Tune CuTile GPU kernels without repeat searches.
What it does
Guides changes to CuTile autotuning configurations, producing cached kernel launches validated against a fixed configuration.
When to use it
Use when adding, modifying, optimizing, or debugging CuTile autotuning code, including correctness or performance problems in autotuned kernels.
How to use it
It changes how the agent adds or adjusts CuTile autotuning in a project, limiting changes to configuration, caching, and tuned launches rather than kernel code.
What you provide
- An existing project
What you need · 3
Requires CuTile with the cuda.tile.tune.exhaustive_search API for searching configurations and launching tuned kernels.
Uses torch to obtain the CUDA stream and GPU device properties in the launch wrapper.
Running correctness tests and benchmarks requires a CUDA GPU; the example explicitly requires CUDA tensors and queries GPU properties.
Access · 1
This skill
GPU clock frequency
Locks GPU clock frequency