File size: 56,243 Bytes
70b4af3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 | """
Block-sparse runtime utilities for CUTE DSL kernels.
This module contains runtime execution functions for block-sparse attention kernels.
These utilities are used by CUTE DSL kernels to produce and consume block-sparse loads.
"""
from typing import Callable, Optional
from functools import partial
import math
import cutlass
import cutlass.cute as cute
from cutlass import Float32, Int32, const_expr
from .quack import copy_utils
# Import data structures from block_sparsity
from .block_sparsity import BlockSparseTensors
from .named_barrier import NamedBarrierBwd
# NOTE [SM100 block-sparse empty tiles: mbarrier contract]
#
# For block-sparse SM100 forward, a given (m_block, stage) Q tile can have zero active
# KV blocks (total_block_cnt == 0). In that case there is no seqlen_kv iteration, so
# the softmax warp-group has no row stats to publish.
#
# The correction warp-group seeds fully-masked-row stats and runs the usual correction
# epilogue so output/LSE have well-defined values. Both warp-groups must still perform
# the softmax<->correction mbarrier handshake so phases advance correctly across
# empty->empty and empty->non-empty tile sequences.
#
# In the no-sink case, this corresponds to the usual fully-masked-row convention:
# output is zero and LSE is -inf.
#
# Barrier contract (each is `mbar_ptr + <offset> + stage`):
#
# Producer/consumer pairs:
# - `mbar_softmax_corr_full` : softmax arrive -> correction wait
# - `mbar_softmax_corr_empty` : correction arrive -> softmax wait
# - `mbar_P_full_O_rescaled` : softmax arrive (+ correction arrive) -> MMA wait
# - `mbar_P_full_2` : softmax arrive -> MMA wait
# - `mbar_corr_epi_full_/empty` : correction <-> epilogue (only when epilogue is separate)
#
# Empty tile (`total_block_cnt == 0`):
# - Softmax: skips the seqlen_kv softmax path entirely (no P stores, no `mbar_P_full_*`).
# It only arrives `mbar_softmax_corr_full` once per stage as a synthetic "no work" signal.
# At the `softmax_loop` level, softmax unconditionally waits `mbar_softmax_corr_empty`
# before each tile (when block-sparse) to drain a prior correction arrival and keep
# phases aligned across non-empty -> empty transitions.
# - Correction: waits `mbar_softmax_corr_full`, seeds stats + runs `correction_epilogue(scale=0)`,
# and arrives `mbar_softmax_corr_empty` (and `mbar_corr_epi_full_/empty` when applicable).
# - No `mbar_P_full_*` barriers are arrived (no P, no MMA O); only the softmax<->correction
# (and correction<->epilogue) handshakes advance phases.
#
# Non-empty tile:
# - Softmax: runs `softmax_step` (produces P) and uses `mbar_softmax_corr_full/empty` to
# publish row_max (during seqlen_kv) and final row stats (once per tile), and to advance phases;
# arrives `mbar_P_full_*` when P is stored.
# - Correction: waits `mbar_softmax_corr_full`, may rescale/release O, arrives `mbar_softmax_corr_empty`
# to ack/advance, and arrives `mbar_P_full_O_rescaled` when MMA can proceed.
#
# Backward (SM100):
# - Empty KV tile: for a given `n_block`, `total_m_block_cnt == 0` means no Q tiles contribute.
# - Both the load and compute loops guard all pipeline work on `process_tile`, so empty tiles
# skip producer/consumer operations entirely (no per-tile mbarrier phase handshake like forward).
# - In the `not dKV_postprocess` path, dK/dV for empty KV tiles are explicitly written as zeros
# even when `process_tile == False` (see `flash_bwd_sm100.py` `should_zero_dKV`).
@cute.jit
def load_block_list(
block_indices: cute.Tensor,
block_count,
load_q_with_first: cutlass.Constexpr,
first_block_preloaded: cutlass.Constexpr,
kv_producer_state,
load_Q,
load_K,
load_V,
pipeline_k,
pipeline_v,
use_tma_q: cutlass.Constexpr,
tma_q_bytes: cutlass.Constexpr,
intra_wg_overlap: cutlass.Constexpr,
):
"""Iterate over the sparse blocks and load K, V (and Q) into the pipeline.
for the intra_wg_overlap case, we overlap the loads of K and V. And this
means we need to pipeline the last V load from the partial block case,
with the loads for the full blocks. Set first_block_preloaded when the
caller has already issued the first K load for the list.
Note:
we iterate along the block_n indices in reverse.
Returns:
Updated kv_producer_state after processing the block list.
"""
if block_count > 0:
if const_expr(not intra_wg_overlap):
# Peel first iteration: the first block may need to load Q alongside K,
# Parameters are already Constexpr, so no need to wrap in const_expr()
n_block_first = block_indices[block_count - 1]
extra_tx = tma_q_bytes if const_expr(load_q_with_first) and const_expr(use_tma_q) else 0
pipeline_k.producer_acquire(kv_producer_state, extra_tx_count=extra_tx)
if const_expr(load_q_with_first and use_tma_q):
load_Q(tma_bar_ptr=pipeline_k.producer_get_barrier(kv_producer_state))
load_K(src_idx=n_block_first, producer_state=kv_producer_state)
pipeline_v.producer_acquire(kv_producer_state)
load_V(src_idx=n_block_first, producer_state=kv_producer_state)
kv_producer_state.advance()
for offset in cutlass.range(1, block_count):
n_block = block_indices[block_count - 1 - offset]
pipeline_k.producer_acquire(kv_producer_state)
load_K(src_idx=n_block, producer_state=kv_producer_state)
pipeline_v.producer_acquire(kv_producer_state)
load_V(src_idx=n_block, producer_state=kv_producer_state)
kv_producer_state.advance()
else:
n_block_first = block_indices[block_count - 1]
if const_expr(not first_block_preloaded):
extra_tx = (
tma_q_bytes if const_expr(load_q_with_first) and const_expr(use_tma_q) else 0
)
pipeline_k.producer_acquire(kv_producer_state, extra_tx_count=extra_tx)
if const_expr(load_q_with_first and use_tma_q):
load_Q(tma_bar_ptr=pipeline_k.producer_get_barrier(kv_producer_state))
load_K(src_idx=n_block_first, producer_state=kv_producer_state)
for idx in cutlass.range(block_count - 1, unroll=1):
n_block_prev = block_indices[block_count - 1 - idx]
n_block = block_indices[block_count - 2 - idx]
kv_producer_state_prev = kv_producer_state.clone()
kv_producer_state.advance()
pipeline_k.producer_acquire(kv_producer_state)
load_K(src_idx=n_block, producer_state=kv_producer_state)
pipeline_v.producer_acquire(kv_producer_state_prev)
load_V(src_idx=n_block_prev, producer_state=kv_producer_state_prev)
return kv_producer_state
@cute.jit
def finish_overlap_v_load(
block_indices: cute.Tensor,
block_count,
load_V,
pipeline_v,
kv_producer_state,
):
"""Load the final V block after overlapped K/V loads."""
