yuguo-Jack/cuda-optimized-skill

hygon-triton-kernel-optimizer

Capture, triage, benchmark, and optimize TorchInductor or hand-written Triton kernels on Hygon DCU gfx936/gfx938. Use when investigating torch.compile or TorchInductor Triton performance on DCU, parsing Triton autotune logs, saving generated kernels and…

remote-ssh-docker-workflow

Run remote compile, test, profiling, and debug tasks through SSH plus docker exec while keeping code edits local and synced to the remote node. Use when Codex must validate environment readiness, check ROCm/DTK/Hygon GPU card status, inspect Python packages,…

hygon-hip-kernel-optimizer

Iteratively optimize Hygon DCU HIP / CK Tile kernels against a Python reference using hipprof, DTK tools, dccobjdump ISA verification, roofline-style budgeting, branch selection, ablation attribution, and gfx936/gfx938-aware optimization references. Use when…

hygon-hip-baseline-generator

Generate a Hygon DCU HIP/C++ baseline kernel and correctness harness from a Torch, Triton, TileLang, Python, or CUDA/C++ reference plus shape JSON, including evidence-backed CUDA-to-HIP/DCU conversion, then hand the validated baseline to the Hygon HIP kernel…