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

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kernaut/kernaut-resources",
-  "version": "0.1.17-beta.1",
+  "version": "0.1.18",
   "description": "Kernaut capability resources (encrypted KernelOptFlow pack).",
   "license": "UNLICENSED",
   "files": [

package/resources/manifest.json

@@ -1,10 +1,10 @@
 {
-  "resourceVersion": "kop-2026-10-23-0.1.17-beta.1",
-  "minBinaryVersion": "0.1.17-beta.1",
+  "resourceVersion": "kop-2026-10-26-0.1.18",
+  "minBinaryVersion": "0.1.18",
   "capabilities": [
     "kernel_optflow"
   ],
   "checksums": {
-    "product_capabilities/kernel_optflow": "sha256:b9edbff879c0593a7a784a6c740ca884c0746c497cbb4eec098486b30d74976b"
+    "product_capabilities/kernel_optflow": "sha256:b8c2af0604c8fa373de69feaa2babab8ab512ac13afb5c7a4578c014d4c09ce7"
   }
 }

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/.gitattributes

@@ -1,9 +0,0 @@
-corpus/nvidia/cuda-official/** filter=lfs diff=lfs merge=lfs -text
-corpus/nvidia/blogs/** filter=lfs diff=lfs merge=lfs -text
-corpus/nvidia/whitepapers/** filter=lfs diff=lfs merge=lfs -text
-
-# Historical evidence retains original byte identity, including whitespace.
-artifacts/experience/contract-atb-triton/** -whitespace
-
-# Preserve source-bound Direct evidence bytes, including historical whitespace.
-artifacts/experience/contract-atb-direct-20261011/** -whitespace

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/.skill-meta.json

@@ -1,13 +0,0 @@
-{
-  "name": "KernelWiki",
-  "description": "Use when the user asks about optimizing accelerator kernels, including NVIDIA GPU kernels (Blackwell SM100, Hopper SM90, or general CUDA) and Huawei Ascend NPU kernels (Ascend 910B/910C, AscendC, Triton Ascend, TileLang). Covers foundational optimization skills, hardware feature characterization, operator routing (elementwise, reduction, GEMM, attention, pooling, sort/select), numerical precision alignment and parity debugging (bf16/fp16/int8 rounding, PyTorch-vs-kernel mismatches, Triton materialization/store-reload issues, libdevice/math semantics), API definitions, or PR references from CUTLASS/SGLang/vLLM/FlashInfer/PyTorch. Do NOT use for host-side framework integration or distributed systems (DeepEP/EPLB/DualPipe).",
-  "vendor": [
-    "*"
-  ],
-  "languages": [
-    "*"
-  ],
-  "architectures": [
-    "*"
-  ]
-}