if block_count > 0:
n_block_last = block_indices[0]
pipeline_v.producer_acquire(kv_producer_state)
load_V(src_idx=n_block_last, producer_state=kv_producer_state)
kv_producer_state.advance()
return kv_producer_state
@cute.jit
def sparse_tensor_m_block(
m_block,
qhead_per_kvhead: cutlass.Constexpr[int],
q_subtile_factor: cutlass.Constexpr[int],
):
"""Map packed m_block indices to block-sparse tensor indices."""
block = m_block
if const_expr(qhead_per_kvhead != 1):
block = block // qhead_per_kvhead
if const_expr(q_subtile_factor != 1):
block = block // q_subtile_factor
return block
@cute.jit
def produce_block_sparse_loads(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
m_block,
kv_producer_state,
load_Q,
load_K,
load_V,
pipeline_k,
pipeline_v,
use_tma_q: cutlass.Constexpr,
tma_q_bytes: cutlass.Constexpr,
intra_wg_overlap: cutlass.Constexpr,
qhead_per_kvhead: cutlass.Constexpr[int] = 1,
q_subtile_factor: cutlass.Constexpr[int] = 1,
):
"""Iterate over the mask and full block lists for a single tile.
The masked (partial) list may leave the last V load pending when intra-warp-group
overlap is enabled. The first full block must consume that pending V while
issuing its own K load on the next pipeline stage.
In the intra-wg-overlap path, the last masked block leaves its V copy in flight
while we advance the producer state to start the next full K. Either the full list
overlaps that pending V load, or, if no full blocks exist, we explicitly drain it.
Args:
qhead_per_kvhead: Pack-GQA factor. When > 1, m_block is in packed space and
must be converted to unpacked for sparse tensor indexing.
"""
mask_block_cnt, mask_block_idx, full_block_cnt, full_block_idx = blocksparse_tensors
m_block_sparse = sparse_tensor_m_block(m_block, qhead_per_kvhead, q_subtile_factor)
curr_mask_block_cnt = mask_block_cnt[batch_idx, head_idx, m_block_sparse]
curr_mask_block_idx = mask_block_idx[batch_idx, head_idx, m_block_sparse, None]
if const_expr(full_block_cnt is not None):
curr_full_block_cnt = full_block_cnt[batch_idx, head_idx, m_block_sparse]
curr_full_block_idx = full_block_idx[batch_idx, head_idx, m_block_sparse, None]
else:
curr_full_block_cnt = Int32(0)
curr_full_block_idx = None
mask_empty = curr_mask_block_cnt == 0
full_empty = curr_full_block_cnt == 0
if mask_empty:
# No masked blocks: the full list owns the initial Q+K load.
kv_producer_state = load_block_list(
curr_full_block_idx,
curr_full_block_cnt,
load_q_with_first=True,
first_block_preloaded=False,
kv_producer_state=kv_producer_state,
load_Q=load_Q,
load_K=load_K,
load_V=load_V,
pipeline_k=pipeline_k,
pipeline_v=pipeline_v,
use_tma_q=use_tma_q,
tma_q_bytes=tma_q_bytes,
intra_wg_overlap=intra_wg_overlap,
)
if const_expr(intra_wg_overlap) and curr_full_block_cnt > 0:
kv_producer_state = finish_overlap_v_load(
curr_full_block_idx,
curr_full_block_cnt,
load_V,
pipeline_v,
kv_producer_state,
)
else:
# Masked blocks present: load Q together with the first masked K so consumers can
# start immediately. When overlap is disabled this fully drains the list.
kv_producer_state = load_block_list(
curr_mask_block_idx,
curr_mask_block_cnt,
load_q_with_first=True,
first_block_preloaded=False,
kv_producer_state=kv_producer_state,
load_Q=load_Q,
load_K=load_K,
load_V=load_V,
pipeline_k=pipeline_k,
pipeline_v=pipeline_v,
use_tma_q=use_tma_q,
tma_q_bytes=tma_q_bytes,
intra_wg_overlap=intra_wg_overlap,
)
if full_empty:
if const_expr(intra_wg_overlap):
kv_producer_state = finish_overlap_v_load(
curr_mask_block_idx,
curr_mask_block_cnt,
load_V,
pipeline_v,
kv_producer_state,
)
else:
if const_expr(intra_wg_overlap):
# Bridge the masked list to the full list by overlapping the pending masked V
# with the first full K load.
n_block_mask_last = curr_mask_block_idx[0]
n_block_full_first = curr_full_block_idx[curr_full_block_cnt - 1]
kv_producer_state_prev = kv_producer_state.clone()
kv_producer_state.advance()
pipeline_k.producer_acquire(kv_producer_state)
load_K(src_idx=n_block_full_first, producer_state=kv_producer_state)
pipeline_v.producer_acquire(kv_producer_state_prev)
load_V(src_idx=n_block_mask_last, producer_state=kv_producer_state_prev)
kv_producer_state = load_block_list(
curr_full_block_idx,
curr_full_block_cnt,
load_q_with_first=False,
first_block_preloaded=True,
kv_producer_state=kv_producer_state,
load_Q=load_Q,
load_K=load_K,
load_V=load_V,
pipeline_k=pipeline_k,
pipeline_v=pipeline_v,
use_tma_q=use_tma_q,
tma_q_bytes=tma_q_bytes,
intra_wg_overlap=intra_wg_overlap,
)
kv_producer_state = finish_overlap_v_load(
curr_full_block_idx,
curr_full_block_cnt,
load_V,
pipeline_v,
kv_producer_state,
)
else:
