Provide expert guidance for optimizing NVIDIA GPU kernels on Blackwell and Hopper architectures.
npx skills add https://github.com/mit-han-lab/KernelWiki --skill kernelwiki
Please help me install this Agent Skill. Skill: KernelWiki Source: https://github.com/mit-han-lab/KernelWiki/tree/main Command: npx skills add https://github.com/mit-han-lab/KernelWiki --skill kernelwiki
This Skill provides expert knowledge for optimizing NVIDIA Blackwell (SM100, B200) and Hopper (SM90, H100) GPU kernels, enabling users to improve performance and efficiency.
Use the KernelWiki skill to search for information on optimizing a specific kernel, such as python3 scripts/query.py "optimization for FlashAttention-4".
High-intent search queries and answers about installing and using this skill.
To optimize NVIDIA GPU kernels for Blackwell (SM100, B200) and Hopper (SM90, H100) architectures, you need expert knowledge of performance patterns, techniques, and specific kernel case studies to improve performance and efficiency.
Low SM utilization, memory-bound scenarios, register pressure, and compute-bound bottlenecks in CUDA kernels are common performance patterns that require specific optimization techniques and case studies to resolve effectively.
Yes, KernelWiki provides guidance on using CuTe DSL, CUDA C++ with PTX inline, and Triton on Blackwell architecture, offering language and DSL-specific insights for kernel optimization.
To migrate Hopper GPU kernel features to Blackwell, you can query migration guides and specific PR references within KernelWiki to find implementation details and architecture adaptation techniques.
Yes, PyYAML is a required dependency to run the query scripts and search for specific kernel optimization patterns like FlashAttention-4 within the KernelWiki knowledge base.
linkedinProfile, generate variants, and validate Liger Kernel Triton kernels for NVIDIA GPUs.
huggingfaceOptimizes NVIDIA GPU AI inference and training with custom CUDA kernels for HuggingFace libraries.
plurigridGenerate deterministic colors across CPU and GPU with SplitMix64 PRNG.
alibabaAnalyzes GPU kernel performance using Roofline models and profiling data.
