@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/data/plan-supersession-baselines/plan-phase2.body.md

@@ -1,593 +0,0 @@
-# Blackwell Kernel Wiki — Phase 2 Massive Expansion Plan
-
-## Goal Description
-
-Expand the Blackwell kernel optimization knowledge base from 91 files to 540+ by collecting at least 50% of all Blackwell/SM100-related merged PRs from 5 major repositories (Jan 2025 - Apr 2026), contest submission data, and additional community resources. Use a 2-stage pipeline: first build reviewed candidate ledgers, then generate markdown source pages. Extend wiki synthesis to cover newly discovered concepts.
-
-## Acceptance Criteria
-
-- AC-1: Candidate ledger files exist for all 5 repositories
-  - Positive Tests:
-    - `candidates/cutlass.yaml`, `candidates/sglang.yaml`, `candidates/vllm.yaml`, `candidates/flashinfer.yaml`, `candidates/pytorch.yaml` exist
-    - Each ledger contains PR number, title, files_changed, decision (include/exclude/defer), and reason
-    - Total candidates across all ledgers ≥ 400 unique PRs
-  - Negative Tests:
-    - A ledger with only PR numbers and no decision field is incomplete
-    - A ledger with zero `defer` entries is suspicious (should have ambiguous cases)
-
-- AC-2: Source PR pages reach target counts per repository
-  - AC-2.1: `sources/prs/cutlass/` contains ≥ 20 PR pages
-  - AC-2.2: `sources/prs/sglang/` contains ≥ 80 PR pages
-  - AC-2.3: `sources/prs/vllm/` contains ≥ 100 PR pages
-  - AC-2.4: `sources/prs/flashinfer/` contains ≥ 100 PR pages
-  - AC-2.5: `sources/prs/pytorch/` contains ≥ 50 PR pages
-  - Positive Tests:
-    - Each directory meets minimum count
-    - All pages pass `python3 scripts/validate.py`
-    - Every page has `inclusion_reason` in frontmatter
-  - Negative Tests:
-    - A page missing `inclusion_reason` is rejected
-    - A page for a non-merged PR is rejected (status must be "merged")
-
-- AC-3: All new source pages have valid frontmatter and consistent metadata
-  - Positive Tests:
-    - `python3 scripts/validate.py` reports 0 errors on the full corpus
-    - Every PR page's `kernel_types` includes all kernel-type tags from its `tags` field
-    - Every PR page's `hardware_features` includes all hw tags from its `tags` field
-  - Negative Tests:
-    - A page with `tags: [tcgen05]` but `hardware_features: []` fails validation
-    - A page with duplicate values in any list field fails validation
-
-- AC-4: Contest data expanded with submission details
-  - Positive Tests:
-    - GPU Mode pages include top-3 solutions per problem with rank, score, technique analysis
-    - FlashInfer contest pages include analysis of notable fork submissions
-    - Each contest page has a `submissions` list in frontmatter
-  - Negative Tests:
-    - A contest page claiming "top solutions" without specific technique descriptions is incomplete
-
-- AC-5: Additional blogs and docs collected
-  - Positive Tests:
-    - `sources/blogs/` contains ≥ 20 pages (up from 10)
-    - `sources/docs/` contains ≥ 10 pages (up from 6)
-    - Each new page has valid frontmatter and language tags where applicable
-  - Negative Tests:
-    - A blog page without `retrieved_at` date is rejected
-
-- AC-6: Wiki layer expanded to cover new concepts
-  - Positive Tests:
-    - At least 10 new wiki pages created where 3+ new sources support an uncovered concept
-    - New wiki technique/kernel pages include code snippets (reproducibility ≥ snippet)
-    - All new wiki pages pass validation with architecture coverage expressed via `architectures`
-  - Negative Tests:
-    - A new wiki page with fewer than 3 supporting sources is premature
-
-- AC-7: Query indices regenerated and complete
-  - Positive Tests:
-    - `python3 scripts/generate-indices.py` runs successfully on the expanded corpus
-    - `queries/by-repo.md` shows all 5 repos with updated PR counts
-    - Link integrity check reports 0 broken links
-  - Negative Tests:
-    - A query page linking to a non-existent source file is broken
-
-- AC-8: Schema extended for new metadata fields
-  - Positive Tests:
-    - `data/schemas.yaml` includes `inclusion_reason` and `changed_paths` as optional fields for `source-pr`
-    - `data/schemas.yaml` includes `submissions` as optional field for `source-contest`
-    - `candidates/` directory is committed with ledger files
-  - Negative Tests:
-    - Using `inclusion_reason` in frontmatter without the schema update causes validation to ignore it
-
-## Path Boundaries
-
-### Upper Bound (Maximum Acceptable Scope)
-Complete ingestion of every Blackwell-related merged PR from all 5 repos (potentially 500+ PRs), full contest leaderboard data with detailed solution analysis, 25+ additional blogs/docs, 25+ new wiki pages with comprehensive code examples, and updated validation/generation tooling.
