agentsclimarketplace

Cmake cuda extension with git submodule vendoring

Skill kjuhwa/skills-hub/skills/build-system/cmake-cuda-extension-with-git-submodule-vendoring

Self-correcting knowledge corpus for Claude Code — 9 stable shape clusters, bias-correction pipeline baked into contribution flow. 47 papers, 45 techniques, 1.1k skills.

Install
npx -y skills add kjuhwa/skills-hub --skill cmake-cuda-extension-with-git-submodule-vendoring

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Build a PyTorch CUDA extension with CMake, vendoring CUTLASS and fmt as git submodules, plus a pre-built-wheel fast path for CI via GitHub releases.

SKILL.md

2.7 KB, 533 tokens by cl100k_base, as published. Nobody here has run it

CMake CUDA Extension With Git-Submodule Vendoring

What / Why

CUTLASS / CUTE / fmt are tightly coupled to kernel code — version skew = silent miscompiles. Vendoring via git submodule pins exact commits and avoids every user becoming a CUTLASS debugger. A CMake build keeps the CUDA compilation flags centralized. For CI, a DG_FORCE_BUILD env switch lets users download a pre-built wheel from GitHub releases when a matching Python × PyTorch × CUDA tag exists, skipping the 15-minute local compile.

Procedure

  1. Add submodules. 
    git submodule add https://github.com/nvidia/cutlass.git third-party/cutlass
git submodule add https://github.com/fmtlib/fmt.git     third-party/fmt
  2. Pin exact commits via git submodule update --init --recursive in install.sh. Lock CUTLASS to a known-good tag.
  3. setup.py branching.
    • If DG_FORCE_BUILD unset AND matching wheel exists at the releases URL: download, extract, install.
    • Else: call into CMake via CUDAExtension / torch.utils.cpp_extension.BuildExtension.
  4. CMakeLists.txt essentials.
    • include_dirs += third-party/cutlass/include, third-party/cutlass/include/cute, third-party/fmt/include, deep_gemm/include.
    • libraries += cudart nvrtc.
    • Flags: -std=c++20 -O3 -fPIC, PyTorch C++11 ABI alignment flag, CUDA arch list from torch.utils.cpp_extension.CUDA_ARCHITECTURES.
  5. develop.sh symlinks the include tree for in-place development; install.sh is the end-user wrapper.

Key design points

  • Pre-built wheel path is a cache, not a different code path. The same sources build the wheel and a local build — just shorter.
  • Header-only vendoring (CUTLASS, fmt) avoids static-lib ABI issues. If you need a compiled third-party lib, vendor source and build under your CMake, don't link externally.
  • Keep CUDA arch selection dynamic (from PyTorch), not hardcoded — users run SM80-SM100.

References

  • setup.py — wheel fast path + build invocation.
  • CMakeLists.txt — compile flags and include wiring.
  • install.sh / develop.sh — user-facing entry points.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.