537 ],
1538 "loop_orders": [
1539 [
@@ -1545,7 +1545,7 @@
1545 true
1546 ],
1547 "l2_groupings": [
1548 - 1
1548 + 16
1549 ],
1550 "range_unroll_factors": [
1551 0
@@ -1561,12 +1561,12 @@
1561 null
1562 ],
1563 "load_eviction_policies": [
1564 + "last",
1565 "",
1565 - "",
1566 - ""
1566 + "last"
1567 ],
1568 - "num_warps": 8,
1569 - "num_stages": 2,
1568 + "num_warps": 32,
1569 + "num_stages": 1,
1570 "indexing": [
1571 "pointer",
1572 "pointer",
@@ -1577,8 +1577,8 @@
1577 },
1578 "intermediate_8192_numtokens_16": {
1579 "block_sizes": [
1580 - 4,
1581 - 32
1580 + 16,
1581 + 64
1582 ],
1583 "loop_orders": [
1584 [
@@ -1587,7 +1587,7 @@
1587 ]
1588 ],
1589 "flatten_loops": [
1590 - false
1590 + true
1591 ],
1592 "l2_groupings": [
1593 1
@@ -1610,11 +1610,11 @@
1610 "",
1611 ""
1612 ],
1613 - "num_warps": 1,
1613 + "num_warps": 4,
1614 "num_stages": 1,
1615 "indexing": [
1616 "pointer",
1617 - "tensor_descriptor",
1617 + "pointer",
1618 "pointer",
1619 "pointer"
1620 ],
@@ -1622,8 +1622,8 @@
1622 },
1623 "intermediate_11008_numtokens_16": {
1624 "block_sizes": [
1625 - 8,
1626 - 32
1625 + 1,
1626 + 2048
1627 ],
1628 "loop_orders": [
1629 [
@@ -1653,9 +1653,9 @@
1653 "load_eviction_policies": [
1654 "",
1655 "",
1656 - "first"
1656 + ""
1657 ],
1658 - "num_warps": 16,
1658 + "num_warps": 8,
1659 "num_stages": 1,
1660 "indexing": [
1661 "pointer",
@@ -1667,20 +1667,20 @@
1667 },
1668 "intermediate_14336_numtokens_16": {
1669 "block_sizes": [
1670 - 16,
1671 - 32
1670 + 2,
1671 + 256
1672 ],
1673 "loop_orders": [
1674 [
1675 - 0,
1676 - 1
1675 + 1,
1676 + 0
1677 ]
1678 ],
1679 "flatten_loops": [
1680 true
1681 ],
1682 "l2_groupings": [
1683 - 1
1683 + 64
1684 ],
1685 "range_unroll_factors": [
1686 0
@@ -1696,24 +1696,24 @@
1696 null
1697 ],
1698 "load_eviction_policies": [
1699 - "",
1700 - "",
1701 - ""
1699 + "last",
1700 + "last",
1701 + "last"
1702 ],
1703 - "num_warps": 1,
1704 - "num_stages": 1,
1703 + "num_warps": 16,
1704 + "num_stages": 7,
1705 "indexing": [
1706 "tensor_descriptor",
1707 "pointer",
1708 - "pointer",
1709 - "pointer"
1708 + "tensor_descriptor",
1709 + "tensor_descriptor"
1710 ],
1711 "pid_type": "flat"
1712 },
1713 "intermediate_2048_numtokens_24": {
1714 "block_sizes": [
1715 - 16,
1716 - 8
1715 + 1,
1716 + 1024
1717 ],
1718 "loop_orders": [
1719 [
@@ -1725,7 +1725,7 @@
1725 true
1726 ],
1727 "l2_groupings": [
1728 - 1
1728 + 64
1729 ],
1730 "range_unroll_factors": [
1731 0
@@ -1742,35 +1742,35 @@
1742 ],
1743 "load_eviction_policies": [
1744 "",
1745 - "last",
1746 - "last"
1745 + "",
1746 + "first"
1747 ],
1748 - "num_warps": 1,
1749 - "num_stages": 1,
1748 + "num_warps": 4,
1749 + "num_stages": 8,
1750 "indexing": [
1751 + "tensor_descriptor",
1752 "pointer",
1753 "pointer",
1753 - "pointer",
1754 - "pointer"
1754 + "tensor_descriptor"
1755 ],
1756 "pid_type": "flat"
1757 },
1758 "intermediate_2880_numtokens_24": {
1759 "block_sizes": [
1760 - 32,
1761 - 64
1760 + 4,
1761 + 1024
1762 ],
1763 "loop_orders": [
1764 [
1765 - 1,
1766 - 0
1765 + 0,
1766 + 1
1767 ]
1768 ],
1769 "flatten_loops": [
1770 true
1771 ],
1772 "l2_groupings": [
1773 - 1
1773 + 16
1774 ],
1775 "range_unroll_factors": [
1776 0
@@ -1786,24 +1786,24 @@
1786 null
1787 ],
1788 "load_eviction_policies": [
1789 + "first",
1790 "",
1790 - "",
1791 - ""
1791 + "first"
1792 ],
1793 - "num_warps": 4,
1794 - "num_stages": 1,
1793 + "num_warps": 16,
1794 + "num_stages": 3,
1795 "indexing": [
1796 "tensor_descriptor",
1797 "pointer",
1798 - "pointer",
1799 - "tensor_descriptor"
1798 + "tensor_descriptor",
1799 + "pointer"
1800 ],
1801 "pid_type": "flat"
1802 },
1803 "intermediate_4096_numtokens_24": {
1804 "block_sizes": [
1805 - 32,
1806 - 32
1805 + 16,
1806 + 64
1807 ],
1808 "loop_orders": [
1809 [
@@ -1815,7 +1815,7 @@
1815 true
1816 ],
1817 "l2_groupings": [
1818 - 2
1818 + 1
1819 ],
1820 "range_unroll_factors": [
1821 0
@@ -1832,23 +1832,23 @@
1832 ],
1833 "load_eviction_policies": [
1834 "",
1835 - "first",
1835 + "",
1836 ""
1837 ],
1838 - "num_warps": 1,
1838 + "num_warps": 4,
1839 "num_stages": 1,
1840 "indexing": [
1841 "pointer",
1842 "pointer",
1843 "pointer",
1844 - "tensor_descriptor"
1844 + "pointer"
1845 ],
1846 "pid_type": "flat"
1847 },
1848 "intermediate_8192_numtokens_24": {
1849 "block_sizes": [
1850 - 16,
1851 - 32
1850 + 1,
1851 + 2048
1852 ],
1853 "loop_orders": [
1854 [
@@ -1857,10 +1857,10 @@
1857 ]
1858 ],
1859 "flatten_loops": [
1860 - false
1860 + true
1861 ],
1862 "l2_groupings": [
1863 - 1
1863 + 32
1864 ],
1865 "range_unroll_factors": [
1866 0
@@ -1877,35 +1877,35 @@
1877 ],
1878 "load_eviction_policies": [
1879 "",
1880 - "",
1881 - ""
1880 + "last",
1881 + "last"
1882 ],
1883 - "num_warps": 2,
1884 - "num_stages": 1,
1883 + "num_warps": 32,
1884 + "num_stages": 5,
1885 "indexing": [
1886 "pointer",
1887 "pointer",
1888 "pointer",
1889 - "pointer"
1889 + "tensor_descriptor"
1890 ],
1891 "pid_type": "flat"
1892 },
1893 "intermediate_11008_numtokens_24": {
1894 "block_sizes": [
1895 - 32,
1896 - 8
1895 + 1,
1896 + 1024
1897 ],
1898 "loop_orders": [
1899 [
1900 - 0,
1901 - 1
1900 + 1,
1901 + 0
1902 ]
1903 ],
1904 "flatten_loops": [
1905 false
1906 ],
1907 "l2_groupings": [
1908 - 1
1908 + 64
1909 ],
1910 "range_unroll_factors": [
1911 0
@@ -1921,12 +1921,12 @@
1921 null
1922 ],
1923 "load_eviction_policies": [
1924 - "",
1924 + "last",
1925 "",
1926 ""
1927 ],
1928 - "num_warps": 1,
1929 - "num_stages": 1,
1928 + "num_warps": 2,
1929 + "num_stages": 2,
1930 "indexing": [
1931 "pointer",
1932 "pointer",
@@ -1938,19 +1938,19 @@
1938 "intermediate_14336_numtokens_24": {
1939 "block_sizes": [
1940 8,
1941 - 32
1941 + 512
1942 ],
1943 "loop_orders": [
1944 [
1945 - 1,
1946 - 0
1945 + 0,
1946 + 1
1947 ]
1948 ],
1949 "flatten_loops": [
1950 true
1951 ],
1952 "l2_groupings": [
1953 - 1
1953 + 4
1954 ],
1955 "range_unroll_factors": [
1956 0
@@ -1966,24 +1966,24 @@
1966 null
1967 ],
1968 "load_eviction_policies": [
1969 - "",
1970 - "",
1971 - "last"
1969 + "last",
1970 + "first",
1971 + "first"
1972 ],
1973 - "num_warps": 2,
1974 - "num_stages": 1,
1973 + "num_warps": 32,
1974 + "num_stages": 3,
1975 "indexing": [
1976 "pointer",
1977 + "tensor_descriptor",
1978 "pointer",
1978 - "pointer",
1979 - "pointer"
1979 + "tensor_descriptor"
1980 ],
1981 "pid_type": "flat"
1982 },
1983 "intermediate_2048_numtokens_32": {
1984 "block_sizes": [
1985 32,
1986 - 64
1986 + 16
1987 ],
1988 "loop_orders": [
1989 [
@@ -2015,7 +2015,7 @@
2015 "",
2016 ""
2017 ],
2018 - "num_warps": 1,
2018 + "num_warps": 4,
2019 "num_stages": 1,
2020 "indexing": [
2021 "pointer",
@@ -2027,8 +2027,8 @@
2027 },
2028 "intermediate_2880_numtokens_32": {
2029 "block_sizes": [
2030 - 32,
2031 - 64
2030 + 4,
2031 + 256
2032 ],
2033 "loop_orders": [
2034 [
@@ -2040,7 +2040,7 @@
2040 true
2041 ],
2042 "l2_groupings": [
2043 - 2
2043 + 1
2044 ],
2045 "range_unroll_factors": [
2046 0
@@ -2060,7 +2060,7 @@
2060 "",
2061 ""
2062 ],
2063 - "num_warps": 1,
2063 + "num_warps": 4,
2064 "num_stages": 1,
2065 "indexing": [
2066 "pointer",
@@ -2072,8 +2072,8 @@
2072 },
2073 "intermediate_4096_numtokens_32": {
2074 "block_sizes": [
2075 - 32,
2076 - 16
2075 + 4,
2076 + 4096
2077 ],
2078 "loop_orders": [
2079 [
@@ -2085,7 +2085,7 @@
2085 true
2086 ],
2087 "l2_groupings": [
2088 - 1
2088 + 16
2089 ],
2090 "range_unroll_factors": [
2091 0
@@ -2101,15 +2101,15 @@
2101 null
2102 ],
2103 "load_eviction_policies": [
2104 - "",
2105 - "",
2104 + "first",
2105 + "last",
2106 ""
2107 ],
2108 - "num_warps": 1,
2109 - "num_stages": 2,
2108 + "num_warps": 32,
2109 + "num_stages": 5,
2110 "indexing": [
2111 "pointer",
2112 - "pointer",
2112 + "tensor_descriptor",
2113 "pointer",
2114 "pointer"
2115 ],
@@ -2117,8 +2117,8 @@
2117 },
2118 "intermediate_8192_numtokens_32": {
2119 "block_sizes": [
2120 - 32,
2121 - 128
2120 + 4,
2121 + 1024
2122 ],
2123 "loop_orders": [
2124 [
@@ -2130,7 +2130,7 @@
2130 true
2131 ],
2132 "l2_groupings": [
2133 - 1
2133 + 2
2134 ],
2135 "range_unroll_factors": [
2136 0
@@ -2147,28 +2147,28 @@
2147 ],
2148 "load_eviction_policies": [
2149 "",
2150 - "",
2151 - "last"
2150 + "last",
2151 + ""
