[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AVBVF39A2KGKYFMG96XRWW":3,"skill-same:01M7AVBVF39A2KGKYFMG96XRWW":60},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":33,"commitSha":34,"skillDirectory":35,"forksCount":14,"repositoryHtmlUrl":36,"repositoryHomepageUrl":6,"ownerHtmlUrl":37,"ownerBlogUrl":37,"ownerDescription":38,"updatedAt":39,"install":40,"resources":53},"01M7AVBVF39A2KGKYFMG96XRWW","vllm-source-analysis",null,"Use when analyzing vLLM source or comparing vLLM versions.","peterlau123\u002Fskills-you-need","peterlau123","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F19188200?v=4","01M7AV0Y55C30Y4E2TGRQ8X7WR","Sloth","01M7AV0Y5SNAJQ4E3FG1EEK8S4",0,"MIT","https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fmlops\u002Fvllm-source-analysis\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: vllm-source-analysis\ndescription: \"Use when analyzing vLLM source or comparing vLLM versions.\"\nversion: 1.0.0\nlicense: MIT\nmetadata:\n  hermes:\n    tags: [vllm, source-analysis, operator-integration, architecture, version-comparison]\n---\n\n# vLLM Source Architecture Analysis\n\nUse when the user asks to analyze the vLLM codebase (`\u002Fdata\u002Flx\u002Fvllm`), understand how operators\nintegrate, compare versions (e.g. v0.13.0 vs v0.22.0), or assess third-party dependencies. The\nuser is a vLLM + MetaX GPU adaptation engineer — these analyses feed technical reports, papers,\nand adaptation work (see memory). Reply in Chinese.\n\n## Repo facts (mx001)\n\n- Repo: `\u002Fdata\u002Flx\u002Fvllm`. Tags available locally: `v0.13.0`, `v0.13.0rc1-4`, plus a sync branch —\n  **no v0.2x tags are fetched locally**, so a git-diff across major versions is impossible;\n  compare by checking out the tag and re-counting.\n- Check out: `git stash && git checkout v0.13.0`, verify with `git describe --tags`.\n- Scale: v0.13.0 = 1131 Python files \u002F ~429k lines; v0.22.0 = ~1752 files (+55%).\n\n## The core question: how do C++ extensions integrate?\n\nNot all vLLM extensions go through `torch.ops._C::*`. Three invocation classes (verified v0.13.0,\nstill the model for later versions):\n\n1. **torch.ops-registered namespaces** — `vllm._C` → `torch.ops._C::*` (~125 refs in\n   `vllm\u002F_custom_ops.py`); `vllm._moe_C` → its OWN namespace `torch.ops._moe_C::*` (independent\n   `TORCH_LIBRARY` block in `csrc\u002Fmoe\u002Ftorch_bindings.cpp`, 13 ops). Don't lump _moe_C under _C.\n2. **Python import + direct function call** — `vllm.vllm_flash_attn._vllm_fa2\u002Ffa3_C` (call\n   `flash_attn_varlen_func` directly), `vllm._flashmla_C` (`import vllm._flashmla_C` then\n   `flash_mla_with_kvcache`), `vllm._rocm_C`. The `.so` loads via `REGISTER_EXTENSION` →\n   `PyInit_\u003Cname>` (csrc\u002Fcore\u002Fregistration.h), NOT the ops registry.\n3. **Non-ops interfaces** — `vllm.cumem_allocator` implements PyTorch `CUDAPluggableAllocator`\n   (registered as a memory allocator, `torch.cuda.memory.CUDAPluggableAllocator`), and\n   `vllm.triton_kernels` is a C++ wrapper of Triton kernels imported as a plain module.\n\nBuild-time extension list: `grep 'CMakeExtension(name=' setup.py` (9 modules in v0.13.0:\n`_C`, `_moe_C`, `cumem_allocator`, `triton_kernels`, `_rocm_C`, `_vllm_fa2_C`, `_vllm_fa3_C`,\n`_flashmla_C`, `_flashmla_extension_C`). External kernels come via CMake FetchContent\n(cutlass v4.2.1, vllm-flash-attn, FlashMLA, triton_kernels, qutlass) — NOT vendored in git\n(e.g. `vllm\u002Fvllm_flash_attn\u002F` is a `.gitkeep` placeholder).