-
-### Lower Bound (Minimum Acceptable Scope)
-At least 350 new PR source pages (20 CUTLASS + 80 SGLang + 100 vLLM + 100 FlashInfer + 50 PyTorch), contest pages with top-3 solutions, 10 new blogs/docs, 10 new wiki pages, and regenerated query indices.
-
-### Allowed Choices
-- Can use: GitHub REST API (authenticated), `curl`, Python scripts, parallel sub-agents
-- Can use: Candidate YAML ledgers as intermediate artifacts
-- Fixed: Only merged PRs collected (per user decision)
-- Fixed: Include dispatch/selector PRs that enable Blackwell kernels (per user decision)
-- Fixed: Candidate ledgers committed to repo (per user decision)
-- Fixed: Contest top-3 + highlights (per user decision)
-- Cannot use: Unauthenticated API (rate limits too restrictive)
-
-## Feasibility Hints and Suggestions
-
-### Conceptual Approach
-
-```
-Milestone 1: Schema Extension + Candidate Collection
-  1a. Extend data/schemas.yaml with inclusion_reason, changed_paths, submissions
-  1b. For each repo, search GitHub API with multiple keywords by time period
-  1c. Deduplicate by PR number, classify include/exclude/defer
-  1d. Write candidates/{repo}.yaml
-
-Milestone 2: Source Page Generation (parallel per repo)
-  For each repo:
-    For each included candidate:
-      Fetch PR details → create sources/prs/{repo}/PR-{N}.md
-      Ensure all frontmatter fields populated and consistent
-
-Milestone 3: Contest + Blog/Doc Expansion
-  3a. Fetch GPU Mode leaderboards, analyze top-3 solutions
-  3b. Fetch FlashInfer contest forks, analyze notable submissions
-  3c. Collect additional blogs and docs
-
-Milestone 4: Wiki Synthesis
-  Scan new sources for uncovered concepts
-  Create new wiki pages where 3+ sources support a concept
-  Update existing wiki pages with better examples from new sources
-
-Milestone 5: Index Regeneration + Validation
-  Run validate.py, fix errors
-  Run generate-indices.py
-  Run link integrity check
-```
-
-### Relevant References
-- `scripts/validate.py` — Existing validator (441 lines)
-- `scripts/generate-indices.py` — Existing index generator (346 lines)
-- `data/schemas.yaml` — Page-type schemas to extend
-- `data/tags.yaml` — Controlled vocabulary
-- `.humanize/bitlesson.md` — Lessons from Phase 1 (tag validation, merge-vs-fallback)
-
-## Dependencies and Sequence
-
-### Milestones
-
-1. **Schema Extension**: Update schemas and create candidates/ directory
-   - Must complete before any candidate collection
-
-2. **Candidate Collection**: Parallel across 5 repos
-   - Each repo independently searchable
-   - Produces candidates/{repo}.yaml
-   - Depends on: Milestone 1
-
-3. **Source Page Generation**: Parallel across 5 repos
-   - Each repo's pages generated from its candidate ledger
-   - Depends on: Milestone 2 (that repo's ledger)
