@kernaut/kernaut-resources 0.1.17-beta.1 → 0.1.18

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/nvidia/api-definitions/runtime/cudaOccupancyMaxActiveBlocksPerMultiprocessor.md

@@ -1,56 +0,0 @@
----
-func_name: cudaOccupancyMaxActiveBlocksPerMultiprocessor
-namespace: runtime
-header: cuda_runtime.h
-signature: TBD — agent will fill during probe (see api-probing.md Step 2)
-since_cuda: ''
-status: draft
-has_end_to_end_example: true
-source:
-- path: spec
-  anchor: Reference
-probed_by: sources/experience/api-probes/2026-04-17-runtime-cuda-occupancy-max-active-blocks-per-multiprocessor.md
-id: api-cudaOccupancyMaxActiveBlocksPerMultiprocessor
-type: api-definition
-vendor: nvidia
-title: Cudaoccupancymaxactiveblockspermultiprocessor
-source_refs:
-- source_id: cuda-official/toolkit-docs-13.2
-  path: CUDA API References/cuda-runtime-api/cuda_cuda-runtime-api_index.html.md
-  anchor: L3862-L3862
-- source_id: cuda-official/toolkit-docs-13.2
-  path: CUDA API References/cuda-runtime-api/cuda_cuda-runtime-api_index.html.md
-  anchor: L3865-L3865
-- source_id: cuda-official/toolkit-docs-13.2
-  path: CUDA API References/cuda-runtime-api/cuda_cuda-runtime-api_index.html.md
-  anchor: L3890-L3890
-- source_id: source-code/cuda-samples
-  path: Samples/0_Introduction/simpleAWBarrier/README.md
-  anchor: L26-L26
-- source_id: source-code/cuda-samples
-  path: Samples/0_Introduction/simpleAWBarrier/simpleAWBarrier.cu
-  anchor: L204-L204
-- source_id: source-code/cuda-samples
-  path: Samples/0_Introduction/simpleIPC/README.md
-  anchor: L26-L26
-architectures:
-- sm90
-- sm90a
-languages:
-- cuda-cpp
-confidence: inferred
-evidence_level: measured
-tags:
-- cuda-cpp
----
-# cudaOccupancyMaxActiveBlocksPerMultiprocessor
-
-<!-- Auto-seeded stub from `probe_loop seed-stubs`. Agent must probe this API following `reasoning/api-probing.md` and fill the `signature`, body sections, and artifacts. -->
-
-## Semantics (TBD)
-
-Agent: replace this section with parameter list + return semantics + typical use case grounded in source corpus hits listed in frontmatter.
-
-## End-to-End Example
-
-See the probe record at sources/experience/api-probes/2026-04-17-runtime-cuda-occupancy-max-active-blocks-per-multiprocessor.md for a complete end-to-end example, kernel source, build command, and measurement on H200 (sm_90a, CUDA 12.9).

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/nvidia/code-walkthroughs/cutlass-cute/AGENTS.md

@@ -1,93 +0,0 @@
----
-id: code-cutlass-cute-AGENTS
-type: code-walkthrough
-vendor: nvidia
-title: Agents
-upstream_repo: NVIDIA/cutlass-cute
-architectures:
-- sm90
-- sm90a
-languages:
-- ptx
-- cuda-cpp
-- cute-dsl
-hardware_features:
-- wgmma
-- tma
-techniques:
-- warp-specialization
-- persistent-kernel
-- kernel-fusion
-kernel_types:
-- gemm
-- fused-kernel
-confidence: inferred
-evidence_level: inferred
-tags:
-- wgmma
-- tma
-- warp-specialization
-- persistent-kernel
-- kernel-fusion
-- gemm
-- fused-kernel
-- ptx
-- cuda-cpp
-- cute-dsl
----
-# 60-code / cutlass-cute
-
-This directory holds **CUTLASS / CuTe source-reading notes on Hopper**: how specific upstream examples and template fragments express TMA, WGMMA, warp-specialized mainloops, tails, and fused epilogues. The prose half (what the technique *is*) lives in `wiki/nvidia/hardware/<feature>/`, `wiki/nvidia/techniques/<algo>/`, and `wiki/nvidia/foundations/compute/<gemm-variant>/`.
-
-`tools/lint_knowledge.py` skips the entire `wiki/nvidia/code-walkthroughs/` subtree by default (see SKIP_DIRS), so the markdown skeletons, READMEs, and tuning logs are not subjected to the `skill / api-raw / experience` frontmatter contracts. If a topic directory introduces a knowledge-prose document with frontmatter (e.g. an embedded skill), it must move to the right `wiki/nvidia/foundations/` / `wiki/nvidia/hardware/` / `wiki/nvidia/techniques/` location or be flagged for the linter explicitly.
