We built an open-source GPU profiler you point an AI agent at, instead of reading traces yourself

Developers have created an open-source GPU profiler designed for AI agents, eliminating the need for human interpretation of traces. This tool addresses the challenges of profiling vLLM/SGLang/llama.cpp setups, which typically involve manual analysis of nsys traces. It operates without code changes or imports, supporting any CUDA or ROCm process. The profiler can break out kernels hidden within CUDA-graph decode using the "--cuda-graph-trace" node and allows for detailed kernel analysis via a lock-free probe header. It runs locally, binding to 127.0.0.1, and requires an API key for uploads.