-
-4. **Contest + Blog/Doc Expansion**: Independent of PR collection
-   - Can run in parallel with Milestones 2-3
-   - Contest analysis may reference some PR pages
-
-5. **Wiki Synthesis**: Requires source data
-   - Depends on: Milestones 3-4 complete
-   - Scans all new sources to find uncovered concepts
-
-6. **Index Regeneration + Validation**: Final step
-   - Depends on: All previous milestones
-   - Regenerates all query pages, validates entire corpus
-
-## Task Breakdown
-
-| Task ID | Description | Target AC | Tag | Depends On |
-|---------|-------------|-----------|-----|------------|
-| task1 | Extend data/schemas.yaml with inclusion_reason, changed_paths, submissions | AC-8 | coding | - |
-| task2 | Create candidates/ directory structure | AC-1 | coding | task1 |
-| task3 | Collect CUTLASS candidate ledger (all keywords, 2025-01 to 2026-04) | AC-1, AC-2.1 | coding | task2 |
-| task4 | Collect SGLang candidate ledger (split by time periods) | AC-1, AC-2.2 | coding | task2 |
-| task5 | Collect vLLM candidate ledger (split by time periods) | AC-1, AC-2.3 | coding | task2 |
-| task6 | Collect FlashInfer candidate ledger (split by time periods) | AC-1, AC-2.4 | coding | task2 |
-| task7 | Collect PyTorch candidate ledger (split by time periods) | AC-1, AC-2.5 | coding | task2 |
-| task8 | Generate CUTLASS PR source pages from ledger | AC-2.1, AC-3 | coding | task3 |
-| task9 | Generate SGLang PR source pages from ledger | AC-2.2, AC-3 | coding | task4 |
-| task10 | Generate vLLM PR source pages from ledger | AC-2.3, AC-3 | coding | task5 |
-| task11 | Generate FlashInfer PR source pages from ledger | AC-2.4, AC-3 | coding | task6 |
-| task12 | Generate PyTorch PR source pages from ledger | AC-2.5, AC-3 | coding | task7 |
-| task13 | Expand GPU Mode contest pages (top-3 solutions + highlights) | AC-4 | coding | task2 |
-| task14 | Expand FlashInfer contest pages (fork analysis + submissions) | AC-4 | coding | task2 |
-| task15 | Collect additional blogs (10+ new) | AC-5 | coding | - |
-| task16 | Collect additional docs (4+ new) | AC-5 | coding | - |
-| task17 | Wiki synthesis: create new pages for uncovered concepts | AC-6 | coding | task8-task16 |
-| task18 | Regenerate all query indices | AC-7 | coding | task17 |
-| task19 | Run validate.py + link integrity check, fix all errors | AC-3, AC-7 | coding | task18 |
-| task20 | Audit: verify coverage targets met per repo | AC-2 | analyze | task19 |
-| task21 | Audit: verify contest data completeness | AC-4 | analyze | task19 |
-| task22 | Audit: verify wiki expansion quality | AC-6 | analyze | task19 |
-
-Tasks 3-7 can run in parallel. Tasks 8-12 can run in parallel (each depends on its own ledger). Tasks 13-16 can run in parallel with 3-12.
