vllm.model_executor.layers.sparse_attn_indexer
¶
Custom Sparse Attention Indexer layers.
Classes:
-
SparseAttnIndexer–Sparse Attention Indexer Custom Op Layer. This layer is extracted as a
Functions:
-
as_fp32_logits–Upcast BF16 MQA logits for the FP32 top-k kernels.
-
dcp_gather_kv_rows–All-gather this rank's KV shard and return it in global token order.
-
fuse_indexer_wk_weights_proj–Fill the fused BF16
wk_weights_projweight in place (captured CUDA -
init_indexer_wk_weights_proj–Build the indexer's checkpoint layers
wk(BF16, or FP8 as the -
kv_cache_as_quant_view–4D
[num_blocks, block_size, 1, head_width]view expected by -
should_skip_indexer_scoring–Whether the indexer can skip scoring because the main MLA runs a pure
SparseAttnIndexer
¶
Bases: CustomOp
Sparse Attention Indexer Custom Op Layer. This layer is extracted as a
separate custom op since it involves heavy custom kernels like mqa_logits,
paged_mqa_logits and top_k_per_row, etc. Those kernels maybe requires
specific memory layout or implementation for different hardware backends to
achieve optimal performance.
For now, the default native path will use CUDA backend path. Other platform
may requires add the corresponding Custom Op name sparse_attn_indexer to
custom_ops in CompilationConfig to enable the platform specific path.
Methods:
-
forward_cpu–CPU sparse attention indexer: cache write stays eager Python glue
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
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 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 | |
forward_cpu(hidden_states, q_quant, k, weights)
¶
CPU sparse attention indexer: cache write stays eager Python glue
(own K-cache layout, not shared with the main attention cache
write). PREFILL and DECODE both call the ported
fp8_paged_mqa_logits_cpu/topk_transform_512_cpu kernels,
which read the paged K-cache directly via page_table -- no
eager gather step, no per-request Python loop.
prefill_metadata.chunks always has exactly one entry here:
DeepseekV4CPUIndexerMetadataBuilder overrides the base chunk
split to always return the whole step's prefill batch as one
chunk, since the base chunking only bounds CUDA/XPU's dense M*N
logits tensor and flat K-gather workspace, neither of which this
paged kernel allocates.
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 | |
_decode_chunk_ranges(rows, width, budget, itemsize)
¶
Partition independent Q rows under a payload budget of logits whose
elements take itemsize bytes.
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
_gather_workspace_shapes(total_seq_lens, head_dim, fp8_dtype, use_fp4_cache)
¶
Return ((values_shape, values_dtype), (scales_shape, scales_dtype)) for the K-gather workspace. FP8 path: (T, head_dim) fp8 + (T, 4) uint8 fp32 scales. MXFP4 path: (T, head_dim // 2) uint8 packed mxfp4 + (T, head_dim // MXFP4_BLOCK_SIZE) uint8 ue8m0 scales.
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
_merge_dcp_topk_global(logits, topk_indices, topk_tokens, dcp_rank, dcp_world_size, cp_interleave, row_starts=None, logits_dtype=torch.float32, num_positions=None, direct_max_rows=0)
¶
Merge each DCP rank's local top-K into the global top-K.
topk_indices are this rank's local top-K positions into its 1/N KV
shard. A token in the global top-K must also be in its owning rank's local
top-K (at most topk_tokens - 1 tokens rank globally above it, hence at
most that many on its own rank), so exchanging only the per-rank local
candidates is exact -- equivalent to all-gathering the full logit matrix,
but it ships dcp_world_size * topk_tokens candidates instead of the whole
score row. Overwrites topk_indices with global token ids (-1 for
padding); the attention backend localizes them back to physical slots per
rank.
The selector keys only the score bits that logits_dtype (the dtype the
MQA logits were computed in, before the FP32 upcast) carries and the id
bits that positions below num_positions use.
Batches of at most direct_max_rows rows exchange the candidates over
symmetric memory instead of an NCCL all-gather (the workspace was created
by SparseAttnIndexer).
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
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 | |
_varlen_decode_rows(q, q_scale, weights, kv_cache, decode, max_model_len, max_seq_len, topk_backend, topk_tokens, topk_indices_buffer, candidate_blocks, candidate_block_size, candidate_write, logits_dtype, start, end)
¶
The varlen (SM100) decode logits, candidate selection and top-k of rows
[start, end), with the slice's own DeepGEMM schedule.
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
as_fp32_logits(logits, row_ends)
¶
Upcast BF16 MQA logits for the FP32 top-k kernels.
Only columns below row_ends (int32, one per row) are converted, the
span the consumers read; DeepGEMM's aligned row stride is kept, with
storage for whole rows (gvr_2 reads the logits as [rows, stride]).
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
dcp_gather_kv_rows(local_padded, deinterleave_idx, gathered_buf, out_buf)
¶
All-gather this rank's KV shard and return it in global token order.
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
fuse_indexer_wk_weights_proj(indexer)
¶
Fill the fused BF16 wk_weights_proj weight in place (captured CUDA
graphs keep its address) from wk and weights_proj, copying only a
part that is not already a view of its rows (e.g. dots3's own
weights_proj).
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
init_indexer_wk_weights_proj(indexer, hidden_size, quant_config, prefix)
¶
Build the indexer's checkpoint layers wk (BF16, or FP8 as the
checkpoint stores it and run in BF16: bf16_linear) and
weights_proj (BF16), and wk_weights_proj, the BF16 linear that runs
both as one GEMM. Its weight is not a parameter: it holds wk and
weights_proj's rows and is refilled after every load and every in-place
reload of the runtime weights (fuse_indexer_wk_weights_proj from the
indexer's post-load hook).
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
kv_cache_as_quant_view(kv_cache, head_dim, use_fp4_cache)
¶
4D [num_blocks, block_size, 1, head_width] view expected by
DeepGEMM, from the 3D indexer kv-cache allocation.
Source code in vllm/model_executor/layers/sparse_attn_indexer.py
should_skip_indexer_scoring(mla_metadata, cudagraph_runtime_mode)
¶
Whether the indexer can skip scoring because the main MLA runs a pure prefill on dense MHA, which does not read the top-k indices.
The indexer and main MLA may classify the same short extend differently because they use independent decode thresholds, so this takes the main MLA metadata. The indexer K cache must still be written.