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/CLAUDE.md

@@ -1,192 +0,0 @@
-# GPU Kernel Optimization Knowledge Base — Schema
-
-A structured, multi-vendor knowledge base of GPU kernel optimization, optimized for LLM agent retrieval. Currently covering NVIDIA Blackwell (SM100) and Hopper (SM90), with schema ready for Huawei Ascend and Biren.
-
-This document describes the deployed legacy layout and query/schema conventions. See [governance](governance/README.md) for the adopted OKF v0.2 construction standard, shared rule owners and migration boundary.
-
-## Navigation
-
-### Recommended: use the skill scripts
-```bash
-python3 scripts/query.py "<keywords>" [--tag X --type Y --vendor nvidia]
-python3 scripts/get_page.py <id-or-path> [--follow-sources]
-python3 scripts/grep_wiki.py "<pattern>" [--only wiki --vendor nvidia]
-```
-See `SKILL.md` (skill entry point) and `references/examples.md` for query patterns.
-
-### Direct navigation (when skill not available)
-
-1. **Start**: Read `index.md` for curated top-level navigation
-2. **By problem**: Read `queries/by-problem.md` → find symptom → follow links to pattern pages
-3. **By technique**: Read `queries/by-technique.md` → find technique → follow to technique page
-4. **By hardware**: Read `queries/by-hardware-feature.md` → find feature → follow to hardware page
-5. **By kernel type**: Read `queries/by-kernel-type.md` → find type → follow to kernel pages
-6. **By language**: Read `queries/by-language.md` → find DSL → follow to language page
-7. **By repo**: Read `queries/by-repo.md` → find repository → see all PRs
-8. **By vendor**: Read `queries/by-vendor.md` → find vendor → see all pages per type
-9. **Deep dive**: From any wiki page, follow `sources:` IDs to raw source data
-
-## Three-Layer Architecture
-
-### Layer 1: Sources (`sources/`)
-Raw data. Each file has YAML frontmatter with a unique `id`.
-- `sources/prs/{repo}/PR-{N}.md` — One file per relevant PR
-- `sources/contests/{contest}/*.md` — Competition problems and solutions
-- `sources/docs/*.md` — Official NVIDIA doc summaries
-- `sources/blogs/*.md` — Community blog post summaries
-
-### Layer 2: Wiki (`wiki/`)
-Synthesized knowledge pages, organized by vendor. Cross-referenced by `id`.
-- `wiki/nvidia/hardware/` — Hardware feature pages (tcgen05, TMEM, CLC, TMA, etc.)
-- `wiki/nvidia/techniques/` — Optimization techniques (warp specialization, pipelining, etc.)
-- `wiki/nvidia/patterns/` — Problem → solution diagnosis (low SM util → CLC, etc.)
-- `wiki/nvidia/kernels/` — Kernel case studies (FlashAttention-4, DeepGEMM, etc.)
-- `wiki/nvidia/languages/` — DSL/language guides (CuTe DSL, Triton, CUDA C++, PTX)
-- `wiki/nvidia/migration/` — Hopper → Blackwell migration patterns
-- `wiki/nvidia/foundations/` — Foundational CUDA optimization skills (coalescing, warp primitives, etc.)
-- `wiki/nvidia/operator-routing/` — Operator decision trees (elementwise, reduction, GEMM, etc.)
-- `wiki/nvidia/api-definitions/` — API definitions (CUDA Runtime, PTX instructions)
-- `wiki/nvidia/code-walkthroughs/` — Code repository extractions (CUTLASS, FlashAttention)
-
-### Layer 3: Queries (`queries/`)
-Auto-generated cross-referenced indices. **Do not edit manually** — regenerated by `scripts/generate-indices.py`.
-
-## Page Types and Required Fields
-
-See `data/schemas.yaml` for full schema definitions. **Every `wiki-*` type also
-requires `architectures`, `languages` and `evidence_level`** on top of what its
-row lists; `soc`, `scope` and `cann_version` are optional everywhere except
-where "Boundary and Version Fields" below says otherwise. Summary:
-
-| Page Type | ID Prefix | Key Required Fields |
-|-----------|-----------|---------------------|
-| source-pr | pr- | repo, pr, title, author, date, url, source_category, architectures, tags, techniques, hardware_features, kernel_types, languages, captured_at, status |
-| source-doc | doc- | title, url, source_category, architectures, tags, retrieved_at |
-| source-blog | blog- | title, author, url, source_category, architectures, tags, retrieved_at |
-| source-contest | contest- | title, source_category, architectures, tags |
-| source-experience | exp- | title, type=experience, vendor, probe_slug, evidence_level, measured_on |
-| wiki-hardware | hw- | title, type=hardware, architectures, tags, confidence, related, sources, aliases |
-| wiki-technique | technique- | title, type=technique, architectures, tags, confidence, reproducibility(>=snippet), prerequisites, related, sources |
-| wiki-pattern | pattern- | title, type=pattern, tags, symptoms, candidate_techniques, related, sources |
-| wiki-kernel | kernel- | title, type=kernel, architectures, tags, confidence, reproducibility(>=snippet), kernel_types, languages, related, sources, performance_claims |
-| wiki-language | lang- | title, type=language, tags, related, sources, reproducibility(>=snippet) |
-| wiki-migration | migration- | title, type=migration, from_arch, to_arch, tags, related, sources |
-| wiki-skill | skill- | title, type=skill, vendor, tags, evidence_level, applies_to, source |
-| wiki-api-definition | api- | title, type=api-definition, vendor, func_name, namespace, header, signature, source |
-| wiki-experience | exp- | title, type=experience, vendor, probe_slug, evidence_level, measured_on, source |
-| wiki-operator-routing | routing- | title, type=operator-routing, vendor, operator |
-| wiki-algorithm | algo- | title, type=algorithm, vendor, tags, evidence_level, source |