-
-## Source-driven layout
-
-The second-level directory always names a **source repository**:
-
-```
-wiki/nvidia/code-walkthroughs/
-├── cutlass-cute/        # this directory — extracted from cutlass / CuTeDSL
-│   ├── example48-hopper-warp-specialized-gemm/
-│   │                         # ex48: TMA + WGMMA + cooperative WS mainloop
-│   ├── wgmma-atom-decoding/ # CUTLASS WGMMA atom decoding
-│   ├── gemm-aligned/        # aligned cooperative GEMM template parameters
-│   ├── gemm-tail/           # tail-handling template selection by shape
-│   ├── persistent-kernel/   # pingpong / cooperative persistent schedules
-│   └── gemm-fused/          # fused epilogue / prologue patterns
-└── ptx-gemm/                # cutlass-free PTX track — minimal compilable repro lives here
-```
-
-`wiki/nvidia/code-walkthroughs/ptx-gemm/` is a different shape (cutlass-free track keeps its working `.cu` here, since it is the single source of truth for that path — there is no upstream cutlass example to point at). All other topics under `cutlass-cute/` are documentation-only.
-
-## Where the reproducible artifacts actually live
-
-| Topic | Canonical artifact path |
-|---|---|
-| Example 48 Hopper WS GEMM | `artifacts/experience/api-probes/gemm/gemm_aligned.cu` |
-| TMA cutlass-free | `artifacts/experience/hw-probes/tma-ptx/{tma_hello.cu, tma_throughput_probe.cu}` |
-| wgmma cutlass-API | `artifacts/experience/api-probes/gemm/gemm_aligned.cu` (same harness) |
-| wgmma cutlass-free | `artifacts/experience/hw-probes/wgmma-ptx/{wgmma_hello.cu, wgmma_zoo.cu}` |
-| Aligned GEMM | `artifacts/experience/api-probes/gemm/gemm_aligned.cu` |
-| Tail GEMM | `artifacts/experience/api-probes/gemm/{gemm_tail.cu, gemm_tail_small.cu}` |
-| Plain WS schedule | `artifacts/experience/api-probes/gemm/gemm_compare_ws.cu` |
-| Pingpong schedule | `artifacts/experience/api-probes/gemm/gemm_compare_pingpong.cu` |
-| Cooperative schedule | `artifacts/experience/api-probes/gemm/gemm_aligned.cu` (auto-selects cooperative at large aligned shapes) |
-| Cutlass-free WS | `artifacts/experience/kernel-records/2026-04-29-gemm-ws-ptx/gemm_ws_ptx.cu` |
-| Fused-GEMM (ex50) | upstream `examples/50_hopper_gemm_with_epilogue_swizzle/` (built in place by `artifacts/experience/api-probes/gemm/run_fused.sh`) |
-
-To rebuild any topic's binary: `cd artifacts/experience/<api-probes|hw-probes>/<topic> && bash build.sh && bash run.sh`.
-
-## Current contents
-
-```
-wiki/nvidia/code-walkthroughs/cutlass-cute/
-├── README.md                 (this file)
-├── example48-hopper-warp-specialized-gemm/
-│                              (ex48 TMA + WGMMA + cooperative WS mainloop)
-├── wgmma-atom-decoding/      (CUTLASS WGMMA atom decoding + skeleton)
-├── gemm-aligned/             (aligned-GEMM template params + tuning)
-├── gemm-tail/                (tail-handling template selection + tuning)
-├── persistent-kernel/        (persistent-schedule selection + tuning)
-└── gemm-fused/               (fused epilogue / prologue patterns + tuning)
-```

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/nvidia/code-walkthroughs/cutlass-cute/AUDIT.md

@@ -1,218 +0,0 @@
----
-id: code-cutlass-cute-AUDIT
-type: code-walkthrough
-vendor: nvidia
-title: Audit
-upstream_repo: NVIDIA/cutlass-cute
-architectures:
-- sm90
-- sm90a
-- sm100
-languages:
-- ptx
-- cuda-cpp
-- cute-dsl
-- python
-hardware_features:
-- wgmma
-- tma
-- ldmatrix
-- tcgen05
-techniques:
-- warp-specialization
-- persistent-kernel
-- pipeline-stages
-- kernel-fusion
-- shared-memory-optimization
-- swizzling
-- tma-multicast
-- software-exp
-kernel_types:
-- gemm
-- attention
-- flash-attention
-- fused-kernel
-confidence: inferred
-evidence_level: inferred
-tags:
-- wgmma
-- tma
-- ldmatrix
-- tcgen05
-- warp-specialization
-- persistent-kernel
-- pipeline-stages
-- kernel-fusion
-- shared-memory-optimization
-- swizzling
-- tma-multicast
-- software-exp
-- gemm
-- attention
-- flash-attention
-- fused-kernel
-- ptx
-- cuda-cpp
-- cute-dsl
-- python
----
-<!--
-lint-exempt: this file lives under wiki/nvidia/code-walkthroughs/ which reasoning/AGENTS.md exempts from frontmatter
-requirements. A YAML block is included below for human readers; lint will skip it.