2152 ],
2153 "num_warps": 2,
2154 - "num_stages": 2,
2154 + "num_stages": 3,
2155 "indexing": [
2156 + "tensor_descriptor",
2157 + "tensor_descriptor",
2158 "pointer",
2157 - "pointer",
2158 - "pointer",
2159 - "pointer"
2159 + "tensor_descriptor"
2160 ],
2161 "pid_type": "flat"
2162 },
2163 "intermediate_11008_numtokens_32": {
2164 "block_sizes": [
2165 - 16,
2166 - 8
2165 + 2,
2166 + 2048
2167 ],
2168 "loop_orders": [
2169 [
2170 - 1,
2171 - 0
2170 + 0,
2171 + 1
2172 ]
2173 ],
2174 "flatten_loops": [
@@ -2195,20 +2195,20 @@
2195 "",
2196 ""
2197 ],
2198 - "num_warps": 16,
2198 + "num_warps": 8,
2199 "num_stages": 1,
2200 "indexing": [
2201 "pointer",
2202 "pointer",
2203 "pointer",
2204 - "tensor_descriptor"
2204 + "pointer"
2205 ],
2206 "pid_type": "flat"
2207 },
2208 "intermediate_14336_numtokens_32": {
2209 "block_sizes": [
2210 - 32,
2211 - 8
2210 + 1,
2211 + 512
2212 ],
2213 "loop_orders": [
2214 [
@@ -2220,7 +2220,7 @@
2220 true
2221 ],
2222 "l2_groupings": [
2223 - 1
2223 + 64
2224 ],
2225 "range_unroll_factors": [
2226 0
@@ -2236,15 +2236,15 @@
2236 null
2237 ],
2238 "load_eviction_policies": [
2239 - "",
2239 + "first",
2240 "",
2241 ""
2242 ],
2243 - "num_warps": 8,
2244 - "num_stages": 2,
2243 + "num_warps": 4,
2244 + "num_stages": 3,
2245 "indexing": [
2246 "pointer",
2247 - "pointer",
2247 + "tensor_descriptor",
2248 "pointer",
2249 "pointer"
2250 ],
@@ -2252,7 +2252,7 @@
2252 },
2253 "intermediate_2048_numtokens_40": {
2254 "block_sizes": [
2255 - 64,
2255 + 32,
2256 32
2257 ],
2258 "loop_orders": [
@@ -2265,7 +2265,7 @@
2265 true
2266 ],
2267 "l2_groupings": [
2268 - 2
2268 + 1
2269 ],
2270 "range_unroll_factors": [
2271 0
@@ -2285,7 +2285,7 @@
2285 "",
2286 ""
2287 ],
2288 - "num_warps": 2,
2288 + "num_warps": 4,
2289 "num_stages": 1,
2290 "indexing": [
2291 "pointer",
@@ -2297,20 +2297,20 @@
2297 },
2298 "intermediate_2880_numtokens_40": {
2299 "block_sizes": [
2300 - 64,
2301 - 32
2300 + 1,
2301 + 4096
2302 ],
2303 "loop_orders": [
2304 [
2305 - 0,
2306 - 1
2305 + 1,
2306 + 0
2307 ]
2308 ],
2309 "flatten_loops": [
2310 - true
2310 + false
2311 ],
2312 "l2_groupings": [
2313 - 1
2313 + 32
2314 ],
2315 "range_unroll_factors": [
2316 0
@@ -2327,13 +2327,13 @@
2327 ],
2328 "load_eviction_policies": [
2329 "",
2330 - "",
2330 + "last",
2331 "last"
2332 ],
2333 "num_warps": 8,
2334 - "num_stages": 2,
2334 + "num_stages": 4,
2335 "indexing": [
2336 - "pointer",
2336 + "tensor_descriptor",
2337 "pointer",
2338 "pointer",
2339 "pointer"
@@ -2343,12 +2343,12 @@
2343 "intermediate_4096_numtokens_40": {
2344 "block_sizes": [
2345 32,
2346 - 32
2346 + 64
2347 ],
2348 "loop_orders": [
2349 [
2350 - 1,
2351 - 0
2350 + 0,
2351 + 1
2352 ]
2353 ],
2354 "flatten_loops": [
@@ -2378,7 +2378,7 @@
2378 "num_warps": 8,
2379 "num_stages": 1,
2380 "indexing": [
2381 - "tensor_descriptor",
2381 + "pointer",
2382 "pointer",
2383 "pointer",
2384 "pointer"
@@ -2387,8 +2387,8 @@
2387 },
2388 "intermediate_8192_numtokens_40": {
2389 "block_sizes": [
2390 - 64,
2391 - 128
2390 + 2,
2391 + 256
2392 ],
2393 "loop_orders": [
2394 [
@@ -2400,7 +2400,7 @@
2400 true
2401 ],
2402 "l2_groupings": [
2403 - 2
2403 + 16
2404 ],
2405 "range_unroll_factors": [
2406 0
@@ -2417,23 +2417,23 @@
2417 ],
2418 "load_eviction_policies": [
2419 "",
2420 - "",
2421 - ""
2420 + "last",
2421 + "last"
2422 ],
2423 - "num_warps": 4,
2424 - "num_stages": 1,
2423 + "num_warps": 2,
2424 + "num_stages": 2,
2425 "indexing": [
2426 "pointer",
2427 "pointer",
2428 "pointer",
2429 - "pointer"
2429 + "tensor_descriptor"
2430 ],
2431 "pid_type": "flat"
2432 },
2433 "intermediate_11008_numtokens_40": {
2434 "block_sizes": [
2435 - 64,
2436 - 32
2435 + 16,
2436 + 256
2437 ],
2438 "loop_orders": [
2439 [
@@ -2465,7 +2465,7 @@
2465 "",
2466 ""
2467 ],
2468 - "num_warps": 4,
2468 + "num_warps": 8,
2469 "num_stages": 1,
2470 "indexing": [
2471 "pointer",
@@ -2477,8 +2477,8 @@
2477 },
2478 "intermediate_14336_numtokens_40": {
2479 "block_sizes": [
2480 - 32,
2481 - 32
2480 + 4,
2481 + 1024
2482 ],
2483 "loop_orders": [
2484 [
@@ -2497,7 +2497,7 @@
2497 ],
2498 "range_warp_specializes": [],
2499 "range_num_stages": [
2500 - 0
2500 + 1
2501 ],
2502 "range_multi_buffers": [
2503 null
@@ -2506,24 +2506,26 @@
2506 null
2507 ],
2508 "load_eviction_policies": [
2509 + "first",
2510 "",
2510 - "",
2511 - ""
2511 + "last"
2512 ],
2513 - "num_warps": 2,
2514 - "num_stages": 1,
2513 + "num_warps": 16,
2514 + "num_stages": 5,
2515 "indexing": [
2516 - "pointer",
2517 - "pointer",
2518 - "pointer",
2519 - "pointer"
2516 + "tensor_descriptor",
2517 + "tensor_descriptor",
2518 + "tensor_descriptor",
2519 + "tensor_descriptor"
2520 ],
2521 - "pid_type": "flat"
2521 + "pid_type": "persistent_interleaved",
2522 + "num_sm_multiplier": 32,
2523 + "maxnreg": 32
2524 },
2525 "intermediate_2048_numtokens_48": {
2526 "block_sizes": [
2527 32,
2526 - 8
2528 + 32
2529 ],
2530 "loop_orders": [
2531 [
@@ -2535,7 +2537,7 @@
2537 true
2538 ],
2539 "l2_groupings": [
2538 - 2
2540 + 1
2541 ],
2542 "range_unroll_factors": [
2543 0
@@ -2552,23 +2554,23 @@
2554 ],
2555 "load_eviction_policies": [
2556 "",
2555 - "first",
2557 + "",
2558 ""
2559 ],
2558 - "num_warps": 16,
2559 - "num_stages": 2,
2560 + "num_warps": 4,
2561 + "num_stages": 1,
2562 "indexing": [
2563 "pointer",
2564 "pointer",
2563 - "tensor_descriptor",
2565 + "pointer",
2566 "pointer"
2567 ],
2568 "pid_type": "flat"
2569 },
2570 "intermediate_2880_numtokens_48": {
2571 "block_sizes": [
2570 - 8,
2571 - 8
2572 + 16,
2573 + 64
2574 ],
2575 "loop_orders": [
2576 [
@@ -2600,12 +2602,12 @@
2602 "",
2603 ""
2604 ],
2603 - "num_warps": 1,
2605 + "num_warps": 4,
2606 "num_stages": 1,
2607 "indexing": [
2608 "pointer",
2609 "pointer",
2608 - "tensor_descriptor",
2610 + "pointer",
2611 "pointer"
2612 ],
2613 "pid_type": "flat"
@@ -2645,20 +2647,20 @@
2647 "",
2648 ""
2649 ],
2648 - "num_warps": 1,
2650 + "num_warps": 8,
2651 "num_stages": 1,
2652 "indexing": [
2653 "pointer",
2654 "pointer",
2655 "pointer",
2654 - "tensor_descriptor"
2656 + "pointer"
2657 ],
2658 "pid_type": "flat"
2659 },
2660 "intermediate_8192_numtokens_48": {
2661 "block_sizes": [
2660 - 64,
2661 - 32
2662 + 1,
2663 + 4096
2664 ],
2665 "loop_orders": [
2666 [
@@ -2670,7 +2672,7 @@
2672 true
2673 ],
2674 "l2_groupings": [
2673 - 1
2675 + 2
2676 ],
2677 "range_unroll_factors": [
2678 0
@@ -2686,16 +2688,16 @@
2688 null
2689 ],
2690 "load_eviction_policies": [
2691 + "first",
2692 "",
2690 - "",
2691 - ""
2693 + "last"
2694 ],
2695 "num_warps": 4,
2694 - "num_stages": 1,
2696 + "num_stages": 2,
2697 "indexing": [
2698 "pointer",
2697 - "pointer",
2698 - "pointer",
2699 + "tensor_descriptor",
2700 + "tensor_descriptor",
2701 "pointer"
2702 ],
2703 "pid_type": "flat"
@@ -2715,7 +2717,7 @@
2717 true
2718 ],
2719 "l2_groupings": [
2718 - 2
2720 + 1
2721 ],
2722 "range_unroll_factors": [
2723 0
@@ -2735,20 +2737,20 @@
2737 "",
2738 ""
2739 ],
2738 - "num_warps": 32,
2740 + "num_warps": 8,
2741 "num_stages": 1,
2742 "indexing": [
2743 "pointer",
2744 "pointer",
2745 "pointer",
2744 - "tensor_descriptor"
2746 + "pointer"
2747 ],
2748 "pid_type": "flat"
2749 },
2750 "intermediate_14336_numtokens_48": {
2751 "block_sizes": [
2750 - 64,
2751 - 4
2752 + 32,
2753 + 256
2754 ],
2755 "loop_orders": [
2756 [
@@ -2780,10 +2782,10 @@
2782 "",
2783 ""
2784 ],
2783 - "num_warps": 2,
2785 + "num_warps": 8,
2786 "num_stages": 1,
2787 "indexing": [
2786 - "tensor_descriptor",
2788 + "pointer",
2789 "pointer",
2790 "pointer",
2791 "pointer"
@@ -2792,8 +2794,8 @@
2794 },
2795 "intermediate_2048_numtokens_56": {
2796 "block_sizes": [
2795 - 2,
2796 - 8
2797 + 32,
2798 + 32
2799 ],
2800 "loop_orders": [
2801 [
@@ -2825,20 +2827,20 @@
2827 "",
2828 ""
2829 ],
2828 - "num_warps": 1,
2830 + "num_warps": 4,
2831 "num_stages": 1,
2832 "indexing": [
2831 - "tensor_descriptor",
2833 "pointer",
2833 - "tensor_descriptor",
2834 + "pointer",
2835 + "pointer",
2836 "pointer"
2837 ],
2838 "pid_type": "flat"
2839 },
2840 "intermediate_2880_numtokens_56": {
2841 "block_sizes": [
2840 - 8,
2841 - 32
2842 + 1,
2843 + 2048
2844 ],
2845 "loop_orders": [
2846 [