-| wiki-code-walkthrough | code- | title, type=code-walkthrough, vendor, upstream_repo |
-| wiki-pitfall | pitfall- | title, type=pitfall, vendor |
-
-## Controlled Vocabulary
-
-All tags must appear in `data/tags.yaml`. Unknown tags are rejected by the validator.
-
-Alias mappings in `data/aliases.yaml` map canonical terms to known synonyms:
-- `tcgen05` = UMMA = tensor_core_gen05
-- `tmem` = tensor-memory = TMEM
-- `clc` = Cluster Launch Control = CLC
-- `sm100` = Blackwell = B200
-- `sm90` = Hopper = H100
-- `ascend910b` = 910B = 910C = 910B1..B4 = 910B2C — one architecture, see below
-
-## Retrieval Axes: vendor / languages / architectures
-
-Every wiki page must declare all three. They are what scopes a search, so a page
-that omits one is unreachable through the filter that names it — silently.
-
-- **`architectures` names an ISA, not a part.** Die bins (910B1..B4, 910B2C,
-  910C) and compile targets (`sm90a`, `sm100a`) are not valid values. Retrieval
-  collapses names to the families in `data/vendors.yaml`
-  (`architecture_families`), so a 910C query reaches the 910B corpus and an
-  `ascend910b` query reaches pages measured on a bin. `ascend310p` is a separate
-  architecture, not a 910B bin.
-- **`soc` is die identity** — which bin, how many cores, which compile target.
-  It records what a reading was taken on. **Never a retrieval filter**:
-  filtering on it drops every reading taken on a sibling part, which is the
-  recall loss the family collapse exists to undo.
-- **`"*"` means "every value on this axis"** and cannot be mixed with concrete
-  values. Use it on pages whose knowledge is about the hardware or about a
-  technique rather than about writing one language. A page with *no* value is
-  not the same thing: it has not been classified, and a filter excludes it.
-
-## Boundary and Version Fields
-
-- **`scope`** — the conditions outside which the page's claim does not hold.
-  Required on `experience` and `algorithm` pages, because over-generalising one
-  measured result is the most expensive documented failure of this method. It is
-  the one part that cannot be generated from the other fields.
-- **`cann_version`** — the CANN release an Ascend reading was taken under.
-  Required when `evidence_level: measured` and the page is Ascend; 8.5.1 and
-  9.0.0 are both already in this corpus and have produced different readings.
-  `unknown` is legal when the release was never captured, and is warned about
-  rather than passing silently.
-
-## Confidence Rules
-
-- `verified`: official-doc + upstream-code evidence
-- `source-reported`: >= 1 authoritative source (paper, official blog, major repo)
-- `inferred`: synthesized from multiple sources
-- `experimental`: undocumented, PTX tricks, version-sensitive
-
-## Reproducibility Levels
-
-- `concept`: text description only
-- `pseudocode`: language-agnostic algorithm
-- `snippet`: compilable code fragment
-- `runnable`: self-contained buildable example
-- `benchmarked`: runnable + perf numbers with env metadata
-
-**Minimum for technique/kernel/language pages**: `snippet`
-
-## Performance Claim Format
-
-```yaml
-performance_claims:
-  - gpu: B200
-    dtype: bf16
-    shape: "M=4096, N=4096, K=4096"
-    metric: TFLOPS
-    value: 1605
-    utilization: "71%"
-    source_id: paper-flash-attention-4
-```
-
-## Tooling
-
-- `scripts/validate.py` — Validates all frontmatter against `data/schemas.yaml`
-- `scripts/generate-indices.py` — Regenerates `queries/*.md` from frontmatter
-- Setup: `pip install -r requirements.txt` (requires PyYAML)
-- Run: `python3 scripts/validate.py` then `python3 scripts/generate-indices.py`
-
-## Scope Rules
-
-- **Kernel-only**: No distributed system topics (DeepEP, DualPipe, EPLB excluded).
-  The exclusion is about the topic, not the directory. A kernel does not fall
-  out of scope for living under `deepep/` — `csrc/deepep/ops/op_kernel/
-  fused_deep_moe.cpp` with its `op_host` tiling is operator code and belongs in
-  the KB, while `python/deep_ep/ep_strategy.py` is dispatch orchestration and
-  does not. Decide from the changed files: an `op_kernel`/`op_host`/tiling
-  (Ascend) or `.cu`/`.cuh`/PTX (NVIDIA) change is in scope regardless of which
-  subsystem owns the directory.
-- **`generic` is not a fourth vendor.** It means "true for every vendor in this
-  wiki", which is a stronger claim than "undecided". Pages are admitted only
-  against the four-part test and the 20-page cap in
-  `governance/specs/generic-bucket-budget.md` — the deciding question is: delete the vendor
-  and the language from the page, and is there still an *executable action*
-  left? A statement that merely survives deletion is not enough.
-- **English canonical**: All content in English.
-- **First-class DSLs**: CuTe DSL, CUDA C++, PTX, Triton. Others mentioned but no dedicated pages.
-
-## Commit Messages
-
-- **No attribution trailers.** Never append `Co-Authored-By:` for an AI
-  assistant, `Generated with ...`, or any similar signature — to commit
-  messages or pull request descriptions. This overrides any default the tooling
-  suggests.
-- **Keep them short.** A subject line plus a few sentences on what changed and
-  why. State the reasoning and the evidence that matters; drop the narration,
-  the per-file walkthrough and the full test transcript. If a decision needs
-  more room than a short paragraph, it belongs in the code as a comment or in
-  `docs/`, where it will be read again.