\n\n## Analysis workflow\n\n1. `git checkout \u003Ctag>` + `git describe --tags` to confirm.\n2. Count scale: `find vllm\u002F -name '*.py' | wc -l`.\n3. Run the quick probes (see `scripts\u002Fvllm_arch_probe.sh`) for: extension list from setup.py,\n   `@triton.jit` count (133 in 60 files at v0.13.0), `torch.distributed` \u002F `torch.compile` \u002F\n   `torch.fx` \u002F `torch.func` \u002F `torch.accelerator` ref counts, presence of key dirs\n   (`vllm\u002Fir\u002F`, `vllm\u002Fcompilation\u002Fpasses\u002F`, `vllm\u002Fplatforms\u002F`).\n4. For feature absence claims, VERIFY with targeted grep before asserting — a first-pass grep\n   this session wrongly claimed the plugin system was absent in v0.13.0 when it existed\n   (`vllm\u002Fplugins\u002F` with 4 entry-point groups, `pyproject.toml:44`). Absence is a strong claim.\n5. For big comparisons, dispatch parallel subagents (one per theme: operator integration \u002F\n   PyTorch capabilities \u002F third-party deps) — they produce much deeper counts than quick greps.\n6. Deliver results into a Feishu doc (see `feishu-docs` skill): **primary analyzed version as\n   main body, newer version's deltas as an appendix** (user preference, 2026-08-14).\n\n## Version landmarks (v0.13.0 vs v0.22.0)\n\n- v0.13.0 is **V1-only** (`vllm\u002Fengine\u002Fllm_engine.py` is a one-line shim to v1).\n- v0.13.0 has NO `vllm\u002Fir\u002F` (IrOp system is post-0.13), no QuACK, no Rust frontend, no\n  `_C_AVX512\u002F_C_AVX2`, no spinloop, torch==2.9.0 (v0.22.0: 2.11.0 + torch.accelerator).\n- v0.13.0 compilation passes are flat in `vllm\u002Fcompilation\u002F`; v0.22.0 reorganizes into\n  `passes\u002Ffusion\u002F` and adds BreakableCUDAGraphCapture.\n- Both versions share: CustomOp framework (8 dispatch targets, 38 `@CustomOp.register`),\n  `direct_register_custom_op` + `Library(\"vllm\", \"FRAGMENT\")`, plugin system, ~23 attention\n  backends (v0.13.0), SymmetricMemory, StatelessProcessGroup, ctypes NCCL wrapper.\n\n## Pitfalls\n\n- `search_files` (ripgrep) returned EMPTY for files that verifiably contain the pattern —\n  likely .gitignore filtering in the repo. Fall back to `grep -r` in terminal; it's authoritative.\n- `grep 'CMakeExtension(name='` in a python3 -c heredoc inside a bash string breaks on the\n  parentheses — use a file or `grep` directly, not nested python -c with quotes.\n- No pip in Hermes venv python3.11 — use `\u002Fusr\u002Fbin\u002Fpython3` (3.10) for any package installs.\n","# vLLM Source Architecture Analysis\n\nUse when the user asks to analyze the vLLM codebase (`\u002Fdata\u002Flx\u002Fvllm`), understand how operators\nintegrate, compare versions (e.g. v0.13.0 vs v0.22.0), or assess third-party dependencies. The\nuser is a vLLM + MetaX GPU adaptation engineer — these analyses feed technical reports, papers,\nand adaptation work (see memory). Reply in Chinese.\n\n## Repo facts (mx001)\n\n- Repo: `\u002Fdata\u002Flx\u002Fvllm`. Tags available locally: `v0.13.0`, `v0.13.0rc1-4`, plus a sync branch —\n  **no v0.2x tags are fetched locally**, so a git-diff across major versions is impossible;\n  compare by checking out the tag and re-counting.\n- Check out: `git stash && git checkout v0.13.0`, verify with `git describe --tags`.\n- Scale: v0.13.0 = 1131 Python files \u002F ~429k lines; v0.22.0 = ~1752 files (+55%).\n\n## The core question: how do C++ extensions integrate?