# Non-overlap path with both lists: run the full list normally (skipping the Q
# reload because the masked list already issued it).
kv_producer_state = load_block_list(
curr_full_block_idx,
curr_full_block_cnt,
load_q_with_first=False,
first_block_preloaded=False,
kv_producer_state=kv_producer_state,
load_Q=load_Q,
load_K=load_K,
load_V=load_V,
pipeline_k=pipeline_k,
pipeline_v=pipeline_v,
use_tma_q=use_tma_q,
tma_q_bytes=tma_q_bytes,
intra_wg_overlap=intra_wg_overlap,
)
return kv_producer_state
@cute.jit
def consume_block_sparse_loads(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
m_block,
seqlen,
kv_consumer_state,
mma_pv_fn,
mma_one_n_block,
process_first_half_block,
process_last_half_block,
mask_fn,
score_mod_fn,
O_should_accumulate,
mask_mod,
fastdiv_mods,
intra_wg_overlap: cutlass.Constexpr,
warp_scheduler_barrier_sync: Callable,
warp_scheduler_barrier_arrive: Callable,
qhead_per_kvhead: cutlass.Constexpr[int] = 1,
q_subtile_factor: cutlass.Constexpr[int] = 1,
):
"""Consume the mask and full block lists for a single tile on the consumer side.
Mirrors `produce_block_sparse_loads` so that the consumer pipeline uses
the same sparse tensor indexing.
Args:
qhead_per_kvhead: Pack-GQA factor. When > 1, m_block is in packed space and
must be converted to unpacked for sparse tensor indexing.
"""
mask_block_cnt, mask_block_idx, full_block_cnt, full_block_idx = blocksparse_tensors
m_block_sparse = sparse_tensor_m_block(m_block, qhead_per_kvhead, q_subtile_factor)
curr_mask_block_cnt = mask_block_cnt[batch_idx, head_idx, m_block_sparse]
curr_mask_block_idx = mask_block_idx[batch_idx, head_idx, m_block_sparse, None]
curr_full_block_cnt = full_block_cnt[batch_idx, head_idx, m_block_sparse]
curr_full_block_idx = full_block_idx[batch_idx, head_idx, m_block_sparse, None]
processed_any = curr_mask_block_cnt + curr_full_block_cnt > 0
if const_expr(not intra_wg_overlap):
if curr_mask_block_cnt > 0:
mask_n_block = curr_mask_block_idx[curr_mask_block_cnt - 1]
warp_scheduler_barrier_sync()
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=mask_n_block,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(
mask_fn,
mask_mod=mask_mod,
mask_seqlen=True,
fastdiv_mods=fastdiv_mods if cutlass.const_expr(mask_mod is not None) else None,
),
is_first_n_block=True,
)
O_should_accumulate = True
for i in cutlass.range(1, curr_mask_block_cnt):
mask_n_block = curr_mask_block_idx[curr_mask_block_cnt - 1 - i]
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=mask_n_block,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(mask_fn, mask_mod=mask_mod, mask_seqlen=False),
is_first_n_block=False,
)
O_should_accumulate = True
if curr_full_block_cnt == 0:
warp_scheduler_barrier_arrive()
if curr_full_block_cnt > 0:
full_n_block = curr_full_block_idx[curr_full_block_cnt - 1]
if curr_mask_block_cnt == 0:
warp_scheduler_barrier_sync()
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=full_n_block,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(mask_fn, mask_seqlen=True),
is_first_n_block=True,
)
O_should_accumulate = True
for i in cutlass.range(1, curr_full_block_cnt):
full_n_block = curr_full_block_idx[curr_full_block_cnt - 1 - i]
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=full_n_block,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(mask_fn, mask_seqlen=False),
is_first_n_block=False,
)
O_should_accumulate = True
else:
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=full_n_block,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(mask_fn, mask_mod=None, mask_seqlen=True),
is_first_n_block=False,
)
O_should_accumulate = True
for i in cutlass.range(1, curr_full_block_cnt):
full_n_block = curr_full_block_idx[curr_full_block_cnt - 1 - i]
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=full_n_block,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(mask_fn, mask_mod=None, mask_seqlen=False),
is_first_n_block=False,
)
O_should_accumulate = True
warp_scheduler_barrier_arrive()
else:
if curr_mask_block_cnt > 0:
mask_n_block = curr_mask_block_idx[curr_mask_block_cnt - 1]
kv_consumer_state = process_first_half_block(
n_block=mask_n_block,
seqlen=seqlen,
kv_consumer_state=kv_consumer_state,
mask_fn=partial(
mask_fn,
mask_mod=mask_mod,
mask_seqlen=True,
fastdiv_mods=fastdiv_mods if cutlass.const_expr(mask_mod is not None) else None,
),
score_mod_fn=score_mod_fn,
is_first_block=True,
)
for i in cutlass.range(1, curr_mask_block_cnt):
mask_n_block = curr_mask_block_idx[curr_mask_block_cnt - 1 - i]
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=mask_n_block,
seqlen=seqlen,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(mask_fn, mask_mod=mask_mod, mask_seqlen=False),
)
O_should_accumulate = True
if curr_full_block_cnt > 0:
full_n_block = curr_full_block_idx[curr_full_block_cnt - 1]
if curr_mask_block_cnt == 0:
kv_consumer_state = process_first_half_block(
n_block=full_n_block,
seqlen=seqlen,
kv_consumer_state=kv_consumer_state,
mask_fn=partial(mask_fn, mask_mod=None, mask_seqlen=True),
score_mod_fn=score_mod_fn,
is_first_block=True,
)
else:
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=full_n_block,
seqlen=seqlen,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(mask_fn, mask_mod=None, mask_seqlen=True),
)
O_should_accumulate = True
for i in cutlass.range(1, curr_full_block_cnt):
full_n_block = curr_full_block_idx[curr_full_block_cnt - 1 - i]
kv_consumer_state = mma_one_n_block(
kv_consumer_state,
n_block=full_n_block,
seqlen=seqlen,
mma_pv_fn=partial(mma_pv_fn, zero_init=not O_should_accumulate),
mask_fn=partial(mask_fn, mask_mod=None, mask_seqlen=False),
)
O_should_accumulate = True
if curr_mask_block_cnt + curr_full_block_cnt > 0:
kv_consumer_state = process_last_half_block(
kv_consumer_state=kv_consumer_state,
zero_init=not O_should_accumulate,
)
O_should_accumulate = True
return kv_consumer_state, O_should_accumulate, processed_any
@cute.jit
def load_block_list_sm100(
block_indices: cute.Tensor,
block_count,
load_q_with_first: cutlass.Constexpr,
q_stage: cutlass.Constexpr,
kv_producer_state,
load_Q,
load_K,
load_V,
pipeline_kv,
):
"""SM100 version of load_block_list (no intra_wg_overlap, no extra_tx_count)."""
