agentic-cuda-optimizer
An experimental optimization tool that finds CUDA kernel candidates through repeated generation, validation and performance measurement.
SITE CAPTUREBased on information published on the service's website. Selection and review policy
WHAT IT DOES
What does it do?
This tool uses a LangGraph workflow to experiment with changes to CUDA kernel code and execution settings. A C++ tool compiles and runs kernels on a GPU, while Python compares the results and selects candidates. It records each experiment and exports the best kernel file after a successful run.
KEY FEATURES
Key features
- Generates CUDA kernel implementation candidates from task descriptions.
- Validates candidates' correctness and measures execution latency.
- Adjusts kernel code and task-specific execution settings together.
- Uses GPU properties or optional Nsight Compute measurements in subsequent experiments.
- Saves experiment records and exports the fastest validated result.