--->
-
----
-title: "Gap audit: CUTLASS GEMM/Attention coverage in kernel-kb-mvp/knowledge"
-audit_round: rlcr-2026-05-08-round-2
-target_plan: docs/cutlass_explore_update_01.md
-classification_keys: [complete, partial, missing, redundant]
-upstream_repo_pinned: cutlass@f74fea9c
-auditor_inputs:
-  - wiki/nvidia/hardware/
-  - wiki/nvidia/foundations/compute/
-  - wiki/nvidia/foundations/memory/
-  - wiki/nvidia/foundations/sync/
-  - wiki/nvidia/foundations/meta/
-  - wiki/nvidia/techniques/
-  - wiki/nvidia/code-walkthroughs/cutlass-cute/
-  - sources/experience/api-probes/
-  - sources/experience/hw-probes/
-  - corpus/MANIFEST.yaml
-  - templates/frontmatter/
-upstream_root: "{{CUTLASS_REPO_REF}}"
----
-
-# Gap Audit: CUTLASS GEMM/Attention Coverage
-
-This file is the deliverable for the gap-audit task in `docs/cutlass_explore_update_01.md`. It catalogs every existing GEMM/wgmma/TMA/attention-relevant entry in ``, classifies each, and lists concrete additions needed. All "missing | partial" rows cite a specific cutlass source path under `{{CUTLASS_REPO_REF}}/` that justifies the gap, and propose a target KB layer path.
-
-The classification rubric is:
-- **complete** — entry exists, satisfies its layer's contract (skill+pitfalls+optional apis for skill layer; README+tuning+optional skeleton for `wiki/nvidia/code-walkthroughs/`), and its referenced sources/artifacts are present.
-- **partial** — entry exists but is missing a required companion file, has thin content (<≈50 lines for skill body), or its `tuning.md` is absent where convention requires one.
-- **missing** — no entry exists for a topic the plan requires; cite cutlass file(s) justifying the gap.
-- **redundant** — overlapping coverage that should be consolidated; cite both paths.
-
-All paths are relative to `` unless prefixed with `<path-removed> (absolute upstream paths).
-
----
-
-## Section A — `wiki/nvidia/hardware/` (Hardware primitives)
-
-| Entry | Files | Status | Evidence | Classification | Notes / Gap |
-|-------|-------|--------|----------|----------------|-------------|
-| `wiki/nvidia/hardware/tma/` | skill.md, pitfalls.md | verified | measured | complete | TMA cute end-to-end on Hopper; pitfalls list 8 items; could add a measured 3D TMA case but not blocking. |
-| `wiki/nvidia/hardware/tma-ptx/` | skill.md, pitfalls.md | verified | measured | complete | Raw PTX TMA (cutlass-free); has 4 hw-probe records (hello, throughput, multicast v1, multicast v2). |
-| `wiki/nvidia/hardware/wgmma/` | skill.md, pitfalls.md | verified | measured | complete | **Round 1 closed**: Added "## Atomic Usage (cute)" section with 11-row mapping table (MMA_Atom → PTX mnemonic → probe config → probe record) plus cute-API snippet. Cutlass anchor: `{{CUTLASS_REPO_REF}}/include/cute/atom/mma_traits_sm90_gmma.hpp`. |
-| `wiki/nvidia/hardware/wgmma-ptx/` | skill.md, pitfalls.md, atomic_skeleton.md | verified | measured | complete | **Round 1 closed**: Added "## Atomic Usage (PTX)" section with raw PTX template + build/run commands, and created `atomic_skeleton.md` documenting the full single-atom sequence (fence→mma→commit→wait) with descriptor construction and fragment layout. |
-| `wiki/nvidia/hardware/mma-sync-ptx/` | skill.md, pitfalls.md | draft | spec | complete | **Round 1 closed**: Created skill.md with full atom shape table (12 families), PTX mnemonic template, and 5 pitfalls. `status: draft`, `evidence_level: spec`, `requires_sm: ">=8.0"`. |
-| `wiki/nvidia/hardware/ldmatrix-ptx/` | skill.md, pitfalls.md | draft | spec | complete | **Round 1 closed**: Created skill.md with all 6 ldmatrix variants (x1/x2/x4 non-transposed + x2/x4/x8 transposed) plus movmatrix, and 5 pitfalls. Uses corrected paths (`copy_traits_sm75.hpp`, not `mma_traits_sm80*.hpp`). |
-| `wiki/nvidia/hardware/tcgen05-ptx/` | skill.md, pitfalls.md | draft | spec | complete | **Round 1 closed**: Created skill.md with `not_measured_on_target: true`, `requires_sm: ">=10.0"`. Documents tcgen05.cp, tcgen05.ld, and SM100_MMA_* traits. Cites Blackwell whitepaper and cutlass sm100 headers. 3 pitfalls. |