if block_count > 0:
# First iteration: load Q alongside K if requested
n_block_first = block_indices[block_count - 1]
if const_expr(load_q_with_first):
# SM100 loads Q0 and optionally Q1
load_Q(block=0, stage=0)
if const_expr(q_stage == 2):
load_Q(block=1, stage=1)
# SM100 doesn't use producer_acquire for pipeline_kv in load path
# The pipeline barriers are handled inside load_KV
load_K(block=n_block_first, producer_state=kv_producer_state, page_idx=None)
kv_producer_state.advance()
load_V(block=n_block_first, producer_state=kv_producer_state, page_idx=None)
kv_producer_state.advance()
# Remaining blocks
for offset in cutlass.range(1, block_count):
n_block = block_indices[block_count - 1 - offset]
load_K(block=n_block, producer_state=kv_producer_state, page_idx=None)
kv_producer_state.advance()
load_V(block=n_block, producer_state=kv_producer_state, page_idx=None)
kv_producer_state.advance()
return kv_producer_state
# SM100-specific tile processor using SM100 helpers
@cute.jit
def produce_block_sparse_loads_sm100(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
m_block,
kv_producer_state,
load_Q,
load_K,
load_V,
pipeline_kv,
q_stage: cutlass.Constexpr,
q_producer_phase: Int32,
qhead_per_kvhead: cutlass.Constexpr,
q_subtile_factor: cutlass.Constexpr,
):
"""SM100 entry point for sparse block iteration.
SM100 uses PipelineTmaUmma which doesn't support extra_tx_count, so we use
simplified block processing that just calls producer_acquire without extras.
Args:
m_block: which tile of m we are processing
qhead_per_kvhead: Constexpr pack factor
"""
m_block_sparse = sparse_tensor_m_block(m_block, qhead_per_kvhead, q_subtile_factor)
mask_block_cnt, mask_block_idx, full_block_cnt, full_block_idx = blocksparse_tensors
curr_mask_block_cnt = mask_block_cnt[batch_idx, head_idx, m_block_sparse]
curr_mask_block_idx = mask_block_idx[batch_idx, head_idx, m_block_sparse, None]
if const_expr(full_block_cnt is not None):
curr_full_block_cnt = full_block_cnt[batch_idx, head_idx, m_block_sparse]
curr_full_block_idx = full_block_idx[batch_idx, head_idx, m_block_sparse, None]
else:
curr_full_block_cnt = Int32(0)
curr_full_block_idx = None
mask_empty = curr_mask_block_cnt == 0
full_empty = curr_full_block_cnt == 0
q_phase_flipped = False
if mask_empty:
# No masked blocks: process full list with Q loading
kv_producer_state = load_block_list_sm100(
curr_full_block_idx,
curr_full_block_cnt,
load_q_with_first=True,
q_stage=q_stage,
kv_producer_state=kv_producer_state,
load_Q=load_Q,
load_K=load_K,
load_V=load_V,
pipeline_kv=pipeline_kv,
)
q_phase_flipped = not full_empty
else:
# Process masked blocks with Q loading
kv_producer_state = load_block_list_sm100(
curr_mask_block_idx,
curr_mask_block_cnt,
load_q_with_first=True,
q_stage=q_stage,
kv_producer_state=kv_producer_state,
load_Q=load_Q,
load_K=load_K,
load_V=load_V,
pipeline_kv=pipeline_kv,
)
q_phase_flipped = True
if not full_empty:
# Process full blocks without Q loading
kv_producer_state = load_block_list_sm100(
curr_full_block_idx,
curr_full_block_cnt,
load_q_with_first=False,
q_stage=q_stage,
kv_producer_state=kv_producer_state,
load_Q=load_Q,
load_K=load_K,
load_V=load_V,
pipeline_kv=pipeline_kv,
)
if q_phase_flipped:
q_producer_phase ^= 1
return kv_producer_state, q_producer_phase
@cute.jit
def get_total_block_count(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
m_block,
qhead_per_kvhead: cutlass.Constexpr,
q_subtile_factor: cutlass.Constexpr,
):
m_block_sparse = sparse_tensor_m_block(m_block, qhead_per_kvhead, q_subtile_factor)
mask_block_cnt, mask_block_idx, full_block_cnt, full_block_idx = blocksparse_tensors
if const_expr(full_block_cnt is not None):
return (
mask_block_cnt[batch_idx, head_idx, m_block_sparse]
+ full_block_cnt[batch_idx, head_idx, m_block_sparse]
)
else:
return mask_block_cnt[batch_idx, head_idx, m_block_sparse]
@cute.jit
def handle_block_sparse_empty_tile_correction_sm100(
tidx: Int32,
q_stage: cutlass.Constexpr,
m_block_size: cutlass.Constexpr,
qhead_per_kvhead,
pack_gqa: cutlass.Constexpr,
is_split_kv: cutlass.Constexpr,
learnable_sink,
mLSE,
seqlen,
m_block: Int32,
head_idx: Int32,
batch_idx: Int32,
split_idx: Int32,
sScale: cute.Tensor,
stats: list,
correction_epilogue: Callable,
thr_mma_pv: cute.core.ThrMma,
tOtO: cute.Tensor,
sO: cute.Tensor,
pipeline_sm_stats: cutlass.pipeline.PipelineAsync,
sm_stats_barrier: cutlass.pipeline.NamedBarrier,
pipeline_o_epi: cutlass.pipeline.PipelineAsync,
sm_stats_consumer_phase: Int32,
o_corr_consumer_phase: Int32,
corr_epi_producer_phase: Int32,
softmax_scale_log2: Float32,
mO_cur: Optional[cute.Tensor] = None,
gO: Optional[cute.Tensor] = None,
gmem_tiled_copy_O: Optional[cute.TiledCopy] = None,
):
"""Handle SM100 forward block-sparse tiles with no active KV blocks.
This path is taken when `total_block_cnt == 0`. The softmax warp-group still
arrives `mbar_softmax_corr_full` (synthetic "no work") so the correction
warp-group can:
- seed fully-masked-row stats (row_sum=1; row_max=-inf when tracked) for LSE
- run `correction_epilogue` with `scale=0` so the output tile is written as zeros
(independent of any prior tmem contents)
- wait on `mbar_softmax_corr_full` and arrive `mbar_softmax_corr_empty`
(and `mbar_corr_epi_*` when applicable) so phases stay aligned across tiles
This helper intentionally does not touch `mbar_P_full_*` since no P is produced.