-
-## Claude-Codex Deliberation
-
-### Agreements
-- 2-stage pipeline (candidate ledger → page generation) is the right architecture
-- Candidate ledgers should be committed for auditability
-- Only merged PRs in scope
-- Dispatch/selector PRs included when they enable Blackwell kernels
-- Wiki expansion should be a separate stage with explicit entry criteria (3+ sources)
-- validate.py and generate-indices.py remain the quality gates
-- Deferred PRs should be tracked, not silently excluded
-
-### Resolved Disagreements
-- **Include/exclude strictness**: Claude proposed `≥1 hard + ≥1 semantic`. Codex argued this is too strict — misses generic-title bugfixes. Resolution: strong file evidence alone can `include`; semantic-only → `defer`. Dispatch/selector PRs explicitly included per user decision.
-- **Agent count**: Draft proposed 22 parallel agents. Codex suggested 7 (5 repo + 1 contest + 1 blog). Resolution: use 7 primary workers, each handling full time range internally with sequential API calls.
-- **Coverage metric**: Claude proposed "≥50% of API matches." Codex argued raw matches are noisy. Resolution: measure against deduped reviewed candidates (included + deferred), not raw API count. Target remains high per user requirement.
-
-### Convergence Status
-- Final Status: `converged`
-- Rounds: 1
-
-## Pending User Decisions
-
-- DEC-1: PR scope (dispatch PRs)
-  - Decision Status: `Include if enables Blackwell` (user confirmed)
-
-- DEC-2: Candidate ledger persistence
-  - Decision Status: `Commit as audit trail` (user confirmed)
-
-- DEC-3: Contest depth
-  - Decision Status: `Top-3 only + highlights` (user confirmed)
-
-- DEC-4: PR status filter
-  - Decision Status: `Merged only` (user confirmed)
-
-## Implementation Notes
-
-### Code Style Requirements
-- Implementation code and comments must NOT contain plan-specific terminology such as "AC-", "Milestone", "Step", "Phase", or similar workflow markers
-- These terms are for plan documentation only, not for the resulting codebase
-- Use descriptive, domain-appropriate naming in code instead
-
-### Inclusion Rules (Deterministic)
-
-**Include** a merged PR if ANY of these apply:
-1. Touches kernel files (`.cu`, `.cuh`, `.ptx`, Triton `.py`, CuTe DSL `.py`) AND title/body references SM100/Blackwell/tcgen05/TMEM/NVFP4
-2. Modifies scheduler/pipeline/epilogue/dispatch code for SM100 path
-3. Adds or changes kernel backend selection that routes to Blackwell kernels
-4. Contains Blackwell-specific performance benchmarks with kernel changes
-5. Fixes correctness/precision bug in a Blackwell kernel path
-
-**Exclude** a merged PR if ALL of these apply:
-- Title matches: `[CI]`, `[Doc]`, `[Docs]`, `bump`, `typo`, `format`, `lint`, `nit`, `revert`, `readme`, `changelog`
-- No kernel-bearing or dispatch file changes
-- No SM100/Blackwell-specific behavior change
-
-**Defer** when the agent cannot confidently classify. Record `defer_reason`.