-
----
-
-## Section B — `wiki/nvidia/foundations/compute/gemm*` and `wiki/nvidia/foundations/compute/attention/` (Operator skills)
-
-| Entry | Files | Status | Evidence | Classification | Notes / Gap |
-|-------|-------|--------|----------|----------------|-------------|
-| `wiki/nvidia/foundations/compute/gemm/aligned/` | skill.md, pitfalls.md, tuning.md | verified | measured | complete | **Round 1 closed**: Created sibling `tuning.md` with compile-time and runtime parameter tables. Added back-link from `skill.md` body. |
-| `wiki/nvidia/foundations/compute/gemm/non-aligned-tail/` | skill.md, pitfalls.md, tuning.md | partial | measured | partial | **Round 1 partial**: Created sibling `tuning.md` with tail-strategy knobs. Added back-link from `skill.md`. Status remains partial (coverage incomplete). |
-| `wiki/nvidia/foundations/compute/gemm-fused/cutlass-epilogue-prologue/` | skill.md, pitfalls.md, tuning.md | partial | measured | partial | **Round 1 partial**: Created sibling `tuning.md` with EVT and prologue fusion knobs. Added back-link from `skill.md`. Status remains partial (prologue coverage incomplete). |
-| `wiki/nvidia/foundations/compute/gemm-ptx/` | skill.md, pitfalls.md | partial | measured | partial | Cutlass-free PTX GEMM track. No apis.md (acceptable — pure PTX). Cross-link gap to any tuning.md is absent because no `wiki/nvidia/code-walkthroughs/.../gemm-ptx/` partner exists. Out of plan scope unless the Round-N tuning sweep targets PTX-GEMM. |
-| `wiki/nvidia/foundations/compute/attention/mvp-minimal/` | skill.md, pitfalls.md, tuning.md | verified | measured | complete | **Round 9 verified**: TMA+wgmma kernel passes correctness on H200 (max_abs_err=0.000061, 0/16384 mismatches, sanitizer 0 errors). Uses canonical CUTLASS smem layout (Layout_K_SW128_Atom + Swizzle<3,4,3> + CU_TENSOR_MAP_SWIZZLE_128B). Evidence logs checked in. |
-| `wiki/nvidia/foundations/compute/attention/cutlass-fmha/` | skill.md, pitfalls.md, tuning.md | verified | measured | complete | **Round 13 measured**: H200-verified via repo-local wrapper. Example: 376.1 TFLOPS (B2 H16 S1024 D128 full). Collective: 348.5 TFLOPS (B2 H16 S2048 D128 causal). |
-
-Other compute skills (`branch-elimination`, `compiler-hints`, `fast-math`, `half-precision-math`, `occupancy-tuning`, `warp-divergence`, `warp-primitives`, `ilp`) are out of plan scope; only `ilp` is missing apis.md and that is unrelated to this plan.
-
----
-
-## Section C — `wiki/nvidia/techniques/` (End-to-end algorithm patterns)
-
-| Entry | Files | Status | Evidence | Classification | Notes / Gap |
-|-------|-------|--------|----------|----------------|-------------|
-| `wiki/nvidia/techniques/warp-specialization/` | skill.md, pitfalls.md | verified | measured | complete | Producer/consumer warpgroup pattern; pinned to `cutlass@f74fea9c`. Pitfalls is short (<50 lines) but covers the canonical hazards. |
-| `wiki/nvidia/techniques/persistent-kernel/` | skill.md, pitfalls.md | verified | measured | complete | Persistent-kernel ping-pong; same depth/notes as above. |
-| `wiki/nvidia/techniques/online-softmax/` | — | — | — | out-of-scope | Online-softmax is documented inline within `wiki/nvidia/foundations/compute/attention/mvp-minimal/skill.md` (algorithmic skeleton) and `artifacts/experience/api-probes/attention/flash_attn_tma_wgmma.cu` (implementation). A standalone `wiki/nvidia/techniques/` entry is optional — the algorithm is already covered by the attention skill entries. Not required by any plan AC. |
-| `wiki/nvidia/techniques/split-k/` | — | — | — | absent (out of plan scope) | Listed in `reasoning/AGENTS.md` as a future classical algo; not required by current plan ACs. Recorded for reference only. |
-
----
-
-## Section D — `wiki/nvidia/code-walkthroughs/cutlass-cute/` (Code-extraction layer)
-
-This subtree is lint-exempt per reasoning/AGENTS.md; entries carry README.md, tuning.md, and optional `*_skeleton.md`.
