Knowledge Graph Explorer

The tabs below show how concepts in the Torch-Spyre codebase connect. Each view filters the graph to a specific domain so you can explore without noise from unrelated subsystems.

Click a node to see its source location and connections. Double-click a node to jump straight to its code.

Views

Operations — Each PyTorch op and its Spyre implementation path: decomposition, lowering, custom op, CPU fallback, or direct eager kernel. Use this view to check whether a specific op is supported and how the backend handles it.

Compiler Passes — Pass groups and their constituent transformation functions, laid out top-to-bottom in pipeline order.

Architecture — Module dependencies, class inheritance, and dataclass definitions across the torch_spyre package.

Configuration — Environment variables and the modules that read them, showing which runtime knobs control which subsystems.

How the graph is built

A Sphinx extension runs docs/source/_ext/extract_graph.py at build time. The script parses the torch-spyre source tree with Python’s ast module and writes a graph.json into _static/js/. Because extraction is purely syntactic, no imports of torch or torch_spyre are required.

The extractors cover:

  • Op registration decorators (@register_spyre_decomposition, @register_spyre_lowering, @torch.library.custom_op, register_fallback_default, register_torch_compile_kernel)

  • Custom*Passes class definitions and their pass function lists

  • Class definitions with base classes

  • @dataclass-decorated structs and their typed fields

  • Intra-package import statements

  • os.environ and os.getenv call sites