@@ -2870,7 +2872,7 @@
2872 "",
2873 ""
2874 ],
2873 - "num_warps": 4,
2875 + "num_warps": 8,
2876 "num_stages": 1,
2877 "indexing": [
2878 "pointer",
@@ -2883,7 +2885,7 @@
2885 "intermediate_4096_numtokens_56": {
2886 "block_sizes": [
2887 32,
2886 - 4
2888 + 64
2889 ],
2890 "loop_orders": [
2891 [
@@ -2912,15 +2914,15 @@
2914 ],
2915 "load_eviction_policies": [
2916 "",
2915 - "first",
2917 + "",
2918 ""
2919 ],
2918 - "num_warps": 4,
2920 + "num_warps": 8,
2921 "num_stages": 1,
2922 "indexing": [
2923 "pointer",
2924 "pointer",
2923 - "tensor_descriptor",
2925 + "pointer",
2926 "pointer"
2927 ],
2928 "pid_type": "flat"
@@ -2960,7 +2962,7 @@
2962 "",
2963 ""
2964 ],
2963 - "num_warps": 1,
2965 + "num_warps": 8,
2966 "num_stages": 1,
2967 "indexing": [
2968 "pointer",
@@ -2972,8 +2974,8 @@
2974 },
2975 "intermediate_11008_numtokens_56": {
2976 "block_sizes": [
2975 - 32,
2976 - 8
2977 + 1,
2978 + 8192
2979 ],
2980 "loop_orders": [
2981 [
@@ -3002,10 +3004,10 @@
3004 ],
3005 "load_eviction_policies": [
3006 "",
3005 - "first",
3006 - "last"
3007 + "",
3008 + ""
3009 ],
3008 - "num_warps": 1,
3010 + "num_warps": 8,
3011 "num_stages": 1,
3012 "indexing": [
3013 "pointer",
@@ -3017,8 +3019,8 @@
3019 },
3020 "intermediate_14336_numtokens_56": {
3021 "block_sizes": [
3020 - 64,
3021 - 64
3022 + 2,
3023 + 4096
3024 ],
3025 "loop_orders": [
3026 [
@@ -3027,10 +3029,10 @@
3029 ]
3030 ],
3031 "flatten_loops": [
3030 - true
3032 + false
3033 ],
3034 "l2_groupings": [
3033 - 2
3035 + 32
3036 ],
3037 "range_unroll_factors": [
3038 0
@@ -3046,29 +3048,29 @@
3048 null
3049 ],
3050 "load_eviction_policies": [
3049 - "",
3050 - "",
3051 + "first",
3052 + "first",
3053 ""
3054 ],
3053 - "num_warps": 16,
3054 - "num_stages": 1,
3055 + "num_warps": 2,
3056 + "num_stages": 4,
3057 "indexing": [
3058 + "tensor_descriptor",
3059 "pointer",
3057 - "pointer",
3058 - "pointer",
3060 + "tensor_descriptor",
3061 "tensor_descriptor"
3062 ],
3063 "pid_type": "flat"
3064 },
3065 "intermediate_2048_numtokens_64": {
3066 "block_sizes": [
3065 - 16,
3066 - 128
3067 + 64,
3068 + 32
3069 ],
3070 "loop_orders": [
3071 [
3070 - 1,
3071 - 0
3072 + 0,
3073 + 1
3074 ]
3075 ],
3076 "flatten_loops": [
@@ -3098,7 +3100,7 @@
3100 "num_warps": 8,
3101 "num_stages": 1,
3102 "indexing": [
3101 - "tensor_descriptor",
3103 + "pointer",
3104 "pointer",
3105 "pointer",
3106 "pointer"
@@ -3107,13 +3109,13 @@
3109 },
3110 "intermediate_2880_numtokens_64": {
3111 "block_sizes": [
3110 - 4,
3111 - 64
3112 + 1,
3113 + 2048
3114 ],
3115 "loop_orders": [
3116 [
3115 - 1,
3116 - 0
3117 + 0,
3118 + 1
3119 ]
3120 ],
3121 "flatten_loops": [
@@ -3140,7 +3142,7 @@
3142 "",
3143 ""
3144 ],
3143 - "num_warps": 4,
3145 + "num_warps": 8,
3146 "num_stages": 1,
3147 "indexing": [
3148 "pointer",
@@ -3152,8 +3154,8 @@
3154 },
3155 "intermediate_4096_numtokens_64": {
3156 "block_sizes": [
3155 - 2,
3156 - 16
3157 + 64,
3158 + 32
3159 ],
3160 "loop_orders": [
3161 [
@@ -3181,11 +3183,11 @@
3183 null
3184 ],
3185 "load_eviction_policies": [
3184 - "first",
3186 + "",
3187 "",
3188 ""
3189 ],
3188 - "num_warps": 1,
3190 + "num_warps": 8,
3191 "num_stages": 1,
3192 "indexing": [
3193 "pointer",
@@ -3197,8 +3199,8 @@
3199 },
3200 "intermediate_8192_numtokens_64": {
3201 "block_sizes": [
3200 - 8,
3201 - 32
3202 + 64,
3203 + 64
3204 ],
3205 "loop_orders": [
3206 [
@@ -3230,8 +3232,8 @@
3232 "",
3233 ""
3234 ],
3233 - "num_warps": 16,
3234 - "num_stages": 2,
3235 + "num_warps": 8,
3236 + "num_stages": 1,
3237 "indexing": [
3238 "pointer",
3239 "pointer",
@@ -3242,8 +3244,8 @@
3244 },
3245 "intermediate_11008_numtokens_64": {
3246 "block_sizes": [
3245 - 32,
3246 - 32
3247 + 1,
3248 + 8192
3249 ],
3250 "loop_orders": [
3251 [
@@ -3252,7 +3254,7 @@
3254 ]
3255 ],
3256 "flatten_loops": [
3255 - false
3257 + true
3258 ],
3259 "l2_groupings": [
3260 1
@@ -3275,20 +3277,20 @@
3277 "",
3278 ""
3279 ],
3278 - "num_warps": 2,
3280 + "num_warps": 8,
3281 "num_stages": 1,
3282 "indexing": [
3283 "pointer",
3284 "pointer",
3283 - "tensor_descriptor",
3285 + "pointer",
3286 "pointer"
3287 ],
3288 "pid_type": "flat"
3289 },
3290 "intermediate_14336_numtokens_64": {
3291 "block_sizes": [
3290 - 32,
3291 - 8
3292 + 16,
3293 + 512
3294 ],
3295 "loop_orders": [
3296 [
@@ -3300,7 +3302,7 @@
3302 true
3303 ],
3304 "l2_groupings": [
3303 - 1
3305 + 16
3306 ],
3307 "range_unroll_factors": [
3308 0
@@ -3317,23 +3319,23 @@
3319 ],
3320 "load_eviction_policies": [
3321 "",
3320 - "",
3322 + "first",
3323 ""
3324 ],
3323 - "num_warps": 16,
3325 + "num_warps": 4,
3326 "num_stages": 1,
3327 "indexing": [
3328 "pointer",
3327 - "pointer",
3328 - "pointer",
3329 - "pointer"
3329 + "tensor_descriptor",
3330 + "tensor_descriptor",
3331 + "tensor_descriptor"
3332 ],
3333 "pid_type": "flat"
3334 },
3335 "intermediate_2048_numtokens_72": {
3336 "block_sizes": [
3335 - 4,
3336 - 16
3337 + 64,
3338 + 32
3339 ],
3340 "loop_orders": [
3341 [
@@ -3363,9 +3365,9 @@
3365 "load_eviction_policies": [
3366 "",
3367 "",
3366 - "last"
3368 + ""
3369 ],
3368 - "num_warps": 1,
3370 + "num_warps": 8,
3371 "num_stages": 1,
3372 "indexing": [
3373 "pointer",
@@ -3377,8 +3379,8 @@
3379 },
3380 "intermediate_2880_numtokens_72": {
3381 "block_sizes": [
3380 - 64,
3381 - 32
3382 + 32,
3383 + 64
3384 ],
3385 "loop_orders": [
3386 [
@@ -3408,9 +3410,9 @@
3410 "load_eviction_policies": [
3411 "",
3412 "",
3411 - "first"
3413 + ""
3414 ],
3413 - "num_warps": 16,
3415 + "num_warps": 8,
3416 "num_stages": 1,
3417 "indexing": [
3418 "pointer",
@@ -3422,20 +3424,20 @@
3424 },
3425 "intermediate_4096_numtokens_72": {
3426 "block_sizes": [
3425 - 64,
3426 - 16
3427 + 1,
3428 + 1024
3429 ],
3430 "loop_orders": [
3431 [
3430 - 1,
3431 - 0
3432 + 0,
3433 + 1
3434 ]
3435 ],
3436 "flatten_loops": [
3437 true
3438 ],
3439 "l2_groupings": [
3438 - 1
3440 + 4
3441 ],
3442 "range_unroll_factors": [
3443 0
@@ -3451,24 +3453,24 @@
3453 null
3454 ],
3455 "load_eviction_policies": [
3454 - "last",
3456 "",
3456 - ""
3457 + "",
3458 + "first"
3459 ],
3458 - "num_warps": 32,
3459 - "num_stages": 1,
3460 + "num_warps": 16,
3461 + "num_stages": 2,
3462 "indexing": [
3463 + "tensor_descriptor",
3464 "pointer",
3465 "pointer",
3463 - "pointer",
3464 - "pointer"
3466 + "tensor_descriptor"
3467 ],
3468 "pid_type": "flat"
3469 },
3470 "intermediate_8192_numtokens_72": {
3471 "block_sizes": [
3470 - 32,
3471 - 8
3472 + 64,
3473 + 128
3474 ],
3475 "loop_orders": [
3476 [
@@ -3500,7 +3502,7 @@
3502 "",
3503 ""
3504 ],
3503 - "num_warps": 32,
3505 + "num_warps": 8,
3506 "num_stages": 1,
3507 "indexing": [
3508 "pointer",
@@ -3512,8 +3514,8 @@
3514 },
3515 "intermediate_11008_numtokens_72": {
3516 "block_sizes": [
3515 - 32,
3516 - 32
3517 + 4,
3518 + 256
3519 ],
3520 "loop_orders": [
3521 [
@@ -3525,7 +3527,7 @@
3527 true
3528 ],
3529 "l2_groupings": [
3528 - 1
3530 + 2
3531 ],
3532 "range_unroll_factors": [
3533 0
@@ -3541,24 +3543,24 @@
3543 null
3544 ],
3545 "load_eviction_policies": [
3544 - "",
3545 - "",
3546 - ""
3546 + "first",
3547 + "last",
3548 + "last"
3549 ],
3548 - "num_warps": 1,
3549 - "num_stages": 2,
3550 + "num_warps": 32,
3551 + "num_stages": 5,
3552 "indexing": [
3553 "pointer",
3554 "pointer",
3555 "tensor_descriptor",
3554 - "pointer"
3556 + "tensor_descriptor"
3557 ],
3558 "pid_type": "flat"
3559 },
3560 "intermediate_14336_numtokens_72": {
3561 "block_sizes": [
3560 - 128,
3561 - 128
3562 + 32,
3563 + 32
3564 ],
3565 "loop_orders": [
3566 [
@@ -3570,7 +3572,7 @@
3572 true
3573 ],
3574 "l2_groupings": [
3573 - 1
3575 + 2
3576 ],
3577 "range_unroll_factors": [
3578 0
@@ -3586,24 +3588,24 @@
3588 null
3589 ],
3590 "load_eviction_policies": [
3591 + "last",
3592 "",
3590 - "",
3591 - ""
3593 + "last"
3594 ],
3593 - "num_warps": 16,
3594 - "num_stages": 1,
3595 + "num_warps": 8,
3596 + "num_stages": 3,
3597 "indexing": [
3598 "pointer",
3599 "pointer",
3598 - "pointer",
3600 + "tensor_descriptor",
3601 "pointer"
3602 ],
3603 "pid_type": "flat"
3604 },
3605 "intermediate_2048_numtokens_80": {
3606 "block_sizes": [
3605 - 32,
3606 - 64
3607 + 64,
3608 + 32
3609 ],
3610 "loop_orders": [
3611 [