See NOTE [SM100 block-sparse empty tiles: mbarrier contract].
"""
LOG2_E = Float32(math.log2(math.e))
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx()) % 4
for stage in cutlass.range_constexpr(q_stage):
row_sum_value = Float32(1.0)
row_max_value = (
-Float32.inf if const_expr(mLSE is not None or learnable_sink is not None) else None
)
if const_expr(learnable_sink is not None):
sink_val = -Float32.inf
if const_expr(not pack_gqa):
sink_val = Float32(learnable_sink[head_idx])
elif tidx < m_block_size:
q_head_idx = (
(q_stage * m_block + stage) * m_block_size + tidx
) % qhead_per_kvhead + head_idx * qhead_per_kvhead
sink_val = Float32(learnable_sink[q_head_idx])
if sink_val != -Float32.inf and (const_expr(not is_split_kv) or split_idx == 0):
if row_max_value == -Float32.inf:
row_max_value = sink_val * (LOG2_E / softmax_scale_log2)
row_sum_value = Float32(1.0)
else:
row_sum_value = row_sum_value + cute.math.exp2(
sink_val * LOG2_E - row_max_value * softmax_scale_log2, fastmath=True
)
if tidx < m_block_size:
scale_row_idx = tidx + stage * m_block_size
sScale[scale_row_idx] = row_sum_value
if const_expr(mLSE is not None or learnable_sink is not None):
sScale[scale_row_idx + q_stage * m_block_size] = row_max_value
acc_flag = row_sum_value == Float32(0.0) or row_sum_value != row_sum_value
stats[stage] = (row_sum_value, row_max_value, acc_flag)
# See NOTE [SM100 block-sparse empty tiles: mbarrier contract].
# pipeline_sm_stats.consumer_wait_w_index_phase(stage, sm_stats_consumer_phase)
sm_stats_barrier.arrive_and_wait_w_index(index=stage * 4 + warp_idx)
pipeline_sm_stats.consumer_release_w_index(stage)
if const_expr(gmem_tiled_copy_O is None):
pipeline_o_epi.producer_acquire_w_index_phase(stage, corr_epi_producer_phase)
correction_epilogue(
thr_mma_pv,
tOtO[None, None, None, stage],
tidx,
stage,
m_block,
seqlen.seqlen_q,
Float32(0.0), # zero scale ensures empty tile writes zeros into staged outputs
sO[None, None, stage],
mO_cur,
gO[None, None, stage],
gmem_tiled_copy_O,
)
if const_expr(gmem_tiled_copy_O is None):
pipeline_o_epi.producer_commit_w_index(stage)
sm_stats_consumer_phase ^= 1
corr_epi_producer_phase ^= 1
return (
sm_stats_consumer_phase,
o_corr_consumer_phase,
corr_epi_producer_phase,
)
@cute.jit
def softmax_block_sparse_sm100(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
m_block,
softmax_step: Callable,
mask_fn: Callable,
mask_fn_none: Callable,
mma_si_consumer_phase: Int32,
si_corr_producer_phase: Int32,
s0_s1_sequence_phase: Int32,
pipeline_sm_stats: cutlass.pipeline.PipelineAsync,
sm_stats_barrier: cutlass.pipeline.NamedBarrier,
q_stage: cutlass.Constexpr,
stage_idx: Int32,
check_m_boundary: bool,
qhead_per_kvhead: cutlass.Constexpr,
q_subtile_factor: cutlass.Constexpr[int] = 1,
):
warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx()) % 4
m_block_sparse = sparse_tensor_m_block(m_block, qhead_per_kvhead, q_subtile_factor)
mask_block_cnt, mask_block_idx, full_block_cnt, full_block_idx = blocksparse_tensors
curr_mask_block_cnt = mask_block_cnt[batch_idx, head_idx, m_block_sparse]
curr_mask_block_idx = mask_block_idx[batch_idx, head_idx, m_block_sparse, None]
if const_expr(full_block_cnt is not None):
curr_full_block_cnt = full_block_cnt[batch_idx, head_idx, m_block_sparse]
curr_full_block_idx = full_block_idx[batch_idx, head_idx, m_block_sparse, None]
else:
curr_full_block_cnt = Int32(0)
curr_full_block_idx = None
total_block_cnt = curr_mask_block_cnt + curr_full_block_cnt
if total_block_cnt == 0:
# See NOTE [SM100 block-sparse empty tiles: mbarrier contract].
# pipeline_sm_stats.producer_commit_w_index(stage_idx)
sm_stats_barrier.arrive_w_index(index=stage_idx * 4 + warp_idx)
else:
if curr_mask_block_cnt > 0:
mask_n_block = curr_mask_block_idx[curr_mask_block_cnt - 1]
(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
) = softmax_step(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
mask_n_block,
is_first=True,
mask_fn=partial(mask_fn, mask_seqlen=True, check_q_boundary=check_m_boundary),
)
for i in cutlass.range(1, curr_mask_block_cnt):
mask_n_block = curr_mask_block_idx[curr_mask_block_cnt - 1 - i]
(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
) = softmax_step(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
mask_n_block,
mask_fn=partial(mask_fn, mask_seqlen=False, check_q_boundary=check_m_boundary),
)
if curr_full_block_cnt > 0:
full_n_block = curr_full_block_idx[curr_full_block_cnt - 1]
if curr_mask_block_cnt == 0:
(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
) = softmax_step(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
full_n_block,
is_first=True,
mask_fn=partial(
mask_fn_none, mask_seqlen=True, check_q_boundary=check_m_boundary
),
)
else:
(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
) = softmax_step(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
full_n_block,
is_first=False,
mask_fn=partial(
mask_fn_none, mask_seqlen=False, check_q_boundary=check_m_boundary
),
)
for i in cutlass.range(1, curr_full_block_cnt):
full_n_block = curr_full_block_idx[curr_full_block_cnt - 1 - i]
(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
) = softmax_step(
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
full_n_block,
mask_fn=partial(
mask_fn_none, mask_seqlen=False, check_q_boundary=check_m_boundary
),
)
return (
mma_si_consumer_phase,
si_corr_producer_phase,
s0_s1_sequence_phase,
total_block_cnt == 0,
)
# =============================================================================
# Backward-specific block-sparse helpers (SM100)
# =============================================================================
#
# In backward, iteration is transposed compared to forward:
# - Forward: outer loop over m_blocks (Q tiles), inner loop over n_blocks (KV tiles)
# - Backward: outer loop over n_blocks (KV tiles), inner loop over m_blocks (Q tiles)
#
# The backward block-sparse tensors use "Q direction" indexing:
# - q_block_cnt[batch, head, n_block] → count of m_blocks to process for this KV tile
# - q_block_idx[batch, head, n_block, :] → indices of m_blocks to process
#
@cute.jit
def get_total_q_block_count_bwd(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
n_block,
subtile_factor: cutlass.Constexpr = 1,
m_block_max: int = 0,
):
"""Count total tile iterations for given n_block (KV tile) in backward."""