-
-| Entry | Files | Classification | Notes / Gap |
-|-------|-------|----------------|-------------|
-| `wiki/nvidia/code-walkthroughs/cutlass-cute/example48-hopper-warp-specialized-gemm/` | README.md, mainloop_skeleton.md | complete | Reading guide for cutlass example 48; pairs with `wiki/nvidia/foundations/compute/gemm/aligned/`. Verify: `{{CUTLASS_REPO_REF}}/examples/48_hopper_warp_specialized_gemm/`. |
-| `wiki/nvidia/code-walkthroughs/cutlass-cute/gemm-aligned/` | README.md, tuning.md | complete | Cross-linked from this audit; needs back-link from `wiki/nvidia/foundations/compute/gemm.md` body (closes AC-9 cross-link). |
-| `wiki/nvidia/code-walkthroughs/cutlass-cute/gemm-fused/` | README.md, tuning.md | complete | Same cross-link note. |
-| `wiki/nvidia/code-walkthroughs/cutlass-cute/gemm-tail/` | README.md, tuning.md | complete | Same cross-link note. |
-| `wiki/nvidia/code-walkthroughs/cutlass-cute/persistent-kernel/` | README.md, tuning.md | complete | Documents persistent-kernel knobs; pairs with `wiki/nvidia/techniques/persistent-kernel/`. |
-| `wiki/nvidia/code-walkthroughs/cutlass-cute/wgmma-atom-decoding/` | README.md, wgmma_skeleton.md | complete | Skeleton walkthrough for wgmma atom decoding; pairs with `wiki/nvidia/hardware/wgmma/`. |
-| `wiki/nvidia/code-walkthroughs/cutlass-cute/attention-fmha-example/` | README.md, fmha_skeleton.md, tuning.md | complete | **Round 13**: H200-measured 88_hopper_fmha (376.1 TFLOPS ping-pong at B2 H16 S1024 D128). Build/run scripts + probe record. 41_fused fallback context documented. |
-| `wiki/nvidia/code-walkthroughs/cutlass-cute/attention-fmha-collective/` | README.md, tuning.md | complete | **Round 13**: H200-measured collective config (348.5 TFLOPS ping-pong at B2 H16 S2048 D128 causal). Probe record with collective settings cited. |
-| `wiki/nvidia/code-walkthroughs/flash-attention-v3/` | README.md, tuning.md | complete | **Round 1 closed**: Registered in MANIFEST.yaml (commit `bbda031f`, BSD-3-Clause, local path `{{FLASH_ATTENTION_REPO_REF}}`). Created README (source files, algorithmic comparison vs cutlass FMHA and MVP minimal) and tuning.md (compile-time + runtime knobs, differences from cutlass FMHA). |
-
----
-
-## Section E — `sources/experience/hw-probes/` (Microbench records)
-
-| Entry | Records | Classification | Notes / Gap |
-|-------|---------|----------------|-------------|
-| `sources/experience/hw-probes/wgmma-ptx/` | hello, zoo (11 configs) | complete | Already exercises the canonical zoo. AC-2.1 needs explicit 1:1 mapping table in `wiki/nvidia/hardware/wgmma/skill.md`, not new probe records. |
-| `sources/experience/hw-probes/tma-ptx/` | hello, throughput, multicast-v1, multicast-v2 | complete | No gap. |
-| `sources/experience/hw-probes/canonical/` | — (empty) | partial | Empty placeholder; clarify with maintainers whether this is a stub for future canonical microbenches or stale. Out of plan scope; flagged for visibility. |
-| `sources/experience/hw-probes/attention-tuning/` | 2026-05-08-attention-tile-sweep.md, artifacts/profiles/ | complete | **Round 0 created**: 6-config BLOCK_M x BLOCK_N tile-shape sweep at fixed B=1,H=2,S=256,D=64 on H200. CSV includes throughput_gflops column. |
-| `sources/experience/hw-probes/gemm-tuning/` (or operator-specific) | — | missing-or-deferred | AC-9 requires ≥1 measured sweep across operators; if attention-tuning is the chosen sweep, this stays optional. Plan lower bound permits one measured sweep total. |
-
-Other probe directories (smem-bank-conflict, occupancy-sweep, warp-divergence-cost, fast-math, register-pressure, etc.) are out of plan scope.
-
----
-
-## Section F — `sources/experience/api-probes/` (API-probe records)
-
-| Entry | Files | Classification | Notes / Gap |
-|-------|-------|----------------|-------------|
-| `sources/experience/api-probes/gemm/` | 10 dated records, artifacts/ (14 .cu, build/run scripts, device.json, profiles/) | complete | Hot path; existing artifacts are the reference for AC-9 cross-linking. |
-| `sources/experience/api-probes/attention/` | artifacts/{flash_attn_tma_wgmma.cu, flash_attn_minimal.cu, wgmma_desc_test.cu, cutlass_88_hopper_fmha.cu, build*.sh, run*.sh, device.json, profiles/}, 4 dated probe records | complete | **Round 14**: repo-local FMHA wrapper added. TMA+wgmma (0.35 TFLOPS), FMHA example (376.1 TFLOPS), FMHA collective (348.5 TFLOPS). All with H200 evidence. |
-
----
-
-## Section G — `corpus/MANIFEST.yaml` (Source registry)
-
-| Source | Present? | Classification | Notes / Gap |
-|--------|----------|----------------|-------------|
-| cuda-official/toolkit-docs-13.2 | yes | complete | — |
-| blogs/colfax | yes | complete | Tags: tma, wgmma, async-pipeline. |
-| whitepapers/gpu-wite-paper | yes | complete | Hopper + Blackwell whitepapers (covers tcgen05 spec citation). |
-| source-code/cutlass | yes | complete | Pinned at `cutlass@f74fea9c`. |
-| source-code/cuda-samples | yes | complete | — |
-| legacy-knowledge | yes | complete | KernelPilot pre-refactor; not truth source. |
-| FlashAttention v3 (Tri Dao) | yes | complete | **Round 1 closed**: Registered as `source-code/flash-attention` with commit `bbda031f`, BSD-3-Clause license, local path `{{FLASH_ATTENTION_REPO_REF}}`. |
-
----
-
-## Summary
-
-### Targeted closures by AC
-- **AC-1**: this file.
