Concepts

The ideas behind eagle, one page each. Read Foundations first; the rest can be read in any order. The glossary defines the words.

• Foundations: the words the documentation uses, with an example each.

• Parallel execution: one body of code run across many items, on a GPU or a CPU.

• CUDA graphs: record once, replay many times; the execution model.

• Graph model: a step as nodes, and how eagle works out what depends on what.

• Graph view: draw any graph as ASCII, Mermaid or DOT.

• Gradients through a graph: reverse mode over a whole graph, on the CPU or the GPU.

• Same bits, every time: the fixed-order sum behind reproducible reductions.

• Interoperability: numpy, cupy and PyTorch arrays in and out through DLPack.

• eagle in your PyTorch program: one changed line in a training loop.

• eagle in your JAX program: one changed line in a jitted training step.

• eagle in your NumPy and CuPy pipeline: the same call on either, with streams and the memory pool.

• eagle in your NVIDIA Warp program: swap one wp.launch, keep the rest.

The library’s modules are toured on the Library structure page, and the measured speed-ups are on the Performance page.