Optimize Triton and CUDA kernels using roofline-based trial-loop benchmarking.
npx skills add https://github.com/IsNoobgrammer/skills-for-agents --skill tritonify
Please help me install this Agent Skill. Skill: tritonify Source: https://github.com/IsNoobgrammer/skills-for-agents/tree/main/skills/tritonify Command: npx skills add https://github.com/IsNoobgrammer/skills-for-agents --skill tritonify
This skill addresses the common failure mode of accepting sub-optimal performance from standard libraries by providing a rigorous, roofline-targeted optimization framework for Triton and CUDA kernels.
Invoke the tritonify skill to analyze the current kernel and begin the trial-loop optimization process for the specified hardware target.
High-intent search queries and answers about installing and using this skill.
To optimize Triton and CUDA kernels to the hardware limit, apply a roofline-based trial-loop methodology to systematically identify bottlenecks and maximize GPU hardware utilization. This ensures optimization efforts target the actual speed-of-light bound.
The roofline model helps with GPU kernel optimization by systematically identifying whether a kernel is bottlenecked by HBM bandwidth or compute capacity. This framework ensures optimization efforts target the actual hardware limit rather than guessing at performance constraints.
To eliminate HBM round-trips when optimizing custom attention kernels, apply operator fusion strategies within a trial-loop methodology. This process measures the speed-of-light bound and systematically removes redundant memory access patterns to improve LLM operation performance.
Yes, you can use this roofline-based trial-loop methodology to optimize fused loss functions for LLM operations. It targets custom fused loss functions alongside GEMM, MLP, and attention kernels across various GPU architectures to beat framework baselines.
The best way to benchmark GPU kernel performance and account for HBM traffic is through rigorous performance benchmarking combined with HBM traffic accounting. This access-aware optimization strategy identifies memory bottlenecks to maximize hardware utilization.
Your custom GPU kernel underperforms compared to standard framework baselines likely due to unoptimized HBM round-trips or compute bottlenecks. A structured, tree-based trial loop can iteratively improve kernel performance by measuring against the roofline bound.
TheBushidoCollectiveProfile bottlenecks and apply validated optimizations with repeatable metrics.
xchang1121Implement Triton CUDA kernels for PyTorch neural network layers.
alibabaAnalyzes GPU kernel performance using Roofline models and profiling data.
yzlnewDevelop, optimize, and debug TileLang GPU kernels for AI workloads.