-
-### Search Keywords Per Repo
-
-```
-Common: blackwell, sm100, sm_100, sm_100a, tcgen05, tmem, nvfp4, fp8, B200
-CUTLASS: 2sm, cta_group, clc, cluster_launch, umma, persistent, warp_special, fmha, mla, epilogue, tile_scheduler, grouped_gemm, sub_byte, block_scale
-SGLang: cutlass, deepgemm, flashmla, mla, flashinfer, triton, attention, moe, sparse, gated_delta, sgl-kernel, trtllm, decode, prefill, quantiz
-vLLM: cutlass, mla, deepseek, flashmla, attention, moe, fused_moe, sparse_attention, gated_delta, quantiz, block_scale, triton, flashinfer, grouped_gemm, decode, prefill
-FlashInfer: tcgen05, tmem, cutlass, cute_dsl, fmha, mla, attention, moe, fp4, block_scale, trtllm, decode, prefill, sparse, gated_delta, splitk, allreduce, head_dim, gemm, gemv
-PyTorch: sm_100a, cuda_13, nvfp4, inductor, triton, cutlass, flash_attention, vec128, launch_bounds, sm120, compute_10
-```
-
-### Rate Limiting
-- Single coordinator queues API requests
-- Max 30 requests/minute globally (well under 5000/hour authenticated limit)
-- Exponential backoff on 403/429 responses
-- Each sub-agent processes its repo sequentially, not in parallel bursts
-
---- Original Design Draft Start ---
-
-# Blackwell Kernel Wiki — 大规模扩展计划 (Phase 2)
-
-## 现状
-
-Phase 1 建立了知识库骨架（91 files），但PR覆盖严重不足：
-- CUTLASS: 7/~41+ Blackwell PRs
-- SGLang: 5/~194+ sm100 PRs (466 提到blackwell)
-- vLLM: 6/~254+ sm100 PRs (574 提到blackwell, 555 nvfp4)
-- FlashInfer: 6/~300+ sm100 PRs (247 提到blackwell)
-- PyTorch: 7/~141+ sm100 PRs (159 提到blackwell)
-- 比赛: 仅7个概述页，无具体提交方案
-
-注意：上述数字包含大量仅mention关键词的PR（如CI/doc/依赖更新），实际kernel相关PR需要过滤。
-
-## 目标
-
-从2025-01-01至今，完整覆盖每个仓库中**与Blackwell/Hopper kernel编程直接相关**的PR。同时收录比赛的具体提交方案和排名。
-
-## 过滤标准
-
-PR必须满足以下至少一条才收录：
-1. **新增/修改kernel代码**（CUDA C++, CuTe DSL, Triton, PTX, TileLang）
-2. **kernel性能优化**（tile size tuning, schedule改进, 新的pipeline策略）
-3. **新增硬件特性支持**（tcgen05, TMEM, CLC, 2-SM, NVFP4, FP8 block scale）
-4. **kernel算法创新**（新的attention variant, MoE routing, sparse pattern）
-5. **重要bug fix影响kernel正确性**（sync issues, precision bugs, launch bounds）
-
-不收录的PR类型：
-- 纯CI/CD改动
-- 文档/README更新（除非包含kernel技术细节）
-- 依赖版本bump
-- 代码格式化/lint
-- 纯Python binding（无kernel逻辑改动）
-
-## 搜索策略
-
-### NVIDIA/cutlass（预计收录 30-50 PRs）
-
-GitHub API显示 ~41 PRs提到"blackwell"。CUTLASS是Blackwell kernel的核心实现。
-
-**搜索关键词**（每个单独搜索，去重合并）：
-```
-blackwell, sm100, sm_100, sm_100a, tcgen05, tmem, tensor_memory,
-nvfp4, fp4, mxfp4, block_scale, 2sm, cta_group, clc,
-cluster_launch, umma, persistent, warp_special, fmha, mla,
-epilogue, tile_scheduler, grouped_gemm, sub_byte
-```
-
-**按时间段分批搜索**（避免API限制）：
-- 2025-01 ~ 2025-03（早期Blackwell支持）
-- 2025-04 ~ 2025-06（CUTLASS 4.0发布期）
