CUDA Kernel Optimizer

Optimize the complete GPU execution path, not just an isolated kernel. CUDA Kernel Optimizer helps ChatGPT analyze bottlenecks, implement candidates, check correctness, and measure changes against your real workload.

What it does

• Checks the workload, benchmark, GPU environment, dependencies, and profiler access.
• Analyzes CUDA, CUTLASS, and Triton kernels, plus scheduling, transfers, communication, and serving behavior in PyTorch, vLLM, and TensorRT-LLM workloads.
• Uses Nsight Compute, Nsight Systems, PyTorch Profiler, compiler artifacts, and paired measurements to test performance hypotheses.
• Keeps reproducible experiment records and a validated best variant for review and handoff.
• Includes Python execution tools and an offline reference library.

How it works

ChatGPT chooses the optimization direction and the next useful test. Deterministic tools execute explicit operations and record evidence. Low-cost checks come before expensive validation; a faster microbenchmark is not presented as a faster production service.

Getting started

Install the skill in your coding agent, then provide a representative test workload, correctness checks, a target GPU environment, and the paths you allow it to modify. You can start with a read-only ten-minute readiness check if the environment is incomplete. No GPU compute or hosted inference service is included.

Source and license

Free and open source under the MIT License. The package retains its LICENSE and NOTICE, including upstream acknowledgements.

Source, installation instructions, releases, and issue tracker: https://github.com/troycheng/cuda-kernel-optimizer

This package is built from public main commit beb239b. Performance gains depend on the workload and must be established by correctness checks and repeatable measurements; no fixed speedup is promised.