Tenzr¶

A differentiable n-dimensional array (tensor) library for heterogeneous computing¶

Tenzr is a tensor library for heterogeneous hardware. It traces ordinary host code into an intermediate representation, applies functional transforms (jit, grad, vmap, jvp), and executes the result through PJRT on CPUs, GPUs, and accelerators.

Start Here¶

• API Reference: the C++ classes, functions, and source files.
• Install: pip install tenzr, apt install tenzr, or dnf install tenzr.
• Backends: tenzr[cpu], tenzr[cuda12], ... install only the PJRT plugins you need; the CPU plugin is the baseline of every accelerator extra.