-- 2025-07 ~ 2025-09（SM100 attention kernel）
-- 2025-10 ~ 2025-12（sub-byte GEMM, NVFP4）
-- 2026-01 ~ 2026-04（CuTe-DSL, bugfix）
-
-**重点关注**：
-- examples/cute/tutorial/blackwell/ 相关PR
-- SM100 FMHA forward/backward
-- Grouped GEMM / MoE support
-- CuTe-DSL SM100 atoms
-- Block-scale GEMM (NVFP4, MXFP4)
-- PersistentTileSchedulerSm100
-- 2-CTA cooperative MMA
-
-### sgl-project/sglang（预计收录 20-30 PRs）
-
-SGLang有大量Blackwell集成工作。~194 PRs提到sm100。
-
-**搜索关键词**：
-```
-blackwell, sm100, cutlass, deepgemm, flashmla, fp8, nvfp4,
-mla, flashinfer, triton, attention, moe, sparse, gated_delta,
-sgl-kernel, trtllm, decode, prefill, quantiz
-```
-
-**重点关注**：
-- DeepGEMM集成（FP8 GEMM, grouped GEMM）
-- FlashMLA集成（dense/sparse MLA decode）
-- CUTLASS MLA backend
-- FP8 blockwise quantization
-- NSA (Native Sparse Attention) 支持
-- GatedDeltaNet 支持
-- Blackwell-specific kernel选择逻辑
-
-### vllm-project/vllm（预计收录 25-40 PRs）
-
-vLLM有最多的Blackwell适配PR。~254 PRs提到sm100。
-
-**搜索关键词**：
-```
-blackwell, sm100, cutlass, fp8, nvfp4, mxfp4, B200,
-mla, deepseek, flashmla, attention, moe, fused_moe,
-sparse_attention, gated_delta, quantiz, block_scale,
-triton, flashinfer, grouped_gemm, decode, prefill
-```
-
-**重点关注**：
-- CUTLASS FP8 GEMM on SM100
-- NVFP4/MXFP4 GEMM和fused MoE
-- CUTLASS MLA for Blackwell
-- DeepSeek V3.2 sparse attention支持
-- GatedDeltaNet/Qwen3-Next支持
-- FP8 block-scale quantization kernels
-- Attention backend selection for SM100
-
-### flashinfer-ai/flashinfer（预计收录 25-35 PRs）
-
-FlashInfer是Blackwell attention kernel的核心库。~300 PRs提到sm100。
-
-**搜索关键词**：
-```
-blackwell, sm100, sm_100, tcgen05, tmem, cutlass, cute_dsl,
-fmha, mla, attention, moe, fp8, fp4, nvfp4, block_scale,
-trtllm, decode, prefill, sparse, gated_delta, splitk,
-allreduce, head_dim, gemm, gemv
-```
-
-**重点关注**：
-- CuTe-DSL FMHA kernels for SM100
-- SM100 MLA kernels (forward/backward)
-- FP4/FP8 fused MoE on Blackwell
-- FlashInfer-Bench kernel definitions
-- TGV GEMM和alternative backends
-- head_dim variants for SM100
-- GEMM+allreduce fusion
-
-### pytorch/pytorch（预计收录 15-25 PRs）
-
-PyTorch的Blackwell支持更侧重build/runtime层面，但也有kernel相关改动。
-
-**搜索关键词**：
-```
-blackwell, sm100, sm_100, sm_100a, cuda_13, nvfp4,
-B200, inductor, triton, cutlass, flash_attention,
-vec128, launch_bounds, sm120, compute_10
-```
-
-**重点关注**：
-- TorchInductor SM100 code generation
-- Triton kernel templates for Blackwell
-- CUTLASS integration for SM100
-- NVFP4 support in torch
-- Flash attention Blackwell path
-- Launch bounds fixes for SM100/SM120
-- 新的vectorized ops (vec128)
-
-## 比赛数据扩展
-
-### GPU Mode NVFP4 Blackwell Hackathon
-
-**当前**：4个problem概述页
-**需要新增**：
