vllm.v1.attention.backends.rocm_rdna_attn
¶
Attention for gfx1151 (opt-in).
Decodes of up to MAX_M query tokens per sequence run csrc/rocm/rdna_decode_attn.cu, other batches csrc/rocm/rdna_prefill_attn.cu, on TRITON_ATTN's KV cache. Anything the kernels do not support runs TRITON_ATTN.
Classes:
-
RocmRdnaAttentionBackend–TRITON_ATTN with the RDNA3.5 kernels.
-
RocmRdnaAttentionImpl–Runs the RDNA3.5 kernels when they support the call, else Triton.
-
RocmRdnaMetadata–Triton's metadata plus the host's copy of the batch for the prefill
-
RocmRdnaMetadataBuilder–Triton's builder, decodes first and counted. CUDA graphs only for
Functions:
-
mm_decode_ranges–Per request the decode kernel may serve (at most MAX_M query tokens),
-
mm_query_overlap–How the step's query tokens sit in its multimodal ranges.
RocmRdnaAttentionBackend
¶
Bases: TritonAttentionBackend
TRITON_ATTN with the RDNA3.5 kernels.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
RocmRdnaAttentionImpl
¶
Bases: TritonAttentionImpl
Runs the RDNA3.5 kernels when they support the call, else Triton.
Methods:
-
forward–The kernels when they can serve the call, else Triton's forward.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
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 | |
_plan(md, block_size, win, first, q)
¶
This step's build, plan, head group and device query starts for the
sequences from first on; made by the first layer, shared by the
rest.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
_prepare(kv_cache, **kwargs)
¶
The decode kernel and its scratch, or None to fall back.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
_prepare_prefill(kv_cache, md, kwargs, first=0)
¶
The prefill launch for the sequences from first on, or None.
Not under CUDA-graph capture: the launch depends on the host's M and S.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
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 | |
_reject(reason)
¶
Log why this call falls back to Triton.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
_serve_mixed(kv_cache, md, kwargs)
¶
Leading decodes on the decode kernel, the rest on the prefill kernel; False if a kernel cannot take its part.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
_serve_prefill(kv_cache, md, kwargs)
¶
Run the prefill kernel; False if it cannot serve the call.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
_served_tensors(kv_cache, kwargs, head_sizes)
¶
Whether the kernels support these tensors.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
_served_window(kwargs)
¶
The sliding window in keys (0: none), or None if unsupported.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
forward(layer, query, key, value, kv_cache, attn_metadata, output, output_scale=None, output_block_scale=None)
¶
The kernels when they can serve the call, else Triton's forward.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
RocmRdnaMetadata
dataclass
¶
Bases: TritonAttentionMetadata
Triton's metadata plus the host's copy of the batch for the prefill plan.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
RocmRdnaMetadataBuilder
¶
Bases: TritonAttentionMetadataBuilder
Triton's builder, decodes first and counted. CUDA graphs only for uniform decode batches: a prefill plan follows the host.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
_mm_ranges_decode(md)
¶
Write mm_decode_ranges to mm_ranges_buffer, a row per request.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
mm_decode_ranges(md)
¶
Per request the decode kernel may serve (at most MAX_M query tokens), its non-empty ranges that may hold one of its query tokens; none for the others, which run on the prefill kernel.
The kernel only bounds the mask of query tokens, so the others change nothing. On async spec-decode rows seq_lens_cpu_upper_bound may overshoot by the rejected drafts, M - 1 while the draft count holds, so the span is widened by M - 1; at most 2M - 1 disjoint ranges then remain. None if the model has no ranges.
Source code in vllm/v1/attention/backends/rocm_rdna_attn.py
mm_query_overlap(md)
¶
How the step's query tokens sit in its multimodal ranges.
Returns:
-
bool–(whether a query token's range has later keys, which only the
-
int–bidirectional mask shows; how far back the farthest in-range query
-
tuple[bool, int]–token reaches into its range, -1 if none is in one). Ranges are
-
tuple[bool, int]–inclusive; start >= end is empty.