KernelWiki/data/version-claims.yaml at master · mit-han-lab/KernelWiki · GitHub

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
## Version-claims registry — central side of the DEC-1 hybrid registry.
##
## Schema: data/schemas.yaml :: version-claims-registry.
##
## Each claim entry has:
##   id:                       (string, prefix vs-)
##   tool:                     (one of triton/cutlass/cuda/ptx/cute-dsl/cutile/tilelang/tilus/jax-pallas)
##   claim_valid_for:          (semver range, e.g. ">=3.6")
##   last_verified_release:    (must match a release name in data/tool-versions.yaml)
##   last_verified_at:         (ISO date)
##   applies_to:               (list of file paths or YAML JSON-pointers)
##   source_ids:               (list; each must resolve to a sources/**/*.md frontmatter id)
##   dev_branch_sha:           (required when claim_valid_for contains "dev")
##   notes:                    (optional)
##
## Wiki/reference pages and parsed YAML scalars in scope cite an entry by id
## via their `version_sensitive: <id>` pointer. The validator (task5) enforces
## that every per-page pointer resolves here AND every entry's `applies_to`
## paths exist + carry the matching pointer.

claims:
  - id: vs-triton-3.6-blackwell-tcgen05
    tool: triton
    claim_valid_for: ">=3.6"
    last_verified_release: "3.6.0"
    last_verified_at: 2026-04-27
    applies_to:
      - wiki/languages/triton-blackwell.md
      - references/primer.md
      - references/examples.md
      - data/inclusion-policy.yaml::triton.description
    source_ids:
      - doc-triton-3.6-blackwell
      - pr-vllm-34597
      - pr-vllm-29339
      - pr-sglang-22079
      - pr-sglang-21019
      - pr-sglang-5390
      - pr-sglang-21595
      - pr-pytorch-175826
    notes: |
      Tracks the claim that Triton 3.6+ has native Blackwell (SM100) lowering
      paths through tcgen05.mma + TMEM (descriptor/TMA + warp_specialize on
      tl.dot, plus Gluon multi-CTA + 2CTA on tcgen05_mma_scaled), replacing
      the obsolete 3.5-era "no tcgen05 / no TMEM" framing. Supporting evidence:
      data/triton-3.6-evidence.md memo plus the official release notes
      (doc-triton-3.6-blackwell). AC-1.1 "new tracked-repo PR page demonstrates
      a kernel" anchor: pr-vllm-34597 (post-refresh; vLLM "[Kernel] Add FP8 KV
      cache support to Triton MLA decode attention", merged 2026-02-16, sm100,
      directly modifying @triton.jit kernels in vllm/v1/attention/ops/
      triton_decode_attention.py with tl.dot matmul + vllm/v1/attention/
      backends/mla/triton_mla.py; Triton kernel shipped verbatim under
      artifacts/prs/vllm/PR-34597/ pinned at SHA a1257fd1). Supplementary
      post-refresh anchor: pr-vllm-29339 (Triton 3.6 triton_kernels library
      scoped to [sm100, sm90] for MXFP4 dispatch-gate, merged 2025-11-24).
      Pre-refresh historical anchors retained as supplementary context (each
      appears in data/refresh-cutoff.yaml::previous_pages_manifest):
      pr-sglang-22079 (Gemma4 NVFP4 Triton attention kernel doing real tl.dot
      matmul on [sm100, sm90], merged 2026-04-03), pr-sglang-21019
      (memory-rearrangement Triton kernel on SM100, merged 2026-03-20, no
      tl.dot), pr-sglang-5390, pr-sglang-21595, pr-pytorch-175826 (ecosystem
      context). The strongest possible demonstrations of the 3.6 Blackwell-
      native paths (descriptor/TMA + warp_specialize, tl.dot_scaled, Gluon
      multi-CTA/2CTA) have NOT been found in any tracked-downstream merged
      PR locally; future refresh rounds should backfill such an anchor when
      one becomes available.