-1. **排名数据**：每个problem的完整leaderboard（top 10+排名、用户名、成绩）
-2. **获奖方案分析**：
-   - Problem 1 (GEMV): Top 3方案的具体实现策略和代码模式
-   - Problem 2 (GEMM): Simon/yue/currybab的kernel设计
-   - Problem 3 (Gated Dual GEMM): 前几名的fusion策略
-   - Problem 4 (Grouped GEMM): 前几名的CLC/tile策略
-3. **参赛者博客和代码**：
-   - 搜索GitHub repos: `nvfp4 hackathon`, `gpu-mode blackwell`
-   - 搜索博客: GPU Mode Discord讨论摘要
-4. **Reward hack分析**：已有概述，扩展技术细节
-
-**搜索方法**：
-```
-# GitHub repos
-curl "https://api.github.com/search/repositories?q=nvfp4+hackathon"
-curl "https://api.github.com/search/repositories?q=gpu-mode+blackwell+kernel"
-
-# GPU Mode leaderboard API (if available)
-# https://www.gpumode.com/leaderboard/595 (GEMV)
-# https://www.gpumode.com/leaderboard/597 (GEMM)
-
-# Participant blogs
-WebSearch: "blackwell nvfp4 hackathon solution"
-WebSearch: "gpu mode hackathon blackwell winner"
-WebSearch: "nvfp4 kernel optimization blog"
-```
-
-### FlashInfer AI Kernel Generation Contest (MLSys 2026)
-
-**当前**：3个track概述页
-**需要新增**：
-1. **提交方案仓库**：搜索所有flashinfer-bench-starter-kit的forks（17个fork）
-2. **每个fork的solution分析**：
-   - Track A (Fused MoE): 参赛者的kernel实现
-   - Track B (Sparse Attention): 参赛者的indexer+attention kernel
-   - Track C (GatedDeltaNet): 参赛者的prefill/decode kernel
-3. **Agent baseline分析**：
-   - flashinfer-ai/mlsys26-agent-baseline 的evolve/iterative agent策略
-   - K-Search论文的co-evolving world model方法
-4. **FlashInfer-Bench leaderboard数据**：
-   - bench.flashinfer.ai 的AI-generated kernel排名
-   - 每个model (Gemini, GPT-5, Claude) 的per-kernel性能
-
-**搜索方法**：
-```
-# Contest forks
-curl "https://api.github.com/repos/flashinfer-ai/flashinfer-bench-starter-kit/forks?per_page=100"
-
-# Contest submissions tagged on GitHub
-curl "https://api.github.com/search/repositories?q=flashinfer+mlsys+2026"
-curl "https://api.github.com/search/repositories?q=flashinfer-bench"
-
-# Leaderboard
-WebFetch: https://bench.flashinfer.ai/
-WebSearch: "flashinfer mlsys 2026 contest results"
-WebSearch: "flashinfer bench leaderboard kernel"
-```
-
-## 执行策略
-
-### 并行度
-
-大仓库需要更多sub-agent，每个agent负责一个时间段内50-80个PR的搜索+过滤+创建：
-
-**CUTLASS**（41 PRs, 小规模）：2 agents
-- Agent: 2025-01 ~ 2025-09
-- Agent: 2025-10 ~ 2026-04
-
-**SGLang**（194 PRs）：4 agents
-- Agent: 2025-01 ~ 2025-03
-- Agent: 2025-04 ~ 2025-06
-- Agent: 2025-07 ~ 2025-09
-- Agent: 2025-10 ~ 2026-04
-
-**vLLM**（254 PRs）：5 agents
-- Agent: 2025-01 ~ 2025-03
-- Agent: 2025-04 ~ 2025-06
-- Agent: 2025-07 ~ 2025-09
-- Agent: 2025-10 ~ 2025-12
-- Agent: 2026-01 ~ 2026-04
-
-**FlashInfer**（300 PRs）：5 agents
-- Agent: 2025-01 ~ 2025-03
-- Agent: 2025-04 ~ 2025-06
-- Agent: 2025-07 ~ 2025-09
-- Agent: 2025-10 ~ 2025-12