@@ -3635,7 +3637,7 @@
3637 "",
3638 ""
3639 ],
3638 - "num_warps": 4,
3640 + "num_warps": 8,
3641 "num_stages": 1,
3642 "indexing": [
3643 "pointer",
@@ -3648,7 +3650,7 @@
3650 "intermediate_2880_numtokens_80": {
3651 "block_sizes": [
3652 32,
3651 - 128
3653 + 64
3654 ],
3655 "loop_orders": [
3656 [
@@ -3680,7 +3682,7 @@
3682 "",
3683 ""
3684 ],
3683 - "num_warps": 16,
3685 + "num_warps": 8,
3686 "num_stages": 1,
3687 "indexing": [
3688 "pointer",
@@ -3692,8 +3694,8 @@
3694 },
3695 "intermediate_4096_numtokens_80": {
3696 "block_sizes": [
3695 - 32,
3696 - 32
3697 + 64,
3698 + 64
3699 ],
3700 "loop_orders": [
3701 [
@@ -3725,7 +3727,7 @@
3727 "",
3728 ""
3729 ],
3728 - "num_warps": 2,
3730 + "num_warps": 8,
3731 "num_stages": 1,
3732 "indexing": [
3733 "pointer",
@@ -3737,8 +3739,8 @@
3739 },
3740 "intermediate_8192_numtokens_80": {
3741 "block_sizes": [
3740 - 32,
3741 - 8
3742 + 4,
3743 + 128
3744 ],
3745 "loop_orders": [
3746 [
@@ -3766,14 +3768,14 @@
3768 null
3769 ],
3770 "load_eviction_policies": [
3769 - "",
3770 - "",
3771 - ""
3771 + "last",
3772 + "last",
3773 + "last"
3774 ],
3773 - "num_warps": 1,
3774 - "num_stages": 1,
3775 + "num_warps": 16,
3776 + "num_stages": 2,
3777 "indexing": [
3776 - "tensor_descriptor",
3778 + "pointer",
3779 "pointer",
3780 "pointer",
3781 "pointer"
@@ -3782,8 +3784,8 @@
3784 },
3785 "intermediate_11008_numtokens_80": {
3786 "block_sizes": [
3785 - 64,
3786 - 8
3787 + 1,
3788 + 4096
3789 ],
3790 "loop_orders": [
3791 [
@@ -3795,7 +3797,7 @@
3797 true
3798 ],
3799 "l2_groupings": [
3798 - 1
3800 + 32
3801 ],
3802 "range_unroll_factors": [
3803 0
@@ -3811,24 +3813,24 @@
3813 null
3814 ],
3815 "load_eviction_policies": [
3814 - "",
3815 - "",
3816 + "last",
3817 + "last",
3818 ""
3819 ],
3818 - "num_warps": 8,
3819 - "num_stages": 1,
3820 + "num_warps": 16,
3821 + "num_stages": 2,
3822 "indexing": [
3823 "pointer",
3822 - "pointer",
3823 - "pointer",
3824 - "pointer"
3824 + "tensor_descriptor",
3825 + "tensor_descriptor",
3826 + "tensor_descriptor"
3827 ],
3828 "pid_type": "flat"
3829 },
3830 "intermediate_14336_numtokens_80": {
3831 "block_sizes": [
3830 - 32,
3831 - 512
3832 + 2,
3833 + 2048
3834 ],
3835 "loop_orders": [
3836 [
@@ -3840,7 +3842,7 @@
3842 true
3843 ],
3844 "l2_groupings": [
3843 - 2
3845 + 4
3846 ],
3847 "range_unroll_factors": [
3848 0
@@ -3856,24 +3858,24 @@
3858 null
3859 ],
3860 "load_eviction_policies": [
3859 - "",
3860 - "first",
3861 + "last",
3862 + "last",
3863 ""
3864 ],
3863 - "num_warps": 16,
3864 - "num_stages": 1,
3865 + "num_warps": 4,
3866 + "num_stages": 6,
3867 "indexing": [
3866 - "pointer",
3868 "tensor_descriptor",
3869 "pointer",
3869 - "pointer"
3870 + "pointer",
3871 + "tensor_descriptor"
3872 ],
3873 "pid_type": "flat"
3874 },
3875 "intermediate_2048_numtokens_88": {
3876 "block_sizes": [
3875 - 32,
3876 - 16
3877 + 64,
3878 + 32
3879 ],
3880 "loop_orders": [
3881 [
@@ -3905,20 +3907,20 @@
3907 "",
3908 ""
3909 ],
3908 - "num_warps": 4,
3909 - "num_stages": 2,
3910 + "num_warps": 8,
3911 + "num_stages": 1,
3912 "indexing": [
3913 "pointer",
3914 "pointer",
3915 "pointer",
3914 - "tensor_descriptor"
3916 + "pointer"
3917 ],
3918 "pid_type": "flat"
3919 },
3920 "intermediate_2880_numtokens_88": {
3921 "block_sizes": [
3920 - 16,
3921 - 128
3922 + 8,
3923 + 256
3924 ],
3925 "loop_orders": [
3926 [
@@ -3950,25 +3952,25 @@
3952 "",
3953 ""
3954 ],
3953 - "num_warps": 4,
3955 + "num_warps": 8,
3956 "num_stages": 1,
3957 "indexing": [
3956 - "tensor_descriptor",
3958 "pointer",
3959 "pointer",
3959 - "tensor_descriptor"
3960 + "pointer",
3961 + "pointer"
3962 ],
3963 "pid_type": "flat"
3964 },
3965 "intermediate_4096_numtokens_88": {
3966 "block_sizes": [
3967 64,
3966 - 32
3968 + 64
3969 ],
3970 "loop_orders": [
3971 [
3970 - 1,
3971 - 0
3972 + 0,
3973 + 1
3974 ]
3975 ],
3976 "flatten_loops": [
@@ -3995,20 +3997,20 @@
3997 "",
3998 ""
3999 ],
3998 - "num_warps": 32,
4000 + "num_warps": 8,
4001 "num_stages": 1,
4002 "indexing": [
4001 - "tensor_descriptor",
4002 - "tensor_descriptor",
4003 - "tensor_descriptor",
4003 + "pointer",
4004 + "pointer",
4005 + "pointer",
4006 "pointer"
4007 ],
4008 "pid_type": "flat"
4009 },
4010 "intermediate_8192_numtokens_88": {
4011 "block_sizes": [
4010 - 128,
4011 - 64
4012 + 64,
4013 + 128
4014 ],
4015 "loop_orders": [
4016 [
@@ -4036,11 +4038,11 @@
4038 null
4039 ],
4040 "load_eviction_policies": [
4039 - "first",
4040 - "first",
4041 - "last"
4041 + "",
4042 + "",
4043 + ""
4044 ],
4043 - "num_warps": 2,
4045 + "num_warps": 8,
4046 "num_stages": 1,
4047 "indexing": [
4048 "pointer",
@@ -4052,8 +4054,8 @@
4054 },
4055 "intermediate_11008_numtokens_88": {
4056 "block_sizes": [
4055 - 32,
4056 - 128
4057 + 16,
4058 + 2048
4059 ],
4060 "loop_orders": [
4061 [
@@ -4065,7 +4067,7 @@
4067 true
4068 ],
4069 "l2_groupings": [
4068 - 4
4070 + 2
4071 ],
4072 "range_unroll_factors": [
4073 0
@@ -4081,14 +4083,14 @@
4083 null
4084 ],
4085 "load_eviction_policies": [
4084 - "",
4085 - "",
4086 + "first",
4087 + "first",
4088 ""
4089 ],
4090 "num_warps": 32,
4089 - "num_stages": 1,
4091 + "num_stages": 2,
4092 "indexing": [
4091 - "pointer",
4093 + "tensor_descriptor",
4094 "tensor_descriptor",
4095 "pointer",
4096 "pointer"
@@ -4097,8 +4099,8 @@
4099 },
4100 "intermediate_14336_numtokens_88": {
4101 "block_sizes": [
4100 - 16,
4101 - 128
4102 + 4,
4103 + 512
4104 ],
4105 "loop_orders": [
4106 [
@@ -4110,7 +4112,7 @@
4112 true
4113 ],
4114 "l2_groupings": [
4113 - 1
4115 + 16
4116 ],
4117 "range_unroll_factors": [
4118 0
@@ -4126,12 +4128,12 @@
4128 null
4129 ],
4130 "load_eviction_policies": [
4129 - "first",
4130 - "last",
4131 + "",
4132 + "",
4133 ""
4134 ],
4133 - "num_warps": 8,
4134 - "num_stages": 1,
4135 + "num_warps": 32,
4136 + "num_stages": 5,
4137 "indexing": [
4138 "pointer",
4139 "pointer",
@@ -4142,8 +4144,8 @@
4144 },
4145 "intermediate_2048_numtokens_96": {
4146 "block_sizes": [
4145 - 128,
4146 - 4
4147 + 64,
4148 + 32
4149 ],
4150 "loop_orders": [
4151 [
@@ -4155,7 +4157,7 @@
4157 true
4158 ],
4159 "l2_groupings": [
4158 - 2
4160 + 1
4161 ],
4162 "range_unroll_factors": [
4163 0
@@ -4172,10 +4174,10 @@
4174 ],
4175 "load_eviction_policies": [
4176 "",
4175 - "last",
4177 + "",
4178 ""
4179 ],
4178 - "num_warps": 4,
4180 + "num_warps": 8,
4181 "num_stages": 1,
4182 "indexing": [
4183 "pointer",
@@ -4188,7 +4190,7 @@
4190 "intermediate_2880_numtokens_96": {
4191 "block_sizes": [
4192 32,
4191 - 128
4193 + 64
4194 ],
4195 "loop_orders": [
4196 [
@@ -4220,11 +4222,11 @@
4222 "",
4223 ""
4224 ],
4223 - "num_warps": 16,
4225 + "num_warps": 8,
4226 "num_stages": 1,
4227 "indexing": [
4228 "pointer",
4227 - "tensor_descriptor",
4229 + "pointer",
4230 "pointer",
4231 "pointer"
4232 ],
@@ -4232,20 +4234,20 @@
4234 },
4235 "intermediate_4096_numtokens_96": {
4236 "block_sizes": [
4235 - 16,
4236 - 256
4237 + 1,
4238 + 4096
4239 ],
4240 "loop_orders": [
4241 [
4240 - 0,
4241 - 1
4242 + 1,
4243 + 0
4244 ]
4245 ],
4246 "flatten_loops": [
4245 - true
4247 + false
4248 ],
4249 "l2_groupings": [
4248 - 2
4250 + 64
4251 ],
4252 "range_unroll_factors": [
4253 0
@@ -4261,24 +4263,24 @@
4263 null
4264 ],
4265 "load_eviction_policies": [
4264 - "",
4265 - "",
4266 - ""
4266 + "first",
4267 + "last",
4268 + "last"
4269 ],
4268 - "num_warps": 1,
4269 - "num_stages": 3,
4270 + "num_warps": 4,
4271 + "num_stages": 6,
4272 "indexing": [
4273 "pointer",
4274 "pointer",
4273 - "pointer",
4274 - "tensor_descriptor"
4275 + "tensor_descriptor",
4276 + "pointer"
4277 ],
4278 "pid_type": "flat"
4279 },
4280 "intermediate_8192_numtokens_96": {
4281 "block_sizes": [
4282 64,
4281 - 64
4283 + 128
4284 ],
4285 "loop_orders": [
4286 [
@@ -4290,7 +4292,7 @@
4292 true
4293 ],
4294 "l2_groupings": [
4293 - 2
4295 + 1
4296 ],
4297 "range_unroll_factors": [
4298 0
@@ -4306,11 +4308,11 @@
4308 null
4309 ],
4310 "load_eviction_policies": [
4309 - "first",
4311 "",
4311 - "last"
4312 + "",
4313 + ""
4314 ],
4313 - "num_warps": 1,
4315 + "num_warps": 8,
4316 "num_stages": 1,
4317 "indexing": [
4318 "pointer",
@@ -4322,20 +4324,20 @@
4324 },
4325 "intermediate_11008_numtokens_96": {
4326 "block_sizes": [
