Retrieve evidence-backed GPU kernel documentation from a SQLite corpus.
npx skills add https://github.com/fmh66/kernel-opt-agent --skill kernel-kbs
Please help me install this Agent Skill. Skill: kernel-KBS Source: https://github.com/fmh66/kernel-opt-agent/tree/main/skills/kernel-KBS Command: npx skills add https://github.com/fmh66/kernel-opt-agent --skill kernel-kbs
kernel-KBS eliminates slow, unfocused searching when you need evidence-backed GPU kernel implementation patterns across CUDA, Triton, CuTe, CUTLASS, and Blackwell/Hopper research. It helps you find merged-PR provenance, curated synthesis pages, and performance-related notes without mixing in environment checks, correctness testing, or profiling/benchmarking workflows.
Ask kernel-KBS to find evidence for your topic by querying the SQLite corpus for the most relevant pages, e.g. "Retrieve the top 10 evidence-backed pages for nvfp4 GEMV on sm100 and include the sources and any code."
High-intent search queries and answers about installing and using this skill.
You can retrieve canonical documents with provenance links by querying a SQLite-centered GPU kernel knowledge base. It surfaces source artifacts and curated synthesis pages for CUDA, Triton, CuTe, and CUTLASS without mixing in environment checks or profiling workflows.
You can filter GPU kernel knowledge by architecture, language, repo, tags, symptoms, and confidence levels. This targets specific hardware-feature or technique-based implementation searches across CUDA, Triton, CuTe, and CUTLASS research documents.
Yes, the knowledge base includes code assets and merged-PR provenance. You can optionally fetch documents following sources and including code, citing stable IDs and canonical paths for evidence-backed GPU kernel retrieval.
No, this tool explicitly excludes Nsight Compute profiling, benchmarking, correctness checks, and optimization bookkeeping. It enforces read-only safe retrieval of GPU kernel knowledge to focus solely on evidence-backed implementation pattern searches.
Query the SQLite-centered corpus to retrieve top evidence-backed pages filtered by architecture like sm100. Fetch selected documents to include sources and code assets, retrieving canonical pages with provenance for your specific hardware research questions.
thistleknotManage project-local SQLite3 feature records, file mappings, and architectural decisions.
fmh66Profile CUDA-C++ kernels with Nsight Compute metrics and classify bottlenecks.
cheetah-aloEnforce repository commit message format and pre-commit safety checks.
mindspore-aiProvide reference documentation for Triton's core CUDA APIs including tl.load, tl.store, tl.dot, tl.ammo, tl.atomic and tl.async_copy with usage examples for building GPU kernels.
