Static graphs in a world gone eager

What it does

PyTensor is a symbolic mathematics library for multi-dimensional arrays. You define expressions as a static, mutable graph; it then optimizes that graph—folding constants, eliminating dead operations, and substituting BLAS calls—before compiling to C, JAX, or Numba. It is the current computational backend for PyMC, the Bayesian modeling library.

The interesting bit

In an era where PyTorch and TensorFlow chased eager execution, PyTensor deliberately preserves the old Theano philosophy: a static graph that can be modified in-place. That anachronism is the point—advanced optimizations are easier when you can rewrite the whole program before running it.

Key highlights

• Pure-Python codebase described as “hackable” and extensible
• Graph transpilation framework targets C, JAX, and Numba backends
• Static computation graph supports in-place symbolic optimizations
• Direct lineage from Theano via Aesara
• Serves as the computational engine for PyMC

Verdict

Grab it if you are building Bayesian models in PyMC or need optimizable symbolic computation with backend flexibility. Look elsewhere if you want an eager, general-purpose deep-learning framework.