This repository implements libkineto, a high-performance GPU/CPU profiling library that serves as the underlying profiler for PyTorch. It provides timeline tracing of GPU activities (kernels, memory copies, synchronizations) and CPU operations, correlating them to produce Chrome Trace JSON output for visualization in Perfetto. The architecture is a C++ library with multiple backends (CUDA via CUPTI, AMD ROCm, Intel XPU) selected at build time, plus a generic CPU activity profiler. Key entry points include `libkineto_init()` for setup, and a config-driven lifecycle supporting both synchronous (iteration-bound) and asynchronous (daemon-triggered) profiling. The library also includes hardware counter sampling (PM sampling for NVIDIA), a typed metadata system for structured field-level data, and stress/benchmarking harnesses. Though development is frozen with upstreaming to PyTorch, the codebase demonstrates strong engineering craftsmanship with well-abstracted interfaces, factory patterns, and platform-agnostic designs. It is not a data analysis project; its benchmarks measure trace analysis performance, not produce analytics.