How-to guides#
Short, task-shaped pages: you know what you want to do, here is how.
Launch a kernel and use the registry: load a compiled kernel by name, then call, stream or graph it.
C++ examples: complete, tested C++ programs and how their gate runs, including embedding a compiled kernel in your own program.
Device properties and a two-plugin manifest: read the device, then combine two compiled force plugins from one manifest.
Early termination: a batch that finishes at different times, with a conditional graph node that skips the work once every sample is done.
Zero-copy interop: one kernel fed from and handed back into NumPy, CuPy and PyTorch arrays with no copy.