Profile TileOPs GPU kernel latency and bottlenecks with CUDA-event timings.
npx skills add https://github.com/tile-ai/TileOPs --skill tune
Please help me install this Agent Skill. Skill: tune Source: https://github.com/tile-ai/TileOPs/tree/main/.claude/skills/tune Command: npx skills add https://github.com/tile-ai/TileOPs --skill tune
This guide provides a repeatable methodology to measure, analyze, and tune GPU kernel performance in TileOPs, ensuring reported metrics reflect GPU-only execution and that autotune trial runs do not contaminate profiler traces.
Run the benchmark to collect median GPU-only latency, disable autotune to fix the config, and then run nsys or ncu to capture clean kernel traces for analysis.
High-intent search queries and answers about installing and using this skill.
To profile GPU kernel performance, run benchmark scripts to measure CUDA-event median latency, disable autotune to fix the best config, and use Nsight Compute to inspect per-kernel metrics like memory throughput and stall reasons.
Autotune contaminates profiler traces by running multiple trial configurations, skewing aggregated kernel stats. Prevent this by disabling autotune and fixing the best configuration before profiling with nsys to capture steady-state behavior.
When benchmarking TileOPs kernels, analyze memory throughput, L1/L2/HBM hit rates, occupancy, and stall reasons using Nsight Compute to identify GPU kernel performance bottlenecks accurately.
Yes, you can use nsys and ncu for regression analysis on GEMV, MHA/GQA, and other TileOPs operators by comparing autotuned configurations and measuring GPU-only latency through benchmark scripts.
To ensure benchmark metrics reflect GPU-only execution, use benchmark scripts that report CUDA-event median timings and apply practical TMPDIR and launch-skip/launch-count workarounds during Nsight Compute profiling.
Yes, you need to disable autotune and fix the best config before profiling with nsys so aggregated kernel stats reflect steady-state behavior without contamination from trial runs.
graphsignalInstrument Python AI inference workloads with Graphsignal SDK for tracing and profiling.
alexsun07Analyze SGLang torch-profiler output to generate kernel, overlap, and fuse-pattern reports.
mlc-aiQuery Nsight Systems profiles to measure compute/communication overlap and inspect per-rank streams.
fsx950223Profile GPU kernels with rocprofv3 and generate optimization plans.