-- **AC-2.1**: extend `wiki/nvidia/hardware/wgmma/skill.md` with `## Atomic Usage (cute)` + 1:1 mapping table to `sources/experience/hw-probes/wgmma-ptx/zoo/`.
-- **AC-2.2**: extend `wiki/nvidia/hardware/wgmma/skill-wgmma-ptx.md` with `## Atomic Usage (PTX)` + create `wiki/nvidia/hardware/wgmma-ptx/atomic_skeleton.md` (companion `.cu` referenced from `sources/experience/api-probes/`).
-- **AC-3**: create `wiki/nvidia/hardware/mma-sync-ptx/{skill.md,pitfalls.md}` and `wiki/nvidia/hardware/ldmatrix-ptx/{skill.md,pitfalls.md}`.
-- **AC-4**: create `wiki/nvidia/hardware/tcgen05-ptx/{skill.md,pitfalls.md}` (spec-only).
-- **AC-5**: create `wiki/nvidia/foundations/compute/attention/mvp-minimal/{skill.md,pitfalls.md,tuning.md}` + artifacts at `artifacts/experience/api-probes/attention/`.
-- **AC-6**: create `wiki/nvidia/code-walkthroughs/cutlass-cute/attention-fmha-example/{README.md,fmha_skeleton.md,tuning.md}` + **measured artifact** for `88_hopper_fmha` on H200 (at least one correctness/timing datapoint).
-- **AC-7**: create `wiki/nvidia/code-walkthroughs/cutlass-cute/attention-fmha-collective/{README.md,tuning.md}` + **measured collective datapoint** on H200 (at least one configuration with measured_on + artifact paths).
-- **AC-8**: register FAv3 in `corpus/MANIFEST.yaml` and create `wiki/nvidia/code-walkthroughs/flash-attention-v3/{README.md,tuning.md}`.
-- **AC-9**: cross-link `wiki/nvidia/code-walkthroughs/cutlass-cute/<op>/tuning.md` from each `wiki/nvidia/foundations/compute/gemm*/skill.md`; add `tuning.md` to new attention skill (AC-5); land ≥1 measured sweep at `sources/experience/hw-probes/<operator>-tuning/`.
-- **AC-10**: run `python -m tools.lint_knowledge` and confirm exit 0 across full tree and per-layer.
-- **AC-11**: grep produced artifacts to confirm no plan-document terminology leaked.
-
-### Out-of-scope items flagged for visibility (not closed by this plan)
-- `wiki/nvidia/foundations/compute/ilp/` missing apis.md.
-- `wiki/nvidia/foundations/memory/bank-conflict/` missing apis.md.
-- `wiki/nvidia/foundations/sync/memory-ordering/` is draft + spec-only.
-- `sources/experience/hw-probes/canonical/` empty placeholder.
-- `wiki/nvidia/api-definitions/{ptx,math-intrinsics,cccl/cub,_classified}/` empty namespaces.
-
-### Reproducibility note
-Re-running this audit against the current `` tree should reproduce the same classification provided no entries are added or moved. The classification uses observable filesystem state only (presence of `skill.md`, `pitfalls.md`, `tuning.md`, frontmatter `status`/`evidence_level`, hw-probe record counts) and verified upstream paths under `{{CUTLASS_REPO_REF}}/`.

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/nvidia/code-walkthroughs/cutlass-cute/attention-fmha-collective/README.md

@@ -1,135 +0,0 @@
----
-id: code-cutlass-cute-README
-type: code-walkthrough
-vendor: nvidia
-title: Readme
-upstream_repo: NVIDIA/cutlass-cute
-architectures:
-- sm90
-- sm90a
-languages:
-- cuda-cpp
-- cute-dsl
-hardware_features:
-- wgmma
-- tma
-- mbarrier
-techniques:
-- warp-specialization
-- pipeline-stages
-- kernel-fusion
-- shared-memory-optimization
-- software-exp
-- conditional-rescaling
-kernel_types:
-- gemm
-- attention
-- flash-attention
-- fused-kernel
-confidence: inferred
-evidence_level: inferred
-tags:
-- wgmma
-- tma
-- mbarrier
-- warp-specialization
-- pipeline-stages
-- kernel-fusion
-- shared-memory-optimization
-- software-exp
-- conditional-rescaling
-- gemm
-- attention
-- flash-attention
-- fused-kernel
-- cuda-cpp
-- cute-dsl
----
-# cutlass FMHA Collective Layer
-
-This directory documents the collective-builder layer that implements the core FMHA algorithm in cutlass `examples/88_hopper_fmha/collective/`.