4325 - 64,
4326 - 256
4327 + 1,
4328 + 2048
4329 ],
4330 "loop_orders": [
4331 [
4330 - 0,
4331 - 1
4332 + 1,
4333 + 0
4334 ]
4335 ],
4336 "flatten_loops": [
4335 - true
4337 + false
4338 ],
4339 "l2_groupings": [
4338 - 1
4340 + 64
4341 ],
4342 "range_unroll_factors": [
4343 0
@@ -4355,20 +4357,20 @@
4357 "",
4358 ""
4359 ],
4358 - "num_warps": 2,
4360 + "num_warps": 32,
4361 "num_stages": 2,
4362 "indexing": [
4363 + "tensor_descriptor",
4364 "pointer",
4365 "pointer",
4363 - "pointer",
4364 - "pointer"
4366 + "tensor_descriptor"
4367 ],
4368 "pid_type": "flat"
4369 },
4370 "intermediate_14336_numtokens_96": {
4371 "block_sizes": [
4370 - 32,
4371 - 64
4372 + 4,
4373 + 4096
4374 ],
4375 "loop_orders": [
4376 [
@@ -4377,10 +4379,10 @@
4379 ]
4380 ],
4381 "flatten_loops": [
4380 - true
4382 + false
4383 ],
4384 "l2_groupings": [
4383 - 1
4385 + 32
4386 ],
4387 "range_unroll_factors": [
4388 0
@@ -4396,24 +4398,24 @@
4398 null
4399 ],
4400 "load_eviction_policies": [
4399 - "",
4400 - "",
4401 - "first"
4401 + "last",
4402 + "first",
4403 + ""
4404 ],
4403 - "num_warps": 1,
4405 + "num_warps": 4,
4406 "num_stages": 1,
4407 "indexing": [
4406 - "pointer",
4407 - "pointer",
4408 "tensor_descriptor",
4409 - "pointer"
4409 + "tensor_descriptor",
4410 + "pointer",
4411 + "tensor_descriptor"
4412 ],
4413 "pid_type": "flat"
4414 },
4415 "intermediate_2048_numtokens_104": {
4416 "block_sizes": [
4415 - 32,
4416 - 8
4417 + 64,
4418 + 32
4419 ],
4420 "loop_orders": [
4421 [
@@ -4445,10 +4447,10 @@
4447 "",
4448 ""
4449 ],
4448 - "num_warps": 16,
4449 - "num_stages": 2,
4450 + "num_warps": 8,
4451 + "num_stages": 1,
4452 "indexing": [
4451 - "tensor_descriptor",
4453 + "pointer",
4454 "pointer",
4455 "pointer",
4456 "pointer"
@@ -4457,8 +4459,8 @@
4459 },
4460 "intermediate_2880_numtokens_104": {
4461 "block_sizes": [
4460 - 64,
4461 - 64
4462 + 8,
4463 + 256
4464 ],
4465 "loop_orders": [
4466 [
@@ -4490,7 +4492,7 @@
4492 "",
4493 ""
4494 ],
4493 - "num_warps": 32,
4495 + "num_warps": 8,
4496 "num_stages": 1,
4497 "indexing": [
4498 "pointer",
@@ -4502,8 +4504,8 @@
4504 },
4505 "intermediate_4096_numtokens_104": {
4506 "block_sizes": [
4505 - 32,
4506 - 32
4507 + 64,
4508 + 64
4509 ],
4510 "loop_orders": [
4511 [
@@ -4512,7 +4514,7 @@
4514 ]
4515 ],
4516 "flatten_loops": [
4515 - false
4517 + true
4518 ],
4519 "l2_groupings": [
4520 1
@@ -4547,8 +4549,8 @@
4549 },
4550 "intermediate_8192_numtokens_104": {
4551 "block_sizes": [
4550 - 8,
4551 - 8
4552 + 64,
4553 + 128
4554 ],
4555 "loop_orders": [
4556 [
@@ -4560,7 +4562,7 @@
4562 true
4563 ],
4564 "l2_groupings": [
4563 - 2
4565 + 1
4566 ],
4567 "range_unroll_factors": [
4568 0
@@ -4586,14 +4588,14 @@
4588 "pointer",
4589 "pointer",
4590 "pointer",
4589 - "tensor_descriptor"
4591 + "pointer"
4592 ],
4593 "pid_type": "flat"
4594 },
4595 "intermediate_11008_numtokens_104": {
4596 "block_sizes": [
4595 - 128,
4596 - 16
4597 + 2,
4598 + 8192
4599 ],
4600 "loop_orders": [
4601 [
@@ -4622,23 +4624,23 @@
4624 ],
4625 "load_eviction_policies": [
4626 "",
4625 - "",
4626 - ""
4627 + "last",
4628 + "first"
4629 ],
4628 - "num_warps": 2,
4630 + "num_warps": 8,
4631 "num_stages": 1,
4632 "indexing": [
4633 "pointer",
4634 "tensor_descriptor",
4633 - "pointer",
4635 + "tensor_descriptor",
4636 "pointer"
4637 ],
4638 "pid_type": "flat"
4639 },
4640 "intermediate_14336_numtokens_104": {
4641 "block_sizes": [
4640 - 32,
4641 - 16
4642 + 8,
4643 + 512
4644 ],
4645 "loop_orders": [
4646 [
@@ -4650,7 +4652,7 @@
4652 true
4653 ],
4654 "l2_groupings": [
4653 - 1
4655 + 2
4656 ],
4657 "range_unroll_factors": [
4658 0
@@ -4666,15 +4668,15 @@
4668 null
4669 ],
4670 "load_eviction_policies": [
4669 - "",
4670 - "",
4671 + "last",
4672 + "last",
4673 ""
4674 ],
4673 - "num_warps": 2,
4675 + "num_warps": 8,
4676 "num_stages": 1,
4677 "indexing": [
4678 "pointer",
4677 - "pointer",
4679 + "tensor_descriptor",
4680 "pointer",
4681 "tensor_descriptor"
4682 ],
@@ -4682,8 +4684,8 @@
4684 },
4685 "intermediate_2048_numtokens_112": {
4686 "block_sizes": [
4685 - 32,
4686 - 1024
4687 + 64,
4688 + 32
4689 ],
4690 "loop_orders": [
4691 [
@@ -4711,15 +4713,15 @@
4713 null
4714 ],
4715 "load_eviction_policies": [
4714 - "first",
4716 + "",
4717 "",
4718 ""
4719 ],
4718 - "num_warps": 32,
4720 + "num_warps": 8,
4721 "num_stages": 1,
4722 "indexing": [
4723 "pointer",
4722 - "tensor_descriptor",
4724 + "pointer",
4725 "pointer",
4726 "pointer"
4727 ],
@@ -4727,8 +4729,8 @@
4729 },
4730 "intermediate_2880_numtokens_112": {
4731 "block_sizes": [
4730 - 32,
4731 - 32
4732 + 2,
4733 + 2048
4734 ],
4735 "loop_orders": [
4736 [
@@ -4737,7 +4739,7 @@
4739 ]
4740 ],
4741 "flatten_loops": [
4740 - false
4742 + true
4743 ],
4744 "l2_groupings": [
4745 1
@@ -4760,7 +4762,7 @@
4762 "",
4763 ""
4764 ],
4763 - "num_warps": 1,
4765 + "num_warps": 8,
4766 "num_stages": 1,
4767 "indexing": [
4768 "pointer",
@@ -4772,8 +4774,8 @@
4774 },
4775 "intermediate_4096_numtokens_112": {
4776 "block_sizes": [
4775 - 32,
4776 - 128
4777 + 64,
4778 + 64
4779 ],
4780 "loop_orders": [
4781 [
@@ -4801,11 +4803,11 @@
4803 null
4804 ],
4805 "load_eviction_policies": [
4804 - "last",
4806 + "",
4807 "",
4808 ""
4809 ],
4808 - "num_warps": 16,
4810 + "num_warps": 8,
4811 "num_stages": 1,
4812 "indexing": [
4813 "pointer",
@@ -4817,8 +4819,8 @@
4819 },
4820 "intermediate_8192_numtokens_112": {
4821 "block_sizes": [
4820 - 32,
4821 - 128
4822 + 4,
4823 + 512
4824 ],
4825 "loop_orders": [
4826 [
@@ -4830,7 +4832,7 @@
4832 true
4833 ],
4834 "l2_groupings": [
4833 - 1
4835 + 16
4836 ],
4837 "range_unroll_factors": [
4838 0
@@ -4846,24 +4848,24 @@
4848 null
4849 ],
4850 "load_eviction_policies": [
4849 - "",
4850 - "",
4851 - ""
4851 + "first",
4852 + "last",
4853 + "last"
4854 ],
4853 - "num_warps": 8,
4854 - "num_stages": 1,
4855 + "num_warps": 16,
4856 + "num_stages": 3,
4857 "indexing": [
4858 "pointer",
4859 "pointer",
4858 - "pointer",
4859 - "tensor_descriptor"
4860 + "tensor_descriptor",
4861 + "pointer"
4862 ],
4863 "pid_type": "flat"
4864 },
4865 "intermediate_11008_numtokens_112": {
4866 "block_sizes": [
4865 - 16,
4866 - 64
4867 + 4,
4868 + 1024
4869 ],
4870 "loop_orders": [
4871 [
@@ -4875,7 +4877,7 @@
4877 true
4878 ],
4879 "l2_groupings": [
4878 - 1
4880 + 2
4881 ],
4882 "range_unroll_factors": [
4883 0
@@ -4892,23 +4894,23 @@
4894 ],
4895 "load_eviction_policies": [
4896 "",
4895 - "",
4896 - ""
4897 + "last",
4898 + "last"
4899 ],
4898 - "num_warps": 8,
4899 - "num_stages": 1,
4900 + "num_warps": 32,
4901 + "num_stages": 3,
4902 "indexing": [
4903 "pointer",
4904 "pointer",
4903 - "pointer",
4904 - "pointer"
4905 + "tensor_descriptor",
4906 + "tensor_descriptor"
4907 ],
4908 "pid_type": "flat"
4909 },
4910 "intermediate_14336_numtokens_112": {
4911 "block_sizes": [
4910 - 32,
4911 - 8
4912 + 64,
4913 + 256
4914 ],
4915 "loop_orders": [
4916 [
@@ -4920,7 +4922,7 @@
4922 true
4923 ],
4924 "l2_groupings": [
4923 - 2
4925 + 1
4926 ],
4927 "range_unroll_factors": [
4928 0
@@ -4940,8 +4942,8 @@
4942 "",
4943 ""
4944 ],
4943 - "num_warps": 1,
4944 - "num_stages": 2,
4945 + "num_warps": 8,
4946 + "num_stages": 1,
4947 "indexing": [
4948 "pointer",
4949 "pointer",
@@ -4952,8 +4954,8 @@
4954 },
4955 "intermediate_2048_numtokens_120": {
4956 "block_sizes": [
4955 - 32,
4956 - 64
4957 + 8,
4958 + 256
4959 ],
4960 "loop_orders": [
4961 [
@@ -4965,7 +4967,7 @@
4967 true
4968 ],
4969 "l2_groupings": [
4968 - 1
4970 + 16
4971 ],
4972 "range_unroll_factors": [
4973 0
@@ -4981,33 +4983,33 @@
4983 null
4984 ],
4985 "load_eviction_policies": [
4984 - "",
4985 - "",
4986 + "first",
4987 + "last",
4988 ""
4989 ],
4988 - "num_warps": 2,
4989 - "num_stages": 1,
4990 + "num_warps": 4,
4991 + "num_stages": 6,
4992 "indexing": [
4993 + "tensor_descriptor",
4994 + "tensor_descriptor",
4995 "pointer",
4992 - "pointer",
4993 - "pointer",
4994 - "pointer"
4996 + "tensor_descriptor"
4997 ],
4998 "pid_type": "flat"
4999 },
5000 "intermediate_2880_numtokens_120": {
5001 "block_sizes": [
5000 - 32,
5001 - 16
5002 + 2,
5003 + 2048
5004 ],
5005 "loop_orders": [
5006 [
5005 - 1,
5006 - 0
5007 + 0,
5008 + 1
