---
title: Modernizing a CUDA library from 2015
tldr: Using an AI agent to modernize a 2015 CUDA library into a modern PyTorch package
author: Samuel Vaiter
date: 2026-01-10
status: finished
tags:
  - optimization
  - agent
---

In 2022, we released (led by [Thomas Moreau](https://tommoral.github.io)) [benchopt](https://github.com/benchopt/benchopt), a benchmarking suite tailored for machine learning workflows.
Among the initial benchmarks, there was one dedicated to [2D Total Variation](https://github.com/benchopt/benchmark_tv_2d).
Total Variation is a regularization introduced in 1992 by Leonid Rudin, Stanley Osher and Emad Fatemi that was (is) quite popular in inverse problems for imaging, before the use of deep priors.

I opened [an issue](https://github.com/benchopt/benchmark_tv_2d/issues/6) to add one of my solvers, that I implemented in 2015.
It was a CUDA library that I developed when I was a postdoc with [Antonin Chambolle](https://www.ceremade.dauphine.fr/~chambolle/) in 2014-2015. It was designed to be a fast implementation of a [Total Variation solver](https://en.wikipedia.org/wiki/Total_variation_denoising) based on [our joint paper](https://link.springer.com/article/10.1007/s10851-017-0724-6) with [Pauline Tan](https://sites.google.com/view/paulinetan) and Antonin.
But I never took the time to make it easily pip/conda/uv-friendly, and the idea kinda died due to the usual (pretext of) lack of time.

Fast-forward to 2026, agents are quite powerful to manipulate existing codebases, especially simple ones like `ftvp`.
I tried today [amp](https://ampcode.com) using Claude Opus 4.5 with the following prompt.

> My goal is to modernize this repository. `ftvp` is a CUDA library dedicated to the computation of the proximal operator of the isotropic Total Variation in 2D and 3D on Nvidia GPU. I want to drop the C-first aspect, the goal is to package a Python algorithm. It should use modern CUDA as of 2026. Do not change the algorithm, but feel free to change everything. The use is described in README.md. A reference implementation in C is in TV4colorCPTV.c. The other files are the CUDA implementation. Ultimately I want a PyTorch-aware function proxtv(img, lambda) that works on BW and color images where img is the torch array and lamba the regularization parameter of TV.

Following this prompt, Amp used \$1.20 to generate this [commit](https://github.com/svaiter/ftvp/pull/2/commits/abfdf9ebc8c5cfc7d75b3cd01fbfe4a972c3522a).
It forgot to add ninja as a requirement, and the examples were only trying the prox on a black square. So I asked,

> 1. You need to add ninja to the requirements to be able to JIT.
> 2. Write me an additional examples that perform the denoising on a real image

The generated script was correct, but again it missed two dependencies required to run the examples.

> Add numpy and pillow as dependency for running this example.

Finally, I generated the PR request as follow

> Generate a summary of the changes so that I can copy it in the PR interface of githb. Be explicit on the fact that is was driven by a AI agent.

Overall, the total cost was \$2.15 (of free credits) for something that I am convinced I would never have done otherwise.
The full thread is [here](https://ampcode.com/threads/T-019ba740-5029-745f-8915-0d21d531d394).
I *still* have to check that this is correct to merge it, and then submit a pull request on [benchopt/benchmark_tv_2d](https://github.com/benchopt/benchmark_tv_2d).

I will update this feedback note when this is done.