TorchLean

4. Semantics and Graphs🔗

Autograd records one execution, and then that tape can be released. A verification claim must survive longer. We first recover the model as a mathematical function, then preserve its architecture in GraphSpec, and finally lower it to the ordinary node array used by importers and verification passes. This is where “the same model” stops being a slogan and becomes a sequence of explicit correspondence questions.

  1. 4.1. The Mathematical Specification
  2. 4.2. GraphSpec: One Architecture, Several Meanings
  3. 4.3. The Canonical Graph IR