5009 ]
5010 ],
5011 "flatten_loops": [
5010 - false
5012 + true
5013 ],
5014 "l2_groupings": [
5015 1
@@ -5030,7 +5032,7 @@
5032 "",
5033 ""
5034 ],
5033 - "num_warps": 4,
5035 + "num_warps": 8,
5036 "num_stages": 1,
5037 "indexing": [
5038 "pointer",
@@ -5042,8 +5044,8 @@
5044 },
5045 "intermediate_4096_numtokens_120": {
5046 "block_sizes": [
5045 - 32,
5046 - 16
5047 + 64,
5048 + 64
5049 ],
5050 "loop_orders": [
5051 [
@@ -5055,7 +5057,7 @@
5057 true
5058 ],
5059 "l2_groupings": [
5058 - 2
5060 + 1
5061 ],
5062 "range_unroll_factors": [
5063 0
@@ -5075,7 +5077,7 @@
5077 "",
5078 ""
5079 ],
5078 - "num_warps": 2,
5080 + "num_warps": 8,
5081 "num_stages": 1,
5082 "indexing": [
5083 "pointer",
@@ -5088,7 +5090,7 @@
5090 "intermediate_8192_numtokens_120": {
5091 "block_sizes": [
5092 64,
5091 - 32
5093 + 128
5094 ],
5095 "loop_orders": [
5096 [
@@ -5117,10 +5119,10 @@
5119 ],
5120 "load_eviction_policies": [
5121 "",
5120 - "last",
5122 + "",
5123 ""
5124 ],
5123 - "num_warps": 2,
5125 + "num_warps": 8,
5126 "num_stages": 1,
5127 "indexing": [
5128 "pointer",
@@ -5132,8 +5134,8 @@
5134 },
5135 "intermediate_11008_numtokens_120": {
5136 "block_sizes": [
5135 - 64,
5136 - 32
5137 + 1,
5138 + 1024
5139 ],
5140 "loop_orders": [
5141 [
@@ -5142,10 +5144,10 @@
5144 ]
5145 ],
5146 "flatten_loops": [
5145 - true
5147 + false
5148 ],
5149 "l2_groupings": [
5148 - 1
5150 + 4
5151 ],
5152 "range_unroll_factors": [
5153 0
@@ -5161,24 +5163,24 @@
5163 null
5164 ],
5165 "load_eviction_policies": [
5164 - "",
5165 - "",
5166 - ""
5166 + "first",
5167 + "last",
5168 + "last"
5169 ],
5168 - "num_warps": 16,
5169 - "num_stages": 1,
5170 + "num_warps": 1,
5171 + "num_stages": 6,
5172 "indexing": [
5171 - "pointer",
5172 - "pointer",
5173 - "pointer",
5174 - "pointer"
5173 + "tensor_descriptor",
5174 + "tensor_descriptor",
5175 + "tensor_descriptor",
5176 + "tensor_descriptor"
5177 ],
5178 "pid_type": "flat"
5179 },
5180 "intermediate_14336_numtokens_120": {
5181 "block_sizes": [
5180 - 128,
5181 - 32
5182 + 32,
5183 + 128
5184 ],
5185 "loop_orders": [
5186 [
@@ -5190,7 +5192,7 @@
5192 true
5193 ],
5194 "l2_groupings": [
5193 - 1
5195 + 2
5196 ],
5197 "range_unroll_factors": [
5198 0
@@ -5206,24 +5208,24 @@
5208 null
5209 ],
5210 "load_eviction_policies": [
5209 - "",
5210 - "",
5211 - ""
5211 + "last",
5212 + "first",
5213 + "last"
5214 ],
5213 - "num_warps": 8,
5215 + "num_warps": 16,
5216 "num_stages": 1,
5217 "indexing": [
5218 "pointer",
5219 "pointer",
5218 - "pointer",
5219 - "pointer"
5220 + "tensor_descriptor",
5221 + "tensor_descriptor"
5222 ],
5223 "pid_type": "flat"
5224 },
5225 "intermediate_2048_numtokens_128": {
5226 "block_sizes": [
5225 - 32,
5226 - 64
5227 + 128,
5228 + 16
5229 ],
5230 "loop_orders": [
5231 [
@@ -5255,8 +5257,8 @@
5257 "",
5258 ""
5259 ],
5258 - "num_warps": 2,
5259 - "num_stages": 2,
5260 + "num_warps": 8,
5261 + "num_stages": 1,
5262 "indexing": [
5263 "pointer",
5264 "pointer",
@@ -5267,8 +5269,8 @@
5269 },
5270 "intermediate_2880_numtokens_128": {
5271 "block_sizes": [
5270 - 128,
5271 - 64
5272 + 2,
5273 + 2048
5274 ],
5275 "loop_orders": [
5276 [
@@ -5297,10 +5299,10 @@
5299 ],
5300 "load_eviction_policies": [
5301 "",
5300 - "last",
5302 + "",
5303 ""
5304 ],
5303 - "num_warps": 4,
5305 + "num_warps": 8,
5306 "num_stages": 1,
5307 "indexing": [
5308 "pointer",
@@ -5343,10 +5345,10 @@
5345 "load_eviction_policies": [
5346 "",
5347 "",
5346 - "last"
5348 + ""
5349 ],
5348 - "num_warps": 32,
5349 - "num_stages": 2,
5350 + "num_warps": 8,
5351 + "num_stages": 1,
5352 "indexing": [
5353 "pointer",
5354 "pointer",
@@ -5357,7 +5359,7 @@
5359 },
5360 "intermediate_8192_numtokens_128": {
5361 "block_sizes": [
5360 - 32,
5362 + 128,
5363 64
5364 ],
5365 "loop_orders": [
@@ -5390,7 +5392,7 @@
5392 "",
5393 ""
5394 ],
5393 - "num_warps": 32,
5395 + "num_warps": 8,
5396 "num_stages": 1,
5397 "indexing": [
5398 "pointer",
@@ -5402,8 +5404,8 @@
5404 },
5405 "intermediate_11008_numtokens_128": {
5406 "block_sizes": [
5405 - 128,
5406 - 128
5407 + 2,
5408 + 1024
5409 ],
5410 "loop_orders": [
5411 [
@@ -5415,7 +5417,7 @@
5417 true
5418 ],
5419 "l2_groupings": [
5418 - 1
5420 + 64
5421 ],
5422 "range_unroll_factors": [
5423 0
@@ -5431,15 +5433,15 @@
5433 null
5434 ],
5435 "load_eviction_policies": [
5434 - "",
5436 + "first",
5437 "",
5438 "last"
5439 ],
5438 - "num_warps": 8,
5440 + "num_warps": 2,
5441 "num_stages": 1,
5442 "indexing": [
5441 - "pointer",
5442 - "pointer",
5443 + "tensor_descriptor",
5444 + "tensor_descriptor",
5445 "pointer",
5446 "pointer"
5447 ],
@@ -5447,8 +5449,8 @@
5449 },
5450 "intermediate_14336_numtokens_128": {
5451 "block_sizes": [
5450 - 16,
5451 - 128
5452 + 4,
5453 + 4096
5454 ],
5455 "loop_orders": [
5456 [
@@ -5460,7 +5462,7 @@
5462 true
5463 ],
5464 "l2_groupings": [
5463 - 2
5465 + 1
5466 ],
5467 "range_unroll_factors": [
5468 0
@@ -5480,8 +5482,8 @@
5482 "",
5483 ""
5484 ],
5483 - "num_warps": 4,
5484 - "num_stages": 3,
5485 + "num_warps": 8,
5486 + "num_stages": 1,
5487 "indexing": [
5488 "pointer",
5489 "pointer",
@@ -5493,7 +5495,7 @@
5495 "intermediate_2048_numtokens_136": {
5496 "block_sizes": [
5497 128,
5496 - 16
5498 + 32
5499 ],
5500 "loop_orders": [
5501 [
@@ -5505,7 +5507,7 @@
5507 true
5508 ],
5509 "l2_groupings": [
5508 - 2
5510 + 1
5511 ],
5512 "range_unroll_factors": [
5513 0
@@ -5525,10 +5527,10 @@
5527 "",
5528 ""
5529 ],
5528 - "num_warps": 1,
5529 - "num_stages": 3,
5530 + "num_warps": 8,
5531 + "num_stages": 1,
5532 "indexing": [
5531 - "tensor_descriptor",
5533 + "pointer",
5534 "pointer",
5535 "pointer",
5536 "pointer"
@@ -5537,7 +5539,7 @@
5539 },
5540 "intermediate_2880_numtokens_136": {
5541 "block_sizes": [
5540 - 8,
5542 + 64,
5543 64
5544 ],
5545 "loop_orders": [
@@ -5567,7 +5569,7 @@
5569 ],
5570 "load_eviction_policies": [
5571 "",
5570 - "first",
5572 + "",
5573 ""
5574 ],
5575 "num_warps": 8,
@@ -5582,8 +5584,8 @@
5584 },
5585 "intermediate_4096_numtokens_136": {
5586 "block_sizes": [
5585 - 32,
5586 - 16
5587 + 128,
5588 + 64
5589 ],
5590 "loop_orders": [
5591 [
@@ -5615,7 +5617,7 @@
5617 "",
5618 ""
5619 ],
5618 - "num_warps": 2,
5620 + "num_warps": 8,
5621 "num_stages": 1,
5622 "indexing": [
5623 "pointer",
@@ -5627,8 +5629,8 @@
5629 },
5630 "intermediate_8192_numtokens_136": {
5631 "block_sizes": [
5630 - 32,
5631 - 128
5632 + 2,
5633 + 512
5634 ],
5635 "loop_orders": [
5636 [
@@ -5640,7 +5642,7 @@
5642 true
5643 ],
5644 "l2_groupings": [
5643 - 1
5645 + 32
5646 ],
5647 "range_unroll_factors": [
5648 0
@@ -5656,24 +5658,24 @@
5658 null
5659 ],
5660 "load_eviction_policies": [
5659 - "",
5661 + "first",
5662 "",
5663 ""
5664 ],
5663 - "num_warps": 4,
5664 - "num_stages": 1,
5665 + "num_warps": 32,
5666 + "num_stages": 7,
5667 "indexing": [
5668 + "tensor_descriptor",
5669 "pointer",
5670 "pointer",
5668 - "pointer",
5669 - "pointer"
5671 + "tensor_descriptor"
5672 ],
5673 "pid_type": "flat"
5674 },
5675 "intermediate_11008_numtokens_136": {
5676 "block_sizes": [
5675 - 16,
5676 - 8
5677 + 4,
5678 + 8192
5679 ],
5680 "loop_orders": [
5681 [
@@ -5682,10 +5684,10 @@
5684 ]
5685 ],
5686 "flatten_loops": [
5685 - true
5687 + false
5688 ],
5689 "l2_groupings": [
5688 - 1
5690 + 8
5691 ],
5692 "range_unroll_factors": [
5693 0
@@ -5702,14 +5704,14 @@
5704 ],
5705 "load_eviction_policies": [
5706 "",
5705 - "",
5706 - ""
5707 + "last",
5708 + "last"
5709 ],
5708 - "num_warps": 4,
5709 - "num_stages": 3,
5710 + "num_warps": 8,
5711 + "num_stages": 4,
5712 "indexing": [
5713 "pointer",
5712 - "tensor_descriptor",
5714 + "pointer",
5715 "pointer",
5716 "pointer"
5717 ],
@@ -5717,8 +5719,8 @@
5719 },
5720 "intermediate_14336_numtokens_136": {
5721 "block_sizes": [
5720 - 32,
5721 - 8
5722 + 4,
5723 + 16384
5724 ],
5725 "loop_orders": [
5726 [
@@ -5730,7 +5732,7 @@
5732 true
5733 ],
5734 "l2_groupings": [
5733 - 1
5735 + 8
5736 ],
5737 "range_unroll_factors": [
5738 0
@@ -5747,14 +5749,14 @@
5749 ],
5750 "load_eviction_policies": [
5751 "",
5750 - "",
5751 - ""
5752 + "last",
5753 + "last"
5754 ],
5753 - "num_warps": 1,
5754 - "num_stages": 1,
5755 + "num_warps": 32,
5756 + "num_stages": 6,