-- Agent: 2026-01 ~ 2026-04
-
-**PyTorch**（141 PRs）：3 agents
-- Agent: 2025-01 ~ 2025-06
-- Agent: 2025-07 ~ 2025-12
-- Agent: 2026-01 ~ 2026-04
-
-**比赛**：3 agents
-- Agent: GPU Mode hackathon leaderboards + solutions
-- Agent: FlashInfer contest forks + submissions
-- Agent: Additional blogs + docs
-
-总共：**2 + 4 + 5 + 5 + 3 + 3 = 22 个并行 sub-agents**
-
-每批可发6个并行agent，需要 4 批执行。
-
-### 每个sub-agent的工作流
-
-```
-1. 用多个关键词搜索GitHub API，收集所有匹配PR的number和title
-2. 去重合并
-3. 对每个PR：
-   a. 获取PR描述（gh api repos/{repo}/pulls/{number}）
-   b. 根据过滤标准判断是否收录
-   c. 如果收录：提取metadata，创建sources/prs/{repo}/PR-{N}.md
-4. 使用严格的controlled vocabulary（从data/tags.yaml）
-5. 确保frontmatter完整且validate.py通过
-```
-
-### 质量保证
-
-1. 每个sub-agent完成后立即运行 `python3 scripts/validate.py`
-2. 不合格的文件立即修复
-3. 最终运行 `python3 scripts/generate-indices.py` 重建所有索引
-4. 运行link integrity check
-
-### 预期结果（收录至少一半匹配PR）
-
-| 来源 | API匹配数 | 目标（≥50%） | 当前 | 新增 |
-|------|----------|-------------|------|------|
-| CUTLASS PRs | 41 (blackwell) | **20+** | 7 | 13+ |
-| SGLang PRs | 194 (sm100) | **100+** | 5 | 95+ |
-| vLLM PRs | 254 (sm100) | **130+** | 6 | 124+ |
-| FlashInfer PRs | 300 (sm100) | **150+** | 6 | 144+ |
-| PyTorch PRs | 141 (sm100) | **70+** | 7 | 63+ |
-| GPU Mode 比赛 | leaderboards + solutions | **20+** | 4 | 16+ |
-| FlashInfer 比赛 | 17 forks + leaderboard | **15+** | 3 | 12+ |
-| Blogs | identified | **25+** | 10 | 15+ |
-| Docs | identified | **12+** | 6 | 6+ |
-| **Total sources** | **~930+** | **540+** | **54** | **~490+ new** |
-
-加上现有37 wiki pages + 大量新增wiki pages → **总计 600-700+ files**
-
-### 过滤策略说明
-
-GitHub API匹配数中有大量重复（多个关键词匹配同一PR）和低相关PR。实际唯一PR数约为匹配数的60-70%。过滤掉纯CI/doc/bump后，目标是保留**至少50%的唯一匹配PR**。
-
-对于大仓库（SGLang 194, vLLM 254, FlashInfer 300），需要：
-1. 先用API获取全部匹配PR列表（number + title）
-2. 按title快速过滤掉明显无关的（`[CI]`, `[Doc]`, `bump`, `typo`, `format`）
-3. 对剩余PR获取详情并创建source page
-4. 每个sub-agent处理50-100个PR，需要拆更多时间段
-
-## 额外博客和文档待收录
-
-### 博客（新增 10-15 篇）
-- Simon's NVFP4 GEMV Blog (Part 1 + Part 2)
-- TFLOPS Gap: FP4 MoE Kernel Engineering on Blackwell (HuggingFace)
-- NVFP4 Format Details (Harold Benoit)
-- JAX Pallas Blackwell Matmul tutorial
-- Tilus ASPLOS paper summary
-- Blackwell Microbenchmarking papers (arxiv 2512.02189, 2507.10789)
-- vLLM DeepSeek V3.2 Sparse Attention blog
-- Qwen3-Next architecture blog (NVIDIA, Alibaba)
-- K-Search automated kernel generation (arxiv 2602.19128)
-- GPU Mode "Anatomy of a Reward Hack" full analysis
-
-### 文档（新增 4-6 篇）
-- CUTLASS Changelog (SM100 entries)
-- CUTLASS CLC documentation
-- Blackwell Compatibility Guide
-- NVIDIA cuTile Python DSL reference
-- Tilus documentation
-
---- Original Design Draft End ---