Examples

Four standalone notebooks, each one task, built from the scripts in examples/. Unlike the tutorials, these need hawk (kernel authoring) and eagle (execution) installed from PyPI (see Installation), plus pip install "raptor-core[demo]" for numpy/torch. Every one of them also runs as a plain script from a checkout — see each notebook’s first cell for the equivalent command line.

• Hybrid orbit: analytic physics + a learned correction — one traced kernel performs a whole RK4 step with gravity written by hand and drag learned by a small torch-trained network, compiled once and run on CPU threads or as a replayed CUDA graph. The same compile-once, train-with-loss.backward() pattern is covered end to end in eagle’s Train through your kernel with PyTorch and hawk’s capstone.

• Early termination: a batch that finishes at different times — a device-side live count and a conditional graph node let a captured CUDA graph skip replays once every sample in the batch has stopped. For the general trade-off — when thinning a batch like this is worth the overhead — see eagle’s Finished samples: compaction, reorder, and when not to.

• Zero-copy interop — the numpy/cupy/torch crossings the certification matrix declares, exercised for real: a pointer-stable host round-trip and (where a CUDA-capable torch build is available) a zero-copy device alias.

• Autodiff against torch — hawk differentiates the traced kernel itself (reverse- and forward-mode) rather than executing a separate tape, checked against torch.autograd and a float64 finite-difference Jacobian.