5757 "indexing": [
5756 - "pointer",
5757 - "pointer",
5758 + "tensor_descriptor",
5759 + "tensor_descriptor",
5760 "pointer",
5761 "pointer"
5762 ],
@@ -5762,20 +5764,20 @@
5764 },
5765 "intermediate_2048_numtokens_144": {
5766 "block_sizes": [
5765 - 8,
5766 - 16
5767 + 1,
5768 + 1024
5769 ],
5770 "loop_orders": [
5771 [
5770 - 1,
5771 - 0
5772 + 0,
5773 + 1
5774 ]
5775 ],
5776 "flatten_loops": [
5777 true
5778 ],
5779 "l2_groupings": [
5778 - 2
5780 + 16
5781 ],
5782 "range_unroll_factors": [
5783 0
@@ -5791,15 +5793,15 @@
5793 null
5794 ],
5795 "load_eviction_policies": [
5794 - "",
5795 - "",
5796 - ""
5796 + "last",
5797 + "last",
5798 + "first"
5799 ],
5798 - "num_warps": 2,
5799 - "num_stages": 1,
5800 + "num_warps": 16,
5801 + "num_stages": 7,
5802 "indexing": [
5803 "pointer",
5802 - "tensor_descriptor",
5804 + "pointer",
5805 "pointer",
5806 "pointer"
5807 ],
@@ -5807,8 +5809,8 @@
5809 },
5810 "intermediate_2880_numtokens_144": {
5811 "block_sizes": [
5810 - 256,
5811 - 8
5812 + 64,
5813 + 64
5814 ],
5815 "loop_orders": [
5816 [
@@ -5840,8 +5842,8 @@
5842 "",
5843 ""
5844 ],
5843 - "num_warps": 32,
5844 - "num_stages": 2,
5845 + "num_warps": 8,
5846 + "num_stages": 1,
5847 "indexing": [
5848 "pointer",
5849 "pointer",
@@ -5853,7 +5855,7 @@
5855 "intermediate_4096_numtokens_144": {
5856 "block_sizes": [
5857 128,
5856 - 32
5858 + 64
5859 ],
5860 "loop_orders": [
5861 [
@@ -5862,7 +5864,7 @@
5864 ]
5865 ],
5866 "flatten_loops": [
5865 - false
5867 + true
5868 ],
5869 "l2_groupings": [
5870 1
@@ -5885,7 +5887,7 @@
5887 "",
5888 ""
5889 ],
5888 - "num_warps": 4,
5890 + "num_warps": 8,
5891 "num_stages": 1,
5892 "indexing": [
5893 "pointer",
@@ -5897,20 +5899,20 @@
5899 },
5900 "intermediate_8192_numtokens_144": {
5901 "block_sizes": [
5900 - 128,
5901 - 64
5902 + 1,
5903 + 2048
5904 ],
5905 "loop_orders": [
5906 [
5905 - 0,
5906 - 1
5907 + 1,
5908 + 0
5909 ]
5910 ],
5911 "flatten_loops": [
5912 false
5913 ],
5914 "l2_groupings": [
5913 - 2
5915 + 4
5916 ],
5917 "range_unroll_factors": [
5918 0
@@ -5927,23 +5929,23 @@
5929 ],
5930 "load_eviction_policies": [
5931 "",
5930 - "",
5931 - ""
5932 + "last",
5933 + "first"
5934 ],
5933 - "num_warps": 4,
5934 - "num_stages": 1,
5935 + "num_warps": 1,
5936 + "num_stages": 4,
5937 "indexing": [
5936 - "pointer",
5937 - "pointer",
5938 - "pointer",
5939 - "pointer"
5938 + "tensor_descriptor",
5939 + "tensor_descriptor",
5940 + "tensor_descriptor",
5941 + "tensor_descriptor"
5942 ],
5943 "pid_type": "flat"
5944 },
5945 "intermediate_11008_numtokens_144": {
5946 "block_sizes": [
5945 - 32,
5946 - 4
5947 + 256,
5948 + 16
5949 ],
5950 "loop_orders": [
5951 [
@@ -5955,7 +5957,7 @@
5957 true
5958 ],
5959 "l2_groupings": [
5958 - 1
5960 + 2
5961 ],
5962 "range_unroll_factors": [
5963 0
@@ -5972,13 +5974,13 @@
5974 ],
5975 "load_eviction_policies": [
5976 "",
5975 - "last",
5976 - ""
5977 + "first",
5978 + "first"
5979 ],
5978 - "num_warps": 8,
5979 - "num_stages": 1,
5980 + "num_warps": 16,
5981 + "num_stages": 2,
5982 "indexing": [
5981 - "pointer",
5983 + "tensor_descriptor",
5984 "pointer",
5985 "pointer",
5986 "pointer"
@@ -5987,8 +5989,8 @@
5989 },
5990 "intermediate_14336_numtokens_144": {
5991 "block_sizes": [
5990 - 32,
5991 - 8
5992 + 64,
5993 + 128
5994 ],
5995 "loop_orders": [
5996 [
@@ -6000,7 +6002,7 @@
6002 true
6003 ],
6004 "l2_groupings": [
6003 - 1
6005 + 16
6006 ],
6007 "range_unroll_factors": [
6008 0
@@ -6016,24 +6018,24 @@
6018 null
6019 ],
6020 "load_eviction_policies": [
6021 + "last",
6022 "",
6020 - "",
6021 - "first"
6023 + "last"
6024 ],
6023 - "num_warps": 1,
6024 - "num_stages": 2,
6025 + "num_warps": 16,
6026 + "num_stages": 8,
6027 "indexing": [
6028 "pointer",
6029 "tensor_descriptor",
6028 - "pointer",
6030 + "tensor_descriptor",
6031 "pointer"
6032 ],
6033 "pid_type": "flat"
6034 },
6035 "intermediate_2048_numtokens_152": {
6036 "block_sizes": [
6035 - 32,
6036 - 8
6037 + 4,
6038 + 256
6039 ],
6040 "loop_orders": [
6041 [
@@ -6045,7 +6047,7 @@
6047 true
6048 ],
6049 "l2_groupings": [
6048 - 1
6050 + 8
6051 ],
6052 "range_unroll_factors": [
6053 0
@@ -6061,14 +6063,14 @@
6063 null
6064 ],
6065 "load_eviction_policies": [
6064 - "",
6065 - "",
6066 - ""
6066 + "first",
6067 + "last",
6068 + "first"
6069 ],
6068 - "num_warps": 16,
6069 - "num_stages": 1,
6070 + "num_warps": 8,
6071 + "num_stages": 7,
6072 "indexing": [
6071 - "pointer",
6073 + "tensor_descriptor",
6074 "pointer",
6075 "pointer",
6076 "pointer"
@@ -6077,7 +6079,7 @@
6079 },
6080 "intermediate_2880_numtokens_152": {
6081 "block_sizes": [
6080 - 16,
6082 + 64,
6083 64
6084 ],
6085 "loop_orders": [
@@ -6110,7 +6112,7 @@
6112 "",
6113 ""
6114 ],
6113 - "num_warps": 4,
6115 + "num_warps": 8,
6116 "num_stages": 1,
6117 "indexing": [
6118 "pointer",
@@ -6122,8 +6124,8 @@
6124 },
6125 "intermediate_4096_numtokens_152": {
6126 "block_sizes": [
6125 - 64,
6126 - 4
6127 + 128,
6128 + 64
6129 ],
6130 "loop_orders": [
6131 [
@@ -6135,7 +6137,7 @@
6137 true
6138 ],
6139 "l2_groupings": [
6138 - 2
6140 + 1
6141 ],
6142 "range_unroll_factors": [
6143 0
@@ -6155,7 +6157,7 @@
6157 "",
6158 ""
6159 ],
6158 - "num_warps": 2,
6160 + "num_warps": 8,
6161 "num_stages": 1,
6162 "indexing": [
6163 "pointer",
@@ -6167,8 +6169,8 @@
6169 },
6170 "intermediate_8192_numtokens_152": {
6171 "block_sizes": [
6170 - 32,
6171 - 32
6172 + 64,
6173 + 16
6174 ],
6175 "loop_orders": [
6176 [
@@ -6177,10 +6179,10 @@
6179 ]
6180 ],
6181 "flatten_loops": [
6180 - false
6182 + true
6183 ],
6184 "l2_groupings": [
6183 - 2
6185 + 64
6186 ],
6187 "range_unroll_factors": [
6188 0
@@ -6198,22 +6200,22 @@
6200 "load_eviction_policies": [
6201 "first",
6202 "",
6201 - ""
6203 + "first"
6204 ],
6205 "num_warps": 1,
6204 - "num_stages": 1,
6206 + "num_stages": 2,
6207 "indexing": [
6208 + "tensor_descriptor",
6209 + "tensor_descriptor",
6210 "pointer",
6207 - "pointer",
6208 - "pointer",
6209 - "pointer"
6211 + "tensor_descriptor"
6212 ],
6213 "pid_type": "flat"
6214 },
6215 "intermediate_11008_numtokens_152": {
6216 "block_sizes": [
6215 - 32,
6216 - 32
6217 + 1,
6218 + 4096
6219 ],
6220 "loop_orders": [
6221 [
@@ -6225,7 +6227,7 @@
6227 true
6228 ],
6229 "l2_groupings": [
6228 - 2
6230 + 32
6231 ],
6232 "range_unroll_factors": [
6233 0
@@ -6241,24 +6243,24 @@
6243 null
6244 ],
6245 "load_eviction_policies": [
6246 + "first",
6247 "",
6245 - "",
6246 - ""
6248 + "first"
6249 ],
6248 - "num_warps": 8,
6249 - "num_stages": 1,
6250 + "num_warps": 4,
6251 + "num_stages": 6,
6252 "indexing": [
6251 - "pointer",
6252 - "pointer",
6253 - "pointer",
6254 - "pointer"
6253 + "tensor_descriptor",
6254 + "tensor_descriptor",
6255 + "tensor_descriptor",
6256 + "tensor_descriptor"
6257 ],
6258 "pid_type": "flat"
6259 },
6260 "intermediate_14336_numtokens_152": {
6261 "block_sizes": [
6260 - 64,
6261 - 16
6262 + 2,
6263 + 16384
6264 ],
6265 "loop_orders": [
6266 [
@@ -6270,7 +6272,7 @@
6272 true
6273 ],
6274 "l2_groupings": [
6273 - 1
6275 + 64
6276 ],
6277 "range_unroll_factors": [
6278 0
@@ -6286,24 +6288,24 @@
6288 null
6289 ],
6290 "load_eviction_policies": [
6289 - "",
6291 + "first",
6292 "",
6293 ""
6294 ],
6293 - "num_warps": 2,
6294 - "num_stages": 1,
6295 + "num_warps": 16,
6296 + "num_stages": 6,
6297 "indexing": [
6296 - "pointer",
6298 "tensor_descriptor",
6299 "pointer",
6299 - "pointer"
6300 + "tensor_descriptor",
6301 + "tensor_descriptor"
6302 ],
6303 "pid_type": "flat"
6304 },
6305 "intermediate_2048_numtokens_160": {
6306 "block_sizes": [
6305 - 32,
6306 - 16
6307 + 4,
6308 + 256
6309 ],
6310 "loop_orders": [
6311 [
@@ -6315,7 +6317,7 @@
6317 true
6318 ],
6319 "l2_groupings": [
6318 - 2
6320 + 64
6321 ],
6322 "range_unroll_factors": [
6323 0
@@ -6332,23 +6334,23 @@
6334 ],
6335 "load_eviction_policies": [
6336 "",
6335 - "",
6336 - ""
6337 + "last",
6338 + "last"
6339 ],
6338 - "num_warps": 2,
6339 - "num_stages": 2,
6340 + "num_warps": 1,
6341 + "num_stages": 3,
6342 "indexing": [
6343 + "tensor_descriptor",
6344 + "tensor_descriptor",
6345 "pointer",
6342 - "pointer",
6343 - "pointer",
6344 - "pointer"
6346 + "tensor_descriptor"
6347 ],
6348 "pid_type": "flat"