-
-## Source location
-
-```
-{{CUTLASS_REPO_REF}}/examples/88_hopper_fmha/collective/
-├── fmha_collective_tma_warpspecialized.hpp   # Primary forward collective
-├── fmha_collective_bwd_tma_warpspecialized.hpp  # Backward collective
-├── fmha_collective_load.hpp                  # TMA load helpers
-├── fmha_collective_softmax.hpp               # Online softmax implementation
-├── fmha_collective_tma.hpp                   # TMA descriptor helpers
-├── fmha_common.hpp                           # Common types and utilities
-├── fmha_epilogue.hpp                         # Output writeback
-├── fmha_epilogue_bwd.hpp                     # Backward epilogue
-└── fmha_fusion.hpp                           # Fusion interface (causal, etc.)
-```
-
-Upstream commit: `cutlass@f74fea9c`.
-
-## API surface
-
-### Entry types
-
-- `FmhaCollectiveTmaWarpSpecialized<ElementQK, ElementVO, kHeadDim, Options...>` — the primary forward collective. Uses TMA for data movement and wgmma for compute.
-- `FmhaCollectiveBwdTmaWarpSpecialized` — the backward-pass collective.
-
-### Builder dispatch
-
-The collective is instantiated by `FmhaKernelBuilder` in `kernel/fmha_kernel_builder.hpp`. The builder resolves compile-time options (see `fmha_options.hpp`) and passes them through as template parameters.
-
-### Warp-specialization variants
-
-The forward collective uses a **producer/consumer** warp-specialization pattern:
-
-1. **Producer warpgroup** (warpgroup 0): Issues TMA loads for K and V tiles into shared memory pipeline stages. Manages mbarrier-based pipeline synchronization. Optionally loads Q separately (`kLoadsQSeparately`).
-
-2. **Consumer warpgroup(s)** (warpgroup 1+): Execute the attention computation:
-   - `S = Q @ K^T` via wgmma (shared-memory Q × shared-memory K^T)
-   - Online softmax (per-row max tracking, exp, rescale)
-   - `O += P @ V` via wgmma (register-resident P × shared-memory V)
-
-This mirrors the GEMM warp-specialization pattern from `examples/48_hopper_warp_specialized_gemm/`.
-
-### Core methods
-
-| Method | Role |
-|---|---|
-| `load(pipeline, ...)` | Producer: TMA loads K/V tiles into smem stages |
-| `mma(pipeline, ...)` | Consumer: wgmma S=Q@K^T, softmax, wgmma O=P@V |
-| `epilogue(...)` | Final normalization and global store |
-
-## Differences from GEMM collective
-
-| Aspect | GEMM collective | FMHA collective |
-|---|---|---|
-| Inner loop | K-dimension tiles | Sequence-dimension (KV) tiles |
-| Softmax | None | Online softmax with accumulator rescaling |
-| Output accumulator | Single, grows with K tiles | Rescaled every KV tile (FlashAttention-2) |
-| TMA loads per iteration | A + B tiles | K + V tiles (Q loaded once, reused) |
-| Epilogue | Direct store or fused op | Normalize by softmax denominator + store |
-
-## Builder-level vs kernel-level knobs
-
-See `tuning.md` for the full table. Key distinction:
-
-- **Builder-level knobs** (resolved by `FmhaKernelBuilder`): `kNumMmaWarpGroups`, `kIsPersistent`, `kStagesQ`, `kStagesKV`, `kClusterM`. These determine the collective's template instantiation.
-- **Kernel-level knobs** (inside the collective): Tile shapes (BLOCK_M, BLOCK_N derived from warpgroup count and head dim), pipeline depth, softmax precision.
-
-## H200 measured collective datapoint
-
-See sources/experience/api-probes/attention.md.
-
-Configuration: B=2 H=16 Q=2048 K=2048 D=128, causal mask, kNumMmaWarpGroups=2 (BLOCK_M=128).
-
-| Config | TFLOPS/s | Correctness |
-|--------|----------|-------------|
-| cooperative 128x128x128 | 307.1 | OK |
-| ping-pong 128x128x128 | 348.5 | OK |
-
-Measured on H200-SXM, sm_90a, CUDA 12.9.