q_block_cnt, _, full_block_cnt, _ = blocksparse_tensors
total = q_block_cnt[batch_idx, head_idx, n_block]
if const_expr(full_block_cnt is not None):
total = total + full_block_cnt[batch_idx, head_idx, n_block]
return total * subtile_factor
@cute.jit
def produce_block_sparse_q_loads_bwd_sm100(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
n_block,
# Pipeline states (will be returned after advancing)
producer_state_Q_LSE,
producer_state_dO_dPsum,
# Pipelines
pipeline_Q,
pipeline_LSE,
pipeline_dO,
pipeline_dPsum,
# Load functions
load_K,
load_V,
load_Q,
load_dO,
copy_stats,
# Global tensors for LSE/dPsum
gLSE,
sLSE,
gdPsum,
sdPsum,
# TMA copy bytes for extra_tx_count
tma_copy_bytes_K,
tma_copy_bytes_V,
# Flags for which loads to perform
should_load_Q: cutlass.Constexpr,
should_load_dO: cutlass.Constexpr,
# Subtiling factor and bounds
subtile_factor: cutlass.Constexpr = 1,
m_block_max: int = 0,
):
"""SM100 backward block sparse loading with subtiling.
Returns updated (producer_state_Q_LSE, producer_state_dO_dPsum).
First iteration loads K/V alongside Q/dO; subsequent iterations load only Q/dO.
"""
(
curr_q_cnt,
curr_q_idx,
curr_full_cnt,
curr_full_idx,
loop_count,
) = get_block_sparse_iteration_info_bwd(
blocksparse_tensors, batch_idx, head_idx, n_block, subtile_factor, m_block_max
)
for iter_idx in cutlass.range(loop_count, unroll=1):
m_block, _ = get_m_block_from_iter_bwd(
iter_idx,
curr_q_cnt,
curr_q_idx,
curr_full_cnt,
curr_full_idx,
subtile_factor,
m_block_max,
)
m_block_safe = m_block
if m_block_max > 0:
m_block_safe = cutlass.min(m_block, m_block_max - 1)
if iter_idx == 0:
# First block: load K/V alongside Q/dO
if const_expr(should_load_Q):
pipeline_Q.producer_acquire(producer_state_Q_LSE, extra_tx_count=tma_copy_bytes_K)
load_K(tma_bar_ptr=pipeline_Q.producer_get_barrier(producer_state_Q_LSE))
load_Q(m_block_safe, producer_state=producer_state_Q_LSE)
pipeline_Q.producer_commit(producer_state_Q_LSE)
pipeline_LSE.producer_acquire(producer_state_Q_LSE)
with cute.arch.elect_one():
copy_stats(
gLSE[None, m_block_safe],
sLSE[None, producer_state_Q_LSE.index],
mbar_ptr=pipeline_LSE.producer_get_barrier(producer_state_Q_LSE),
)
producer_state_Q_LSE.advance()
if const_expr(should_load_dO):
pipeline_dO.producer_acquire(
producer_state_dO_dPsum, extra_tx_count=tma_copy_bytes_V
)
load_V(tma_bar_ptr=pipeline_dO.producer_get_barrier(producer_state_dO_dPsum))
load_dO(m_block_safe, producer_state=producer_state_dO_dPsum)
pipeline_dO.producer_commit(producer_state_dO_dPsum)
pipeline_dPsum.producer_acquire(producer_state_dO_dPsum)
with cute.arch.elect_one():
copy_stats(
gdPsum[None, m_block_safe],
sdPsum[None, producer_state_dO_dPsum.index],
mbar_ptr=pipeline_dPsum.producer_get_barrier(producer_state_dO_dPsum),
)
producer_state_dO_dPsum.advance()
else:
# Subsequent blocks: just load Q/dO (K/V already loaded)
if const_expr(should_load_Q):
pipeline_Q.producer_acquire(producer_state_Q_LSE)
load_Q(m_block_safe, producer_state=producer_state_Q_LSE)
pipeline_Q.producer_commit(producer_state_Q_LSE)
pipeline_LSE.producer_acquire(producer_state_Q_LSE)
with cute.arch.elect_one():
copy_stats(
gLSE[None, m_block_safe],
sLSE[None, producer_state_Q_LSE.index],
mbar_ptr=pipeline_LSE.producer_get_barrier(producer_state_Q_LSE),
)
producer_state_Q_LSE.advance()
if const_expr(should_load_dO):
pipeline_dO.producer_acquire(producer_state_dO_dPsum)
load_dO(m_block_safe, producer_state=producer_state_dO_dPsum)
pipeline_dO.producer_commit(producer_state_dO_dPsum)
pipeline_dPsum.producer_acquire(producer_state_dO_dPsum)
with cute.arch.elect_one():
copy_stats(
gdPsum[None, m_block_safe],
sdPsum[None, producer_state_dO_dPsum.index],
mbar_ptr=pipeline_dPsum.producer_get_barrier(producer_state_dO_dPsum),
)
producer_state_dO_dPsum.advance()
return producer_state_Q_LSE, producer_state_dO_dPsum
@cute.jit
def get_block_sparse_iteration_info_bwd(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
n_block,
subtile_factor: cutlass.Constexpr = 1,
m_block_max: int = 0,
):
"""Extract block-sparse iteration info for backward pass.
Returns (curr_q_cnt, curr_q_idx, curr_full_cnt, curr_full_idx, total_count).
