Automate tuning of AITER CK GEMM and fused MoE kernels on AMD GPUs.
npx skills add https://github.com/alexsun07/agent_skills --skill aiter-ck-gemm-tune
Please help me install this Agent Skill. Skill: aiter-ck-gemm-tune Source: https://github.com/alexsun07/agent_skills/tree/main/skills/aiter-ck-gemm-tune Command: npx skills add https://github.com/alexsun07/agent_skills --skill aiter-ck-gemm-tune
This Skill automates the tuning of AITER's Composable Kernel (CK) GEMM and fused MoE kernels, improving performance for specific model shapes on AMD GPUs.
Use the aiter-ck-gemm-tune skill to tune the GEMM and MoE kernels for your model shapes.
High-intent search queries and answers about installing and using this skill.
To tune AITER CK GEMM kernels, you can automate the discovery and benchmarking of kernel configurations to optimize performance for specific model shapes on AMD GPUs. This process extracts shapes from inference logs to find the optimal setup.
Optimizing fused MoE kernel performance on ROCm involves automating the tuning process to benchmark and compare configurations before and after adjustments. This ensures optimal LLM inference efficiency on AMD GPUs.
Yes, you need the AITER package installed along with rocminfo. These dependencies are required to automate the kernel configuration and benchmarking process for Composable Kernel GEMM operations.
Yes, you can extract model shapes directly from inference logs. The tuning process uses these extracted shapes to discover, benchmark, and apply the best Composable Kernel GEMM and fused MoE configurations.
Yes, the kernel tuning process includes benchmarking functionality that compares performance metrics before and after tuning. This allows you to verify the optimization gains for your specific GEMM and MoE kernels.
amdCollect OS, Python, GPU, and accelerator facts for Quark installation readiness.
codes1gnEnforce a finite state machine loop for NVIDIA GPU kernel tuning.
alexsun07Benchmarks sglang on AMD GPUs using Python scripts and parallel configurations.
alexsun07Analyze SGLang torch-profiler output to generate kernel, overlap, and fuse-pattern reports.