6349 },
6350 "intermediate_2880_numtokens_160": {
6351 "block_sizes": [
6350 - 128,
6351 - 128
6352 + 64,
6353 + 64
6354 ],
6355 "loop_orders": [
6356 [
@@ -6380,7 +6382,7 @@
6382 "",
6383 ""
6384 ],
6383 - "num_warps": 2,
6385 + "num_warps": 8,
6386 "num_stages": 1,
6387 "indexing": [
6388 "pointer",
@@ -6392,8 +6394,8 @@
6394 },
6395 "intermediate_4096_numtokens_160": {
6396 "block_sizes": [
6395 - 32,
6396 - 8
6397 + 128,
6398 + 64
6399 ],
6400 "loop_orders": [
6401 [
@@ -6425,7 +6427,7 @@
6427 "",
6428 ""
6429 ],
6428 - "num_warps": 2,
6430 + "num_warps": 8,
6431 "num_stages": 1,
6432 "indexing": [
6433 "pointer",
@@ -6438,7 +6440,7 @@
6440 "intermediate_8192_numtokens_160": {
6441 "block_sizes": [
6442 64,
6441 - 4
6443 + 512
6444 ],
6445 "loop_orders": [
6446 [
@@ -6466,24 +6468,24 @@
6468 null
6469 ],
6470 "load_eviction_policies": [
6469 - "first",
6471 + "",
6472 "",
6473 ""
6474 ],
6473 - "num_warps": 4,
6474 - "num_stages": 1,
6475 + "num_warps": 32,
6476 + "num_stages": 4,
6477 "indexing": [
6476 - "pointer",
6477 - "pointer",
6478 - "pointer",
6478 + "tensor_descriptor",
6479 + "tensor_descriptor",
6480 + "tensor_descriptor",
6481 "tensor_descriptor"
6482 ],
6483 "pid_type": "flat"
6484 },
6485 "intermediate_11008_numtokens_160": {
6486 "block_sizes": [
6485 - 32,
6486 - 32
6487 + 128,
6488 + 128
6489 ],
6490 "loop_orders": [
6491 [
@@ -6492,10 +6494,10 @@
6494 ]
6495 ],
6496 "flatten_loops": [
6495 - false
6497 + true
6498 ],
6499 "l2_groupings": [
6498 - 1
6500 + 4
6501 ],
6502 "range_unroll_factors": [
6503 0
@@ -6511,24 +6513,24 @@
6513 null
6514 ],
6515 "load_eviction_policies": [
6514 - "",
6515 - "",
6516 - ""
6516 + "first",
6517 + "first",
6518 + "last"
6519 ],
6518 - "num_warps": 1,
6519 - "num_stages": 1,
6520 + "num_warps": 32,
6521 + "num_stages": 8,
6522 "indexing": [
6523 "pointer",
6524 "pointer",
6525 "pointer",
6524 - "pointer"
6526 + "tensor_descriptor"
6527 ],
6528 "pid_type": "flat"
6529 },
6530 "intermediate_14336_numtokens_160": {
6531 "block_sizes": [
6530 - 128,
6531 - 128
6532 + 1,
6533 + 1024
6534 ],
6535 "loop_orders": [
6536 [
@@ -6537,10 +6539,10 @@
6539 ]
6540 ],
6541 "flatten_loops": [
6540 - true
6542 + false
6543 ],
6544 "l2_groupings": [
6543 - 1
6545 + 8
6546 ],
6547 "range_unroll_factors": [
6548 0
@@ -6558,22 +6560,22 @@
6560 "load_eviction_policies": [
6561 "",
6562 "",
6561 - ""
6563 + "first"
6564 ],
6563 - "num_warps": 2,
6564 - "num_stages": 1,
6565 + "num_warps": 8,
6566 + "num_stages": 8,
6567 "indexing": [
6566 - "pointer",
6568 "tensor_descriptor",
6569 + "pointer",
6570 "tensor_descriptor",
6569 - "pointer"
6571 + "tensor_descriptor"
6572 ],
6573 "pid_type": "flat"
6574 },
6575 "intermediate_2048_numtokens_168": {
6576 "block_sizes": [
6575 - 128,
6576 - 16
6577 + 4,
6578 + 2048
6579 ],
6580 "loop_orders": [
6581 [
@@ -6585,7 +6587,7 @@
6587 true
6588 ],
6589 "l2_groupings": [
6588 - 1
6590 + 32
6591 ],
6592 "range_unroll_factors": [
6593 0
@@ -6601,24 +6603,24 @@
6603 null
6604 ],
6605 "load_eviction_policies": [
6606 + "last",
6607 "",
6605 - "",
6606 - ""
6608 + "last"
6609 ],
6608 - "num_warps": 8,
6609 - "num_stages": 1,
6610 + "num_warps": 32,
6611 + "num_stages": 8,
6612 "indexing": [
6613 + "pointer",
6614 "tensor_descriptor",
6615 "tensor_descriptor",
6613 - "pointer",
6614 - "pointer"
6616 + "tensor_descriptor"
6617 ],
6618 "pid_type": "flat"
6619 },
6620 "intermediate_2880_numtokens_168": {
6621 "block_sizes": [
6620 - 32,
6621 - 16
6622 + 8,
6623 + 512
6624 ],
6625 "loop_orders": [
6626 [
@@ -6650,8 +6652,8 @@
6652 "",
6653 ""
6654 ],
6653 - "num_warps": 4,
6654 - "num_stages": 2,
6655 + "num_warps": 8,
6656 + "num_stages": 1,
6657 "indexing": [
6658 "pointer",
6659 "pointer",
@@ -6662,20 +6664,20 @@
6664 },
6665 "intermediate_4096_numtokens_168": {
6666 "block_sizes": [
6665 - 64,
6666 - 32
6667 + 128,
6668 + 64
6669 ],
6670 "loop_orders": [
6671 [
6670 - 1,
6671 - 0
6672 + 0,
6673 + 1
6674 ]
6675 ],
6676 "flatten_loops": [
6677 true
6678 ],
6679 "l2_groupings": [
6678 - 2
6680 + 1
6681 ],
6682 "range_unroll_factors": [
6683 0
@@ -6693,22 +6695,22 @@
6695 "load_eviction_policies": [
6696 "",
6697 "",
6696 - "last"
6698 + ""
6699 ],
6698 - "num_warps": 1,
6699 - "num_stages": 2,
6700 + "num_warps": 8,
6701 + "num_stages": 1,
6702 "indexing": [
6703 "pointer",
6704 "pointer",
6703 - "tensor_descriptor",
6705 + "pointer",
6706 "pointer"
6707 ],
6708 "pid_type": "flat"
6709 },
6710 "intermediate_8192_numtokens_168": {
6711 "block_sizes": [
6710 - 64,
6711 - 8
6712 + 128,
6713 + 128
6714 ],
6715 "loop_orders": [
6716 [
@@ -6753,7 +6755,7 @@
6755 "intermediate_11008_numtokens_168": {
6756 "block_sizes": [
6757 64,
6756 - 4
6758 + 256
6759 ],
6760 "loop_orders": [
6761 [
@@ -6785,7 +6787,7 @@
6787 "",
6788 ""
6789 ],
6788 - "num_warps": 16,
6790 + "num_warps": 8,
6791 "num_stages": 1,
6792 "indexing": [
6793 "pointer",
@@ -6797,8 +6799,8 @@
6799 },
6800 "intermediate_14336_numtokens_168": {
6801 "block_sizes": [
6800 - 32,
6801 - 512
6802 + 64,
6803 + 32
6804 ],
6805 "loop_orders": [
6806 [
@@ -6810,7 +6812,7 @@
6812 true
6813 ],
6814 "l2_groupings": [
6813 - 2
6815 + 64
6816 ],
6817 "range_unroll_factors": [
6818 0
@@ -6826,36 +6828,36 @@
6828 null
6829 ],
6830 "load_eviction_policies": [
6829 - "",
6830 - "",
6831 - ""
6831 + "first",
6832 + "last",
6833 + "first"
6834 ],
6835 "num_warps": 2,
6834 - "num_stages": 1,
6836 + "num_stages": 6,
6837 "indexing": [
6838 + "tensor_descriptor",
6839 "pointer",
6837 - "pointer",
6838 - "pointer",
6839 - "tensor_descriptor"
6840 + "tensor_descriptor",
6841 + "pointer"
6842 ],
6843 "pid_type": "flat"
6844 },
6845 "intermediate_2048_numtokens_176": {
6846 "block_sizes": [
6845 - 32,
6846 - 128
6847 + 128,
6848 + 32
6849 ],
6850 "loop_orders": [
6851 [
6850 - 1,
6851 - 0
6852 + 0,
6853 + 1
6854 ]
6855 ],
6856 "flatten_loops": [
6857 true
6858 ],
6859 "l2_groupings": [
6858 - 2
6860 + 1
6861 ],
6862 "range_unroll_factors": [
6863 0
@@ -6875,8 +6877,8 @@
6877 "",
6878 ""
6879 ],
6878 - "num_warps": 32,
6879 - "num_stages": 2,
6880 + "num_warps": 8,
6881 + "num_stages": 1,
6882 "indexing": [
6883 "pointer",
6884 "pointer",
@@ -6887,8 +6889,8 @@
6889 },
6890 "intermediate_2880_numtokens_176": {
6891 "block_sizes": [
6890 - 32,
6891 - 32
6892 + 16,
6893 + 256
6894 ],
6895 "loop_orders": [
6896 [
@@ -6920,7 +6922,7 @@
6922 "",
6923 ""
6924 ],
6923 - "num_warps": 2,
6925 + "num_warps": 8,
6926 "num_stages": 1,
6927 "indexing": [
6928 "pointer",
@@ -6932,8 +6934,8 @@
6934 },
6935 "intermediate_4096_numtokens_176": {
6936 "block_sizes": [
6935 - 4,
6936 - 8
6937 + 128,
6938 + 4
6939 ],
6940 "loop_orders": [
6941 [
@@ -6945,7 +6947,7 @@
6947 true
6948 ],
6949 "l2_groupings": [
6948 - 1
6950 + 2
6951 ],
6952 "range_unroll_factors": [
6953 0
@@ -6965,10 +6967,10 @@
6967 "",
6968 "first"
6969 ],
6968 - "num_warps": 8,
6969 - "num_stages": 1,
6970 + "num_warps": 4,
6971 + "num_stages": 6,
6972 "indexing": [
6971 - "pointer",
6973 + "tensor_descriptor",
6974 "pointer",
6975 "pointer",
6976 "pointer"
@@ -6977,8 +6979,8 @@
6979 },
6980 "intermediate_8192_numtokens_176": {
6981 "block_sizes": [
6980 - 8,
6981 - 16
6982 + 1,
6983 + 8192
6984 ],
6985 "loop_orders": [
6986 [
@@ -6987,10 +6989,10 @@
6989 ]
6990 ],
6991 "flatten_loops": [
6990 - true
6992 + false
6993 ],
6994 "l2_groupings": [
6993 - 1
6995 + 16
6996 ],
6997 "range_unroll_factors": [
6998 0
@@ -7006,24 +7008,24 @@
7008 null
7009 ],
7010 "load_eviction_policies": [
7011 + "first",
7012 "",
7010 - "",
7011 - ""
7013 + "first"
7014 ],
7013 - "num_warps": 1,
7014 - "num_stages": 2,
7015 + "num_warps": 16,
7016 + "num_stages": 5,
7017 "indexing": [
7018 + "tensor_descriptor",
7019 + "tensor_descriptor",
7020 "pointer",
7017 - "pointer",
7018 - "pointer",
7019 - "pointer"
7021 + "tensor_descriptor"
7022 ],
7023 "pid_type": "flat"
7024 },
7025 "intermediate_11008_numtokens_176": {
7026 "block_sizes": [
7025 - 8,
7026 - 32
7027 + 64,
7028 + 256
7029 ],
7030 "loop_orders": [
7031 [
@@ -7032,7 +7034,7 @@
7034 ]
7035 ],
7036 "flatten_loops": [
7035 - false
7037 + true
7038 ],

This file is too large to show in full.