"""
q_cnt, q_idx, full_cnt, full_idx = blocksparse_tensors
curr_q_cnt = q_cnt[batch_idx, head_idx, n_block]
curr_q_idx = q_idx[batch_idx, head_idx, n_block, None]
if const_expr(full_cnt is not None):
curr_full_cnt = full_cnt[batch_idx, head_idx, n_block]
curr_full_idx = full_idx[batch_idx, head_idx, n_block, None]
else:
curr_full_cnt = Int32(0)
curr_full_idx = None
sparse_block_count = curr_q_cnt
if const_expr(full_cnt is not None):
sparse_block_count = sparse_block_count + curr_full_cnt
total_count = sparse_block_count * subtile_factor
return curr_q_cnt, curr_q_idx, curr_full_cnt, curr_full_idx, total_count
@cute.jit
def get_m_block_from_iter_bwd(
iter_idx,
curr_q_cnt,
curr_q_idx: cute.Tensor,
curr_full_cnt,
curr_full_idx: Optional[cute.Tensor],
subtile_factor: cutlass.Constexpr = 1,
m_block_max: int = 0,
):
"""Derive m_block index and is_full_block flag from iteration index.
Returns (m_block, is_full_block):
- m_block: The actual Q-tile block index
- is_full_block: True if this is a full block (no mask_mod needed)
"""
sparse_iter_idx = iter_idx // subtile_factor
subtile_offset = iter_idx % subtile_factor
sparse_m_block = Int32(0)
is_full_block = False
if const_expr(curr_full_idx is not None):
if sparse_iter_idx < curr_q_cnt:
sparse_m_block = curr_q_idx[sparse_iter_idx]
else:
sparse_m_block = curr_full_idx[sparse_iter_idx - curr_q_cnt]
is_full_block = True
else:
sparse_m_block = curr_q_idx[sparse_iter_idx]
return sparse_m_block * subtile_factor + subtile_offset, is_full_block
@cute.jit
def _load_q_do_block_sm90(
m_block,
producer_state_Q,
producer_state_dO,
pipeline_Q,
pipeline_dO,
load_K,
load_V,
load_Q,
load_dO,
load_LSE,
load_dPsum,
tma_copy_bytes_K,
tma_copy_bytes_V,
Q_stage_eq_dO_stage: cutlass.Constexpr,
load_kv: bool,
):
"""Load one Q/dO block, optionally loading K/V on first iteration."""
if load_kv:
pipeline_Q.producer_acquire(producer_state_Q, extra_tx_count=tma_copy_bytes_K)
load_K(tma_bar_ptr=pipeline_Q.producer_get_barrier(producer_state_Q))
else:
pipeline_Q.producer_acquire(producer_state_Q)
load_Q(m_block, producer_state=producer_state_Q)
load_LSE(m_block, producer_state=producer_state_Q)
producer_state_dO_cur = (
producer_state_dO if const_expr(not Q_stage_eq_dO_stage) else producer_state_Q
)
if load_kv:
pipeline_dO.producer_acquire(producer_state_dO_cur, extra_tx_count=tma_copy_bytes_V)
load_V(tma_bar_ptr=pipeline_dO.producer_get_barrier(producer_state_dO_cur))
else:
pipeline_dO.producer_acquire(producer_state_dO_cur)
load_dO(m_block, producer_state=producer_state_dO_cur)
load_dPsum(m_block, producer_state=producer_state_dO_cur)
producer_state_Q.advance()
producer_state_dO.advance()
return producer_state_Q, producer_state_dO
@cute.jit
def produce_block_sparse_q_loads_bwd_sm90(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
n_block,
producer_state_Q,
producer_state_dO,
pipeline_Q,
pipeline_dO,
load_K,
load_V,
load_Q,
load_dO,
load_LSE,
load_dPsum,
tma_copy_bytes_K,
tma_copy_bytes_V,
Q_stage_eq_dO_stage: cutlass.Constexpr,
subtile_factor: cutlass.Constexpr,
m_block_max: int,
):
"""SM90 backward block sparse loading with separate partial/full loops.
K/V are loaded with the first valid block. Iterates partial blocks first,
then full blocks, matching consumer order.
Returns updated (producer_state_Q, producer_state_dO).
"""
q_cnt, q_idx, full_cnt, full_idx = blocksparse_tensors
curr_q_cnt = q_cnt[batch_idx, head_idx, n_block]
curr_q_idx = q_idx[batch_idx, head_idx, n_block, None]
if const_expr(full_cnt is not None):
curr_full_cnt = full_cnt[batch_idx, head_idx, n_block]
curr_full_idx = full_idx[batch_idx, head_idx, n_block, None]
else:
curr_full_cnt = Int32(0)
curr_full_idx = None
kv_loaded = False
for iter_idx in cutlass.range(curr_q_cnt * subtile_factor, unroll=1):
sparse_idx = iter_idx // subtile_factor
subtile_offset = iter_idx % subtile_factor
m_block = curr_q_idx[sparse_idx] * subtile_factor + subtile_offset
if m_block < m_block_max:
producer_state_Q, producer_state_dO = _load_q_do_block_sm90(
m_block,
producer_state_Q,
producer_state_dO,
pipeline_Q,
pipeline_dO,
load_K,
load_V,
load_Q,
load_dO,
load_LSE,
load_dPsum,
tma_copy_bytes_K,
tma_copy_bytes_V,
Q_stage_eq_dO_stage,
load_kv=not kv_loaded,
)
kv_loaded = True
if const_expr(full_cnt is not None):
for iter_idx in cutlass.range(curr_full_cnt * subtile_factor, unroll=1):
sparse_idx = iter_idx // subtile_factor
subtile_offset = iter_idx % subtile_factor
m_block = curr_full_idx[sparse_idx] * subtile_factor + subtile_offset
if m_block < m_block_max:
producer_state_Q, producer_state_dO = _load_q_do_block_sm90(
m_block,
producer_state_Q,
producer_state_dO,
pipeline_Q,
pipeline_dO,
load_K,
load_V,
load_Q,
load_dO,
load_LSE,
load_dPsum,
tma_copy_bytes_K,
tma_copy_bytes_V,
Q_stage_eq_dO_stage,
load_kv=not kv_loaded,
)
kv_loaded = True
return producer_state_Q, producer_state_dO
@cute.jit
def consume_block_sparse_mma_bwd_sm90(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
n_block,
consumer_state_Q,
consumer_state_dO,
mma_one_m_block_fn,
mask,
mask_mod,
is_causal: cutlass.Constexpr,
is_local: cutlass.Constexpr,
thr_mma_SdP,
score_mod_fn=None,
score_mod_bwd_fn=None,
subtile_factor: cutlass.Constexpr = 1,
m_block_max: int = 0,
aux_tensors=None,
fastdiv_mods=(None, None),
):
"""SM90 backward block sparse MMA consumption with separate partial/full loops.
