Agent skill · software engineering · aradotso
tilekernels-gpu-kernels
Expert skill for using TileKernels, a library of optimized GPU kernels for LLM operations (MoE routing, quantization, transpose, engram gating, Manifold HyperConnection) built with TileLang.
Why this skill is useful
Adds optimized GPU kernel implementations for LLM operations that are not available in public documentation.
What it needs
Requires cuda, python, tilelang, torch installed locally. About 5k tokens when loaded. Last updated 2026-07-09. 70 stars on the source repository.
What this skill does
TileKernels GPU Kernel Library Skill by ara.so — Daily 2026 Skills collection. TileKernels is a high-performance GPU kernel library for LLM operations (MoE routing, FP8/FP4 quantization, transpose, engram gating, Manifold HyperConnection) written in TileLang — a Python DSL for expressing GPU kernels with automatic optimization. Kernels target NVIDIA SM90/SM100 (Hopper/Blackwell) architectures and approach hardware performance limits. Requirements Python 3.10+ PyTorch 2.10+ TileLang 0.1.9+ NVIDIA SM90 or SM100 GPU (H100/H200/B100/B200) CUDA Toolkit 13.1+ Installation Project Structure Key Modules and Usage MoE Routing Kernels Quantization Kernels Transpose Kernels Engram Gating Kernels Manifold HyperConnection (mHC) Kernels High-Level Modeling Layers Testing Common Patterns Pattern: Fused MoE Forward Pass Pattern: Using PyTorch Reference Implementations for Validation Pattern: Benchmarking a Kernel Pattern: Custom TileLang Kernel (extending the library) Architecture-Specific Notes SM90 (Hopper: H100/H200): Full support, primary target SM100 (Blackwell: B100/B200): Full support Kernels use hardware-specific features (tensor memory accelerator, async copy, warp-specialized pipelines) — do NOT run on older GPUs (Ampere/Ada) Troubleshooting CUDA Architecture Mismatch → You need SM90 or SM100. Check with: python -c "import torch; print(torch.cuda.getdevicecapability())" TileLang Version Mismatch → Ensure TileLang >= 0.1.9: pip install tilelang>=0.1.9 CUDA Toolkit Version → Requires CUDA 13.1+. Check: nvcc --version Out of Shared Memory → Kernels are tuned for specific tile sizes. If you hit shared memory limits, reduce batch size or sequence length, or file an issue. Running Tests Without Benchmark Flag Citation
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aradotso/tilekernels-gpu-kernels