\n\nNot all vLLM extensions go through `torch.ops._C::*`. Three invocation classes (verified v0.13.0,\nstill the model for later versions):\n\n1. **torch.ops-registered namespaces** — `vllm._C` → `torch.ops._C::*` (~125 refs in\n   `vllm\u002F_custom_ops.py`); `vllm._moe_C` → its OWN namespace `torch.ops._moe_C::*` (independent\n   `TORCH_LIBRARY` block in `csrc\u002Fmoe\u002Ftorch_bindings.cpp`, 13 ops). Don't lump _moe_C under _C.\n2. **Python import + direct function call** — `vllm.vllm_flash_attn._vllm_fa2\u002Ffa3_C` (call\n   `flash_attn_varlen_func` directly), `vllm._flashmla_C` (`import vllm._flashmla_C` then\n   `flash_mla_with_kvcache`), `vllm._rocm_C`. The `.so` loads via `REGISTER_EXTENSION` →\n   `PyInit_\u003Cname>` (csrc\u002Fcore\u002Fregistration.h), NOT the ops registry.\n3. **Non-ops interfaces** — `vllm.cumem_allocator` implements PyTorch `CUDAPluggableAllocator`\n   (registered as a memory allocator, `torch.cuda.memory.CUDAPluggableAllocator`), and\n   `vllm.triton_kernels` is a C++ wrapper of Triton kernels imported as a plain module.\n\nBuild-time extension list: `grep 'CMakeExtension(name=' setup.py` (9 modules in v0.13.0:\n`_C`, `_moe_C`, `cumem_allocator`, `triton_kernels`, `_rocm_C`, `_vllm_fa2_C`, `_vllm_fa3_C`,\n`_flashmla_C`, `_flashmla_extension_C`). External kernels come via CMake FetchContent\n(cutlass v4.2.1, vllm-flash-attn, FlashMLA, triton_kernels, qutlass) — NOT vendored in git\n(e.g. `vllm\u002Fvllm_flash_attn\u002F` is a `.gitkeep` placeholder).\n\n## Analysis workflow\n\n1. `git checkout \u003Ctag>` + `git describe --tags` to confirm.\n2. Count scale: `find vllm\u002F -name '*.py' | wc -l`.\n3. Run the quick probes (see `scripts\u002Fvllm_arch_probe.sh`) for: extension list from setup.py,\n   `@triton.jit` count (133 in 60 files at v0.13.0), `torch.distributed` \u002F `torch.compile` \u002F\n   `torch.fx` \u002F `torch.func` \u002F `torch.accelerator` ref counts, presence of key dirs\n   (`vllm\u002Fir\u002F`, `vllm\u002Fcompilation\u002Fpasses\u002F`, `vllm\u002Fplatforms\u002F`).\n4. For feature absence claims, VERIFY with targeted grep before asserting — a first-pass grep\n   this session wrongly claimed the plugin system was absent in v0.13.0 when it existed\n   (`vllm\u002Fplugins\u002F` with 4 entry-point groups, `pyproject.toml:44`). Absence is a strong claim.\n5. For big comparisons, dispatch parallel subagents (one per theme: operator integration \u002F\n   PyTorch capabilities \u002F third-party deps) — they produce much deeper counts than quick greps.\n6. Deliver results into a Feishu doc (see `feishu-docs` skill): **primary analyzed version as\n   main body, newer version's deltas as an appendix** (user preference, 2026-08-14).\n\n## Version landmarks (v0.13.0 vs v0.22.0)\n\n- v0.13.0 is **V1-only** (`vllm\u002Fengine\u002Fllm_engine.py` is a one-line shim to v1).\n- v0.13.0 has NO `vllm\u002Fir\u002F` (IrOp system is post-0.13), no QuACK, no Rust frontend, no\n  `_C_AVX512\u002F_C_AVX2`, no spinloop, torch==2.9.0 (v0.22.0: 2.11.0 + torch.accelerator).\n- v0.13.0 compilation passes are flat in `vllm\u002Fcompilation\u002F`; v0.22.0 reorganizes into\n  `passes\u002Ffusion\u002F` and adds BreakableCUDAGraphCapture.\n- Both versions share: CustomOp framework (8 dispatch targets, 38 `@CustomOp.register`),\n  `direct_register_custom_op` + `Library(\"vllm\", \"FRAGMENT\")`, plugin system, ~23 attention\n  backends (v0.13.0), SymmetricMemory, StatelessProcessGroup, ctypes NCCL wrapper.\n\n## Pitfalls\n\n- `search_files` (ripgrep) returned EMPTY for files that verifiably contain the pattern —\n  likely .gitignore filtering in the repo. Fall back to `grep -r` in terminal; it's authoritative.