Tangent

Source-to-source automatic differentiation for Python.

Tangent takes your Python function and generates a new Python function that computes its gradient — code you can read, print, step through in a debugger, and reason about. No graphs, no tapes hidden inside a C++ runtime, no black boxes.

import tangent

def f(x):
    return x ** 3 - 2 * x ** 2 + 3 * x - 1

df = tangent.grad(f, verbose=1)   # prints the generated gradient code
print(df(2.0))                    # f'(2) = 7.0

# What verbose=1 prints - the gradient is just Python:
def dfdx(x, bf=1.0):
    ...
    bx = 3 * x ** 2 * bf - 4 * x * bf + 3 * bf
    return bx

---

Why Tangent?

• Readable gradients

  ---

  The derivative of your function is ordinary Python source. Inspect it, profile it, or paste it into a code review.

• Debuggable

  ---

  Set a breakpoint inside the backward pass. Print an intermediate adjoint. No other autodiff system lets you do this naturally.

• One API, six backends

  ---

  The same tangent.grad differentiates code written against NumPy, JAX, TensorFlow, PyTorch, Keras 3, or tinygrad.

• Real Python

  ---

  Loops (with break/continue/early return), conditionals, closures, classes, list building, comprehensions, tuple unpacking, containers in and out.

• Higher-order

  ---

  Forward and reverse mode, second and third derivatives, Hessian-vector products, forward-over-reverse.

• √n checkpointing

  ---

  grad(f, checkpoint=True) trades recomputation for memory: 96.8% peak-tape reduction measured, with bit-identical gradients.

---

Reject-or-be-correct

Tangent's guiding principle: a construct either differentiates correctly in every mode — verified against finite differences by a 79,000+ case test suite — or it is rejected with a clear, actionable error. Never a silently wrong gradient.

At a glance

Install	pip install tangent-ad — see Installation
Import	import tangent (the import name is unchanged)
Python	3.9 – 3.13
License	Apache 2.0
Lineage	Created by Google Research (2017); revived, completed, and maintained by @pedronahum

Get started Language support matrix