Partial blocks are processed first (with mask_mod applied), then full blocks
(without mask_mod). This ensures mask_mod is only applied where needed.
Returns updated (consumer_state_Q, consumer_state_dO).
"""
q_cnt, q_idx, full_cnt, full_idx = blocksparse_tensors
curr_q_cnt = q_cnt[batch_idx, head_idx, n_block]
curr_q_idx = q_idx[batch_idx, head_idx, n_block, None]
if const_expr(full_cnt is not None):
curr_full_cnt = full_cnt[batch_idx, head_idx, n_block]
curr_full_idx = full_idx[batch_idx, head_idx, n_block, None]
else:
curr_full_cnt = Int32(0)
curr_full_idx = None
dKV_accumulate = False
mask_fn_partial = partial(
mask.apply_mask,
batch_idx=batch_idx,
head_idx=head_idx,
n_block=n_block,
thr_mma=thr_mma_SdP,
mask_seqlen=True,
mask_causal=is_causal,
mask_local=is_local,
mask_mod=mask_mod,
aux_tensors=aux_tensors,
fastdiv_mods=fastdiv_mods,
)
mask_fn_full = partial(
mask.apply_mask,
batch_idx=batch_idx,
head_idx=head_idx,
n_block=n_block,
thr_mma=thr_mma_SdP,
mask_seqlen=True,
mask_causal=is_causal,
mask_local=is_local,
aux_tensors=aux_tensors,
fastdiv_mods=fastdiv_mods,
)
for iter_idx in cutlass.range(curr_q_cnt * subtile_factor, unroll=1):
sparse_idx = iter_idx // subtile_factor
subtile_offset = iter_idx % subtile_factor
m_block = curr_q_idx[sparse_idx] * subtile_factor + subtile_offset
if m_block < m_block_max:
consumer_state_Q, consumer_state_dO = mma_one_m_block_fn(
m_block,
consumer_state_Q,
consumer_state_dO,
mask_fn=mask_fn_partial,
score_mod_fn=score_mod_fn,
score_mod_bwd_fn=score_mod_bwd_fn,
dKV_accumulate=dKV_accumulate,
)
dKV_accumulate = True
if const_expr(full_cnt is not None):
for iter_idx in cutlass.range(curr_full_cnt * subtile_factor, unroll=1):
sparse_idx = iter_idx // subtile_factor
subtile_offset = iter_idx % subtile_factor
m_block = curr_full_idx[sparse_idx] * subtile_factor + subtile_offset
if m_block < m_block_max:
consumer_state_Q, consumer_state_dO = mma_one_m_block_fn(
m_block,
consumer_state_Q,
consumer_state_dO,
mask_fn=mask_fn_full,
score_mod_fn=score_mod_fn,
score_mod_bwd_fn=score_mod_bwd_fn,
dKV_accumulate=dKV_accumulate,
)
dKV_accumulate = True
return consumer_state_Q, consumer_state_dO
@cute.jit
def _store_one_dQaccum_sm90(
m_block,
sdQaccum: cute.Tensor,
gdQaccum: cute.Tensor,
num_mma_warp_groups: cutlass.Constexpr,
num_threads_per_warp_group: cutlass.Constexpr,
tma_copy_bytes_dQ,
):
"""Store dQaccum for a single m_block."""
for warp_group_idx in cutlass.range_constexpr(num_mma_warp_groups):
cute.arch.cp_async_bulk_wait_group(num_mma_warp_groups - 1 - warp_group_idx, read=True)
cute.arch.barrier_arrive(
barrier_id=int(NamedBarrierBwd.dQEmptyWG0) + warp_group_idx,
number_of_threads=num_threads_per_warp_group + cute.arch.WARP_SIZE,
)
for warp_group_idx in cutlass.range_constexpr(num_mma_warp_groups):
cute.arch.barrier(
barrier_id=int(NamedBarrierBwd.dQFullWG0) + warp_group_idx,
number_of_threads=num_threads_per_warp_group + cute.arch.WARP_SIZE,
)
with cute.arch.elect_one():
copy_utils.cpasync_reduce_bulk_add_f32(
sdQaccum[None, warp_group_idx].iterator,
gdQaccum[None, warp_group_idx, m_block].iterator,
tma_copy_bytes_dQ,
)
cute.arch.cp_async_bulk_commit_group()
@cute.jit
def dQaccum_store_block_sparse_bwd_sm90(
blocksparse_tensors: BlockSparseTensors,
batch_idx,
head_idx,
n_block,
sdQaccum: cute.Tensor,
gdQaccum: cute.Tensor,
subtile_factor: cutlass.Constexpr,
m_block_max: int,
num_mma_warp_groups: cutlass.Constexpr,
num_threads_per_warp_group: cutlass.Constexpr,
tma_copy_bytes_dQ,
):
"""SM90 backward block sparse dQaccum store with separate partial/full loops.
Iterates partial blocks first, then full blocks, matching producer/consumer order.
"""
q_cnt, q_idx, full_cnt, full_idx = blocksparse_tensors
curr_q_cnt = q_cnt[batch_idx, head_idx, n_block]
curr_q_idx = q_idx[batch_idx, head_idx, n_block, None]
if const_expr(full_cnt is not None):
curr_full_cnt = full_cnt[batch_idx, head_idx, n_block]
curr_full_idx = full_idx[batch_idx, head_idx, n_block, None]
else:
curr_full_cnt = Int32(0)
curr_full_idx = None
for iter_idx in cutlass.range(curr_q_cnt * subtile_factor, unroll=1):
sparse_idx = iter_idx // subtile_factor
subtile_offset = iter_idx % subtile_factor
m_block = curr_q_idx[sparse_idx] * subtile_factor + subtile_offset
if m_block < m_block_max:
_store_one_dQaccum_sm90(
m_block,
sdQaccum,
gdQaccum,
num_mma_warp_groups,
num_threads_per_warp_group,
tma_copy_bytes_dQ,
)
if const_expr(full_cnt is not None):
for iter_idx in cutlass.range(curr_full_cnt * subtile_factor, unroll=1):
sparse_idx = iter_idx // subtile_factor
subtile_offset = iter_idx % subtile_factor
m_block = curr_full_idx[sparse_idx] * subtile_factor + subtile_offset
if m_block < m_block_max:
_store_one_dQaccum_sm90(
m_block,
sdQaccum,
gdQaccum,
num_mma_warp_groups,
num_threads_per_warp_group,
tma_copy_bytes_dQ,
)
|