-
-## How this fits in the KB
-
-- Detailed per-knob documentation: `wiki/nvidia/code-walkthroughs/cutlass-cute/attention-fmha-example/tuning.md`
-- Code flow skeleton: `wiki/nvidia/code-walkthroughs/cutlass-cute/attention-fmha-example/fmha_skeleton.md`
-- Hand-built minimal attention kernel (no cutlass dependency): `wiki/nvidia/foundations/compute/attention/mvp-minimal/skill.md`

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/nvidia/code-walkthroughs/cutlass-cute/attention-fmha-collective/tuning.md

@@ -1,95 +0,0 @@
----
-id: code-cutlass-cute-tuning
-type: code-walkthrough
-vendor: nvidia
-title: Tuning
-upstream_repo: NVIDIA/cutlass-cute
-architectures:
-- sm90
-- sm90a
-languages:
-- cuda-cpp
-- cute-dsl
-hardware_features:
-- wgmma
-- tma
-- mbarrier
-- cluster
-techniques:
-- warp-specialization
-- persistent-kernel
-- pipeline-stages
-- register-budgeting
-- shared-memory-optimization
-- tma-multicast
-- software-exp
-- communication-overlap
-kernel_types:
-- gemm
-- attention
-- flash-attention
-confidence: inferred
-evidence_level: inferred
-tags:
-- wgmma
-- tma
-- mbarrier
-- cluster
-- warp-specialization
-- persistent-kernel
-- pipeline-stages
-- register-budgeting
-- shared-memory-optimization
-- tma-multicast
-- software-exp
-- communication-overlap
-- gemm
-- attention
-- flash-attention
-- cuda-cpp
-- cute-dsl
----
-# FMHA Collective Tuning Parameters
-
-## Builder-level knobs (resolved before collective instantiation)
-
-These knobs are passed as template options to `FmhaKernelBuilder` and determine which collective template is instantiated.
-
-| Knob | Default | Range | Effect on collective |
-|---|---|---|---|
-| `kNumMmaWarpGroups` | 2 | 1, 2 | Determines BLOCK_M (64 per warpgroup). Default 2 gives BLOCK_M=128. 1 warpgroup gives BLOCK_M=64 with less register/smem pressure. |
-| `kStagesKV` | auto | 2, 3, 4 | Number of smem pipeline stages for K/V tiles. Each stage holds one (BLOCK_N × HEAD_DIM) K tile + one (BLOCK_N × HEAD_DIM) V tile. More stages hide TMA load latency. |
-| `kStagesQ` | auto | 1, 2 | Q pipeline stages. Usually 1 (Q is loaded once and reused across all KV tiles). 2 enables prefetch overlap for the next Q tile in persistent mode. |
-| `kClusterM` | 1 | 1, 2, 4 | CTA cluster along the Q-tile dimension. Enables TMA multicast — one TMA load distributes K/V tiles to multiple clustered CTAs. Reduces HBM bandwidth per CTA. |
-| `kIsPersistent` | false | true, false | Persistent kernel mode. CTAs loop over work items (batch, head, Q-tile) instead of one-shot. Improves SM utilization for small problem sizes. |
-
-## Kernel-level knobs (inside the collective)
-
-These are derived or fixed inside the collective implementation.
-
-| Knob | Source | Typical value | Effect |
-|---|---|---|---|
-| BLOCK_M | `64 × kNumMmaWarpGroups` | 64, 128 | Q-tile row count. Must be a multiple of wgmma M=64. |
-| BLOCK_N | derived from HEAD_DIM | 64, 128 | KV-tile sequence chunk. Larger BLOCK_N means fewer mainloop iterations but more smem. |
-| HEAD_DIM | template param | 64, 128, 256 | Inner dimension of Q@K^T. Fixed per model. |
-| Softmax precision | `kAccQK` option | f32 | Accumulator type for Q@K^T scores. f16 reduces register pressure but loses precision for long sequences. |
-| Producer warps | 1 warpgroup | 128 threads | Fixed at 1 warpgroup for TMA load; not tunable. |
-| Pipeline barrier type | mbarrier (default) or lock | mbarrier | `kIsMainloopLocked` switches to lock-based sync. |
-
-## Recommended configurations
-
-| Scenario | kNumMmaWarpGroups | kStagesKV | kClusterM | Notes |
-|---|---|---|---|---|
-| Large batch, D=128 | 2 | 2 | 1 | High compute; 2 consumer warpgroups fully utilize SM. |
-| Small batch, D=128 | 1 | 3 | 2 | Multicast saves bandwidth; persistent mode recommended. |
-| D=64 | 1 | 4 | 1 | Smaller tiles → more iterations → more stages to hide latency. |
-| D=256 | 2 | 2 | 1 | Large D already fills smem; minimize stages. |
-
-## Interaction with the rest of the tuning space
-
-The FMHA collective's tile shapes interact with the GEMM-level wgmma atom selection (the collective uses `SM90_64xNxK_*_SS_TN` atoms internally). The atom N-size is determined by BLOCK_N and HEAD_DIM. See `wiki/nvidia/hardware/wgmma/skill.md` for atom shape selection guidance.
-
-## Source reference
-
-Builder-level knobs: `kernel/fmha_options.hpp` (`Tag` enum) + `kernel/fmha_kernel_builder.hpp` (option resolution).
-Collective implementation: `collective/fmha_collective_tma_warpspecialized.hpp`.