Implement and optimize CUTLASS and CuTe C++ kernels with validation.
npx skills add https://github.com/vipshop/cache-dit --skill cutlass-cpp-kernel
Please help me install this Agent Skill. Skill: cutlass-cpp-kernel Source: https://github.com/vipshop/cache-dit/tree/main/.copilot/skills/cutlass-cpp-kernel Command: npx skills add https://github.com/vipshop/cache-dit --skill cutlass-cpp-kernel
Provides a focused workflow to read, implement, debug, and optimize CUTLASS and CuTe C++ kernels so engineers can achieve correct, high-performance GEMM, pipeline, and epilogue implementations across modern NVIDIA architectures.
Open the workspace CUTLASS checkout at /workspace/dev/vipshop/cutlass and ask to analyze or optimize a specific CUTLASS/CuTe kernel by naming the target file, target GPU architecture, data types, and the desired correctness or performance goal.
High-intent search queries and answers about installing and using this skill.
Optimize CUTLASS C++ kernels by analyzing tile shapes, MMA/copy atoms, stage counts, and epilogue choices. Apply architecture-specific smXX optimization guides and interpret Nsight outputs to tune performance for Hopper and Blackwell GPU targets.
Debug CuTe C++ kernel correctness using a stepwise checklist for synchronization and schedule errors. Validate numerical parity by running unit tests and comparing kernel outputs against PyTorch baselines to ensure accuracy.
Port handwritten C++ GEMM code to CuTe DSL workflows using rewrite guidance that preserves operator contracts. Validate both accuracy and performance during the transition to ensure the new implementation matches the original kernel behavior.
You need a workspace CUTLASS checkout path to analyze and tune GEMM kernels. Provide the target file, GPU architecture, data types, and desired correctness or performance goal to begin the optimization process.
Find CuTe C++ headers and example kernels for GEMM implementation using source navigation. This maps key CUTLASS header locations and example kernels to enable fast discovery and code reuse across template-based pipelines.
CUTLASS epilogue pipeline schedule errors and synchronization issues arise from misconfigured tiling and staging. Diagnose these problems by analyzing collective-builder code and applying a stepwise debugging checklist to validate execution.
NVIDIAAnalyze Nsight Compute .ncu-rep reports to classify CUDA kernel bottlenecks.
linkedinAutomate Triton kernel development for Liger Kernel with a three-stage pipeline.
m0atAnalyze CUDA kernel performance bottlenecks with Nsight Systems and Nsight Compute.
NVIDIAWrite GPU kernels with NVIDIA CuTe DSL using Python decorators.