\n- `grep 'CMakeExtension(name='` in a python3 -c heredoc inside a bash string breaks on the\n  parentheses — use a file or `grep` directly, not nested python -c with quotes.\n- No pip in Hermes venv python3.11 — use `\u002Fusr\u002Fbin\u002Fpython3` (3.10) for any package installs.\n",{"name":5,"license":15,"version":24,"metadata":25,"description":7},"1.0.0",{"hermes":26},{"tags":27},[28,29,30,31,32],"vllm","source-analysis","operator-integration","architecture","version-comparison","main","c0b220603dea88e986e5b31268098ef2f5b07fd0","mlops\u002Fvllm-source-analysis","https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need","https:\u002F\u002Fgithub.com\u002Fpeterlau123","A software engineer in deep learning\r\n","2026-08-19 16:46:00",{"source":8,"skillName":5,"commands":41,"prompt":51,"zipAvailable":52},[42,45,48],{"tool":43,"command":44},"npx","npx skills add peterlau123\u002Fskills-you-need --skill vllm-source-analysis",{"tool":46,"command":47},"bunx","bunx skills add peterlau123\u002Fskills-you-need --skill vllm-source-analysis",{"tool":49,"command":50},"pnpm","pnpm dlx skills add peterlau123\u002Fskills-you-need --skill vllm-source-analysis","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fmlops\u002Fvllm-source-analysis\u002FSKILL.md",true,[54],{"path":55,"resourceType":56,"resourceTypeDesc":57,"sizeBytes":58,"sourceHtmlUrl":59},"scripts\u002Fvllm_arch_probe.sh","SCRIPT","脚本",2013,"https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fmlops\u002Fvllm-source-analysis\u002Fscripts\u002Fvllm_arch_probe.sh",[61,68,75,82,89,96],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M7AVH2D16QBRC8T2P6SVXSTS","jekyll-chirpy-blog","Use when editing the Jekyll\u002FChirpy blog or fixing mermaid.","https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fweb\u002Fjekyll-chirpy-blog\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":19},"01M7AVGYYFDK4GKCSNHBCP59Y6","test-driven-development","TDD: enforce RED-GREEN-REFACTOR, tests before code.","https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fsoftware-development\u002Ftest-driven-development\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M7AVGT9GAVB5EBKZ2616R27Z","systematic-debugging","4-phase root cause debugging: understand bugs before fixing.","https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fsoftware-development\u002Fsystematic-debugging\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":19},"01M7AVGQ23T5EHRQP5GZ45KZC5","spike","Throwaway experiments to validate an idea before build.","https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fsoftware-development\u002Fspike\u002FSKILL.md",[],[],{"skillNo":90,"name":91,"nameZh":6,"description":92,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":19},"01M7AVGKZ0ZD91ATKD9QB44E1B","simplify-code","Parallel 4-agent cleanup of recent code changes.","https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fsoftware-development\u002Fsimplify-code\u002FSKILL.md",[],[],{"skillNo":97,"name":98,"nameZh":6,"description":99,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":100,"categories":101,"keywords":102,"publishedAt":19},"01M7AVGFBR76BWQNG2X4CN5SQT","requesting-code-review","Pre-commit review: security scan, quality gates, auto-fix.","https:\u002F\u002Fgithub.com\u002Fpeterlau123\u002Fskills-you-need\u002Fblob\u002Fmain\u002Fsoftware-development\u002Frequesting-code-review\u002FSKILL.md",[],[]]